Newsletter · · Ashutosh Agarwal

OpenAI Leases 10 Gigawatts and the Hard Part Is Building It - Powering AI: Grid, Gas, Generation & Nuclear - Week of August 28, 2026

Powering AI: Grid, Gas, Generation & Nuclear for the week of August 28, 2026. Podcast synthesis on OpenAI's 20-year 10-gigawatt Ohio lease with SB Energy backed by NVIDIA to 105 billion dollars and a government-owned gas plant, the transformer wall running to five years and 14 million dollars a unit, and Vertiv's CEO calling power a pacing item rather than a ceiling.

Powering AI: Grid, Gas, Generation & Nuclear

Week of August 28, 2026: OpenAI Leases 10 Gigawatts and the Hard Part Is Building It

The money for the AI power buildout has never been easier to find. The electrons and the iron have never been harder. That gap, hundreds of billions in financing chasing transformers that take five years to make, was the whole story on the podcasts this week. One deal captured it perfectly: OpenAI signing the largest data-center lease ever announced, on a site that leans on a gas plant the US government will own. Grand ambition, meet the loading dock.

TL;DR

  • OpenAI locked in a 20-year, 10-gigawatt data-center lease in southern Ohio with SoftBank's SB Energy, the biggest project announced to date, with NVIDIA backing the value up to $105B and a 9.2-gigawatt, government-owned gas plant underneath it.
  • The bottleneck has moved from chips to hardware you can't rush: large power transformers run up to five years and $14M each, the US builds only ~20% of what it buys, and demand could double by next year.
  • Vertiv's CEO guided 2026 revenue to ~$14B (from $10.2B) and called power a "pacing item, not a ceiling," the clearest operator read of the week that demand is still accelerating.

What's new

The biggest data-center deal yet, and it runs on gas the taxpayer owns. On This Week in Cleantech, TigerComm's Mike Casey walked through OpenAI's new 20-year, 10-gigawatt lease in southern Ohio with SoftBank's SB Energy, "the largest data center project announced to date." NVIDIA is backing the completed facility's value "to the tune of up to $105 billion" (a far smaller ask than the ~$250B all-up guarantee it reportedly floated earlier this summer, which knocked its stock down ~5% in late July), taking a $1.5B equity stake in SB Energy and becoming exclusive chip supplier for the first half of the campus. The kicker: the site sits partly on a former Department of Energy uranium-enrichment property, is expected to power seven million homes, and is "backed by a 9.2-gigawatt gas plant that will be owned by the US government and financed through the US-Japan trade deal." As Casey put it, dryly: "it's really good that the hyperscalers are getting underwritten by taxpayers." SB Energy is eyeing a $5-7B IPO as soon as next month. Why it matters: this is the template, hyperscaler demand, chipmaker credit, and public balance sheets stacked behind a single site of unprecedented size.

The transformer wall is the real ceiling. The same episode flagged a New York Times piece that should be required reading for anyone long the theme. Large power transformers, the truck-sized units that step voltage down for local use, "are mostly bespoke and hand-built and can take up to five years and cost as much as $14 million each," and "the US only manufactures 20% of what it buys." The existing fleet was already averaging 38-40 years old over a decade ago. The punchline: "replacing this aging fleet and powering new data centers could double the demand for large power transformers by 2027." Hyperscalers are now "paying premiums to jump the manufacturing queue." A fix is emerging, startups like Ampersand and Huron Power commercializing modular solid-state transformers, with NVIDIA already redesigning data-center blueprints around them, but that's a multi-year story, not a this-year one.

Roughly 750 gigawatts of storage projects are sitting in interconnection queues... the median wait time has jumped from 1.5 years in 2015 to 5 years in 2025. (This Week in Cleantech)

The grid queue is clogging, not clearing. Also from This Week in Cleantech: about 750 GW of battery storage is stuck waiting to plug in, with median interconnection waits now five years. PG&E says some circuit breakers take nearly four years to procure. In New York, a Con Edison rule making developers help pay for grid upgrades has added ~$21M per project and already triggered 25 cancellations. And in a move the hosts called "literally contradictory," the DOE is killing the review of its last three interregional transmission corridors (~3,500 miles) even though its own draft study flags a "dire need" for exactly that kind of transmission, while touting loans instead, including $1.6B and $3.3B to AEP and $26B to Georgia Power and Alabama Power (both Southern Company subsidiaries). Why it matters: the constraint isn't generation on paper, it's the switchgear, breakers, and wires to connect it.

Vertiv's CEO says power is a "pacing item," not a wall. The cleanest operator signal of the week came on At Barron's, where CEO Gio Albertazzi laid out Vertiv's numbers: 2025 revenue of $10.2B, a 2026 guide "recently... a $14 billion top line," 37% year-on-year growth this year (about 45% in the second half after a 24% Q2), and long-term guidance of "between 20 and 22%" five-year growth, "and, oh, by the way, we're going faster than that as we speak." On the power bottleneck everyone worries about, he was pointed: "Access to power seems to be like creating a ceiling that the industry would not ever be able to break. But in fact, we see that a lot of power behind the meter is being developed... what seemed to be a constraint has become just a pacing item." His industry mantra now: "tokens per second per megawatt," every hard-won electron converted into revenue-generating output. This is insider commentary, and it says demand is still accelerating faster than the guide.

The chips-to-power handoff, in one company's numbers. On Schwab Network, Rick Ducat broke down Analog Devices' quarter (reported Aug 19): revenue up 40% year-over-year, with the communications segment, power and networking for data centers, up 84%, and data centers now roughly 80% of that segment. Management guided optical circuit switching revenue to "approximately double this year," with a similar 2027 target. The framing number for the whole complex: global AI capex from tech giants and hyperscalers "projected to exceed $700 billion to $1 trillion in 2026."

The debate

The bull case: this is a multi-year supercycle, and the physical world can't keep up with demand. The most vivid version came from an operator. On Motley Fool Hidden Gems Investing, ExoWatt co-founder and CEO Hannan Happy, an ex-GE and Tesla engineer, backed by Sam Altman and Andreessen Horowitz, laid out the scale of the gap. Data-center "building blocks" have gone from 10-20 megawatts a few years ago to "300 to 500 to even 700 megawatts," the average site is now about a gigawatt ("consuming the equivalent of 1 million US households"), and some are pushing 10 gigawatts. Data centers are headed toward "9 to 10 percent of total energy produced in the US." The urgency is economic: he pegged recent SpaceX compute deals with Anthropic and Google at "a billion dollars a month," so "a year delay... is about $12 billion of missed revenue." Demand like that doesn't wait for a five-year transformer queue, it pays up. Marko Papic on Geopolitical Cousins made the investable version explicit: "the picks and shovels for what you're talking about, like somebody who makes transformers, get ready. It's going to be fun."

The bear case: the money is levered and hidden, and it always overbuilds. Two threads carried the skeptics. On Big Technology Podcast, Alex Kantrowitz and Ranjan Roy dissected how the spending is being kept off balance sheets: Meta's Hyperion data center in Louisiana, and the $27B of debt financing it, don't appear on Meta's books because Blue Owl Capital-managed funds own most of the joint venture that owns the campus, with Meta as minority partner, tenant, and guarantor. Meta has disclosed ~$347 billion in total lease obligations that haven't begun yet, against a market cap around $1.39 trillion. "This is bet-the-company type activity," Kantrowitz said. Papic's cousin Jacob added the historical warning on Geopolitical Cousins: every capex cycle built on a new technology "ended in tears," because "we always overbuilt", and hyperscalers "are competing... they won't [collude], so they will eventually overbuild."

The honest read of the week: even the bears aren't calling the top yet. Papic's own tell was that the demand data is "so strong" the turn "has been pushed off into late 27, if not 28." The debate right now isn't whether real demand exists, cloud revenue "jumped 43% in Q2 of this year... off of a base of $350 billion", it's whether the financing structures and the supply chain can survive the wait. As a reminder, both sides here are pundits and journalists reading the tea leaves; the only operators who spoke this week (Vertiv, ExoWatt) were unambiguously bullish on demand.

The names in play

Vertiv (VRT) is the one stock a guest genuinely moved this week, and it was the CEO doing the moving, see the guide above. Worth noting his caution to investors, too: the stock went from ~$10 at its 2020 listing to ~$370 and back to ~$270, a ~$100B company now, and he pointedly wouldn't talk about "equilibrium." A rocket ship, priced like one.

Uranium turned on the In it to Win it podcast: spot around $89.40, the cheapest 12-month term price ~$95.50, and "summer seasonality officially over." The Sprott Physical Uranium Trust holds 81.7 million pounds off the market, its discount narrowing from 12.5% to ~10% as spot rises, with new Kazatomprom news treated as setting "the new price." The host's discipline is instructive, he'd start trimming above term, into $100-105. It's a price-and-flows call, not a company thesis.

Freeport (FCX) got a passing nod as a chart breakout on Squawk on the Street, with copper printing an all-time high ($6.87) on the In it to Win it episode and the copper-miners ETF up 10.4% on the week, momentum rather than a fresh fundamental case.

Read-throughs

  • Transformers and switchgear are the choke point, so they're also the trade. The same shortage that threatens blackouts (per the NYT piece the cleantech hosts cited) and clogs the battery queue is the one Papic says to lean into. If large-transformer demand really doubles by 2027 against ~20% domestic supply, pricing power sits with whoever can build the iron, and with the solid-state disruptors (Ampersand, Huron Power) NVIDIA is designing around.
  • Gas is the default fill-in, but it has its own five-to-seven-year wall. ExoWatt's Happy was blunt that turbines "have been backlogged five to seven years in some cases," and "just having a gas line doesn't mean you actually have access to that gas." That's the read-through to turbine OEMs and gas-weighted supply, scarce capacity and long lead times, and the reason a government-owned 9.2 GW plant is underwriting the Ohio deal at all.
  • The financing plumbing is now a systemic variable. With hyperscalers "borrowing hundreds of billions" and crowding the bond market (Geopolitical Cousins), and $347B of Meta lease obligations sitting off balance sheet (Big Technology), the same vendor-financing engine powering the buildout, the Apollo-led package for Broadcom's AI chips that Squawk pegged at $70-100B, enabling 20+ GW for Anthropic, is also the fault line. If credit tightens, the electrons stop before the demand does.
  • Backlash is now a real permitting cost, everywhere. WIRED's Molly Taft, on the cleantech show, argued data-center opposition has gone "cross-partisan," Sierra Club to Marjorie Taylor Greene, and Squawk noted states like Texas and Pennsylvania cooling on projects. Even Vertiv's CEO acknowledged "some jurisdictions" tightening rules. The demand goes to friendlier counties, but the friction is a tax on timelines.