Who pays for the watts

The industry did not run out of money. It ran into the grid. Five hyperscalers are prepared to spend on a scale that used to describe national infrastructure programs. The constraint that will price the cycle is not whether the capital can be raised. It is whether the watts show up — and, if they do not, whose balance sheet, whose ratepayers, and whose offtake contract is left holding the residual.

This is a power story wearing a technology costume. Capex guidance is the bid. Interconnection, transformers, and large-load tariffs are the settlement. Vendor financing is the mechanism that decides who is long the difference.

The bid is not a delivery schedule

Microsoft, Alphabet, Amazon, Meta, and Oracle are guiding $660–690 billion of capital expenditure in 2026, almost all of it data centers. Amazon is near $200 billion. Alphabet has guided $175–185 billion. Meta, $115–135 billion.Worth, Aug 2026 That is the demand signal the market is trading. It is not a construction calendar, and it is not a megawatt count.

The OpenAI-led Stargate program, announced in January 2025, planned roughly 7 GW across Texas, New Mexico, and Ohio by September 2025. The plan later moved past 9 GW and added Wisconsin and Michigan. Bloom Energy’s January 2026 power report put U.S. data-center demand at about 80 GW in 2025, heading toward about 150 GW by 2028.Bloom Energy, Jan 2026

Those figures describe appetite. They do not describe steel in the ground, and they do not describe a signed interconnection agreement. A gigawatt in a keynote is a press-release gigawatt until a utility, an RTO, and a transformer vendor agree it is real.

Announcements are not energized megawatts. The argument over the denominator is the finding.

Sightline Climate’s 2026 outlook is the uncomfortable slide. Of 16 GW slated for this year, only about 5 GW is physically under construction. Sightline’s read is that 30–50% of the pipeline may miss the year.Sightline Climate SemiAnalysis disputes the denominator — what counts as slated, what counts as under construction, what a hyperscaler can build without showing up in a public tracker. Do not pick a winner. The disagreement is the point.

Local politics is now a line item, not a footnote. Data Center Watch, a 10a Labs project, counted 75 projects blocked or delayed in the first quarter of 2026, about $130 billion of project value. Of 63 local moratorium actions, 54 passed.Data Center Watch / 10a Labs

The industry can raise the capital. It cannot vote the substation into existence, and it cannot manufacture a high-voltage transformer on a software cycle. That gap — announced GW versus energized MW — is where the bag is being assembled.

The vendor-financing loop

Money is not the scarce input. Structure is. The chip vendor has become a credit intermediary. The buyer has become a deployment covenant. The lender is underwriting a circle.

Reuters reported a $100 billion-scale Nvidia plan that pairs chip supply with a financial stake, to help OpenAI build at least 10 GW of Nvidia systems.Reuters Vendor financing plus equity. The seller helps fund the buyer who has committed to take the product. Chip supply is not independent of the financing that makes the order possible.

Separate talks — not signed; the parties declined comment — go further. One structure under discussion would have Nvidia guarantee roughly $250 billion so OpenAI can lease a 10 GW SoftBank / SB Energy campus in southern Ohio. Project cost has been cited above $500 billion. Separate talks run as high as $350 billion to finance OpenAI chip purchases. Treat those figures as talks until there is a filing, a term sheet, or a confirmation. They are not a closed book.

What Nvidia has said officially is different, and the distinction is the whole story. The company announced partnerships with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR to mobilize over $500 billion of third-party capital over time for AI compute financing.Nvidia, 10 Aug 2026 That is other people’s money, assembled against Nvidia-backed demand. It is not $500 billion of Nvidia cash. Memorandums of understanding are not funded commitments. The number is a mobilization target.

Google is running a quieter version of the same idea on the other side of the stack. The Information reported that Google agreed to backstop as much as about $44 billion of third-party data-center lease payments — the Anthropic / Broadcom side of the build.The Information A lease backstop is not ownership. It is a promise that if the tenant does not pay, a larger balance sheet will. Lenders price that promise. The building gets built. The residual sits off to the side until it does not.

The chip vendor sells the GPUs, takes a stake or guarantee in the buyer, the buyer commits to deploy those GPUs, lenders underwrite against the guarantee. Demand is circular.

Circular demand is not automatically fraud, and it is not automatically a bubble. It is a residual-value problem. If utilization holds, the loop is a working capital market for compute. If utilization does not hold, the same documents that made the project bankable become the documents that allocate the loss: vendor, hyperscaler, lessor, or — if the tariff is loose — the residential class that never signed the offtake.

One thing the loop cannot do: it cannot shorten an interconnection queue, a transformer lead time, or a FERC review. Capital rearranges who is long the wait. It does not compress the wait.

Who pays for the watts

PJM’s capacity price is the cleanest public invoice in the United States. It moved from $28.92 per MW-day in 2024/25 to $329.17 in 2026/27 — more than a 1,000% increase.PJM 2026/27 BRA That is not a rounding error in a household bill. It is a capacity market telling you the reserve margin got tight, and that large new load is a primary reason the curve moved.

PJM’s independent market monitor attributed 63% of the 2025/26 auction jump to data centers, about $9.3 billion recovered from customers.Monitoring Analytics via IEEFA Industry-funded work by E3 puts a smaller share on load.E3 Both numbers can be true in their own frames. One isolates data-center load inside the auction. The other spreads the jump across supply-side and design effects. Cite both. Do not pick a team.

The tariff response is already here, and it is faster than the rhetoric. At least 38 large-load tariffs were filed or adopted between 2018 and 2026; 30 of them in the last two years. The design is converging on the same idea: the load that causes the upgrade pays for the upgrade, and pays whether the servers ever spin.

Virginia is the template. Loads of 25 MW and above pay at least 85% of contracted transmission and distribution and 60% of generation whether they use the power or not, on minimum 14-year contracts, with collateral per megawatt.Virginia SCC Ohio pushes the same logic harder: large loads pay up front and 100% of the build-out. A letter of agreement that assigns construction cost to the customer is not a slogan. It is a credit event if the campus slips.

On 4 March 2026, Amazon, Google, Meta, Microsoft, OpenAI, Oracle, and xAI signed the Ratepayer Protection Pledge: build, bring, or buy generation; pay for delivery upgrades; pay whether they use the power or not.Ratepayer Protection Pledge It is nonbinding. A pledge cannot override a tariff. Harvard’s Ari Peskoe put the constraint cleanly: utilities and regulators hold the pen.

FERC followed in June 2026 with a show-cause to six regional operators — PJM, MISO, SPP, CAISO, ISO-NE, and NYISO.FERC, Jun 2026 The principle is not subtle. If a data center does not show up, residentials should not eat stranded transmission. That is the federal version of the same sentence the state commissions are already writing into large-load classes.

Costs that land on the project do not vanish. They pass to whoever buys the compute.

Follow the invoice one more step. A take-or-pay tariff, a lease backstop, a vendor guarantee — each of those is a cost of goods for a token, an API call, a training run. The residential class is one possible residual. The compute buyer is the other. If the industry succeeds in keeping households off the hook, the watts still get paid for. They get paid for in the price of the cluster.

What to look at

The Iguana method is a checklist, not a narrative. If a deck cannot answer these six, it is not a power story. It is a press release.

1. 01 Energized megawatts

   Count energized MW and executed interconnection, not press-release GW. A campus that is “slated” for 2026 and not under construction is a probability, not an asset. Sightline and SemiAnalysis are arguing about this number because it is the only number that clears.

2. 02 The backstop

   Who is the residual-value, lease, or utilization backstop — the vendor, the hyperscaler, or nobody? A $44 billion lease guarantee and a $100 billion-scale vendor-plus-equity plan are not the same instrument, but they answer the same question: if the tenant walks, whose paper is live?

3. 03 The tariff in that jurisdiction

   Read the large-load tariff and the collateral. Virginia’s 85 / 60 / 14-year structure and Ohio’s up-front, 100% build-out assignment change project IRRs before a single GPU is racked. A national pledge does not replace a state tariff.

4. 04 Water and cooling

   Design in a stressed basin is redesign risk. A cooling scheme that assumed abundant water, or a once-through plan that a county will not permit, is not an ESG slide. It is a delay, a capex revision, and sometimes a dead site.

5. 05 Offtake concentration

   Customer concentration on the offtake is Oracle-style single-name risk. A campus that works if one buyer stays current, and does not work if that buyer slips, is a credit on a technology story. Price it as credit.

6. 06 Chip life versus shell life

   GPUs are short-lived. The building depreciates on a longer clock. The financing often pretends those lives match. They do not. Residual-value guarantees exist because the hardware will be stale before the concrete is.

The invoice is the product

Unlimited capital found a limited grid. The vendor-financing loop can fund the chips and paper the leases. It cannot energize a queue. The watts will be paid for — by a hyperscaler, a chip vendor, a private-credit fund, a compute customer, or, if the tariff is written badly, by a household that never asked for the cluster. The work is to see which of those names is actually on the invoice before the campus is sold as inevitable.

This brief is the product. Week two: crypto take-rate.