How Do Hyperscalers Buy Power? The PPA and Own-Plant Model

A hyperscaler buys megawatts the way it buys chips: it signs first, pays before the plant is live, and only then can a hall earn.

HYPERSCALERS · Data as at 20 Oct 2026 · Next update: after late-October 2026 hyperscaler results and the North Carolina large-load order
Affects: Hyperscalers · US · Neoclouds

This is the power-contract mechanism on the hyperscaler layer — a long offtake, a development cheque, or an own-plant — used to pull watts that speed-to-power has not yet delivered. Amazon, Google, Microsoft and Meta already have company pages. See What Does Amazon Do? and AI capex.

In short: A hyperscaler power contract is a promise to pay for megawatts that do not yet sit on the bus. Named prints to 19 October 2026: Amazon’s 960 MW Indiana study; Google’s 890 MW / 20-year Constellation uprate PPA (first 2028) plus 2,700 MW of existing supply; AEP’s 702 MW Cheyenne PPA for Google; Duke’s 50 MW+ North Carolina settlement. Who gets paid first: the plant owner and the vendor. Money sticks at commercial operation.

Education only, not advice to buy or sell any security. Figures are from company filings and company releases dated as shown. Outlook and studies are labelled.

Numbers that matter

Sources: AEP, 19 Oct 2026; Amazon, 19 Oct 2026; X-energy, 19 Oct 2026; Constellation / Google, 6 Oct 2026; AEP, 14 Oct 2026; NRC Federal Register, 25 Sep 2026; Duke Energy, 7 Oct 2026.

Our arithmetic. Twelve 80 MW units are 960 MW. Cascade plus Indiana is 1,920 MW, 38% of a 5,000 MW target (1,920 ÷ 5,000 = 0.384), leaving 3,080 MW unnamed. Amazon’s named offtake — Cascade’s first 320 MW plus Indiana 960 — is 1,280 MW, 26% (1,280 ÷ 5,000 = 0.256). Keep Google’s 890 MW and 2,700 MW on separate rows (3,590 MW if added). AEP’s two customer-backed prints in six days are 1,662 MW (702 + 960). Constellation’s “more than $4.3 billion” against 890 MW is about $4.8 million per new megawatt (4,300 ÷ 890 ≈ 4.83) — intensity, not a PPA price.

What does it actually do?

Three steps, in order.

Sign. Fund. Wait for COD.

1. It signs for watts that are not yet on the bus. The instrument is a power purchase agreement, a development letter, a large-load electric service agreement, or an own-plant. Amazon’s 19 October 2026 print with AEP and X-energy is a joint development agreement to study 12 Xe-100 reactors at the retired Breed station in Sullivan County, Indiana. Amazon “intends to purchase the power under a long-term contract, covering the costs of the facility and its operations.” That is not a permit, a price or a commercial-operation date. See speed-to-power.
2. It funds the work that might make those watts. Amazon is the “project development funder.” X-energy calls the money “milestone-based development capital” and names Indiana the second site under a collaboration “aimed at bringing more than 5 GW of new nuclear capacity to the U.S. grid by 2039,” after Energy Northwest’s Cascade plant in Washington. Amazon led X-energy’s 2024 Series C-1 round, which closed at $700 million. That cheque paid for design and fuel work, not concrete at Breed. See cash capex.
3. It waits on the clock speed-to-power already named. Google’s 6 October 2026 Constellation deal is the nearer version: a 20-year PPA for 890 MW of new output from uprates at 11 existing units, “with the first uprate expected to be delivered by 2028,” plus a 15-year supply agreement for 2,700 MW from the operating PJM fleet. Uprates reuse a licensed plant. Indiana does not. Microsoft’s Crane restart — formerly Three Mile Island Unit 1 — is the restart version: the NRC’s 25 September 2026 assessment cites an 835 MW machine and a 20-year PPA signed in 2024; remaining licence actions were still open. See the US door and PJM.

How does it make money?

The hyperscaler does not get paid on the power contract. It gets paid later, when a hall those watts light bills cloud or ads. The people who get paid on the contract sit one hop upstream.

• The plant owner is paid when the offtake is bankable. Constellation spends “more than $4.3 billion” on the Google uprates and books the 20-year PPA plus 15 years on 2,700 MW of existing fleet. AEP would own and run Indiana through Indiana Michigan Power, and it owns about 92% of Cheyenne Fuel Cell, LLC, the vehicle for the 702 MW Google PPA.
• The technology vendor is paid first, on studies and kit. X-energy is paid as Amazon-funded feasibility work is done around the Xe-100. Bloom Energy is the named technology on AEP’s Cheyenne plant. Neither announcement bills a megawatt-hour to a hall.
• The hyperscaler pays now and is paid last. Amazon would be “the first customer to purchase power from the plant once operational.” The cloud invoice still waits on remaining performance obligation converting as halls light. See What Does Google Do? and What Does Amazon Do?.

Who gets paid. X-energy and Bloom, on development and kit. AEP and Constellation, on a plant or an uprate once it runs. Duke, on substations and security before the hall switches on. Money sticks at commercial operation — not at a 960 MW study. Google’s 890 MW line is new output from existing machines; the 2,700 MW line is a financial offtake on the existing fleet. Keep them on separate rows. See cash capex.

What limits it?

A signed megawatt is not a live megawatt. Four clocks sit under that sentence.

1. A study is not a permit. AEP, Amazon and X-energy have not published a cost, a final investment decision, a licensing route or a build date for Breed. Our Indiana Wire records Cascade as still in pre-application talks on the NRC page last updated 23 September 2026. Ready is not paid.
2. An uprate is faster than a greenfield, and still dated 2028. Constellation’s own release is the clock: first of the 890 MW “by 2028.” The 2,700 MW supply line does not add a turbine. Crane’s 835 MW still needed NRC actions after the 25 September 2026 finding of no significant impact. See PJM.
3. The grid now bills before it builds. Duke Energy’s 7 October 2026 North Carolina settlement — Amazon, Google, Meta, Microsoft among the signers — requires an upfront payment for customer-specific grid kit, security for shared upgrades, and a High Load Factor rate. If approved, it applies to new loads of 50 MW or more signing after 1 June 2026. A decision is “expected by mid-November.” The 75% minimum-demand floor inside the filed settlement is the Duke Wire. See What Does Meta Do?.
4. Own-plant still has a COD. AEP’s Cheyenne fuel-cell project is the nearest dated own-build: construction planned in 2026, “fully operational by the end of 2028,” about $4 billion. That is a target, not a print.

The honest uncertainty is conversion. Do not add a study, a 2028 uprate, a fuel-cell target and a restart into one “hyperscaler nuclear GW.”

Who else does this?

Roles, not a league table.

See neocloud vs hyperscaler.

What would change this view?

The read on the mechanism weakens if:

• Indiana files a construction-permit application and a dated final investment decision with a published cost
• Cascade files at the NRC, or Seadrift receives a construction permit
• Constellation slips the “first by 2028” uprate line
• The North Carolina commission (mid-November 2026) rejects or rewrites Duke’s large-load settlement
• AEP’s Cheyenne plant misses the end-2028 target, or Crane’s remaining NRC actions are denied
• Late-October 2026 results print a material cut in infrastructure spend
• Next GPU attach is late because HBM is short — the stuck step would then sit in semiconductors

It stays intact while named nuclear sites remain studies or 2028 clocks, Duke’s order is still pending, and cash leaves before a megawatt-hour is billed.

Practical takeaways

• The model is Sign → Fund → Wait for COD: the announcement is not the watt
• Amazon–AEP–X-energy (19 Oct 2026): 960 MW study; Amazon funds development and intends a full-cost offtake
• Our arithmetic: two named sites 1,920 MW (38% of the >5 GW / 2039 target); Amazon’s named offtake 1,280 MW
• Google–Constellation (6 Oct 2026): 890 MW of new uprates, first by 2028, plus 2,700 MW existing supply; Constellation spends more than $4.3bn
• AEP–Google Cheyenne (14 Oct 2026): 702 MW, ~$4bn, full ops targeted end-2028
• Crane: 835 MW, 20-year Microsoft PPA; NRC actions still open after 25 Sep 2026
• Stuck step: commercial operation, not the press release
• Next reads: North Carolina’s mid-November order; late-October hyperscaler prints

FAQ

How do hyperscalers buy power?
They sign a long contract — a power purchase agreement, a development-funding letter, a large-load utility agreement, or an own-plant — that commits them to pay for megawatts a hall will need. The plant owner and the equipment vendor get paid on that contract. The hyperscaler gets paid later, when the hall those watts light bills cloud or ads.

What is a hyperscaler power PPA?
A power purchase agreement is a multi-year promise to buy a plant’s output at a contracted price and volume. Google’s 6 October 2026 Constellation deal is a 20-year PPA on 890 MW of new nuclear uprates. Amazon’s 19 October 2026 Indiana print is not yet a priced PPA; it is a development agreement plus an intention to buy the power on a contract that AEP says would cover the plant’s costs.

Is a signed megawatt the same as a live megawatt?
No. Indiana’s 960 MW is a feasibility study. Google’s first 890 MW uprate is dated 2028. Crane’s 835 MW still sat in front of remaining NRC actions after 25 September 2026. A signed megawatt starts the clock. A live megawatt is commercial operation at the rack.

Who gets paid first on these contracts?
The technology vendor, on studies and kit — X-energy on Amazon-funded feasibility, Bloom on AEP’s Cheyenne cells. Then the plant owner, once the offtake makes the spend financeable and the machine runs. The hyperscaler is paid last, in cloud or advertising, after the watt arrives. See speed-to-power.

How is this different from a cloud remaining performance obligation?
Remaining performance obligation is contracted cloud revenue the operator has not yet recognised. A power PPA is contracted supply the buyer has not yet received. One is a sale waiting on a hall. The other is a purchase waiting on a plant. See remaining performance obligation.

Should I buy shares on the back of a power deal?
We do not make buy or sell calls, and nothing here is a recommendation. What we can tell you is what the contract does, who gets paid on it, and which dated prints would change the picture. The decision is yours.

Keep reading on the map

• The layer: Hyperscalers — signed megawatts, late plants
• The term: Hyperscaler, explained
• The mechanism: Speed-to-power · AI capex · Cash capex · How long it takes to build an AI data center
• The contrast: Remaining performance obligation · Neocloud vs hyperscaler
• The buyers: What Does Amazon Do? · What Does Google Do? · What Does Microsoft Do? · What Does Meta Do? · What Does Oracle Do?
• The Wire: Amazon’s 960 MW Indiana study · Google–Constellation 890 MW · Duke’s 75% minimum bill
• Where it happens: US · PJM capacity short · Virginia first power

The members cut on the long clock is already live: Where does the stuck step move?.

How we check this

Figures come from company filings and company releases, each dated below. Indiana terms that are not in the three 19 October releases — cost, Amazon’s dollar commitment, offtake price, build date — are treated as undisclosed, not as zero. The 75% Duke floor is from the filed settlement via our Wire. Crane’s 835 MW and 20-year Microsoft PPA are from the NRC’s 25 September 2026 notice. Cascade’s NRC status is from our Wire citing the regulator page updated 23 September 2026. Arithmetic is shown in full.

Last reviewed: 21 Oct 2026 · Next review: after the North Carolina large-load order (mid-November 2026) and the late-October 2026 Alphabet / Microsoft / Amazon / Meta prints

Spot an error? Tell us; we will correct it here. Editorial standards →

Sources

1. American Electric Power, Indiana advanced nuclear JDA, 19 Oct 2026 — 960 MW; 12 × 80 MW; Amazon development funding; intended full-cost offtake
2. Amazon, Nearly 1 GW Indiana nuclear project, 19 Oct 2026 — first power customer; >5 GW by 2039; Series C-1 $700 million
3. X-energy, Indiana SMR project, 19 Oct 2026 — milestone-based development capital; second Amazon site
4. Constellation Energy / Google, 890 MW PJM uprates, 6 Oct 2026 — 20-year PPA; first by 2028; >$4.3bn; 15-year 2,700 MW supply
5. American Electric Power, 702 MW Bloom fuel cells for Google, 14 Oct 2026 — ~$4bn; 20-year PPA; ~92% AEP; full ops end-2028
6. U.S. NRC, Crane EA and FONSI, 25 Sep 2026 — 835 MW; 20-year Microsoft PPA
7. Duke Energy, Large-load customer protections, 7 Oct 2026 — 50 MW floor; mid-November decision
8. 75% minimum-demand floor, via Duke Wire, 6–7 Oct 2026
9. Cascade NRC status, via Indiana Wire, NRC page updated 23 Sep 2026