This earnings season, one of the most revealing numbers sits in a lease footnote: $359.3 million.

That is the first layer of costs above the original Barber Lake budget that Cipher has agreed to bear under its amended lease arrangements. Above that threshold, the tenant will reimburse half of the additional costs through extra rent over the combined 20-year contracted term, with a contracted return on those reimbursed amounts.

The figure is a contractual threshold, rather than a final tally of the project’s overrun. But it puts a useful number against a broader question: once an AI lease is signed, who pays when delivering the facility becomes more expensive?

Cipher’s call offered a candid explanation of how that happens.

Management said both the company and its tenants had been eager to move quickly. Leases were signed before designs were far advanced, and more change orders followed than expected. Changes to cooling infrastructure, for example, can require additional crews and disrupt work elsewhere.

Cipher also described lower-than-expected productivity, late equipment deliveries and crews waiting for them. A tight labor market compounds the problem. Equipment-price inflation itself, he said, was not a major driver.

A project can become more expensive even when equipment prices barely change. Deliveries, crews and designs still need to line up.

Cipher says it is responding by doing more engineering earlier. For its next generation of leases, management has held off on signing until it has agreed design sets that are 30%–60% developed. The Barber Lake amendment also adds another ten-year lease term and additional rent, so its cost exposure needs to be assessed alongside what Cipher receives for taking it on.

Riot’s call on Wednesday picked up almost exactly where Cipher’s left off.

Asked about industry overruns and its Rockdale budget, management explained that construction prices become binding as defined work packages enter guaranteed maximum-price contracts. Appointing a general contractor does not, by itself, lock the whole bill.

Riot said it was finalizing its contractor arrangement and expected roughly 70% of Rockdale’s total budget to fall under that framework. The balance covers equipment Riot purchases directly. Management maintained its estimate of $11 million–$12 million per critical IT megawatt.

That protection is still being finalized, and some risks remain. Tenant-requested changes are reimbursed or incorporated into rent. Contractors bear costs above the agreed maximum for covered scope. Riot retains exposure to uncovered scope, site conditions and equipment it buys itself, with reserves and contractual provisions addressing labor, tariffs and material escalation.

Those details tell investors more than a headline capex estimate. How developed was the design at signing? Which prices are binding? And does the developer recover extra spending immediately, or finance it until future rent arrives?

The last question becomes more pressing when credit tightens.

Riot’s CFO described wider credit spreads since September and deeper lender diligence on tenant credit, power already interconnected and energized, and a demonstrated delivery record. Construction uncertainty affects both the capital a project needs and lenders’ willingness to provide it.

Riot’s illustrative funding model shows the sensitivity. At 90% loan-to-cost, management estimated a cash-equity requirement of about $800 million. At 80%, it rises to roughly $1.5 billion—a $700 million difference.

That is a scenario range, not evidence lenders have cut Riot’s leverage. It shows why the share of a project lenders will finance can matter as much as the interest rate.

Core Scientific shows how that exposure can differ even within one company. Its filing says tariffs raised costs for certain equipment and materials purchased directly for non-customer-funded projects. Some equipment has lead times exceeding 12 months. Its CoreWeave conversions are almost entirely customer-funded, with Core’s contribution capped; the AMD and Neocloud buildout does not have the same protection.

Core disclosed $4.4 billion in outstanding purchase and construction commitments, substantially all expected within 24 months, and plans additional project-level bond financing. That commitment figure is not a financing shortfall. But Core says unavailable financing could force spending changes, and its existing secured-note financing includes an uncapped completion guarantee for specified projects. Raising project debt can still leave the parent carrying construction risk.

Riot’s new Corsicana arrangement is worth watching in that context.

The tenant will procure and own the modular deployment, while Riot provides powered pads and power infrastructure. Riot estimates approximately $6 million per critical IT MW, including tenant reimbursements and site work, and says the structure reduces procurement and delivery exposure. Its responsibilities differ from a traditional turnkey development, which matters when comparing costs per megawatt.

Missing the schedule still has consequences. The lease provides escalating rent credits for capacity delivered more than 60 days late, and potential termination rights for an undelivered phase after 270 days, subject to contractual protections and extensions.

These were already live issues at our Energy Investors Forum 2026 in July. LU-VE Group’s Giovanni Mariani warned that equipment demand was growing faster than manufacturers could keep up. William Hossain, representing Project Vanguard, discussed the time needed to develop experienced workers and the risk of losing them to competitors. Galaxy’s Blake King highlighted the sequencing challenges of developing generation alongside data centers.

The latest earnings commentary shows how those constraints feed into budgets, rental terms and financing. After the lease announcement, the question becomes how much uncertainty remains—and who has agreed to carry it.

We’ll continue that conversation at Energy Investors Forum 2027, June 15–16 in Dallas. If you are building, financing or supplying these projects, join us. And if you have firsthand experience managing construction costs, securing equipment or structuring project finance, we’d welcome your perspective on stage.

• Explore EIF 2027 and register

• Express interest in speaking

• Explore branding and partnerships

Regulation News

• AI Data Center Boom Set for Upheaval After US Election Rout

Hardware and Infrastructure News

• Vistra Finalizes Purchase of Cogentrix Energy, Adding 5,500 MW of Natural Gas Power Plants

• Big Tech Holds $3 Trillion in Spending Commitments for AI Boom

• WhiteFiber Closes $60M Data Center Acquisition as NC-1 Delays Full Revenue Ramp

• Oracle cancels $7bn Tencent AI deal under Trump pressure

• Microsoft Passes on Nscale Texas Expansion Option: Reports

• MARA Wins FERC Approval for $1.5 Billion Long Ridge Power Deal

Corporate News

• Cipher Digital Books First HPC Revenue as Data Center Buildout Advances

• OpenAI Plans ‘Compute Campus’ Rebrand as Data Center Opposition Grows

• Two Workers Injured in Fire at Cango’s Georgia Bitcoin Mining Site

• Nvidia Tells Cloud Clients to Reduce Chinese Business: FT

• Microsoft blames data centre delays for lower capital expenditure

Financial News

• Dominion Energy Reports $1.2 Billion Third-Quarter Net Income for 2026

• Hut 8 Achieves $189.5 Million Q3 Profit and Secures $1.07 Billion Credit Facility

• Riot Sold 10,599 Bitcoin in Third Quarter as AI Expansion Accelerated

• TeraWulf Secures Google Credit Backstop for $19 Billion Anthropic AI Lease

• Cipher, Hut 8 Plunge as Texas Election Raises Data Center Risks

• American Bitcoin Reports $702.9 Million Bitcoin Reserve in Q3 2026 Financials

• Ionic Digital Revenue Rises 32% as AI Leasing Income Outpaces Cash Rent

• Keel Revenue Falls 59% as Bitcoin Mining Gives Way to AI

• Core Scientific’s Colocation Revenue Jumps 58% as AI Capacity Expands

• Duke Energy Raises Earnings Forecast Floor as Quarterly Profit Climbs

Feature

• INTERVIEW SERIES: AI’s Infrastructure Race Runs Into a Slower Clock

• What data centre construction delays mean for debt holders - FT

• TEM Explainer Series: Behind the AI Power Rush–One Electricity Bill, Two Competing Narratives

• Nvidia Chief Seen as Target of Chinese Spy Plot - NYT

Email us

If the rebound in AI infrastructure stocks holds, Monday’s Anthropic S-1 may be remembered as the moment sentiment turned. The filing landed on October 26, just as the group began recovering from the recent selloff. TheEnergyMag’s AI Infrastructure Growth Index showed a local low of 4,086 last week, before bouncing back by 13% to 4,618 as of Wednesday.

There is a case for it. Anthropic disclosed approximately $265.6 billion of future fixed or minimum guaranteed purchase commitments as of June 30. Those obligations stretch over years and across providers, but their scale offers support for the demand assumptions behind the infrastructure buildout.

The next test comes from the companies supplying that capacity. Even with a willing customer, construction overruns, delayed handovers and financing costs can erode returns. Earnings calls running from Cipher on November 2 through CoreWeave on November 10 are expected to reveal how much contracted demand is becoming billable capacity, recurring revenue and cash.

Galaxy offers an early indication

Galaxy’s data center adjusted EBITDA rose to $69 million in Q3 from $11 million in Q2, while delivered critical IT capacity remained at 133 MW. Helios Phase I had been handed over to CoreWeave by the end of June, making Q3 the first full quarter of revenue-generating operations. Meanwhile, quarterly data center capex increased to $757 million from $448 million as expansion continued.

That captures two forces shaping this earnings season: completed capacity is starting to contribute, while the next phase still needs substantial capital. Timing matters. A hall delivered late in a quarter may barely affect revenue until the following period.

Which capacity is earning—and getting paid?

Core Scientific ended June with 395 MW of billable capacity and reached 485 MW earlier this month. Its results should show how much further billing capacity increased and how those deliveries translated into colocation revenue. TeraWulf’s revenue-generating critical IT capacity rose from 81 MW at June-end to 102 MW in July. That should support its HPC revenue contribution, alongside progress toward further deliveries previously expected in the second half.

Revenue composition is worth watching too. Riot’s $23.2 million of Q2 data center revenue included $18.3 million of tenant fit-out services and $4.9 million of operating leases. With the initial 25 MW for AMD delivered, Q3 should provide a clearer view of recurring rent. Another 10 MW was targeted for November.

For Ionic Digital, collections deserve attention. Q2 infrastructure leasing revenue reached $43.8 million, while cash payments for the operating 234 MW at Ward County began in August. The effect on receivables and operating cash flow should become clearer. Its additional 466 MW received provisional ERCOT Base Load classification in September, with energization targeted for end-2027.

What remains to be built and financed?

Cipher’s October presentation outlined additional completion costs at Barber Lake and an amended rent commencement schedule spanning Q4 and Q1 2027 for several halls. Black Pearl began generating rent in August. The call should clarify the billing ramp, remaining handovers and cash required to finish construction.

Hut 8’s initial deliveries at River Bend and Beacon Point were targeted for Q2 and Q3 2027, respectively. Construction progress and remaining funding needs will therefore say more about those developments than Q3 revenue. Keel remains earlier in the process: its August update described tenant negotiations at three sites. Any signed lease would need to be assessed alongside its construction budget, delivery obligations and financing.

CoreWeave illustrates the financing burden at scale. Q2 revenue reached $2.575 billion and adjusted EBITDA was $1.510 billion, alongside $640 million of net interest expense and a $626 million net loss. Converting backlog into revenue must be weighed against the spending and borrowing needed to deliver it. Across the sector, capitalized interest, principal repayments and restricted cash also deserve scrutiny.

Bitdeer’s Q2 AI Cloud revenue reached $14 million, but the segment recorded a $2.3 million gross loss. Improving margins and progress toward Tydal’s initial Q4 colocation delivery would strengthen its AI expansion case. American Bitcoin requires a different lens: mining costs, fleet efficiency, capital spending and Bitcoin per diluted share, including how accumulation is financed.

Anthropic’s filing may have restored some confidence in demand. Results showing capacity entering billing, rent becoming cash and construction staying within funded budgets would give that confidence firmer support. If Monday marked the bottom, earnings season is where the case for a lasting recovery starts to take shape.

P.S. Recall our last issue on bitcoin’s hashrate regaining the 1 ZH/s level? Turns out bitcoin’s mining difficulty is set for another 6% jump in a few hours. Who’s staying plugged in—and why? Join our October 29 X Space to hear from the miners sticking with Bitcoin as their peers turn to AI.

Regulation News

• ‘Pure arrogance’: anger in Finland as Google illegally fells 250,000 trees for datacentre

• Trump to roll back methane rules in latest Texas campaign push

Hardware and Infrastructure News

• HIVE Secures $129 Million Loan at 9% for Cohere AI Project

• IREN Closes $363 Million AI Equipment Refinancing With Wingspire

Corporate News

• Nvidia to Invest $3 Billion in Meta as Companies Expand AI Partnership

• Luxor Raises $16 Million From CMT-Led Group to Expand AI Business

• Constellation Secures Conditional $5 Billion DOE Loan Commitment to Expand Nuclear Output

Financial News

• Bloom Energy Reports Q3 2026 Revenue of $1.24 Billion

• TeraWulf Expands Revolving Credit Line 60% to About $400 Million

• Soluna Repays $28.7 Million Credit Facility to Generate Capital

• Anthropic Files for IPO as Quarterly Sales Hit $11.6 Billion

• AI Infrastructure Stocks Rebound as Anthropic Reveals Operating Profit

• Galaxy’s Data Center Earnings Jump as Helios Leasing Revenue Ramps Up

Feature

• OP-ED: Private Miners Have an Evidence Problem

• ENGIE’s North America Chief Says AI Boom Needn’t Raise Power Bills

• TEM Explainer Series: A New Data Center Doesn’t Just Draw Power. It Can Redraw the Grid

Email us

Big public miners are making room for AI. Bitcoin’s hashrate is climbing anyway.

The network’s seven-day average has regained 1 zettahash per second—1,000 EH/s—over the past week and reached its highest level since March 3. That rebound is especially interesting given how much mining capacity the public companies have been shedding.

The recovery is showing up in difficulty. Bitcoin’s mining difficulty rose 5.92% on October 16, and the pace of block production so far in the current difficulty epoch points to another estimated 5.16% increase in about seven days. That estimate will move as more blocks arrive, but for now, the network is on course for two substantial increases in a row.

Who’s staying plugged in—and why? Join our October 29 X Space to hear from the miners sticking with Bitcoin as their peers turn to AI.

That’s hardly the breathing room remaining miners might have expected from the AI migration.

We’ve been following that retreat. Earlier last month, we examined public miners’ declining realized hashrate alongside a 52% quarterly increase in directly reported HPC and AI revenue. Additionally, we looked at the hardware investment behind the capacity being displaced: roughly $1.5 billion equivalent, using an assumed purchase price of $20 per terahash.

Converting a mining site to AI is a lasting change. Once the machines are removed, the facility is rebuilt and its power is committed to AI customers, that capacity isn’t waiting for a better hashprice to switch back on.

So the recent rebound needs to be understood against that backdrop. The network is rebuilding hashrate despite those conversions. The replacement capacity has to come from elsewhere.

Some of it is easy to spot.

Take Bitdeer. Its September update showed self-mining capacity reaching 82.3 EH/s, up from 79.9 EH/s in August and 35 EH/s a year earlier. Its co-mining business, which places Bitdeer-owned machines in third-party data centers, grew to 28.7 EH/s from 21.6 EH/s. Across those two categories, that’s another 9.5 EH/s in a month. Bitdeer is pursuing AI, too. It just hasn’t stopped expanding its mining business along the way.

There may also be a quieter part of the recovery: small and medium-sized operators picking up equipment that the departing giants no longer need.

A mining site can leave Bitcoin for good without its machines sharing the same fate. Move those machines to a buyer with cheaper electricity and available rack space, and they can compete again. A discounted purchase price makes that proposition more attractive.

The resale channels are there. Sazmining, for example, lists pre-owned Antminer S21 XP machines for hosting in Paraguay and Texas. Those listings don’t identify the original owners or tell us how much public-miner inventory is reaching smaller buyers. But they show how equipment can find another route back into service.

The contrast is playing out in new hardware, too.

Core Scientific terminated its equipment agreement with Block and Proto Global during the second quarter, recording a $41.9 million loss as it moved toward high-density colocation. The original agreement covered chips representing 15 EH/s.

Meanwhile, Sazmining is offering individual buyers access to Proto immersion hashboards, with deployment and operation handled at a US site. That doesn’t directly establish that Sazmining is selling Core’s canceled inventory, but the timing makes the connection intriguing.

It’s a striking reversal of the 2021 mining arms race: the giants that once crowded smaller buyers out of the hardware market are stepping aside, while equipment finds its way into retail hands.

With one sharp difficulty increase already recorded and another taking shape, the competitive pressure is coming back.

Regulation News

• Memphis Rejects Longer Data Center Freeze, Approves 135-Day Pause

Hardware and Infrastructure News

• Nscale’s Data Center Delays Put Delivery Risks in Focus Ahead of IPO

• Nscale Faces IPO Setback as Microsoft Reportedly Scraps Essex AI Deal

• Anthropic Signs First Canadian Data Center Lease With EverWind

• Nscale, Anthropic Discuss Second $45 Billion AI Computing Deal: Report

• EverWind’s Anthropic Data Center Faces Backlash Over Power, Water

• Soluna, Applied Digital Plan 50-Megawatt AI Data Center in Texas

• IREN Secures Works License for 300-Megawatt Spanish AI Campus

Corporate News

• IREN Adds Real Madrid Sponsorship After Spanish Data Center Expansion

• Musk Pledges SpaceX Shares for Children After Anthropic Unveils Fund

• Nebius Publishes Live GPU Spot Prices in European AI Push

Financial News

• Bitdeer Mines 1,355 Bitcoin in September as AI Cloud Backlog Surges to $2.9B

• Top hyperscalers likely to have more AI-linked debt than now known, S&P Global warns

• Elice Plunges 37% in Korean Debut as AI Infrastructure Bets Falter

• Anthropic’s $80.7 Billion First-Half Loss Driven by Financing Charges

• Nebius Secures $3.5 Billion Citi Facility to Fund AI Expansion

Feature

• TEM Explainer Series: Large Power Users Can Mean Smaller Energy Bills

Email us

About $74 billion in equity market value disappeared from the 20 companies tracked by TheEnergyMag over the past week, reflecting mounting concern about the debt financing the AI buildout. The question is whether investors are recognizing credit risks they previously overlooked—or whether online alarm is pushing the selloff beyond what the underlying finances justify.

TheEnergyMag’s AI Infrastructure Growth Index was down nearly 20% over the past 7 days and recorded a 10% decline on Wednesday alone.

It unfolded in a sequence familiar to anyone who has followed crypto markets: a worrying chart, a prominent investor’s warning and a short reply that travels further than the financial details. It’s easy to label the resulting anxiety FUD—fear, uncertainty and doubt. But the underlying credit concerns are harder to dismiss.

The Neocloud selloff gets a megaphone

At the center of the recent selloff, Neoclouds like Nebius and CoreWeave accounted for about $26.4 billion of the weekly loss—more than a third of the total from just two companies. CoreWeave had fallen approximately 25.4% from its October 7 close. Nebius was down about 21.6%.

Both businesses depend on getting an expensive sequence right: secure equipment and data center capacity, bring it online, then collect enough revenue to justify the investment. Financing has to remain available while that happens.

In his October 13 post on AI and private credit, VC investor Chamath Palihapitiya questioned whether lenders were adequately pricing the risks. His follow-up highlighted CoreWeave’s 2032 bond yield touching 15% before closing around a record 14.5%. Elon Musk’s reply helped carry the concern beyond the usual credit-market audience.

Crypto investors know how quickly these conversations can compress. Questions about one company’s financing become questions about everyone’s balance sheet. Falling prices then appear to validate the concern, drawing more attention to the original warning.

That is a useful lens here, although it cannot tell us how much selling the posts caused. Chamath’s chart documented a bond move already underway, but sentiment and financing can reinforce each other. Falling bond prices can signal tougher terms for future borrowing. Falling shares make equity fundraising more dilutive.

Dismissing the concern as online noise would miss that connection. Treating every falling stock as confirmation of the worst-case scenario would miss the differences between companies.

We have been watching this bill come due

A previous Miner Weekly issue looked at the pressure already building.

Fourteen companies tracked by TheEnergyMag brought in $48.24 billion of net financing across their latest six-month reporting periods. Net debt supplied $35.09 billion, nearly 73%. CoreWeave, Nebius, Hut 8 and IREN accounted for roughly 76% of the financing inflow – and three of them are neoclouds.

CoreWeave shows how strong growth and financial pressure can coexist. Its second-quarter revenue more than doubled to $2.575 billion. Net interest expense reached $640 million—roughly a quarter of revenue. It also reported approximately $104 billion of backlog, with revenue recognition dependent on delivery and service availability. The challenge is turning them into cash on a schedule that works for the balance sheet.

And that 15% bond yield needs some context. It does not suddenly reset the rate on all of CoreWeave’s debt. It tells us what return investors demand to buy that particular bond at its market price—a warning worth watching when more financing may be needed.

The next dollar is getting more expensive

Wednesday’s inflation report gave investors another reason to worry. September CPI rose 3.7% from a year earlier, up from 3.4% in August and above economists’ 3.6% forecast. Traders increased their bets on another Fed rate hike as Treasury yields rose and stocks came under broader pressure.

That adds a second layer to the AI financing concerns. Investors are questioning individual borrowers’ creditworthiness while also reassessing how expensive money could become across the economy. For companies still funding large construction and equipment commitments, higher benchmark rates and a bigger premium for credit risk can compound the pressure.

The CPI release cannot explain the entire week’s losses, but it helps explain why Wednesday’s selling reached beyond the companies drawing attention on X.

Investors are also looking harder at the revenue assumptions behind AI spending. October 8 reporting put OpenAI’s September annualized revenue near $50 billion, below the roughly $70 billion previously circulated. The difference involved gross-versus-net accounting for partner sales; it does not mean $20 billion of demand vanished.

Anthropic offered a more encouraging picture at its October 14 investor meeting in San Francisco. Attendees told Reuters that third-quarter revenue reached $16.8 billion, up 45% from the previous quarter, and that the company reported an adjusted profit. Analysts who met executives estimated its annual revenue run rate could reach $110 billion by year-end. That is a projection of its revenue pace, rather than $110 billion of sales expected across 2026.

Those figures complicate the idea that AI demand is running out of steam. But strong growth at a model developer does not guarantee attractive returns for every infrastructure supplier. The credit question remains whether cash reaches each borrower fast enough to cover its obligations. Competition, equipment spending and interest costs can absorb much of the upside even as usage keeps climbing.

There is another bill to watch

Political uncertainty may also be adding to the unease. With the midterms approaching and forecasts pointing toward Democratic gains in the House, investors have another question to weigh: how might the rules around the AI buildout change? That is a plausible source of concern, even if its contribution to the selloff is difficult to isolate.

The political pressure is already visible—and crosses party lines. As explored in a previous issue, concerns about household electricity bills, water use and local control are dividing even the Republican coalition behind Trump’s AI ambitions.

Those arguments have financial consequences. Longer approvals can delay revenue, while stricter requirements to fund grid upgrades or pay for reserved capacity can add obligations before a facility starts earning.

At the center of the debate is a deceptively simple question: will a new data center raise the neighbors’ electricity bills—and who should pay if it does?

That is the starting point for a new four-part Explainer series by TheEnergyMag. The first installment, “Your Electricity Bill Is More Than the Electricity You Use,” breaks down what households actually pay for and how data-center demand can change the equation. Understanding those costs helps make sense of both the political backlash and the financial risks investors are trying to price.

Regulation News

• KEEL-Linked 75 MW Data Center Wins Conditional Approval in Allentown

Hardware and Infrastructure News

• Bitdeer Secures 67MW AI Data Center Campus in Malaysia, Bringing Total to 273.5MW

• Bitdeer AI Secures 60MW Data Center Agreement in Malaysia to Reach 95% of Global Capacity Target

• Binance to Shut Mining Pool After Years of Market Share Losses

• Core Scientific’s Billing AI Data Center Capacity Rises to 485 MW

Corporate News

• Duke Energy Reaches Data Center Cost Deal With Tech Giants

• CleanSpark to End Monthly Mining Updates as Data Center Business Expands

Financial News

• Nvidia-Backed Firmus Scraps $5B Australian IPO, Seeks Private Funding

• IREN Rises Overnight After Report of Anthropic AI Cloud Deal

• SWI Raises €800M to Fund AI Push After Genesis Digital Takeover

• Cipher Digital Plans $850 Million Convertible Sale to Fund Data Centers

• CoreWeave Falls 7% Premarket as Bond-Yield Surge Fuels Debt Concerns

Feature

• TEM Explainer Series: Your Electricity Bill is More Than the Electricity You Use

Email us

AI infrastructure doesn’t run on chips alone. It needs power, capital, and the right people to bring it all together.

Following our inaugural edition in July, Energy Investors Forum (EIF) returns to Dallas on June 15–16, 2027. We’re bringing together institutional investors, lenders, energy developers, data center and bitcoin mining operators, and industry leaders for candid discussions and meaningful connections at the intersection of energy, compute and capital.

We’re now shaping the speaker lineup and welcoming discussions with prospective partners and participants. More details below.

Now let’s dive in.

Buying chips for the AI boom is becoming a financing business of its own. On Tuesday, FT reported that SpaceX is looking to raise $40 billion to buy Nvidia AI chips, with $10 billion in bank loans and $30 billion in investment-grade debt.

Just one day later, WSJ reported that Broadcom has been working to arrange an even larger package at $50 billion for the custom AI chip it is developing with OpenAI. Meanwhile, Oracle is separately discussing chip financing with Apollo and Goldman Sachs. And interestingly, Apollo is involved in all three discussions.

While these talks remain preliminary and might not produce a deal, they come at an intricate time as big lenders have become more selective about data center loans and decisions increasingly depend on the customer, the contract and the path to completion.

Oracle’s Project Jupiter in New Mexico illustrates the scrutiny. Oracle’s force-majeure notice over potential power-related delays was prompting bankers and investors to reassess how development risks are allocated.

These reports suggest greater lender selectivity, although company disclosures show that substantial financing remains available. The differences between transactions are revealing.

IREN announced a $3.65 billion investment-grade GPU financing package on June 1 to support its Microsoft cloud contract, with a blended debt cost of 6%. In its August results, IREN disclosed a separate $2.4 billion financing at a 9% fixed rate for its Mackenzie air-cooled expansion.

The three-percentage-point gap is notable, though it can’t be treated as a like-for-like increase in IREN’s borrowing costs since the packages finance different exposures. But their terms demonstrate how widely the price of capital can vary within the same company.

They also highlight the need to distinguish equipment financing from funding for an entire development. A facility covering most GPU purchases does not automatically cover buildings, electrical infrastructure, construction contingencies and working capital.

CoreWeave’s disclosures show how the duration of customer contracts can introduce another financing consideration.

In March, the company announced an $8.5 billion non-recourse, investment-grade financing facility. Its floating-rate tranche carried a margin of 2.25 percentage points above SOFR, the benchmark borrowing rate, while its fixed-rate tranche cost approximately 5.9%.

An August $2.6 billion facility carried a margin of 5.5 percentage points above SOFR, a parent guarantee and a minimum debt-service coverage requirement of 1.35 times once testing begins. CoreWeave said the roughly five-year financing extended beyond the underlying customer contracts’ average three-year duration.

That maturity gap means the borrower may need to renew customers or find new users while debt remains outstanding. The higher margin and parent guarantee illustrate the price and obligations that can accompany more flexible financing.

Cipher Digital offers a counterpoint to any suggestion that the market has broadly closed. Its Stingray subsidiary priced $810 million of senior secured notes in June with a 6% coupon, issued slightly below face value. In August, Cipher said the completed offering funded Stingray through substantial completion and reimbursed some previously incurred project expenditure.

That is a concrete financing milestone: a specific development with funding secured through a defined stage. It gives investors more information than a headline capacity figure or prospective customer agreement.

Even then, the coupon captures only part of the economic bargain.

In Cipher’s separate Barber Lake transaction announced in September 2025, Google agreed to backstop $1.4 billion of Fluidstack’s lease obligations in support of project financing. In exchange, it received warrants representing approximately 5.4% pro forma equity ownership, subject to adjustments and potential cash settlement under certain circumstances.

Credit support helped make the project financeable, while giving Google participation in Cipher’s equity value. For shareholders, warrants, required equity contributions, parent guarantees and restrictions on project cash all belong alongside interest expense when assessing the cost of expansion.

Getting energy and compute infrastructure projects financed takes more than a compelling business case. It takes the right investors, developers, power providers and customers working together. That’s the thinking behind Energy Investors Forum (EIF) 2027, returning to Dallas on June 15–16. Building on an inaugural edition that brought together 300+ senior executives, we’re shaping the next forum around candid discussions, meaningful introductions and opportunities to turn conversations into business outcomes.

SpaceX’s proposed $40 billion financing will be an important transaction to watch: whether it closes, how it is priced and what protections investors obtain. Its scale alone cannot tell us what another developer will pay.

Interested in being part of EIF 2027?
Apply to speak
Apply for a complimentary institutional investor pass
Explore sponsorship opportunities
Secure tickets at the launch rate through November 30
Learn more at energyinvestorsforum.com

Regulation News

• California Man Charged in $300 Million Nvidia AI Chip Smuggling Case

• Google Told to Halt Work at Two Finnish Data Center Sites Pending Review

Hardware and Infrastructure News

• Applied Digital Activates 75 MW of AI Infrastructure at North Dakota Campus

• TeraWulf Expands Muskie AI Data Center Campus to 1 GW

• Fortitude Plans $100M Purchase of Bitmain’s Next-Generation Zcash Miners

• Google and Constellation Lock In 890 MW of New Nuclear for PJM

• Applied Digital Secures Up to 1 GW Power Agreement for Finnish AI Data Center Expansion

• DarkHorse Secures 50 MW Power with Hopkinsville Electric Plant Board

• Black Hills Plans $1.8B Gas-Power Build for Google Data Center

Corporate News

• Nscale Appoints Former Meta Executive Justin Osofsky as Chief Operating Officer

• Amazon Pledges $1B to Data Center Communities, Ends Government NDAs

• Nscale Appoints Former Comcast Executive Amy Banse to Board

• NextEra Energy Reaches $15.5 Million Settlement in Shareholder Derivative Litigation

• CoreWeave Enters India with 240 MW AdaniConneX Collaboration

• Foundry CEO Colyer Steps Down as Bitcoin Pool’s Share Retreats

Financial News

• SpaceX looks to raise $40bn to buy Nvidia chips

• Applied Digital reports fiscal first quarter 2027 financial results

• Oracle, Broadcom and SpaceX Seek Blockbuster Debt Deals to Pay for AI Chips

Feature

• OP-ED: AI Infrastructure Starts at Home

Email us

The hashrate public bitcoin miners shed in the first half of 2026 carries an uncomfortable price tag: roughly $1.5 billion in equivalent hardware investment, even at an assumed acquisition price of just $20 per terahash (TH/s).

A previous Miner Weekly issue estimated that public miners lost 75 EH/s of realized hashrate while their directly reported HPC and AI revenue rose 52% quarter over quarter. Much of that retreat reflected a redirection of power toward AI infrastructure.

The revenue growth showed what operators were gaining. But putting a purchase price on the disappearing hashrate helps illuminate what they had already paid for.

At $20 per TH/s, 75 EH/s corresponds to $1.5 billion of mining machines. That excludes the buildings, electrical equipment, cooling and installation needed to operate them.

That is an often-overlooked cost of the AI transition: investors had already financed the mining capacity being displaced. At some sites, equipment was marked down within months of starting production, raising questions about how much of that investment mining could recover before AI took over.

And the accounting consequences are becoming visible.

A separate review by TheEnergyMag of 12 tracked companies identified approximately $1.1 billion in asset impairments and held-for-sale markdowns during the first half of 2026. IREN and Core Scientific accounted for almost 89% of the total.

The similarity between the two billion-dollar figures may seem coincidental. One estimates hardware investment equivalent to a reduction in mining output; the other measures disclosed accounting charges across different assets and companies.

But both point toward the same investment question: how much value remains in the mining buildout as operators redirect their sites toward another business?

The spending surge was recent. A March 2025 Miner Weekly issue reported nearly $5 billion in equipment and infrastructure spending across the preceding reporting cycle. Companies that separately disclosed mining-hardware purchases accounted for more than $3 billion in 2024 alone.

That expansion helped push Bitcoin’s network into zetahash territory in 2025. Operators bought more efficient machines and added capacity to defend their share of production, while their competitors did the same.

Now, some of those investments are being displaced much sooner than the scale of the spending might suggest.

Cipher’s Black Pearl facility began mining in mid-2025. By year-end, following an agreement to convert the site to high-performance computing, Cipher had recorded a $96.1 million markdown specifically on Black Pearl’s mining machines. Those same machines generated $57.9 million in revenue during 2025—meaning the markdown exceeded the site’s revenue from its first several months of operation.

That comparison does not establish the project’s lifetime return, but it captures the compressed timetable. A newly operating mine was already preparing for a different purpose. Cipher’s charge also occurred before the first-half 2026 write-down total.

IREN illustrates the scale. It reached 50 EH/s in June 2025, then recorded roughly $695 million in impairments and held-for-sale markdowns during January–June 2026, with substantial charges associated with mining assets displaced by AI conversions.

Not every charge is an AI conversion cost. Core Scientific attributed its major mining impairment to deteriorating mining economics. The broader total includes other asset categories as well. The evidence supports a costly reassessment of mining investment, rather than a claim that every dollar was sacrificed for AI.

Interest arrives before the revenue ramp

While write-downs are noncash when recognized, construction and interest payments continue to consume cash when replacement facilities are being completed.

TeraWulf generated approximately $53 million in HPC leasing revenue during the first half while paying $131 million in cash interest across the company. Cash reserves and other income also support those obligations, so the comparison is not a standalone measure of financial distress. It shows the gap between the revenue ramp and the cost of financing it.

Access to the next round of funding may also become harder. According to an excerpt from The Information’s reporting on Thursday, Société Générale and SMBC have become more selective about data center financing, while MUFG is also stepping back.

That would add another constraint for miners midway through conversions: retiring mining capacity reduces the old revenue stream, while completing its replacement still requires capital. More selective lenders could mean tougher terms, additional equity funding or a slower buildout.

At the same time, delays can extend that period after mining revenue has already been reduced. Oracle is not a miner, but the recent news on Project Jupiter offers a concrete example of how power constraints can reach the financial arrangements behind an AI campus.

The underlying setback is documented: New Mexico’s land commissioner again rejected a request in July to route part of the project’s natural-gas pipeline through state trust lands. Jupiter’s planned fuel-cell power system depends on natural gas.

A delayed campus opening is not established. What is clear is that energy delivery risk has become a question of contractual obligations and who pays while a project waits.

Communities are demanding more control

That uncertainty sits within a widening political debate over data centers’ electricity demand, water use and effects on household bills. New York announced a pause on state environmental permits for new hyperscale projects. Massachusetts introduced additional approval requirements, while Chicago’s mayor proposed a year-long moratorium.

Existing mining sites may become more valuable where they offer usable power access. But conversions still need to satisfy technical requirements and obtain applicable approvals.

The cost of waiting is also facing renewed interest-rate pressure with the Federal Reserve raising its benchmark range by a quarter point to 3.75%–4%. Existing fixed-rate debt retains its terms, but new financing, floating-rate borrowing and refinancing face a less forgiving backdrop.

AI may ultimately prove the better use of these sites, but the return on the AI transition will depend on more than the revenue produced by each converted megawatt. It will also depend on the value recovered from the mining equipment, the cost of rebuilding the site, and how long lenders must be paid before the new capacity begins earning.

Hardware and Infrastructure News

• Oracle Shares Fall 4% on Reported Bid to Defer Payments on AI Project

• Applied Digital Confirms Brookwood, Alabama, as Site of Delta Forge 2 AI Factory

• Cipher Secures 20-Year Barber Lake Commitment Worth Over $9 Billion

• Lambda Plans Oklahoma Data Center, Projects $500 Million in Taxes

• Crusoe Details Google’s Texas AI Campus, With Links to 531 MW of Wind

• Bitdeer AI Secures $1.7 Billion in Off-Take Agreements for Malaysia Data Center

Corporate News

• MARA Posts $100M Deposit, Reworks Texas Deal Amid Data Center Audit

• Hut 8 Secures $1B Revolving Credit Facility to Support AI Infrastructure Growth

• NVIDIA Boosts Share Buyback $150 Billion Through 2028

• Soluna and Bitdeer increase capacity at Project Kati 1 to 35 MW

Financial News

• Nscale Secures $3.36 Billion in Pre-IPO Financing Led by Third Point

• Riot Repays Coinbase Loan, Ends $200 Million Bitcoin-Backed Facility

• Anthropic Eyes $2 Trillion IPO Valuation After $7.3 Billion Compute Spend

• OpenAI Seeks at Least $30 Billion at $1.4 Trillion Valuation

Feature

• The AI Build-Out Is Becoming the Biggest Economic Bet in U.S. History -WSJ

• Cracks Emerge in AI’s Debt-Fueled Data Center Boom - The Information

Email us

Texas Governor Greg Abbott has put another obstacle in the path of the data center boom President Donald Trump wants to accelerate.

On September 21, Abbott directed the Texas Commission on Environmental Quality to halt permits sought by data centers until the state’s grid operator completes an audit. His directive also said state agencies should not advance related regulatory approvals before obtaining the information needed to assess the projects.

“Simply put, Texans must come first,” Abbott said.

The move followed an August directive requiring scrutiny of projects seeking connections to the Texas grid and a September crackdown on failures to report water use. Developers face questions about their electricity demand, cooling systems, tax incentives, ownership, and effects on surrounding communities.

In essence, a Republican governor is using state approvals to impose conditions on infrastructure that a Republican president has made a national priority.

Trump’s administration has sought to ease federal permitting and make federal land available for AI development. Its argument is that building computing capacity strengthens American economic and national security. Abbott’s directives reflect another political calculation: the communities supplying the land, electricity, and water want assurances about what they receive in return.

That tension is dividing the coalition behind the AI buildout. It is also expanding a conflict bitcoin miners already know well.

The opposition increasingly draws on concerns that fit comfortably within conservative politics: property rights, local autonomy, household bills, and distrust of large corporations.

Molly McCann Sanders, president of the Eagle Forum Education & Legal Defense Fund, illustrated the divide in a recent WIRED interview. A conservative Trump supporter who describes herself as a “crunchy MAHA mom” said she disagreed with the president on data centers and wanted development to proceed more cautiously.

Polling suggests that division reaches well beyond individual activists. A September New York Times/Siena survey found that 61% of likely voters opposed data center construction. Republicans were almost evenly divided, with 49% supporting construction and 47% opposing it.

Republican campaign advisers are responding. In a September memo reported by Axios, Trump pollster Tony Fabrizio warned that “unqualified support for building more AI data centers is a losing position.” His recommended approach emphasized enforceable protections for voters, rather than leading with promises about jobs or competition with China.

Utah offers evidence of the electoral stakes. State Senate President Stuart Adams lost his June Republican primary after becoming associated with a controversial data center proposal in Box Elder County. Local reporting identified the project as a significant factor in his defeat and those of two county commissioners, although the results cannot be attributed to a single issue.

The conflict extends beyond the Republican Party. Across states, residents and officials are disputing who has the authority to approve projects—and who can challenge the terms.

In Port Washington, Wisconsin, voters approved an April referendum requiring public approval for future tax-increment financing districts with project costs or a base value of at least $10 million. The measure followed controversy over financing for Vantage’s Lighthouse data center campus and passed with 66.4% support.

It did not stop the already-approved campus. It changed the rules for future financing decisions, giving residents a direct role in commitments previously made by local officials.

In Michigan, Attorney General Dana Nessel is challenging the utility commission’s approval of electricity contracts for a 1.4-gigawatt data center in Saline Township. Her appeal seeks a contested hearing, questioning both the approval process and the evidence supporting it. Those arguments remain legal claims, but the dispute shows how data center development can divide institutions within the same state.

West Virginia has taken a different approach. Its 2025 framework for certified microgrids and high-impact data centers curtailed local authority over qualifying developments, including restrictions involving noise, lighting, and zoning. The policy sought to attract investment, while communities around Thomas and Davis pushed back over their diminished influence.

The debate also follows developers beyond the electricity grid.

In June, the Justice Department sought to intervene and dismiss a lawsuit over gas turbines powering xAI infrastructure in Mississippi. The department argued that state authorities had determined no permit was required and invoked the facility’s economic and military importance. Community and environmental groups allege violations of air-pollution requirements.

The implication is uncomfortable for developers. Building dedicated generation may address concerns about drawing power from the grid, while introducing disputes over emissions, noise, and the location of power plants.

But there is an important distinction between the intensity of this political conflict and its effect on construction.

SemiAnalysis’s recent research counted more than 300 adopted local moratoriums and bans. Yet it estimated that only 1.525 GW of capacity was actually delayed by local restrictions, out of roughly 20 GW located within affected boundaries.

Its explanation was practical: many projects already held the necessary approvals, lay outside the jurisdiction imposing a restriction, or would not need the affected permit until after a temporary pause expired. Including New York’s intervention, the firm estimated about 2.3 GW of delays.

Those are proprietary analyst estimates, and the research predates Abbott’s September 21 permit halt. Yet, it provides a useful counterweight to the impression that every new moratorium removes another project from the construction pipeline.

However, political resistance can change the economics of development even when construction continues.

Ohio’s dedicated AEP data center tariff requires large new customers to pay for at least 85% of contracted capacity, with commitments extending up to 12 years. The framework aims to protect other customers from infrastructure costs associated with demand that fails to materialize.

The same argument has reached Congress. On September 16, the House passed the Ratepayer Protection Act, H.R. 9340, by 417–3, demonstrating overwhelming bipartisan support for addressing the infrastructure costs associated with large electricity users.

The bill would require state utility regulators to consider standards under which large-load customers pay the full incremental cost of the generation, transmission, and distribution infrastructure needed to serve them. There’s a catch: requiring regulators to consider a standard does not require them to adopt it.

The measure stalled the following day when Republican Senator Jon Husted of Ohio sought passage by unanimous consent. Democratic Senator Martin Heinrich of New Mexico objected that the protections were too weak, arguing that Congress should require data centers to pay for grid connections rather than leave adoption of the proposed standards to states.

Heinrich proposed his GRID Savings Act instead, but Republican Senator Bernie Moreno of Ohio blocked that request. The competing objections prevented expedited passage of either proposal; they were not Senate votes defeating the bills.

Trump’s administration has separately promoted a Ratepayer Protection Pledge, calling for developers to fund the energy and infrastructure their facilities require. The congressional impasse exposes the unresolved question behind that promise: how much protection should depend on corporate commitments and state decisions, and how much should federal law compel?

For communities facing new projects, a nearly unanimous House vote has yet to produce an enacted federal safeguard. For developers, the debate adds uncertainty over who will ultimately bear the costs of connecting their facilities to the grid.

Regulation News

• Data Center Power-Cost Bill Stalls in Senate Over Enforcement Dispute

• Pushback against AI data centers emerges as central issue in Michigan elections

• Poway City Council unanimously advances AI data center ban

• Virginia Curbs State Support for Large Data Centers, Plans Tighter Rules

• Texas Governor Orders Data Center Permit Halt Pending Grid Review

Hardware and Infrastructure News

• CleanSpark Names Meta as Tenant Behind $6.6 Billion Data Center Lease

• Fluidstack Starts $4 Billion Texas AI Campus Phase Planned for 1.5 GW

• New Era Energy Signs 207 MW Vistra Power Deal for Texas Data Center

• Hut 8 Wins Poolin Auction With $140 Million Bid as AI Fuels Race for Power

Corporate News

• Anthropic, OpenAI hunt for smaller data center deals in race to deploy AI capacity

• GlobalFoundries, Marvell expand chip capacity deal for AI data center connectivity

• Nscale’s IPO Filing Reveals $103B in AI Contracts with Heavy Funding Needs

• Cypherpunk Names Amanda Fabiano to Board as Zcash Mining Commences

• Nscale founder Joshua Payne in line for near $500m IPO pay day

• TeraWulf CEO Moves to Monetize 3 Million Shares Tied to Land Lease

Financial News

• Crusoe Secures $3.9B Series F at $30.9B Valuation to Scale AI Factory Operations

• Oracle-Linked $18B AI Data Center Loans Come Under Pressure, FT Reports

• Big Tech uses guarantees to keep $300bn of AI exposure off balance sheets

• Fortitude Upsizes DCG Credit Facility to $50M, Plans $31M Draw in ZEC to Fund Hardware and Infrastructure Expansion

Email us

Anthropic CEO Dario Amodei’s weekend proposal to pace frontier model development drew support from Sam Altman and Elon Musk, putting the speed of AI advancement under scrutiny from the executives driving it. Amodei explicitly distinguished pacing capabilities from halting model training or technical progress.

Investors nevertheless reassessed the infrastructure trade. The AI Infrastructure Growth Index tracked by TheEnergyMag plunged to 5,774 on Tuesday, down 9% since the market’s close on Friday before Amodei’s remark.

For public miners and AI infrastructure peers, the financial consequences turn on a different clock. The industry has raised and spent substantial capital against expectations of future AI demand. Money already spent cannot be recovered simply by slowing the next project. Interest on outstanding borrowing continues to accrue under its contractual terms, even if the revenue expected to service it takes longer to arrive.

Fourteen companies tracked by TheEnergyMag generated $48.24 billion of net financing cash inflows in their latest six-month periods, exceeding $35.03 billion in the 2025 annual comparison periods. Net debt supplied $35.09 billion, or 72.7%, of that financing. Disclosed cash-interest payments reached $1.5 billion, despite gaps in reporting. The financing wave is concentrated. CoreWeave, Nebius, Hut 8 and IREN together accounted for roughly 76% of the net inflow.

A previous Miner Weekly issue examined the $30 billion capex already being committed to the buildout. Those expenditures have already occurred. Future projects can be deferred, but doing so does not erase existing debt or automatically release companies from signed construction and equipment commitments.

The carrying cost is already visible. CoreWeave paid $982 million in cash interest during the first half, including $176 million capitalized into assets. Excluding capitalized interest, payments more than doubled to $806 million from $362 million a year earlier. Applied Digital paid another $242.8 million during its latest six-month period.

Meanwhile, headline AI revenue does not always represent recurring income. Applied Digital’s $203 million in quarterly HPC Hosting revenue included just $44.1 million in base rent; most of the remainder came from tenant fit-out services. Construction-related revenue can be substantial before recurring rent reaches scale.

That distinction matters when assessing the cash available to service borrowing. Colocation providers need completed facilities to produce tenant receipts; cloud operators need installed GPUs to become billable usage and collections.

CNBC’s reporting this week put the potential consequences of an AI slowdown for the buildout under scrutiny. The financing risk can emerge before physical demand weakens: greater uncertainty about future cash flows could prompt lenders to demand more equity, stronger protections or higher returns on new funding.

The Federal Reserve has now added to the financing pressure, raising its benchmark target range by 25 basis points to 3.75%–4% on Wednesday. The unanimous decision cited elevated inflation and the need to return it to the central bank’s 2% goal. For AI infrastructure builders, the hike can increase costs on floating-rate debt as benchmarks reset and put pressure on new borrowing and refinancing, depending on market rates, loan terms and hedges.

A slower frontier does not necessarily mean declining compute demand. Existing models can keep attracting users, inference workloads can grow, and contracts can protect revenue. Future capex can also be deferred.

But with nearly three-quarters of this financing wave supplied by net debt, the industry has less room for revenue to arrive late. AI developers can debate the pace of advancement. Infrastructure owners still need to turn the capacity they have financed into cash.

Hardware and Infrastructure News

• Global AI Wins Colorado Data Center Permit Amid Power, Water Disputes

• Ethiopia Cuts Bitcoin Miners’ Power Supply as El Niño Hurts Dams

• Wood County residents rally against Jon Husted as $800 million Meta data center rises near Bowling Green

• Constellation agrees to purchase Rhode Island power plant from Shell for $715 million

• Fluidstack Gets Site-Work Permit for $4 Billion Texas AI Data Center

• Bitdeer Secures 65.1MW Malaysia AI Facility as Sales Pipeline Tops $7 Billion

• Hut 8 Files $399 Million Building Project at Beacon Point AI Campus

• Blockfusion Signs CoreWeave to 15-Year Niagara Falls Data Center Lease

• Bitdeer Expands AI Infrastructure with Bhutan Agreement and Increased Bitcoin Output

Corporate News

• Core Scientific Gets Conditional ERCOT Status for 1 GW at Texas Sites

• Nebius Sued Over Alabama AI Land Deal Allegedly Costing Regions $66 Million

• Sangha Texas Campus Gets ERCOT Base Load Status for 110.4 MW Plan

• Soluna Says Dorothy’s 100 MW Gets Conditional ERCOT Base-Load Status

• Cipher, Fluidstack and Anthropic Pledge $10M for Texas Town’s Water System

• MARA’s 2 GW Texas Project Gets Conditional ERCOT Study Status

• HIVE’s BUZZ Partners With ProCogia to Expand Canadian AI Services

• Cipher Details Conditional Texas Grid Status for 3.2 GW of Projects

Financial News

• Pentagon in Talks to Get Into AI Infrastructure Funding With a $5 Billion Loan

• AI Stocks Slide as Anthropic Chief Calls for Slower Development

• OpenAI Discusses Funding at $1.2 Trillion Valuation Ahead of IPO: Report

• Wall Street weighs prospect of an AI slowdown on data center buildout

• CleanSpark Plans $2.2 Billion Debt Sale for Georgia Data Center

• CoreWeave Plans $3B Debt Sale, Opens 35 Million-Share ATM Program

Feature

• In Virginia’s Data Center Hub, a County Fights to Reclaim Control

Email us

In 2021, they competed on mining machine preorders and exahashes promised. In 2024 and 2025, the contest shifted toward HPC contracts: who could sign a major customer and announce the biggest deal. Now the competition reaches further upstream, to the electricity those computing ambitions require.

Over the past week, a flurry of ERCOT Batch Zero announcements shows how a power arms race is playing out, with companies turning their standing in the Texas interconnection process into a public measure of competitive strength.

IREN highlighted conditional Base Load status for its 2 GW Sweetwater hub. Hut 8 flagged the designation for Beacon Point. Soluna announced it for the full 166 MW Kati campus. A technical grid classification has become something to put in a headline—a way to show that a company’s power plans have substance behind them.

But as in the hashrate race, the largest number does not tell the whole story. The classifications distinguish capacity that retains its position based on earlier studies from requests still awaiting allocation. Behind similar announcements, companies—and projects within the same company—can be at very different stages.

Got insights? Message TheEnergyMag

Galaxy’s September 8 disclosure illustrates the distinction. Five projects received conditional classifications covering approximately 4.2 GW. Of that, 1.63 GW at Helios I and II was classified as Base Load. Another 2.6 GW across Caspian, Selene and Helios III was classified as Studied Load, leaving it subject to capacity allocation.

One detail deserves particular attention: Galaxy submitted the 700 MW Caspian project as Base Load. ERCOT provisionally classified it as Studied Load. Even within one company, the headline gigawatts carry different levels of certainty.

The industry’s arms race has acquired a new scoreboard

In December 2021, MARA’s operating update offered a sharp distinction between current output and future capacity. It reported an active fleet producing approximately 3.5 EH/s, alongside purchases expected to bring its fleet to 23.3 EH/s by early 2023, which didn’t fully materialize until mid-2024.

The parallel today is the use of future capacity to establish competitive standing. Power always mattered to mining. What has changed is the prominence of grid access in the expansion story—and the possibility of using that access to serve a different computing market.

A machine order and a grid classification are different commitments, with different risks. Yet both invite the same analytical mistake: treating a milestone toward future production as though it were production already secured.

ERCOT’s predicament explains why a technical classification has become announcement-worthy.

When Texas regulators approved Batch Zero in June, ERCOT reported more than 438 GW of large-load connection requests, nearly 89% from data centers. By August 3, Governor Greg Abbott’s office cited more than 474 GW of requests, roughly 90% from data centers and more than five times ERCOT’s record peak demand. These figures measure proposed connections, not a forecast that every requested megawatt will materialize.

Studying projects individually becomes difficult when multiple campuses rely on the same transmission infrastructure. Batch Zero is the transitional group in ERCOT’s shift toward evaluating large loads together, assessing their combined effects and allocating available transmission capacity. The process generally concerns facilities of at least 75 MW with sufficient prior study work and development commitments.

Within that process, Base Load projects enter the study’s starting model without undergoing a new Batch Zero capacity allocation. Studied Load projects qualify for inclusion but still face evaluation and allocation. Projects that fail to qualify must proceed through a future interconnection process.

That makes Base Load a meaningful advantage. It also explains why “included in Batch Zero” is too broad a description for comparing developers.

CleanSpark, for example, disclosed conditional classifications for 585 MW as Base Load and 300 MW as Studied Load, a split it said was expected. Those numbers belong in separate columns when assessing its Texas portfolio.

Ionic Digital offered a more specific disclosure on Thursday. The company said its transmission provider, Texas-New Mexico Power, informed it that ERCOT had provisionally classified the remaining 466 MW requested at its Ward County campus as Base Load. Energizing that tranche would bring the campus’s total demand to 700 MW. Ionic continues to target completion by the end of 2027, subject to verification, the audit process and ERCOT’s final classification determination.

Cipher’s post on X, on the other hand, shows how much a headline number can leave unanswered. The company said initial feedback from its transmission service providers resulted in 3.2 GW of conditional Batch Zero classifications. It did not disclose how much was Base Load and how much was Studied Load, saying it was reviewing site-specific designations and would provide further details next week.

That missing breakdown is central to interpreting the announcement. Base Load capacity retains its position without a new Batch Zero allocation; Studied Load capacity remains subject to that exercise. Cipher’s 3.2 GW establishes the scale covered by its conditional classifications, but does not tell readers how much falls into each category.

The distinction also captures the dynamics of the new power arms race: companies can announce a large number before investors have the information needed to compare it with competitors’ portfolios. Cipher’s promised follow-up will help establish what sits behind the total.

The word “conditional” adds another layer.

Abbott’s August directive required verification and an audit of data centers advancing through the interconnection process. ERCOT subsequently issued provisional classifications. Its September 3 notice identifies possible conditions including successful verification, corrections to technical modeling data and regulatory exceptions where eligibility requirements were not met. Failure to satisfy an applicable condition can disqualify a project.

The companies’ disclosures do not consistently identify which conditions apply to each site. A project awaiting verification and one requiring an eligibility exception have different unresolved issues, even if both describe their status as conditional.

For readers following the mining industry’s AI expansion, the useful comparison is therefore project by project: capacity already operating, capacity preserved as Base Load, requested capacity still awaiting allocation, and the remaining steps before customers can use it.

The 2021 announcements asked investors to look forward from machine orders to future hashrate. Today’s power announcements ask them to look forward from grid access to future computing revenue. Batch Zero makes some of those paths clearer. The next test is how many of the headline megawatts become operating facilities—and when.

Regulation News

• Weld County officials will vote on whether to approve a controversial data center on Wednesday

Hardware and Infrastructure News

• Zcash Tops $1,000, But Rising Computing Power Blunts Mining Windfall

• IREN’s 2GW Sweetwater Hub Gets Conditional ERCOT Base Load Status

• Soluna Completes Kati 1, Gets Conditional ERCOT Status for 166 MW

• Galaxy’s Texas Expansion Faces Hurdle After Missing Base Load Status

• Ionic Gets Provisional ERCOT Status for 466-MW Texas Expansion

Corporate News

• Argo Blockchain CEO Justin Nolan Steps Down as Firm Shifts Toward AI Infrastructure

• HIVE Tops $1 Million in Daily Revenue, Names BUZZ Sales Chief

• TSMC and Samsung are committing to ASML’s most advanced chip-making machines for AI

• Keel Taps PowerSecure for Backup Power at Moses Lake Data Center

• Sphere 3D Prices $5M Financing, Sells Mining Assets to Fund AI Push

Feature

• Questionable data center forecasts are driving up Ohio power bills

• Pennsylvania residents appeal second data center approval, but developer claims it has ‘no merit’

Email us

Public miners analyzed by TheEnergyMag shed an estimated 56 EH/s of realized hashrate over the first half of 2026—a 15% contraction compared to the 10% drop across the Bitcoin network. Much of that power was not leaving compute altogether. It was being redirected toward the AI retrofit.

The change has appeared in the income statement. Directly reported HPC and AI revenue among the comparable miners rose 52% from the first quarter. For the companies furthest into the transition, Q2 also marked the first time that HPC colocation or AI-cloud revenue overtook the mining revenue being wound down.

The conversion is expensive. A previous Miner Weekly analysis tracked roughly $30 billion of capex by public miners and AI peers. PwC now estimates that the broader AI data-center buildout could require $31.6 trillion through 2050.

The 14 companies in TheEnergyMag’s latest comparison spent $18.6 billion in one aligned quarter. Among the six infrastructure providers already reporting recurring HPC revenue, companywide capex was nearly 15 times their combined period revenue.

That spending-revenue gap leads to the question at the center of this issue:

What is an AI megawatt-hour actually worth—and how does it compare with digital mining compute right now?

For the six comparable AI power shell providers, TheEnergyMag estimated their recurring HPC revenue ranged from approximately $86 to $300 per MWh. The median was about $180.

These figures divide recurring HPC, colocation or base-rent revenue by estimated billable megawatt-hours. Fit-out reimbursements and construction-related revenue are excluded where separable.

The estimates are sensitive to timing. A facility delivered halfway through a quarter cannot be treated as though it operated for all 91 days. Straight-line lease accounting may also recognize revenue before cash is collected.

Even with those limitations, the landlord-like models cluster in a relatively narrow range of approximately $140 to $200 per MWh.

The economics change when a company sells compute instead of powered space.

CoreWeave deployed $6.42 billion in the quarter. Based on its data center cost of revenue and estimated active power, we calculate that it was paying approximately $322 per MWh for hosting, with a range of $268 to $403 depending on the capacity-ramp assumption.

That cost includes more than the rent received by a data center landlord. It helps explain why a neocloud can spend more than $300 per MWh on infrastructure while a miner-turned-landlord initially recognizes only $150 to $200 per MWh.

Full-stack operators report still higher revenue. IREN produced an estimated $807 per MWh of AI-cloud revenue, HIVE approximately $924, WhiteFiber $958, and Bitdeer about $1,213.

Those figures are not comparable to rent. They include the value of GPUs, networking, software and orchestration. They also transfer utilization and hardware-obsolescence risk to the operator. Bitdeer demonstrates the distinction particularly well: its AI-cloud revenue density was the highest in the group, but the segment’s reported costs exceeded its revenue.

How it compares with mining

But still, AI cloud remains in a different revenue league. Its estimated median of $940.74/MWh is more than five times the $179.13/MWh generated by Bitmain’s latest-generation Antminer S23 Hyd. and more than eight times the $113.45/MWh from an S21 Pro.

The more interesting comparison is between bitcoin mining and HPC colocation. TheEnergyMag’s estimated median of $174.90/MWh is almost identical to the S23 Hyd.’s current mining revenue. But the underlying economics are very different: colocation revenue is generally contracted for years and may pass electricity costs through to the customer, while bitcoin revenue changes continuously with bitcoin’s price, network difficulty and transaction fees.

Zcash mining sits between those extremes. A Z15 Pro—rated at 840 KSol/s and 2.78 kW—currently generates an estimated $585.61/MWh. That is roughly three times the revenue density of an S23 Hyd., although it comes with substantially greater volatility. Just a little over a week ago, the revenue was sitting at over $700/MWh for Zcash mining.

That makes Zcash mining temporarily more revenue-dense than most colocation agreements and competitive with lower-end GPU-cloud revenue.

But the word “temporarily” is doing considerable work. A miner can install ASICs quickly and sell production into a liquid market, but the revenue can fall overnight. An HPC campus can require years of capital deployment and construction, but a creditworthy tenant may then provide contracted payments for a decade or longer.

Revenue per MWh explains why miners want AI tenants. Capex explains why only some of them will create attractive returns.

Regulation News

• Texas regulators approve two massive West Texas transmission lines amid outcry from landowners

Hardware and Infrastructure News

• Nvidia Backs MediaTek With $3.5B in Expanded AI Chip Pact

• Bitdeer AI Secures $800M in Offtake Commitments for Malaysia Data Center

• Anthropic’s $35B Lambda Deal Connects Nvidia to Hut 8’s Texas AI Data Center

• Bitdeer Acquires 200 Acres in Texas for $100 Million to Expand AI/HPC Infrastructure

• AI Data Centers to Draw $31.6 Trillion Through 2050, PwC Says

• Cipher Begins Pipeline Development for 2.5 GW On-Site Power Generation

Corporate News

• Cango Targets Bare-Metal and Colocation Revenue at 3 MW Site Following Mining Optimization

• Nscale Secures $3 Billion Debt Package for US AI Data Centers

• NVIDIA Reaches $11.9 Billion Deal to Acquire AI Platform Hugging Face

Financial News

• IREN AI Cloud ARR Hits $1B and Secures $2.8B in GPU Debt

• Bitzero Repays $22.4M Loan as Nordic AI Data Center Plans Advance

• Nvidia-Backed Nscale Touts $103 Billion Contract Book Ahead of IPO

Feature

• Swedish Tax Scrutiny Shadows Crypto Miners’ Pivot to AI

Email us

Bitcoin gained 22.49% over the month through early August 26, according to the TradingView comparison supplied by TheEnergyMag. Among the tracked equities, only Canaan, American Bitcoin and Cango did better, rising 66.99%, 53.71% and 40.96%, respectively. CoreWeave gained 21.45%, Nebius 17.10% and IREN 14.63%, while several miners that have emphasized AI and high-performance computing finished flat or lower.

The divergence is striking because capital had flowed the other way for much of 2026. Investors rewarded contracted AI capacity, power pipelines and data-center development while bitcoin struggled through a bear market. The latest rally reversed that hierarchy: direct exposure to bitcoin suddenly outperformed the promise of future AI revenue - at least for now.

The move began with a macro catalyst, but its speed came from positioning.

On August 19, the U.S. Treasury said it would at least double the maximum size of buybacks for older 10- to 30-year government securities, to $4 billion per operation from $2 billion, starting September 9. The stated aim was to support liquidity in the long end of the bond market. Yields initially fell, helping scarce assets including bitcoin and gold. A White House meeting with crypto executives the same day added a policy tailwind as President Donald Trump urged Congress to advance digital-asset market-structure legislation.

Bitcoin had spent weeks compressed near $62,000 to $65,000 while derivatives open interest climbed. When the catalyst arrived, that crowded setup unwound violently. More than $1.6 billion of crypto positions were liquidated over 24 hours, including more than $800 million of shorts in a single hour, according to Coin Metrics. Bitcoin rose as much as 23% during the week and reached its highest level in three months.

That makes the first leg of the rally easy to explain: it was a short squeeze. The more important question is why the move held initially.

The answer, so far, is spot demand. U.S. spot-bitcoin exchange-traded funds took in a net $1.92 billion from August 17 through August 21, their strongest week in roughly 10 months. Another $651.9 million arrived on August 24 and 25, bringing the seven-session total to about $2.57 billion, though the pace has notably decelerated.

The sub-$1 comeback

The three stocks that beat bitcoin were also three of the most stressed direct bitcoin exposures.

Canaan received a Nasdaq deficiency notice in January after staying below the exchange’s $1 minimum bid requirement. In July, Nasdaq transferred the listing to its Capital Market and granted the mining-machine maker and miner another 180 days, until January 11, 2027, to regain compliance.

Cango received an NYSE notice in March after its average closing price remained below $1 for 30 trading days. The company was given six months to cure the deficiency.

American Bitcoin did not receive the same formal sub-$1 deficiency notice. But the company preempted the move with a 1-for-15 reverse split in July.

Their rebound is therefore both a comeback and a warning about leverage. Low-priced mining stocks can move much faster than bitcoin because a higher bitcoin price can expand mining margins, lift the value of coin holdings and reduce immediate listing pressure at the same time.

What would confirm a cyclical bottom

One month of outperformance does not establish a paradigm shift. The coming weeks and months should provide a clearer test across several areas.

Spot demand must outlast the squeeze. ETF flows should remain positive on a multiweek basis after derivatives positioning normalizes. A rally led by spot buying is more durable than one driven by forced short covering. Renewed exchange inflows and repeated ETF outflow days would weaken the case.

Bitcoin needs to turn resistance into support. The $62,000-to-$65,000 area became a large on-chain cost-basis cluster during the summer. A constructive pattern would be a higher low above that base, followed by sustained trading above the August breakout zone. A rapid return into the old range would suggest the move was mainly mechanical.

On-chain profitability should recover without euphoric distribution. Watch realized capitalization, MVRV, adjusted spent-output profit ratio and the behavior of long-term holders. The healthier sequence is rising realized capital, spending at modest profits and enough new demand to absorb sales from older holders. A surge in coins moving to exchanges alongside extreme unrealized profits would raise late-cycle risk.

Leverage should rebuild slowly, if at all. Futures open interest reached about 220,000 bitcoin before the August squeeze. Funding rates, options skew and the futures-to-spot volume ratio should cool while price holds. If open interest races back to its prior high and perpetual funding becomes expensive, the rally will again be vulnerable to liquidation rather than supported by cash buyers.

For now, the evidence is stronger than it was during bitcoin’s failed rebound in May: ETF demand is greater, exchange balances are moving in a constructive direction and the market has reclaimed levels lost during the spring decline. But the rally began with an official bond-market intervention and a historic short squeeze. That is a catalyst, not yet proof of a cycle change.

Regulation News

• Kentucky Clears 482 MW AI Power Deal for TeraWulf’s Anthropic Campus

Hardware and Infrastructure News

• Zcash Rally Widens Mining Revenue Lead Over Bitcoin and AI HPC

• Bitdeer to Deploy 28 MW at Soluna’s Texas Wind-Powered Bitcoin Mine

• Anthropic Secures 460 MW at Nscale Campus in $45 Billion Deal: Report

• NVIDIA Discloses $105B Liability Cap for 4.25 GW Ohio Data Center Project

• AWS and NVIDIA to Deliver 2 Million Additional GPUs and Next-Generation Infrastructure for Agentic and Physical AI

Corporate News

• Tether’s $120M Uruguay Bitcoin Mining Bet Ends in Power Dispute

• Bitdeer Nears 9% of Bitcoin Block Rewards After Sixfold Share Gain

• Rescale Expands Cloud Ecosystem with CoreWeave for Engineering and AI Workloads

• SpaceXAI Adopts NVIDIA Vera CPU to Accelerate Agentic AI at Massive Scale

Financial News

• Nebius Group Closes $5.75 Billion Private Offering of Convertible Senior Notes

• NVIDIA Q2 Revenue Hits $96.2B as Data Center Surges 117% and Vera Rubin Ramps

• Core Scientific Adds $600M Credit Facilities as AI Buildout Accelerates

Feature

• The multiplying risks of financing data centres - FT

• Data Centers Are Driving an Alarming Gas Power Expansion in the US

• Iconic Bitcoin mine pivots to AI as industry turns back on crypto - BBC

Email us

A dozen public mining companies tracked by TheEnergyMag recorded $6.87 billion of net cash spending on capital assets in the first half of 2026, with IREN contributing only its March quarter because a June-period cash-flow statement was not available. That compares with $6.50 billion for all of 2025.

Adding Applied Digital, AI-native CoreWeave and Nebius to the basket, a total 15-company cohort spent $30.7 billion, which was 42.6% more than the $21.53 billion recorded for the entirety of 2025.

The calculation measures cash purchases and deposits for hardware, property, plant, equipment and other directly identified productive assets, less cash proceeds and refunds received from asset disposals. It excludes finance-lease principal and noncash equipment additions.

AI-native companies set the scale

CoreWeave and Nebius accounted for almost three-quarters of the 15-company total.

CoreWeave paid $14.12 billion for property and equipment, including capitalized internal-use software, during the six months ended June 30. That already exceeded its $10.31 billion expenditure for all of 2025.

Nebius spent $8.13 billion on property, equipment and intangible assets during the first half, net of $5.2 million in disposal proceeds. Its comparable 2025 spending was $4.07 billion. The company said the investment primarily covered GPUs, related hardware and data-center expansion for its AI cloud business. Nebius’s second-quarter results show that its first-half capital purchases were roughly double the entire prior-year figure reported in its Form 20-F.

Among the miners and former miners, TeraWulf spent $1.61 billion, Applied Digital $1.58 billion, Core Scientific $1.18 billion and Cipher $911.5 million.

This concentration illustrates what the AI pivot entails. Power contracts and available land may give miners a starting advantage, but converting those assets into AI-ready capacity requires substations, buildings, cooling systems, networking equipment and, in some business models, GPUs. Much of the cash leaves before the first corresponding dollar of revenue can be recognized.

Revenue is accelerating — from a much smaller base

The revenue side is beginning to respond.

Across nine miners present in both quarters of the supplied revenue schedule, directly reported HPC, AI cloud and colocation revenue increased 52% to $205.8 million in the second quarter from $135.4 million in the first. Including estimates for MARA and Hut 8, the comparable totals were $215.8 million and $140.3 million.

Core Scientific supplied most of the absolute increase. Its colocation revenue rose to $136.7 million from $77.5 million. TeraWulf’s HPC leasing revenue increased to $31.9 million from $21 million, while Bitdeer’s AI Cloud revenue climbed to $14 million from $3.7 million.

The acceleration nevertheless remains small beside the construction bill. Those nine comparable miners generated $341.2 million of directly reported HPC and AI revenue during the first half while spending $5.11 billion on capital assets — about 15 times the revenue.

The same pattern is visible at the AI-native companies, although their revenue bases are considerably larger.

CoreWeave’s revenue increased 24% to $2.58 billion in the second quarter from $2.08 billion in the first. Its $4.65 billion of first-half revenue was equivalent to about one-third of its $14.12 billion of cash capital spending.

Nebius recorded $582.3 million of second-quarter revenue, up 46% from $399 million in the first quarter. First-half revenue reached $981.3 million, compared with $8.13 billion of net capital spending.

These ratios are not conventional measures of project profitability. Capital spending builds assets intended to generate revenue over several years, while quarterly revenue only captures capacity that has been delivered, accepted and placed into service. Customer advances can also finance construction without immediately appearing as revenue.

Still, the gap captures the central risk of the transition: expenditure is immediate and largely irreversible, while revenue conversion depends on construction schedules, grid connections, customer acceptance and sustained demand.

For the rest of 2026, the key question is not whether AI-related revenue will grow. The filings already show that it is growing. The question is whether that revenue can begin catching up with the unprecedented amount of cash already committed to produce it.

Regulation News

• Pennsylvania Enacts Strict Data Center Mandates: Curtailment Priority, Ring-Fenced Grid Costs, and an End to Secret Deals

Hardware and Infrastructure News

• Arkansas AI Data Center Site Emerges With 200 MW Plan

• HIVE Signs $350M AI Cloud Deal for NVIDIA Blackwell GPUs

• NVIDIA, SB Energy Partner for 8 GW AI Data Center Campus in Ohio

• WhiteFiber to Buy Two North Carolina AI Data Centers for $60M

• Bitcoin production slips again in July for CleanSpark, BitFuFu and Canaan

• Cypherpunk Buys $33M Zcash Mining Fleet From Winklevoss Affiliates

• Bitdeer Secures $400M Malaysia AI Contract, Targets 350 MW by 2028

Corporate News

• Soluna Reports Q2 2026 Financial Results and Debt Prepayment Agreement

• Core Scientific Closes Polaris Deal to Expand Oklahoma AI Campus

• With AI Deals in Focus, Keel CEO Buys $200,000 of Stock on Dip

• Bitdeer Strategy Chief Buys $221,000 of Stock After 20% Slide

• Velaura AI Raises $110M After Auradine Rebrand, Bitcoin Miner Pivot

• Ionic Revenue Rises 31% as AI Leasing Replaces Bitcoin Mining

• Fortitude Reports $20.9M Revenue as Zcash Rally Boosts Mining Backdrop

Financial News

• Mike Alfred’s Alpine Fox Adds Soluna, Trims Cipher in Second Quarter

• Riot Secures $573 Million Loan Facility for Rockdale AI Data Center

• HIVE Reports $79M Fiscal Q1 Revenue as AI Cloud Services Scale

• Anthropic-Tied Data Center Inks $1.3 Billion Private Credit Loan

• Nebius Upsizes Convertible Note Sale to $5 Billion for AI Buildout

Feature

• Inside Big Tech’s Frantic Race to Quell the Growing Backlash to AI - WSJ

Email us

The analysis found a notable redistribution beneath an apparently stable aggregate: Core Scientific, IREN, Cipher Digital, TeraWulf and Keel Infrastructure were giving up bitcoin production, while Bitdeer, MARA, Riot Platforms and American Bitcoin absorbed much of the displaced network share.

At the time, the system remained roughly balanced.

The second quarter earnings suggest that balance is beginning to break.

The miners reducing bitcoin exposure continued to unplug machines, but fewer operators expanded enough to compensate. At the same time, colocation revenue rose sharply among the companies furthest along in their conversions.

Core Scientific generated $136.7 million of colocation revenue in Q2, almost five times its $27.5 million of bitcoin mining revenue. Colocation supplied 83% of quarterly sales, up from 67% in Q1, according to the company’s second-quarter results.

TeraWulf followed the same trajectory. HPC lease revenue increased to $31.9 million, or 71% of total revenue, compared with $12.8 million from bitcoin mining. Its HPC share had already reached 62% in Q1 as contracted capacity at Lake Mariner began generating revenue.

For several miners, colocation is becoming a material contributor. At Core Scientific and TeraWulf, it has already overtaken mining.

The rest of the sector remains earlier in the conversion. Riot Platforms reported $23.2 million in data center revenue against $113.7 million from mining. Bitdeer generated $14 million from AI cloud services compared with $197.1 million from mining-related activities. Hut 8 and MARA reported smaller compute contributions, while Cipher and Keel Infrastructure had yet to recognize HPC revenue.

The offsets are no longer keeping up

Using an expanded cohort and updated network series, TheEnergyMag estimates that the public miners tracked for this analysis produced a combined realized hashrate of 368.3 EH/s in Q4 2025, 344.4 EH/s in Q1 2026 and 319.0 EH/s in Q2.

That represents a 13.4% decline in six months.

The Bitcoin network’s quarterly average fell from 1,071 EH/s to 993 EH/s and then to 957 EH/s over the same period, a 10.6% reduction. The public company cohort therefore contracted faster than the network.

In Q1, expansion by a handful of miners largely obscured the scale of the shutdowns elsewhere. By Q2, even those additions were no longer sufficient to keep the cohort stable.

Bitdeer remained the largest offset. Its realized hashrate increased 44% from Q4 to Q2, reaching 63.0 EH/s. Excluding Bitdeer, the rest of the cohort’s realized hashrate fell 21.2%, from 324.6 EH/s to 255.9 EH/s.

Bitdeer’s expansion comes from its own SEALMINER production pipeline. By June, the company reported 73 EH/s of self-mining capacity and 15.9 EH/s of co-mining capacity. It produced 990 bitcoin during the month, 388% more than a year earlier.

MARA and American Bitcoin also continued expanding, but their additions were not enough to counter reductions across Cango, Cipher, Keel, Core Scientific, TeraWulf and IREN.

Cango provides the sharpest example of how quickly mining economics have changed.

The company entered bitcoin mining in late 2024 and reached 50 EH/s of deployed capacity during 2025, only to begin decommissioning inefficient machines, leasing out hashrate and moving capacity toward lower-cost regions.

Its realized hashrate fell from 44.8 EH/s in Q4 2025 to 31.3 EH/s in Q1. TheEnergyMag estimates its Q2 capacity would decline to 16.5 EH/s—a 63% reduction in six months.

Keel Infrastructure has taken the transition further. During Q2, it completed the decommissioning of all U.S. bitcoin mining operations in preparation for data center construction. Mining continues in Canada during the phased transition. Meanwhile, the replacement revenue has not yet arrived.

The post-China arms race unwinds

The longer-term context also looks more striking after Q2.

China’s 2021 mining ban briefly removed roughly half the Bitcoin network’s computing power. Hashrate bottomed at 57.5 EH/s in June 2021, but miners relocated and the network had almost fully recovered by December. The United States emerged as the largest mining hub.

That recovery helped trigger the institutional expansion race TheEnergyMag has tracked over the ensuing years. Public miners raised capital, acquired power sites and ordered successive generations of ASICs, eventually pushing the network beyond one zettahash per second.

Only one halving has occurred since that expansion began.

Now, machines and electrical infrastructure accumulated during the post-China race are being idled, impaired or depreciated faster so their power can be reassigned to GPUs. The industry collectively expanded at enormous cost, only for some of its most visible operators to begin dismantling that capacity after one halving cycle.

Unlike the China ban, the current contraction has no single headline trigger. It is a combination of weak mining economics and a competing use for capital and electricity.

Hardware and Infrastructure News

• Bitcoin ‘Anti-Spam’ Fork Sputters to a Halt After Mining Just Two Blocks

• Nvidia, Wall Street Firms Target $500 Billion for AI Infrastructure

• CenterPoint Ties $5B Savings Forecast to Texas Data Center Power Boom

Corporate News

• MARA Pledges 18,750 BTC for $600 Million New Bitcoin-Backed Loans

• Bitdeer Revenue Jumps 47% as Bitcoin Production Surges Nearly Fivefold

• Keel Falls 10% as Revenue Halves, Shuts Bitcoin Mines for AI Data Centers

• MARA Holdings Reports $611 Million Net Loss for Q2 2026

• CleanSpark Secures $6.6 Billion Lease and Reports Third Quarter 2026 Financials

• Bitdeer Slides 14% on $1B ATM Plan to Fund AI, HPC Expansion

• Riot Jumps 20% on Anthropic AI Deal as Bitcoin Sales Surge 56%

• CoreWeave Closes $2.6 Billion AI Loan at Wider Spread

• CoreWeave Reports $2.58 Billion Q2 2026 Revenue Amid Infrastructure Expansion

• HIVE Digital Delays Quarterly Report Over Swedish Tax Accrual

• WhiteFiber Shares Jump as Contract Exit Payment Lifts Revenue

• Nebius Revenue Jumps Fivefold as AI Cloud Deals Fuel Expansion

• Cerebras Systems Details Q2 2026 Results, Reaching 600 MW of Contracted Data Center Capacity

Email us

Sign a lease. Name-drop an AI tenant. Add a few hundred megawatts. Watch the chart go vertical.

That trade is not dead, but it is getting harder to impress the market.

TheEnergyMag’s analysis of 25 AI and high-performance-computing infrastructure announcements from June 2024 through Aug. 4, 2026 found an interesting split. The deals are arriving faster and carrying slightly richer revenue per megawatt. At the same time, the stocks announcing them are moving less.

The best-fit absolute announcement-day move—the size of the move, regardless of direction—falls from 32.5% at the start of the period to 6.3% at the end. The first eight announcements in the sample produced an average absolute close-to-close move of 24.1%. The median was 14.8%. For the last eight announcements, the average dropped to 10.2% and the median to 7.3%.

Meanwhile, annualized base-term revenue per contracted megawatt for colocation and build-to-suit leases inches up from roughly $1.67 million to $1.90 million.

In other words, the contracts got fatter while the candles got shorter. It is not to suggest that the market is shrugging these deals off. But it appears to be reacting with a polite nod instead of spraying champagne around the room like it used to.

Back when an AI lease could rewrite the whole story

The first wave of deals did something genuinely new. They showed that bitcoin miners and other power-rich operators could turn substations, land and grid connections into credible AI infrastructure businesses.

When Core Scientific announced roughly 200 MW of hosting contracts with CoreWeave in June 2024, the stock jumped 40.2% from the prior close. That agreement did more than add revenue. It gave investors a new way to value the entire company.

Applied Digital’s first 250 MW CoreWeave lease produced a 48.5% move in June 2025. Then TeraWulf’s initial 200-plus MW Fluidstack agreements sent its shares up 59.5% in August.

Those were not ordinary contract announcements. They were corporate identity changes delivered by press releases.

Fast-forward to this summer. TeraWulf signed a $19 billion, 401 MW lease with Anthropic. The stock gained 4.9% on a close-to-close basis. CleanSpark landed a $6.6 billion, 175 MW lease. Its shares rose 8.8%. Bitdeer’s new $4.7 billion, 121 MW Tydal agreement initially pumped its stock by 12% on Tuesday before all of those gains were erased at the market’s close.

These are still enormous contracts. A year earlier, any one of them might have blown the doors off the stock. Now the market wants to see the wiring diagram.

Everyone has an AI pipeline now

Scarcity creates excitement. Familiarity turns excitement into a spreadsheet.

TheEnergyMag’s sample contains just two announcements in 2024, followed by 12 in 2025 and another 11 in the first seven months and four days of 2026. This year’s deals already cover 2,039 MW—slightly more than the 2,001 MW announced during all of last year.

At this point, an AI strategy is no longer a surprise. For many of the miners and power-to-compute companies, it is practically part of the dress code.

That does not mean every announcement is equal. The fitted revenue-per-megawatt line rises over time, but only modestly, and the relationship is noisy. The median 2026 deal is actually below the 2025 median on that measure.

The strongest conclusion is therefore not that lease pricing is marching relentlessly higher. It is that the industry is signing more large deals—and investors have developed a longer checklist for deciding which ones deserve a rerating.

Who is the tenant? Is there an investment-grade backstop? Who pays for the build? When does revenue start? What happens if construction slips? And after debt, equity issuance and project costs, how much of the economics actually reaches each share?

A lease can validate a site while leaving all of those questions open.

The index has a hangover too

The TEM AI Infrastructure Growth Index closed at 6,384 on Aug. 5. It had bounced 23% over the previous seven days after hitting a local low near the end of July.

That sounds healthy until you zoom out. The index remained roughly 28.5% below its June peak and was still down 9% over one month.

That is not a market abandoning the AI infrastructure story. It is a market trying to decide how much of the story it already paid for.

The sector has traveled a long way from the days when access to power and an AI press release were enough. Investors now have a growing stack of signed leases to compare. They can separate scarce sites from ordinary ones, funded projects from financing plans and near-term rent from revenue that may not begin for another two years.

Two years ago, a major AI/HPC agreement could create an entirely new equity story overnight. Today, it is more likely to become another row in the model.

The megawatts are still growing. The easy multiple expansion may be what is running out of fuel.

Regulation News

• Texas pauses AI data center power hookups amid grid fears

Hardware and Infrastructure News

• Crusoe Plans $1B Texas AI Data Center Expansion, Pursues Nuclear Power

• Sphere 3D Plans 50 MW Kentucky Data Center as City Weighs Zoning Rules

• Nexus Seeks $15B for Anthropic Texas AI Data Center With Google Backstop

• DeepSeek Plans 1-Gigawatt AI Data Center in Inner Mongolia

• Bitdeer Signs $4.7 Billion Norway AI Data Center Lease With Volta

• Anthropic Deal Underpins Bitdeer’s $4.7 Billion Norway AI Lease

• CoreWeave Plans 360MW Indonesia Expansion in First Asia AI Data Center Push

Corporate News

• Southern Company Plans $2.15B Raise via Convertible Bond

• American Bitcoin Names Paul Sacks Interim CFO as Matthew Prusak Departs

• Matthew Prusak joins Giga Energy as Chief Business Officer and Interim CFO

• MARA Adds Power, Data Center Veterans to Board Amid AI Push

• NRG Energy Reports $7.48 Billion Q2 2026 Revenue, T.H. Wharton Peaker Plant Begins Operations

• IREN Completes Acquisition of Mirantis

• Nebius appoints Lindsey Irvine as Chief Marketing Officer

• CoreWeave Signs Multi-Year Agreement With Solidigm to Strengthen Its Integrated AI Cloud Platform

Financial News

• Aschenbrenner’s $24B AI Hedge Fund Unwinds Trades After Losses: Report

• Bitzero Secures $25M Through Warrant Issuance for Infrastructure and Debt Reduction

• TeraWulf’s AI Leasing Revenue Reaches 71% as Data Center Shift Accelerates

• Galaxy Digital Reports $85 Million Q2 2026 Net Loss and 526 MW CoreWeave Lease

• Fluence Energy Reports $44.3 Million Net Loss for Q3 Fiscal 2026

• Solaris Energy Infrastructure Reports $219 Million Revenue for Q2 2026

• Canaan Inc. to Repurchase Shares Using Portion of Digital Asset Treasury

Email us

The AI Infrastructure Growth Index tracked by TheEnergyMag fell 11.13% on Wednesday alone, extending its one-month decline to 34%. The index has now retreated about 42% from its record high in June, as selling spreads across chipmakers, neoclouds, power companies and former bitcoin miners recast as data center developers.

The immediate catalyst was a renewed debate over how the AI boom is being financed. Nvidia was reported this week to be considering a guarantee of up to $250 billion to help OpenAI lease a planned 10-gigawatt data-center campus in Ohio, along with separate financing of up to $350 billion for Nvidia chips used at the site.

The discussions revived concerns about "circular financing": chipmakers and hyperscalers invest in AI laboratories or cloud providers, which then spend some of that capital on equipment or computing capacity supplied by their investors.

Such arrangements do not by themselves show that demand is artificial. They can help customers finance infrastructure whose cost and construction timelines would otherwise make development difficult. But they also concentrate risk. If AI laboratories cannot generate enough revenue to honor their commitments, the pressure could flow back through neoclouds, data-center developers, lenders and ultimately chip suppliers.

Other pressures intensified the retreat. Reports of progress in China’s domestic chipmaking industry and the blockbuster debut of memory producer CXMT raised fears of greater competition. Oil prices and Treasury yields also jumped Wednesday, increasing financing costs for a sector whose projected revenue often sits years in the future.

Few examples capture the reversal as clearly as Situational Awareness, the hedge fund founded by former OpenAI employee Leopold Aschenbrenner.

The roughly $20 billion fund recently held discussions with existing investors and lenders about raising fresh capital after suffering losses in the AI-stock rout, the Financial Times reported, citing people briefed on the matter. It also offered some investors the opportunity to buy assets from its portfolio, although one source described the outreach as ad hoc rather than a coordinated fundraising effort.

Situational Awareness had gained 439% on a net basis during the first half of 2026, according to an investor letter reviewed by the FT. The fund used borrowing to magnify its exposure, a strategy that amplified returns during the rally but increased the potential damage when the trade reversed.

Aschenbrenner presented the downturn differently. In the July 24 letter, he argued that the selloff had created some of the most attractive opportunities since early 2025 and invited investors to add capital on Aug. 1. The fund’s fundraising can therefore be read both as a response to pressure and an attempt to buy into the decline.

For mining investors, the broader question is how far the unwind will extend into companies whose valuations increasingly depend on AI rather than bitcoin production.

A 30-day comparison through Wednesday shows bitcoin gaining about 6.84%, while every AI-infrastructure and HPC-related equity included in TheEnergyMag’s comparison declined. CleanSpark, the best performer among the displayed stocks, fell roughly 9%. Cipher Digital and MARA dropped about 16%, while IREN lost 25%, TeraWulf declined 31%, and American Bitcoin slid 46%.

The divergence is striking because mining stocks have historically traded as high-beta bitcoin proxies. That relationship weakened as investors began assigning greater value to miners’ electricity contracts, land, substations and grid connections—all assets that could be redeployed for AI computing.

Multibillion-dollar AI agreements then pushed some miners further away from their original business. Their shares became increasingly sensitive to hyperscaler spending, chip availability, construction financing and the valuations of AI laboratories and neocloud companies.

That shift worked in both directions. Only last week, mining and data-center stocks outperformed during a Nasdaq selloff as investors rewarded new AI contracts. The subsequent rout suggests that the group’s brief insulation from the technology market was less durable than it appeared.

Bitcoin’s relative strength raises the possibility that capital is beginning to rotate back from AI equities to the big orange. Earlier this summer, the movement appeared to run the other way. K33 Research said in June that investors were leaving bitcoin for higher-flying AI stocks, pointing to 62,794 bitcoin of outflows from exchange-traded products over three weeks, the second-largest such streak on record at the time. K33 warned that the opportunity cost of holding bitcoin had become difficult for investors to accept while AI-related equities were soaring.

The latest price action may indicate that dynamic is reversing, but relative performance alone does not establish a capital rotation. Investors selling AI stocks could be moving into cash, bonds or less volatile equities rather than bitcoin. For now, the evidence more clearly supports a decoupling: bitcoin has held its value while the companies that increasingly traded on AI expectations have not.

Mike Alfred, founder of Alpine Fox and a longtime investor in mining infrastructure, offered the countercase at the Energy Investors Forum last week.

“There’s no bubble, so you don’t have to ask,” Alfred said, arguing that demand for power, cooling, land and data-center capacity could expand for another 20 to 30 years. He acknowledged that the buildout would include crashes and downturns, but said those interruptions would not invalidate the longer-term demand for computing.

His exposure is material. Alfred is an IREN director and told the forum that roughly 80% of his fund was concentrated in IREN and Cipher. Both stocks are down around 30% over the past month.

The selloff does not resolve whether AI infrastructure is overbuilt. It does, however, raise the standard for what investors are willing to finance. Signed contracts, creditworthy tenants and funded construction plans are likely to matter more than undeveloped power capacity or prospective AI demand.

Bitcoin may be regaining some independence from the AI trade. For miners that spent the past year persuading investors they were no longer merely bitcoin companies, however, a rotation back to the cryptocurrency would be an awkward kind of victory.

Regulation News

• Russia detains crypto mining firm BitRiver founder on fraud charges

• Hopkinsville weighs limits on data centers as crypto mine proposes expansion

Hardware and Infrastructure News

• Nebius Details Plans for 1.2GW Data Center Campus in Pennsylvania

• NAVER, NVIDIA, and Brookfield to expand South Korea’s AI infrastructure to 200 MW by 2028

• Bitcoin Power Use Jumps 38% with Greener Energy Mix: Report

• Galaxy Digital Acquires 500-Acre Site for Second Texas AI Data Center

• Core Scientific Pays $42M to Exit Block’s Bitcoin Mining Deal as AI Revenue Surges

• NextEra, Brookfield Plan $100B AI Data Center Campus at Former Kentucky Nuclear Site

• EU lays out $11.4 billion for 7 AI gigafactories as it aims to catch up with US and China

• Dynamix-Led Fund Deploys $95M for Texas AI Data Center Grid Deposits

Corporate News

• Ionic Jumps 26% in Nasdaq Debut, Giving Celsius Claimholders Exit Route

• Core Scientific Lands $14 Billion AMD AI Data Center Deal

• Fortitude Invests $45M in Zcash Mining Infrastructure to Drive Vertical Integration

• Nscale Agrees to Buy Anyscale in Push Into AI Software

Financial News

• Bloom Energy Reports $1.06 Billion in Q2 2026 Revenue

• Bitzero Secures $25M Private Placement for AI Infrastructure and Debt Repayment

• Nvidia in Talks With OpenAI to Guarantee $250 Billion Financing for Data Center

• AI Hedge Fund Situational Awareness Seeks Capital Amid July Market Rout

EIF Recap

• The ‘Mullet’ Mining Pitch Has a Catch: Most Bitcoin Mines Aren’t AI-Ready

• AI Infrastructure Isn’t a Bubble, Mike Alfred Says — But Owning GPUs Is a Bigger Bet

• Soluna CEO Sees Stranded Wind Power as a Shortcut to AI Capacity

Email us

Across the EIF panels, generators, utilities, data-center developers, miners, investors and policymakers repeatedly returned to the same physical constraints: generation, interconnection queues, transmission, substations, cooling, equipment lead times, capital, permits and public support. The demand for compute may begin with models and chips, but delivering it ultimately requires electrons at a specific place, on a commercially useful timetable.

No single part of the energy economy can solve that problem alone. Utilities are being asked to accommodate loads at a scale and speed they were not built to handle. Renewable developers need customers for output that would otherwise be curtailed. Power producers need confidence that new demand will support investment in generation. Data-center operators need firm capacity and redundancy. Investors want contracts capable of supporting billions of dollars in infrastructure. Local officials and residents want credible answers about rates, water, noise, taxes and jobs.

Bitcoin mining entered that wider conversation as one of several bridges between energy and compute. Miners have spent years locating underused power, deploying modular loads and curtailing operations when grids become strained. That experience gives some operators a head start as AI developers search for energized land.

But EIF’s broader message was that controlling a grid connection is only the beginning. A parcel with power does not automatically become a financeable AI campus. It may still lack fiber, high-density cooling, transmission capacity, equipment, permits, a creditworthy tenant or community consent.

That distinction is beginning to sort the market. Some energy projects will support hyperscale AI campuses. Others may be better suited to distributed inference, flexible bitcoin mining, grid services or behind-the-meter industrial demand. Some will remain most valuable as power-and-land assets for another developer to build.

EIF therefore offered less a single forecast than a map of the emerging compute-energy stack — and of the bottlenecks that could determine who captures its value.

Mining’s place in the wider power shift

Alexander Neumüller of the Cambridge Centre for Alternative Finance previewed preliminary findings from the next edition of Cambridge’s digital-mining research. He said the underlying survey represented more than half of global bitcoin mining activity.

Among the preliminary findings presented at EIF:

• Estimated annual mining electricity consumption increased to about 190 terawatt-hours, from 138 TWh, between June 2024 and December 2025.

• Estimated emissions rose to roughly 48 million metric tons of CO₂ equivalent, from about 40 million.

• The estimated low-carbon share of mining’s electricity mix increased to *59.4%*, from 52.4%.

• About 10% of respondents said they had already allocated some power to AI or accelerated computing.

• More than 40% of the remaining respondents said they were actively exploring AI or HPC diversification.

• Nearly nine in 10 respondents expected AI/HPC diversification to become a strategic industry theme.

The presentation captured the industry’s ambivalence. Miners want another source of revenue, but many do not want it to destabilize the mining business that gave them their power portfolio in the first place.

The physical layer is the trade

Mike Alfred, founder and managing partner of Alpine Fox, used his EIF fireside chat to explain why he believes power-rich miners are being recast as AI-infrastructure investments.

"I think we have 20 or 30 really good years for building power, data centers and other capabilities and infrastructure for AI," Alfred said. "I think Texas is Mecca. I think Texas is the most important data-center market in the world."

His argument was that miners that secured land and electricity when those assets were undervalued can now use the same positions to pursue contracted data-center revenue. Referring to companies moving from mining into AI infrastructure, Alfred described the transition as going "from a speculative business on the commodity price of bitcoin to now a contractual, repeatable business based on AI."

Alfred also cautioned that the choice of business model will divide the sector’s winners and losers.

"The big divide in the data-center business amongst the folks who started in bitcoin mining and are now in AI is whether or not to own the GPUs yourself," he said. "If you choose to own them, you better have a good reason why. It’s potentially more lucrative, but also potentially more risky."

For miners seeking a less technology-dependent model, Alfred pointed to colocation: the tenant owns the GPUs, servers and racks, while the infrastructure company supplies power, cooling, connectivity and the building.

"That model looks more like a REIT," he said. "It looks more like a real-estate business. It’s much more conservative. It’s easier to finance."

Megawatts have become the strategic reserve

The clearest point of agreement at EIF was that "speed to power" now determines who can compete for AI projects.

John Belizaire, chief executive of Soluna Holdings, described a model built around locating computing facilities beside renewable-generation assets whose output is constrained or curtailed. Rather than entering a new interconnection queue from scratch, Soluna seeks to modify an existing generation interconnection and add data center load.

Belizaire said that approach can take a fraction of the time required to secure a new grid connection. Soluna’s larger thesis is that curtailed wind and solar generation represents "a source of energy that’s hiding in plain sight": electricity that may not reach the grid but can be monetized by colocated compute.

The value is not limited to electrons. A usable AI site needs transmission or distribution access, fiber, water or an alternative cooling design, permitting, equipment and a credible path to operation. That turns an existing interconnection into a form of strategic real estate.

Several panelists also argued that the next phase may be more distributed than the first wave of enormous AI campuses. Clusters of 10-to-20-megawatt facilities could sometimes connect faster, require less upfront capital and offer geographic redundancy. Smaller facilities may also sit closer to inference customers and avoid some of the transmission upgrades associated with gigawatt-scale campuses.

That model would look familiar to miners. The industry spent years deploying modular data centers wherever power was available rather than waiting for ideal locations.

The “mullet” meets operating reality

One panel reiterated the hybrid strategy with a memorable name: the mullet — AI in the front, bitcoin mining in the back.

The premise is straightforward. Mining can monetize a power position while a developer prepares the site for an AI tenant. If a higher-value lease is signed, the operator converts some or all of the location. Until then, the ASIC fleet generates revenue rather than leaving the interconnection idle.

The strategy has helped public mining companies persuade investors to value them less on hashprice and more on controlled megawatts. A panel moderator pointed to TeraWulf’s Kentucky development and a reported long-duration AI lease as evidence that a former mining site could become a major AI asset.

But the panelists disagreed on whether the mullet is a durable operating model or merely a bridge between businesses. Bitcoin mining and AI place different demands on infrastructure:

• Mining can tolerate interruptions and can often earn money by curtailing when the grid is strained.

• AI customers usually require firm power, tight service-level agreements and substantially greater redundancy.

• Mining containers can be deployed quickly and replaced in modules.

• AI facilities require expensive cooling, networking and electrical systems designed for rapidly increasing rack density.

• A mining site can operate in a remote power pocket with limited connectivity.

• AI workloads often need substantial fiber, skilled labor and access to customers or network hubs.

The practical conclusion was that only a portion of miners’ power portfolios will qualify for high-end AI development. The remainder still has value, but it may be worth more as flexible mining load, grid-balancing capacity or a future development option than as a promised hyperscale campus.

The constraint nobody can engineer away

EIF began and ended with a warning about community acceptance.

Data center projects are encountering questions about electricity prices, water, noise, local hiring, tax incentives and the number of permanent jobs they create. Speakers said developers often arrive too late — after residents have encountered a project through social media or a public-hearing notice.

Curtis Harris of Compass Mining argued that developers should show up in communities “early and often,” well before a formal hearing. “It’s not you, it’s us,” Harris said, directing the criticism at the industry. "We’re the ones who need to be better."

He cited a company outreach effort at an Iowa county fair that emphasized three claims residents could evaluate directly: the operation curtails when requested by the local utility, its mining machines use no water, and it hires local contractors and technicians.

Texas state Rep. Jared Patterson similarly argued that lobbying in Austin is not a substitute for local support. Industry participants need to explain the tax base, school funding, grid relationship and actual water design of a project in terms relevant to residents, he said.

The message was uncomfortable but consistent: the industry cannot assume that a technically sound project will receive political approval. Community engagement is not a communications exercise attached at the end. It is part of development.

That lesson applies equally to mining and AI. Residents may distinguish between the two technologies, but they often experience them in the same way — as large industrial loads seeking land, power and public consent.

What we took away from EIF

The forum did not produce a single blueprint for the mining industry’s next phase. It produced a set of tests.

Power is valuable, but only if it can be delivered on a commercially useful timetable. A mining site can become an AI site, but only if its cooling, connectivity, redundancy and location meet a much higher standard. AI contracts can stabilize revenue, but their financeability depends on the customer. Renewable generation can support compute, but intermittency and curtailment must be designed into the operating model. Texas can remain a compute center, but only if generation, transmission, permitting and local consent keep pace.

The miners best positioned for the next cycle may not be those that declare the largest AI pipeline. They may be the ones that classify their megawatts most honestly:

• which sites should remain mines;

• which can support flexible or distributed compute;

• which justify full AI conversion;

• and which are most valuable as power-and-land assets sold or leased to someone else.

That is a less dramatic story than every miner becoming an AI company. It is also a more investable one.

Hardware and Infrastructure News

• Braiins, Enel Launch Energy Tool as Bitcoin Mining Margins Tighten

• AMD, Anthropic Strike AI Chips Deal with up to 2 GW of MI450 GPU Deployments

Corporate News

• Galaxy Digital secures 15-year naming rights for Texas Tech football stadium

• Big Digital, 10NetZero Acquire Texas AI Data Center Site for $10 Million

• Ionic Digital Set to Begin Nasdaq Trading on July 28

• Ionic Digital Forecasts up to $195M in 2026 Revenue Ahead of Nasdaq Debut

• Bitdeer’s Bitcoin Output Climbs to 990 as AI Cloud Run Rate Reaches $76M

• Poolin Files Chapter 11, Lines Up $52M Sale of Texas Bitcoin Mining Assets

Financial News

• Nebius Secures $775M GPU-Backed Loan to Fund AI Cloud Expansion

• Hut 8 Jumps 14% on $9.8 Billion Lease to Fully Commercialize Texas AI Campus

• Greenidge Rebrands as Vulcan After $39.4M Raise to Fund AI Data Center Pivot

• Galaxy Targets $3.5B Notes Sale to Fund Texas AI Data Center Expansion

• Bitcoin Miners Defy Nasdaq Rout as AI Infrastructure Deals Fuel Gains

• Galaxy Locks In 9.875% Rate on $3.5B AI Data Center Bond Sale

Email us

Energy Investors Forum is a high-signal, curated conference convening leaders across energy, compute infrastructure, and capital to map the next decade of power and data center buildout.

Get your tickets today and use code MNRW30 for 30% off exclusively for Miner Weekly subscribers.

For much of the AI infrastructure boom, the industry’s power problem sounded like something that would arrive in the future.

Utilities warned about rising demand. Grid operators pointed to crowded connection queues. Developers talked about waiting years for transmission upgrades.

That future is beginning to arrive.

New York this week became the first US state to impose a statewide moratorium on certain new large data center projects, pausing incomplete state permit applications while regulators develop rules governing how the facilities should be powered, what they should pay for grid access and what benefits they must provide to host communities.

Also this week, PJM Interconnection, which operates the power grid across all or parts of 13 states and the District of Columbia, saw its latest capacity auction reach a federally approved price ceiling of $325 per megawatt-day. Even at that price, PJM still failed to secure enough power resources to meet its reliability target.

Across the Atlantic, Nscale’s planned £2 billion AI data center in Essex reportedly lost its expected grid-connection timeline, forcing the company to explore generating electricity on site.

Taken together, these developments show that the grid constraint is no longer just a risk hanging over the AI infra boom. It is beginning to act as a real brake on it.

New York wants more than jobs

New York Governor Kathy Hochul described the recent rush of data center proposals as an attempt to “flood the zone.”

Her concern is straightforward: the state must keep the lights on while also keeping electricity affordable for households and businesses.

A 50 MW data center, Hochul said in an interview on Bloomberg’s Odd Lots podcast, can consume roughly as much electricity as 50,000 homes. When several large projects arrive at once, they begin competing with residents, manufacturers and other businesses for the same limited power supply.

New York had nearly 12 GW of data center requests in its grid-connection queue as of May, including more than 8 GW added during 2025.

The state’s response is not intended to be a permanent ban, Hochul said. The one-year pause is meant to give regulators time to establish what she repeatedly called the “rules of the road.”

Her summary of those rules was blunt: Data centers should bring their own power or pay a premium to use New York’s grid.

Hochul said the state is considering several requirements. Developers could be asked to fund new generation, contribute toward transmission upgrades and pay into a larger grid-resiliency fund. They may also be expected to provide direct benefits to the towns and counties hosting their facilities.

That last point matters because many proposed data centers are located in smaller communities that do not have experienced planning departments or teams of lawyers negotiating on their behalf.

Hochul, who previously represented a rural part of upstate New York, said those communities may not know what developers are willing to offer.

She suggested that a locality could seek payments based on the size of a project — giving $1 million per megawatt as an example — alongside investments in the local grid and community.

The figure was not presented as a settled statewide fee. But it illustrates the amount of bargaining power Hochul believes host communities may possess.

The state’s goal is to create a standard framework that local governments can use, rather than leaving each town to negotiate alone with some of the world’s largest technology companies.

Who should pay for the grid?

The New York debate is also about who bears the financial risk of the AI buildout.

Utilities may need to spend billions of dollars on power lines, substations and new generation to serve proposed data centers. But developers often examine several locations before deciding which projects to build.

If a utility upgrades the grid for a campus that is later abandoned, ordinary electricity customers could be left paying for infrastructure that is no longer needed.

New York wants more of that risk to sit with the data center developer.

That could mean larger upfront deposits, dedicated power supplies or other financial guarantees before utilities begin major construction.

The policy would amount to a significant change from the early stages of the AI boom, when states and local governments competed to attract data centers with tax breaks, cheap land and promises of abundant electricity.

New York is still competing for technology investment. But it is making clear that access to the grid will not be free or unconditional.

Data centers must compete with factories

Hochul’s interview also revealed a broader political question: how should governments allocate scarce electricity among competing industries?

She contrasted data centers with Micron’s planned semiconductor manufacturing complex in New York, a project expected to create thousands of direct and indirect jobs.

If she had to choose between powering a “largely vacant data center” and a Micron factory employing thousands of people, Hochul said the choice would be obvious.

That comparison gets to the heart of the challenge facing data center developers.

Data centers can produce substantial construction spending, property-tax revenue and technology investment. But after construction is complete, the largest facilities may employ relatively few permanent workers compared with factories consuming a similar amount of electricity.

As power becomes harder to secure, developers may need to show that a project delivers more than electricity demand and a large building.

They may be asked to fund infrastructure, provide community payments, create training programs or supply their own generation.

Hochul even suggested that future data center operators could arrive with their own small modular nuclear reactors, with New York helping to identify suitable land.

The idea remains speculative, but it captures the direction of travel: the biggest new electricity users may increasingly be expected to bring power with them.

PJM shows the shortage is not theoretical

PJM’s latest capacity auction helps explain why states are becoming more cautious.

Capacity markets pay power plants and other resources to be available when electricity demand is highest. The payment is separate from the price customers pay for the electricity they actually consume.

PJM’s auction for the 2028-2029 delivery year reached its approved ceiling of $325 per megawatt-day. Without that cap, PJM estimated the regional price would have risen to about $555 per megawatt-day.

Even at the ceiling, PJM came up 6,831 MW short of the amount of capacity it says is needed to meet its reliability standard.

The price cap limited how high the auction could go. It did not create more power plants.

PJM’s demand forecast continues to rise, in large part because of new data centers. But building dependable generation and transmission takes far longer than announcing an AI campus.

Only 525 MW of new generation and upgrades cleared the auction, far less than the expected growth in demand.

The result shows that the grid bottleneck is no longer just about waiting in a connection queue. It is also beginning to appear in the cost of keeping enough power available for everyone already connected.

When a data center cannot get electricity

Nscale’s proposed AI campus in Loughton, Essex, shows what happens when the grid cannot match a developer’s schedule.

The company was reportedly told that electricity would not be available in time for the planned £2 billion facility to open next year. It is now considering alternatives, including fuel cells from Bloom Energy that could generate power on site using natural gas.

The project highlights a basic mismatch.

A company may be able to secure land, financing, customers and computing equipment within several years. The transmission infrastructure needed to power the campus can take much longer.

On-site generation can help developers avoid connection delays, but it creates new complications involving emissions, gas supply, environmental permits and community opposition.

The power problem does not disappear. It simply moves somewhere else.

What this means for bitcoin miners

The grid bottleneck strengthens the value of mining sites that already have power.

Keel Infrastructure announced this week that it plans to consolidate electricity from three existing bitcoin mines in Québec into a single 96 MW AI and high-performance computing campus in Sherbrooke. The company is not asking for additional power, only approval to transfer its existing allocation.

CleanSpark’s new 20-year lease in Sandersville, Georgia, shows another path. The company is transforming its long-established mining campus into a data center project expected to generate about $6.6 billion in contracted revenue.

Ionic Digital has already leased 234 MW of energized capacity at its former Ward County mining site to Nscale, although its plans to expand the campus further still depend on utility work and regulatory approval.

The examples point to the same advantage: bitcoin miners may already control the land, substations and grid connections that new data center developers are struggling to secure.

But existing power and future power should not be valued equally. A site that can reuse an established allocation, as Keel is proposing, is different from one that still depends on transmission upgrades or new generation.

As grid constraints begin to slow the AI buildout, the most valuable mining assets may be those that allow developers to avoid asking for new power altogether.

Regulation News

• First Statewide Moratorium on New Hyperscale Data Centers Launched by Governor Kathy Hochul

Hardware and Infrastructure News

• Solo Bitaxe Miner Beats 18,000-Year Odds to Mine Full Bitcoin Block

• Meta Expands Louisiana AI Data Center to 5 GW in $50 Billion Buildout

• Nscale’s £2 Billion UK AI Data Center Reportedly Faces Grid Delay

• CleanSpark Signs $6.6B Data Center Lease, Puts Texas Portfolio Under Exclusivity

• Keel Advances 96 MW Quebec AI Data Center With Sherbrooke Approval

Corporate News

• Sphere 3D Adds Tennessee Lobbying, IR Firms to Advance AI Infrastructure Strategy

• BitFuFu Reports June 2026 Bitcoin Production and Hashrate Metrics

• Galaxy Adds Former Xerox CEO to Board as AI Data Center Push Expands

• IREN Appoints Former Nutanix Executive Eric Hammersley as Security Chief

• Soluna Hires Microsoft Data Center Executive Ryan Carver to Lead AI Buildout

Email us

Energy Investors Forum is a high-signal, curated conference convening leaders across energy, compute infrastructure, and capital to map the next decade of power and data center buildout.

Get your tickets today and use code MNRW30 for 30% off exclusively for Miner Weekly subscribers.

For a brief stretch this year, the market seemed willing to believe that every megawatt with a fence around it could become an AI factory.

Bitcoin miners no longer needed to talk only about hashprice, fleet efficiency or the next difficulty adjustment. They could talk about campuses, lease terms, hyperscalers, neoclouds, inference workloads and “critical IT load.” The same substations that once fed racks of ASICs were reintroduced to investors as scarce energy gateways into the artificial intelligence boom. In a power-constrained market, that story worked.

It worked so well that a new question is starting to matter: who got liquidity while the story was working?

That question is moving to the foreground as the TEM AI Infrastructure Growth Index, a basket tracking bitcoin miners, neoclouds, power suppliers and other companies tied to the physical build-out of AI infrastructure, has fallen 16% over the past month. The pullback does not erase the long-term argument for energy-backed compute. Nor does it suggest that recent insider sales or shareholder trims were improper. Many of the transactions were disclosed as prearranged trades under Rule 10b5-1 plans, which are designed to let insiders sell stock according to instructions set in advance.

But market optics change quickly. A planned sale during a rally can look routine. A planned sale followed by a sectorwide drawdown starts to look like a liquidity window.

The recent tape has given investors several examples to digest. Core Scientific’s legal chief sold shares as the company’s AI data center narrative helped lift the stock. Riot Platforms’ chief executive disclosed a prearranged sale after the miner’s stock rebounded. Tether trimmed Bitdeer exposure after buying during an earlier selloff and selling into a recovery. TeraWulf disclosed a new batch of share sales by its chief shortly before one of the most consequential AI lease announcements in the sector.

And at IREN, the controversy is less about insiders selling than about insiders being paid. The company’s board approved more than 18 million restricted stock units for its co-founder co-CEOs, adding a governance and dilution debate to a stock that had become one of the most visible winners of the miner-to-AI pivot.

Together, these episodes mark a shift in the AI infrastructure trade. Investors are no longer only asking which companies have power. They are asking who captures the economics, who absorbs the dilution, who keeps upside exposure and who monetized the rerating before the trade cooled.

TeraWulf Enters the Spotlight

TeraWulf offers the most vivid case study because the company remains one of the sector’s clearest AI-infrastructure rerating stories.

On June 29, Beowulf E&D Holdings, an entity managed by TeraWulf Chairman and Chief Executive Officer Paul Prager, disclosed a sale of 275,000 TeraWulf shares at a weighted average price of $26.596 per share, generating about $7.3 million in gross proceeds. The sale came one week before TeraWulf announced its 20-year AI infrastructure lease with Anthropic.

That June transaction was part of a broader run of disclosed sales by Prager and Beowulf E&D Holdings since late March. In total, Prager and the entity he manages sold about 1.59 million TeraWulf shares for roughly $32.7 million in gross proceeds, implying a weighted average sale price of about $20.55 per share.

Then on July 6, TeraWulf announced a 20-year lease with Anthropic at its Justified Data campus in Hawesville, Kentucky. The lease is expected to generate about $19 billion of contracted revenue over its initial term and support about 401 MW of critical IT load. TeraWulf also agreed to sell its 50.1% interest in the Abernathy joint venture to a Fluidstack-led investor group, monetizing an investment valued at about $450 million and giving the company capital to redeploy into wholly owned AI infrastructure projects.

That is the sort of transaction investors have been waiting for from power-rich miners: a long-term AI customer, a large contracted revenue figure and an argument that legacy mining infrastructure can be upgraded into a higher-multiple asset base.

It is also the kind of moment that makes insider liquidity worth watching.

Cipher, Riot and Core Scientific Show the Same Pattern

Cipher Digital adds the most recent example to the liquidity-window theme.

On July 8, Cipher CEO Tyler Page filed to sell 112,500 CIFR shares with a market value of about $2.38 million, implying an average price of $21.19. The sale was tied to a Rule 10b5-1 trading plan adopted on Dec. 19, 2025. Cipher previously disclosed that Page’s plan covered potential sales of up to 1.5 million shares through Dec. 24, 2026. The 112,500-share notice was 7.5% of the total 1.5 million-share ceiling under the plan.

Riot Platforms had its own version of the story. In May, CEO Jason Les sold 175,000 shares valued at about $4.2 million under a Rule 10b5-1 plan adopted in August 2025. On June 22, he sold another 250,000 shares with a market value of $7.03 million.

Core Scientific has been another focal point for the AI-mining crossover trade. The company emerged from bankruptcy in 2024 and has since repositioned itself around high-density colocation and AI infrastructure, while continuing to report a decline in self-mining revenue.

Core Scientific’s chief legal and administrative officer Todd DuChene filed on July 6 to sell 140,000 shares with a market value of $3.0 million. The planned sale followed 12 prior 10,000-share disposals since April 13, bringing disclosed sales under the plan to about 260,000 shares and $5.9 million in gross proceeds.

These are important caveats. Rule 10b5-1 refers to prearranged trading plans designed to separate insider transactions from later corporate developments, and sales by executives with large equity holdings can reflect diversification, taxes or personal liquidity rather than a negative view of the company. It is not a confession of bearishness. Executives with large stock-heavy compensation packages often sell even when they remain optimistic about a company.

But public markets do not only process legality. They process alignment. When executives sell after a stock has rerated on AI expectations, and the sector then pulls back, investors start asking whether the balance of risk and reward has shifted from insiders to the public float.

The liquidity-window theme is not limited to executives.

Tether’s recent Bitdeer transactions show how strategic holders also used the AI-mining rebound to reduce exposure. As TheEnergyMag reported, Tether trimmed its Bitdeer positions at an average price of around $20 in early June, after buying into Bitdeer for $8.85 apiece during a market selloff earlier this year. While Tether remained one of Bitdeer’s largest shareholders, the trade still fits the pattern: buy into weakness, trim into the AI rerating, and retain a large enough position to keep participating if the story continues.

IREN Adds the Governance Layer

IREN brings a different but related issue into focus.

The company has become one of the most closely watched AI infrastructure names after moving beyond bitcoin mining and pursuing large-scale AI cloud and data center opportunities. But its latest compensation disclosure triggered a backlash among some retail investors and market commentators.

On June 30, IREN’s board approved grants of 9,099,328 restricted stock units each to co-CEOs William Roberts and Daniel Roberts. The awards are subject to a combined six-year vesting and holding period. The company said neither co-CEO will receive another equity incentive grant until fiscal 2031, and that the awards were designed to retain and incentivize the executives through IREN’s next phase of growth.

That explanation did not quiet the debate. Critics focused on the size of the package, its dilution and the fact that the company is still in the middle of proving that its AI infrastructure strategy can generate durable returns. IREN shares fell sharply as governance concerns met a broader selloff in AI-related stocks.

The IREN episode is not an insider-sale story. It is arguably more important: a debate over how much of the AI infrastructure upside founders and executives should receive before the business model has fully matured.

This is where the AI infrastructure trade begins to resemble other capital-intensive booms. The first phase of the rally was about scarcity. The next phase is about governance, capital discipline and execution. In that phase, disclosed insider sales, strategic-holder trims and large founder equity grants become part of the same story. They tell investors where the private incentives sit inside a public-market boom.

Hardware and Infrastructure News

• TeraWulf Surges on $19 Billion Anthropic AI Lease, $450 Million Fluidstack JV Exit

• Bitzero Secures 33-Hectare Land Reservation in Finland for 60MW Data Center

• Galaxy Digital Completes Phase I of Helios Data Center, Delivers 133 MW to CoreWeave

• Bitdeer Breaks Ground on Nevada Factory Despite Bitcoin Hashprice Squeeze

• MARA Buys Texas Site for AI, Bitcoin Mining in Deal Worth Up to $600M

Corporate News

• Keel Hires Digital Realty Veteran Ganesh Aiyer to Drive AI Infrastructure Growth

• Nscale, Nordkraft Form Venture to Run Narvik AI Data Center

Financial News

• Crusoe Reportedly in Talks to Raise $3 Billion at About $30 Billion Valuation

• Core Scientific Legal Chief Sells $5.9 Million Amid AI Rally Boost

• Nscale Secures $900 Million Credit Line to Fund AI Data Center Expansion

• Cipher CEO Files to Sell $2.4M of Stock as AI Rally Triggers Trading Plan

Feature

• EIF Speaker Series: “Bitcoin Is Becoming Increasingly Boring” with Lisa Hough

• This $13 Billion Georgia Family Is Helping Electrify Data Centers - Forbes

Email us

EIF Dallas (July 23) is a high-signal, curated conference convening leaders across energy, compute infrastructure, and capital to map the next decade of power and data center buildout.

Get your tickets today and use code MNRW30 for 30% off exclusively for Miner Weekly subscribers. You can also explore sponsorship opportunities to elevate your brand and connect with leaders across energy, infrastructure, and capital.

Bitcoin miners have long defended their power use with a simple argument: mining may be energy-intensive, but it is unusually interruptible.

A mining facility can curtail quickly when power prices spike or the grid is under stress. It can monetize surplus electricity when power is abundant, then shut down without disrupting households, factories or customer-facing software. That flexibility has become central to the industry’s pitch to utilities, regulators and investors, especially in markets such as Texas where demand response has become part of the mining business model.

But a new line of research raises a provocative question for the AI-infrastructure boom: what if GPU data centers can learn a version of the same trick?

A paper posted to arXiv in late June, titled “Power-Flexible AI Data Centers: A New Paradigm for Grid-Responsive Compute,” argues that AI data centers should not always be modeled as static, inflexible power loads. The authors describe a software architecture that combines grid signals, workload scheduling and power telemetry to reduce, defer or shift AI workloads in response to power-system conditions.

In a real-world deployment on a 130 kW GPU cluster, the paper says the system demonstrated rapid load reduction, sustained curtailment, carbon-aware operation and geographically distributed load shifting while preserving priority service levels.

The headline results were notable for the speed and range of the claimed response. The authors said the system reached 100% compliance across more than 200 National Grid power events in the UK test environment. In a simulated emergency dispatch modeled on a historical grid contingency, the cluster reduced power use by about 30% within 40 seconds. In another fast-response test, it cut load by 40% within roughly one minute. The paper also reported sustained curtailment of 10% to 40% for periods ranging from two to 10 hours, as well as a U.S. geo-shifting test in which 10% of live inference traffic was shifted from Oracle GPUs in Virginia to Illinois.

The paper is not yet peer reviewed. The arXiv version says it has been submitted to IEEE for possible publication, so its findings should be treated as early technical evidence rather than settled academic consensus. Still, the idea is not emerging in isolation. The June paper builds on a related Nature Energy article, “AI data centres as grid-interactive assets,” which reported a field demonstration of software-based demand response at an AI cluster in Phoenix, Arizona. A Nature Reviews Electrical Engineering summary of that work said the researchers showed that data centers can function as flexible, grid-aware loads by adjusting computing workloads in real time rather than relying on new hardware or on-site batteries.

The Phoenix trial is important because it moves the discussion beyond simulation.

According to the preprint version of the Nature Energy work, the test was conducted at an Oracle Phoenix Region Cloud data center on a 256-GPU cluster built with NVIDIA A100 Tensor Core GPUs. The system used Databricks MosaicML for workload orchestration, Weights & Biases for telemetry, and Amperon’s grid-demand forecasting tools. The project involved Emerald AI, NVIDIA, Salt River Project and the Electric Power Research Institute, and was conducted through EPRI’s DCFlex initiative.

The cluster ran representative AI workload mixes, including training, inference and fine-tuning jobs. The researchers classified jobs into flexibility tiers based on how much performance reduction users could tolerate: strict jobs with no performance reduction, and other tiers allowing average throughput reductions of up to 10%, 25% or 50% over a three- to six-hour period. That matters because the system did not simply shut everything off. It selectively slowed, paused or power-capped jobs depending on their tolerance for delay.

The reported result was a 25% reduction in cluster power usage for three hours during peak grid events while maintaining quality-of-service guarantees. The authors said the reduction was achieved without hardware retrofits or battery storage, using workload orchestration and GPU power-limiting techniques such as dynamic voltage and frequency scaling, combined with job pausing.

That does not mean an AI data center is suddenly as flexible as a bitcoin mine.

The Phoenix test involved a 256-GPU cluster, not a 100 MW or 1 GW AI campus. The later arXiv paper’s 130 kW deployment is also small relative to the scale now being discussed by hyperscalers, utilities and public miners. Scaling this approach from a cluster to an entire campus would require coordination across power delivery, cooling, networking, customer contracts and service-level agreements. It would also depend heavily on workload mix.

The researchers acknowledge that limitation. Not all AI workloads are temporally flexible. Batch training, fine-tuning and some inference jobs may be slowed or paused. But strict “Flex 0” jobs — such as real-time inference, streaming and model serving — were not modified in the Phoenix demonstration. Future flexibility may require geographic load shifting, where workloads move between regions to avoid grid stress without creating unacceptable latency or performance penalties.

Even with those caveats, the research cuts into one of bitcoin mining’s most important narratives.

For years, miners have argued that they are uniquely useful to power grids because they can behave less like a conventional industrial load and more like a controllable energy buyer.

AI data centers are more complicated. They are built around expensive GPUs, customer commitments, uptime expectations and high utilization targets. Operators do not want billions of dollars of accelerators sitting idle. But the research suggests the choice may not be binary. An AI facility may not need to run flat out or shut off entirely. It may be able to preserve high-priority workloads while throttling lower-priority jobs, delaying flexible training runs, power-capping GPUs or shifting traffic to another region.

That is the miner-relevant point. Yet there are still reasons for skepticism.

The research has commercial backers like Emerald AI, NVIDIA, Oracle and power-sector partners, with a clear interest in showing that AI infrastructure can be made more grid-friendly. The evidence remains early, and the field tests are far smaller than the campuses now being planned. Utilities will also need market rules, telemetry standards and compensation structures before AI data centers can become meaningful demand-response resources. Customer willingness is another open question: not every AI user will accept slower jobs in exchange for lower power costs or faster interconnection.

But the debate is worth watching.

Bitcoin miners helped popularize the idea that digital infrastructure could act as flexible load. Now AI data centers are beginning to move into that same communication, policy, and engineering conversation.

If the research scales, mining’s grid edge may not disappear, but it may become less unique. The flexibility pitch is no longer just a bitcoin mining story. It is becoming an AI-infrastructure story too.

Regulation News

• DOE Lets PJM Tap AI Data Center Backup Power During Heat Wave

Hardware and Infrastructure News

• Bitdeer Signs Conditional Lease for Norway AI Data Center Site

• Blockfusion Touts Up to $5.4B From AI Lease as SPAC Deal Advances

• Cargo thieves have set their sights on data center supplies

• Bloom, Brookfield Expand AI Power Financing Partnership to $25 Billion

Corporate News

• IREN, Keel Join Russell Indexes as AI Infrastructure Stocks Draw Broader Attention

• Ionic Digital Raises $400M Ahead of Nasdaq Listing as AI Revenue Overtakes Bitcoin Mining

• Sphere 3D Seeks DarkHorse Rebrand as It Pivots Toward AI Infrastructure

• IREN Adds Oracle, Google Veterans to AI Cloud Leadership Team

Financial News

• Applied Digital Expands Credit Facility to $430M as AI Data Center Buildout Continues

• American Bitcoin Sets Reverse Stock Split as Sub-$1 Nasdaq Risk Builds

Email us

EIF Dallas (July 23) is a high-signal, curated conference convening leaders across energy, compute infrastructure, and capital to map the next decade of power and data center buildout.

Get your tickets today and use code MNRW30 for 30% off exclusively for Miner Weekly subscribers. You can also explore sponsorship opportunities to elevate your brand and connect with leaders across energy, infrastructure, and capital.

There are moments in proof-of-work mining when the market briefly reminds everyone that electricity is not tied to a single digital asset, or even a business model.

The same megawatt can secure bitcoin, mine Zcash, chase Kaspa rewards or support an AI/HPC workload. Which one makes the most sense increasingly depends less on ideology than on a moving set of prices: token values, network difficulty, machine efficiency, hosting rates and power costs.

For the past year, public-market investors have largely rewarded the AI side of that equation. Bitcoin miners with large power portfolios have been repriced on the idea that their sites may be more valuable as data center infrastructure than as pure mining operations.

But Zcash’s rally, after years on the sidelines, has complicated that narrative.

The price appreciation of ZEC since late last year, followed by renewed hashrate momentum on the network, has made Equihash mining economics suddenly relevant again. In some cases, the revenue per unit of power now compares favorably not only with other proof-of-work assets, but also with estimated HPC colocation economics.

That backdrop helps explain why Fortitude Mining’s plan to go public is arriving at an interesting time.

Fortitude, a Zcash-focused digital asset mining platform under Digital Curreny Group, announced this week that it plans to list on Nasdaq through an all-stock merger with HeartSciences. The proposed company would operate under the Fortitude brand and is expected to trade under the ticker TUDE, subject to Nasdaq approval.

The transaction is still light on detailed financial disclosure. But Fortitude has released enough operating data to make the broader point clear: this is not just another crypto miner trying to attach itself to the AI trade. It is a test of whether public investors still have an appetite for a crypto-native compute business when the economics of a non-bitcoin proof-of-work network suddenly improve.

A Revenue-per-Megawatt Lens

TheEnergyMag’s latest analysis compares revenue across several proof-of-work mining rigs and a conservative estimated HPC colocation benchmark on a common basis: dollars per megawatt-hour.

The result is a useful snapshot of the compute market’s current hierarchy.

On the high end, a Zcash-focused Z15 Pro generates an estimated $373 per MWh, far above the $223 per MWh conservative estimate for HPC colocation. A next-generation bitcoin miner, the S23 Pro, comes in at about $133 per MWh, while the S21 Pro produces roughly $84 per MWh. By comparison, the DG1+ and Kaspa-focused KS5 and KS3 models generate about $65, $48 and $19 per MWh, respectively, under the assumptions used in the analysis.

That spread matters because it cuts through the usual framing of bitcoin mining versus AI infrastructure. The more precise comparison is not crypto versus AI. It is compute revenue versus power cost, with different machines and workloads competing for the same scarce input: electricity.

Over the past year, the dominant market narrative has been that bitcoin miners should redirect power toward AI and high-performance computing whenever possible. That logic is easy to understand. AI demand has created a scarcity premium for energized land, grid interconnection rights and large-scale power capacity.

But the revenue comparison shows the decision is not always straightforward.

HPC contracts may offer longer duration, lower volatility and stronger visibility than mining, provided that the business development and execution process is done well. That makes them attractive from a financing perspective, especially for public companies trying to reduce dependence on daily crypto-market swings.

Mining, however, retains optionality. When a proof-of-work asset rallies substantially faster than its network difficulty adjusts, mining revenue can temporarily exceed what infrastructure owners might earn from contracted HPC hosting.

Zcash is currently an example of that dynamic. Zcash has historically occupied a smaller niche than bitcoin, with less liquidity, fewer institutional narratives and a narrower base of industrial-scale miners. But the recent ZEC price rally has changed the revenue math for Equihash ASICs.

That does not mean Zcash mining is more valuable than HPC infrastructure on a risk-adjusted basis. Just search Zcash’s price crash and fast rebound earlier this month. Nonetheless, it perfectly shows that crypto-native compute can still produce bursts of exceptional returns, particularly in smaller networks where price and hashrate can move out of sync.

That creates both opportunity and risk for companies like Fortitude. A Zcash-focused platform may enjoy stronger near-term mining economics when ZEC outperforms, but it is also more exposed to the market volatility and the speed at which network difficulty can adjust and erode those gains.

Fortitude’s Timing

Fortitude’s public-market plan lands directly into that debate.

The company has not yet released detailed financials, leaving investors without a full view of its revenue, cost structure, power contracts or margins. But the operating metrics it has disclosed suggest it is not a small exploratory Zcash operation.

Fortitude said it began mining ZEC, the native token of the Zcash network, in 2019 and has since scaled production to an annualized run rate of 157,000 ZEC. Based on ZEC’s current price of about $400, that production rate implies a static annualized revenue of about $62 million.

Either way, the disclosed production rate gives Fortitude a meaningful operating base in a network that remains far smaller than bitcoin but has recently become more economically attractive for specialized miners. The timing is important: ZEC’s price appreciation has improved mining revenue for Equihash ASICs, while rising hashrate suggests more machines are being pointed at the network to capture those returns.

That dynamic gives Fortitude a more differentiated story than public bitcoin miners that are increasingly trying to convince investors their megawatts can be repurposed for AI. Fortitude is entering the market as a crypto-native infrastructure company with direct exposure to Zcash economics.

That identity may help it stand out, but it also narrows the questions investors will ask. Without detailed financials, the key issues will be how much of Fortitude’s revenue depends on current ZEC prices, how quickly network difficulty is rising, what its power costs look like, and whether its fleet can remain competitive as more Equihash capacity comes online.

In that sense, Fortitude’s planned listing could become a useful public-market test case for non-bitcoin proof-of-work exposure. If Zcash economics remain strong, the company may find investor appetite for a miner tied to a differentiated asset. If margins normalize quickly, the listing may instead reinforce how difficult it is for crypto-native miners to turn cyclical revenue spikes into durable infrastructure value.

Regulation News

• US to Fast-Track AI Data Centers as Power Bottleneck Becomes National Priority

• Texas Regulators Approve ERCOT Batch Process for AI Data Center Power Requests

• Thailand seeks to arrest fugitive Chinese businessman over illegal crypto mining

Hardware and Infrastructure News

• Microsoft to Build 2GW AI Data Center Campus in West Texas With Onsite Gas Power

• Constellation, Walmart Sign Nuclear Power Deal for Illinois Supply Chain Expansion

• Bitzero Adds Finland Land Reservation to Nordic AI Data Center Pipeline

• Galaxy Digital Advances $400 Million AI Data Center Plan in McGregor, Texas

• Sphere 3D Signs 30 MW Bitcoin Co-Mining Deal With Bitdeer

Corporate News

• Hut 8 Investors Reach $2.35 Million Settlement Over US Bitcoin Merger Claims

• Riot CEO Sells Another $7M as Bitcoin Miner Briefly Tops $30 on AI Rally

• DOE Announces $17.5 Billion Loan Commitment to Revive American Nuclear Supply Chain

• CoreWeave Expands Swedish Presence with Conapto Data Center Deal

• Bitcoin Miners Move to Defend Exchange Listings as Sub-$1 Shares Persist

Financial News

• DCG’s Fortitude Mining to Go Public Through HeartSciences Merger

• Cerebras Systems Reports Q1 2026 Revenue of $193.4 Million

• HIVE Lines Up Swedish AI Lease as It Seeks $100 Million Note Offering

Feature

• EIF Speaker Series: ‘AI’s Real Buildout Has Barely Begun’ with NG Zhang

• EIF Speaker Series: ‘Inference Could Shift the AI Compute Mix’ with Paul Golding

Email us