The bottleneck for AI isn't chips anymore — it's gigawatts. And the power deals are now in SEC filings

On June 10, Oracle reported one of the most extreme backlog figures the cloud business has ever produced — and the stock fell roughly 10% anyway. That gap between the number and the reaction is the whole story of where the AI build-out is right now: the demand is no longer in doubt. The constraint has moved. For the last two years the binding input was GPUs. In 2026 it is increasingly power — and, unlike the GPU shortage, the power side is producing a paper trail you can read in SEC filings.

This post does what we always do: takes a hot topic and grounds it in primary documents, with explicit labels for what's confirmed versus inferred.

(Financial research, not investment advice. Built from Aether search over SEC filings and earnings exhibits; evidence labels at the end.)

What Oracle's filing actually says

Start with the document, not the headline. From Oracle's FY2026 Q4 earnings exhibit (8-K, SEC source):

Remaining Performance Obligations (RPO) of $638 billion, up 363% year-over-year and up $85 billion sequentially.
FY2027 revenue guidance confirmed at $90 billion, with non-GAAP EPS guided to $8.05.
Record FY2026 operating cash flow of $32.0 billion (up 54%) — against free cash flow of negative $23.7 billion, "as Oracle continued to execute on investments to support the growth of its Cloud Infrastructure business."
A funding wrinkle the market is fixated on: "Most of the RPO increase in both Q3 and Q4 were large scale AI contracts where the customer prepaid Oracle for the purchase of the GPUs, or the customer bought and supplied the GPUs to Oracle. The prepaid and customer supplied hardware portions of our large AI contracts now total $75 billion."

Two things in that filing point straight at power. First, in the same release Oracle states it "is building datacenters that are intended to use clean energy from natural gas fuel cells to generate electricity with minimal emissions." A software company explaining its fuel-cell generation strategy in an earnings release is a tell: power sourcing has become a board-level constraint, not a facilities footnote. Second, Oracle's own risk factors name "our ability to anticipate, plan for, secure and manage datacenter capacity" as a principal risk (ORCL Q3 FY2026 exhibit, SEC source). Datacenter capacity, in 2026, means megawatts before it means square feet.

Evidence label: SEC-filed for every figure and quote above. The interpretation — that the ~10% drop reflects funding/power-execution doubt rather than demand doubt — is inferred (the $638B backlog is the opposite of a demand problem).

The macro constraint, in one paragraph

The framing context, from public research rather than filings: U.S. data-center electricity demand has climbed from roughly 23 GW (2023) toward ~42 GW (2026), grid interconnection queues can run 4–7 years, and a modern AI rack draws 50–100 kW versus 5–10 kW for a traditional rack — so the same floor space needs an order of magnitude more power. Goldman Sachs and others have flagged a multi-gigawatt U.S. capacity shortfall widening through 2028. We label that research-cited, not filing-cited. But here's the point of this post: you no longer have to take the macro number on faith, because the individual deals are now being filed.

The power deals are in the filings — with named gigawatts

This is where the AI-power story stops being narrative and becomes evidence. Searching earnings exhibits and filings surfaces specific, dated, quantified power-offtake commitments:

DTE Energy — Oracle, 1.4 GW (filed, regulator-approved). DTE's earnings exhibit describes its "first hyperscale data center contracts to provide 1.4 gigawatts of load to power Oracle's new data center in Saline Township, MI," with contracts "approved by the Michigan Public Service Commission" (DTE exhibit, SEC source). Critically, the structure is spelled out: the contracts ensure "revenues from Oracle cover the costs associated with powering the data center, so DTE's existing customers will not subsidize data center rates." A separate DTE deck details a 19-year power supply agreement with minimum monthly charges plus a 15-year energy-storage contract (SEC source).

DTE Energy — Google, ~1 GW (filed). In March, DTE submitted contracts to the MPSC to support Google's planned "1 gigawatt (GW) data center in Van Buren Township," under which "Google will pay the full cost of its energy usage, including all expenses related to new generation, storage, transmission, and distribution" — projected to generate "nearly $1.7 billion in affordability benefit over the life of the contract" for existing customers (DTE Q1 FY2026 exhibit, SEC source).

Amazon — gigawatts on both sides of the meter (filed). Amazon's Q1 FY2026 release states it "added 3.8 gigawatts of power capacity in the past 12 months — more than any other cloud provider," and quantifies demand: a commitment from OpenAI to consume "approximately two gigawatts (GW) of Trainium capacity," and Anthropic securing "up to five gigawatts (GW)" of Amazon's chips (AMZN exhibits, SEC source). When a hyperscaler reports added power capacity as a headline operating metric alongside revenue, the constraint has officially moved from chips to electrons.

Talen Energy — the nuclear baseload (filed). Talen describes itself as owning "approximately 10.3 gigawatts of power infrastructure… including 2.2 gigawatts of nuclear power," and states plainly it is "powering the digital infrastructure revolution… as artificial intelligence data centers increasingly demand more reliable, clean power" (TLN Q3 FY2025 exhibit, SEC source).

FuelCell Energy — the gas/fuel-cell answer (filed). FuelCell Energy disclosed an LOI with SDCL "targeting up to 450 MW of potential data center and AI-driven power projects" (FCEL Q1 FY2026 exhibit, SEC source) — the same fuel-cell route Oracle named for its own datacenters.

Oklo — advanced nuclear (SMR) offtake (filed). Oklo ($OKLO) reports a 14 GW customer pipeline for its build-own-operate small modular reactors, with named data-center deals including a "$25M pre-payment for a 20-year PPA, providing up to 500 MW of power to Equinix data centers" and phased campuses scaling "Phase 1: 250 MW → Phase 2: 500 MW → Phase 3: >1,000 MW" (OKLO exhibit, SEC source). The pitch is the cleanest statement of why power, not chips, is now the gating input: "Technology companies don't need to invest in energy operations and can instead simply buy power."

Constellation Energy — nuclear + "powered land" (filed). Constellation ($CEG) signed a 380 MW agreement with CyrusOne at its Freestone, TX site plus an exclusive additional 380 MW Phase 2, and frames "powered land" — shovel-ready, co-located data-center sites — as a repeatable template worth "a $0.20–$0.50 EPS impact per 1,000 MWs" (CEG Q1 FY2026 exhibit, SEC source). The same filings cite utility peer AEP seeing "commitments from customers for more than 15 gigawatts of incremental load by the end of this decade" — the demand showing up across the whole regulated-power complex, not one name.

Vistra — nuclear PPAs to the hyperscalers (filed). Vistra ($VST) signed a 20-year PPA with AWS for up to 1,200 MW of carbon-free power at its Comanche Peak Nuclear Plant, plus "long-term power purchase agreements with Meta at our PJM nuclear sites" (VST Q4 FY2025 exhibit, SEC source) — named hyperscalers underwriting existing nuclear baseload for decades.

GE Vernova — the supply side, in backlog (filed). The clearest single proof that power is the constraint is the equipment order book: GE Vernova ($GEV) now expects "at least 110 GW of combined gas-turbine backlog and slot-reservation agreements by year-end 2026," and booked $2.4 billion of Electrification equipment orders "to support data centers — more than all of last year" (GEV Q1 FY2026 exhibit, SEC source). When turbine slots are reserved years out, the bottleneck isn't silicon.

Bloom Energy — onsite fuel cells (filed). Bloom ($BE) has deployed "1.5 GW of low-carbon power across more than 1,200 installations" and is powering AI data centers directly (an 80 MW fuel-cell project; 20 MW for AWS in Santa Clara), with its CEO stating "demand for electricity is expected to far exceed available supply through the grid" (BE Q2 FY2025 exhibit, SEC source) — the same onsite-generation answer Oracle and FuelCell named.

Evidence label: SEC-filed for every gigawatt figure, counterparty and contract term above. Note the dates: these are point-in-time disclosures, and the deals continue to evolve — re-check the latest exhibit before quoting.

The mechanism nobody is pricing: who pays for the grid

The most underappreciated detail in those filings isn't the gigawatts — it's the cost-allocation structure. DTE's contracts are explicitly written so that the data-center customer "covers the costs associated with powering the data center" and existing ratepayers "will not subsidize data center rates," with minimum monthly charges over 15–19 years. That design choice is doing two things at once:

It de-risks the utility's growth. A 1.4 GW, 19-year take-or-pay-style contract with minimum charges is among the most bankable load a regulated utility can add — and it spreads fixed grid costs across a bigger base, which is why DTE frames it as an affordability benefit to existing customers.
It shifts the cost risk onto the hyperscaler/tenant. The customer underwrites new generation, storage, transmission and distribution. That's a real, contractual, multi-year obligation that lands on the demand side of the build-out — the same side Oracle's negative free cash flow and $90B+ capex plan already represent.

So the AI-power trade has two distinct, filing-visible profiles: regulated utilities and independent power producers signing long-dated, cost-covered load (lower-variance, contracted), versus the AI buyers absorbing escalating energy obligations on top of GPU capex (higher-variance, capital-hungry). One catalyst — the power constraint — splits into two very different income-statement stories. That distinction is the whole point of reading the filing instead of the headline.

Evidence label: SEC-filed for the cost-allocation language and contract tenors; the relative risk framing is inferred from the disclosed terms.

How to research it yourself

This is a follow-the-filings thesis, not a price target. What to do with it:

Trace the dependency. Use the EvidInvest supply-chain map to see the AI build-out from the power side — the power-delivery layer we recently added (SiC substrate → power semis → the rack) sits downstream of the generation deals in this post. Together they're the full electron path from grid to GPU.
Read the utility/IPP exhibits, not just the tech ones. The most concrete AI-demand evidence in 2026 is showing up in the filings of DTE, Talen, and their peers — counterparties naming the hyperscaler and the gigawatts.
Watch the cost-allocation clause. Whether the data-center customer fully covers new generation/transmission is the difference between a clean utility growth story and a future rate-case fight.
Track the AI buyer's free cash flow. Oracle's −$23.7B FCF and the $75B of customer-prepaid/supplied hardware are the demand side's way of saying the build-out is bigger than one balance sheet.

Search the evidence directly with Aether: try data center load gigawatts power supply agreement or OpenAI Anthropic gigawatts capacity to pull these disclosures yourself, then run the financials on the names — ORCL, TLN, AMZN, DTE — on EvidInvest, free.

Bottom line

The AI demand question is settled — a $638B backlog settles it. The open questions are physical and financial: where the gigawatts come from, who signs the long-dated offtake, and who carries the cost of building new generation. For the first time, those questions have filing-grade answers — named utilities, named hyperscalers, named gigawatts, and contract structures written down to the year. The chip shortage was hard to see until it hit the income statement. The power constraint is already on the page. Read it there.

Evidence labels: SEC-filed — Oracle FY2026 Q4 RPO/FCF/fuel-cell/guidance (8-K exhibit 2026-06-10) and datacenter-capacity risk factor (Q3 exhibit 2026-03-10); DTE/Oracle 1.4 GW and DTE/Google 1 GW contracts, terms and cost-allocation language (DTE exhibits 2025-12-31 and 2026-04-30, plus Q3 deck); Amazon 3.8 GW added, OpenAI ~2 GW, Anthropic up to 5 GW (Q1 FY2026 exhibit 2026-04-29); Talen 10.3 GW / 2.2 GW nuclear (Q3 FY2025 exhibit 2025-11-05); FuelCell Energy 450 MW LOI (Q1 FY2026 exhibit 2026-03-09); Oklo 14 GW pipeline / Equinix 500 MW 20-yr PPA / phased 250→500→1,000 MW (exhibit 2025-03-24); Constellation CyrusOne 380 MW + 380 MW Phase 2, "$0.20–$0.50 EPS per 1,000 MWs," AEP >15 GW (Q1 FY2026 exhibit 2026-05-11); Vistra AWS 1,200 MW 20-yr nuclear PPA + Meta PJM nuclear PPAs (Q4 FY2025 exhibit 2026-02-26); GE Vernova ~110 GW gas-turbine backlog/slots by YE2026 + $2.4B data-center electrification orders (Q1 FY2026 exhibit 2026-04-22); Bloom Energy 1.5 GW deployed, AWS/AI fuel-cell projects (Q2 FY2025 exhibit 2025-07-31). Inferred — the read that Oracle's selloff reflects power/funding-execution doubt; relative risk profiles of utilities vs AI buyers. Research-cited — U.S. data-center demand growth, interconnection-queue timelines, per-rack power figures, capacity-shortfall estimates (Goldman Sachs and industry sources). Figures are point-in-time; re-check the latest exhibit before quoting. Not investment advice.