Grid Boiler #002 — Prove the megawatts

Good morning 👋
The AI buildout is getting less theoretical and more financial. Public investors are being asked to fund capacity years before it turns on, while grid operators reject paper demand. Today’s stories rhyme: prove the megawatts, prove the money, then build.

The headline numbers are enormous: roughly $439 billion of contracted backlog and 8.8 GW of data-center capacity contracted or in development. The operating reality is smaller. SB Energy has no operational data centers today. Its first-half 2026 revenue was $138.7 million, generated primarily by its power business, while it recorded a $3.21 billion net loss. The first data-center revenue is expected in the fourth quarter of 2026.

The financing web is just as notable. Nvidia has agreed to invest $1.5 billion through a private placement tied to the IPO, while OpenAI holds warrants valued at about $5.5 billion. OpenAI is also a major customer. At SB Energy’s planned PORTS-Pike campus in Ohio, OpenAI has signed a 20-year lease covering up to 8 GW of critical IT capacity, with Nvidia supplying the AI systems and providing credit support around the project.

Why it matters: public-market investors are being asked to underwrite construction risk, power risk and customer concentration before the data-center business has meaningful revenue. Backlog is not revenue, contracted capacity is not energized capacity, and a strategic customer can also be an investor. The valuation question is whether long-dated, concentrated backlog deserves anything close to the value of operating infrastructure.

What to watch: final IPO pricing, first rent commencement, construction milestones, financing for the Ohio build and whether SB Energy can diversify beyond OpenAI and SoftBank. The registration filing itself does not yet establish the number of IPO shares or the offering price range.

The evidence around “ghost demand” is getting harder to ignore. Reuters reports U.S. data-center power requests above 700 GW—more than ten times estimated current U.S. data-center demand. Texas is auditing projects before they advance through ERCOT; the governor’s office says the queue had grown beyond 474 GW, with roughly 90% tied to data centers.

That does not mean 474 GW will be built. Duplicate sites, speculative requests and underfunded projects can distort transmission planning and obscure real demand.

Why it matters: power access is becoming a credibility test. Developers increasingly need deposits, site control, financing and technical readiness—not just a large MW request—to earn a place in the queue.

Fervo Energy signed a 396 MW power-purchase agreement with Google for its Cape Station enhanced-geothermal project in Utah, expected online in 2028. Google also gets an option for roughly 600 MW more, potentially taking offtake near 1 GW by June 2030.

The distinction matters: 396 MW is contracted; 600 MW is optional. The Google data center this power could support remains potential, dependent on engineering, approvals and commercial conditions.

Why it matters: hyperscalers are moving upstream into firm power procurement because servers are useless without electricity. Enhanced geothermal aims to provide around-the-clock output without wind or solar intermittency. Now Fervo has to execute on drilling, cost, transmission and schedule.

Anthropic has agreed to a reported $35 billion cloud-computing deal with Nvidia-backed Lambda for a roughly 350 MW Texas project, Reuters says, citing a source familiar with the matter. Hut 8 is developing the Nueces County site; the detailed arrangement was not publicly confirmed by the companies when researched.

Hut 8 separately disclosed two 352 MW IT leases at Beacon Point, totaling 704 MW. Reporting identifies Nvidia as the tenant. Do not add Anthropic’s 350 MW to that 704 MW as separate capacity; it appears to sit within the same contractual stack.

Why it matters: Nvidia is increasingly organizing not just chips, but leases and customer access to capacity. That can accelerate deployment while making counterparty exposure harder to untangle.

SLB agreed to acquire thermal-management specialist Kelvion for $3.4 billion in cash and assume about $700 million of debt, for roughly $4.1 billion total. The deal is signed, not closed; completion is expected in the first half of 2027, subject to approvals.

Kelvion expects data centers to contribute $1.2–$1.3 billion of 2026 revenue. SLB says the combined data-center business could reach $4.5–$5 billion in 2028.

Why it matters: higher-density AI racks turn heat removal into core infrastructure. An oilfield-services company paying billions for cooling shows how AI capex is pulling industrial engineering and thermal management into the same supply chain as chips.

Cerebras and Compute Nordic Finland announced a Mikkeli AI data center designed to scale to 165 MW of contracted IT capacity. The status detail: only the initial 50 MW phase is under construction. Seven-year service orders back the capacity; a study cited by the companies estimates €1.0–€1.7 billion of investment at full scale.

The campus is designed for closed-loop cooling and waste-heat recovery, expanding from 50 MW to 80 MW to 165 MW.

Why it matters: Cerebras is building physical capacity around its wafer-scale architecture. More importantly, this is exactly where “contracted” and “under construction” must not become synonyms.

Brazil’s Senate approved legislation creating the REDATA tax regime for data centers and sent it for presidential sanction. The proposal suspends several federal taxes on qualifying data-center equipment.

Brazil’s grid planners already face a large paper pipeline. EPE said projects seeking grid studies reached 26.2 GW by November 2025, while warning that realization depends on multiple conditions.

Why it matters: data-center geography increasingly comes down to tax policy plus power availability. REDATA could improve the economics of imported servers, networking and cooling gear, but tax incentives do not create substations or transmission. It may attract serious builds; it may also swell the queue before projects are financeable.

In plain English:
power demand in utility or grid-connection queues that may never become real load because projects are duplicated, speculative, unfunded or not technically ready. A developer may pursue several sites while intending to build only one.

Why you’re hearing about it:
Texas is auditing data-center requests, while utilities elsewhere use deposits and stricter readiness tests.

Why it matters:
Bbad queue data can trigger expensive grid planning and make genuinely buildable projects harder to identify.

🤯 Did You Know? AI infrastructure is moving from “announce the capacity” to “prove the capacity”—with capital, power contracts and construction milestones becoming the real scoreboard.

Till next time,
Grid Boiler Team