Waiting on megawatts
You can buy the chips. You cannot plug them in. Reciprocating engines buy the years a grid hookup and a nuclear plant cannot.
Ask an American grid for a few hundred megawatts today and, on Ramez Naam’s clock, you are waiting until 2031. He named ERCOT as the fastest queue in the country. Five to seven years is the lag the rest of the record keeps citing. Less than half the gigawatts that have been announced are actually under construction. The Energy Information Administration, in its August 11 Short-Term Energy Outlook, still names data centers as the demand driver and, after the Texas governor’s August 3 pause, cut Texas 2027 load growth from 14 percent to 6. That is a demand restatement, not a campus construction table.
The 1990s telecom boom had a version of this. Carriers could order the routers. Lighting the last mile took longer than the capital plan assumed. This cycle’s last mile is an electron. US data-center load is estimated to go from 176 terawatt-hours to 580 by 2028. Peter Diamandis put the chip-implied ask at 200 to 275 gigawatts against a US grid build-out of roughly 100. The machines that train and serve models do not tolerate a brownout. Someone has to make the watts show up on a construction schedule, not a decade plan.
US data-center electricity demand
Estimate on the nuclear-baseload page. Not a forecast drawn for this article.
The machines that show up in months
A reciprocating-engine genset is not a gas turbine. It is a big diesel or gas engine on a skid, the kind you can site next to a hall and have making backup — or prime — power while the interconnect, the turbine, and the nuclear PPA are still in someone else’s queue. That is the bridge. It is a distinct category, and it is capacity-constrained.
Caterpillar sold $10.2 billion of data-center generators in the prior year, more than a tenth of the company. The large reciprocating-engine backlog is up more than three and a half times since January 2024. Management says it can meet only about 60 percent of 2026 demand. It is tripling that capacity. Payback is penciled for the end of 2030.
Cummins is the other liquid US name. Power Systems printed a record 29.5 percent EBITDA margin in the first quarter of 2026 on data-center backup and prime power — about 39 percent of company EBITDA. The order book runs into 2028. A $450 million program is meant to add about 20 gigawatts. A typical 100-megawatt hall, on Cummins’s own math, wants 120 to 200 megawatts of backup behind it.
GE Vernova sits on the turbine side of the same wait. Jessica Uhl, who ran that business, called the turn in gas-turbine demand “a 180” — from people questioning the future of gas to people who cannot get enough of the machines. The two legs are complementary. A genset covers the months. A turbine covers the larger central plant. Neither is a substitute for the other.
SemiAnalysis’s Energy Model, on September 10, put a size on the behind-the-meter book. Seventy-five gigawatts of firm, binding OEM orders for AI compute sitting off the grid — about twenty of them ordered in the second quarter of 2026 — and about three gigawatts of US datacenter IT capacity expected to be operational behind the meter by year-end. Reciprocating-engine makers with multi-hundred-megawatt off-grid orders went from twelve to twenty-two. The named buyers sit on the same scarcity spine already on this page: Microsoft more than five gigawatts year-to-date behind the meter, 2.7 of them with Joulent and Chevron; Google 930 megawatts of aero-derivatives plus 900 megawatts of Bloom; Anthropic and Meta 300 to 500 megawatts of Enchanted Rock half-megawatt gensets; OpenAI’s Shackelford campus, 1.4 gigawatts of IT on more than 500 Jenbacher J624s. The shop’s inference economics — $100 billion per gigawatt per year at 90 percent-plus gross margin — is why a buyer will take twice the cost, or 30 percent worse efficiency, to get the watts now. Execution cracks, from Oracle Jupiter and Nebius New Jersey permitting pivots to pipeline and labor, sit on the same page as the August secondary-market note. The note is partial. OEM and balance-of-plant winners sit behind the paywall. It does not flip a step, and it is not a reason to re-date Caterpillar or Cummins.
Time to power
Schematic, not a forecast. The five-to-seven-year interconnect lag is from the genset-bridge chain.
What a tenfold capacity price buys
Training and inference want power that does not blink. Solar and wind, without storage that does not yet exist at this scale, do not clear that bar. Nuclear does: baseload, carbon-free, always on. That is why Microsoft paid more than twice the prevailing spot rate to wake Three Mile Island. David Friedberg put the number on All In. The deal is a twenty-year PPA. Restart is aimed at 2027.
Carly Anderson, a Timescale VC who sits on Macro Voices, put the same contract as the thing that made the renaissance financeable. Hyperscalers, she said, are “creating the conditions for a nuclear renaissance because you really do need somebody who’s willing to sign a 20 year PPA for $100 per megawatt hour. And that didn’t exist until a couple of years ago.” That is independent color on why a twenty-year offtake matters. It is not a Constellation re-rate, and it does not name a new reactor.
The price signal that made those contracts urgent sat in PJM, the mid-Atlantic grid. Capacity cleared at $28.92 per megawatt-day for 2024/25. For 2026/27 it cleared at $329.17 — ten times, with data centers 63 percent of the load-growth increment and $9.3 billion of cost recovery headed back to customers. Robin Millican, at Columbia’s Center on Global Energy Policy, calls PJM a special case: a dysfunctional capacity market, and very hard for new projects to come online. Do not read that auction as a national electricity-price story. Read it as the reason a hyperscaler will overpay for a reactor that already exists.
That “ten times” now has to be a collar, not a spike that reversed. The 2027/28 Base Residual Auction, run 17 December 2025, cleared at $333.44 per megawatt-day. The 2028/29 auction, run 14 July 2026, cleared at $325 — 2.5 percent below the prior cap, still at the ceiling, and 6,831 megawatts short of the reliability requirement, with only 525 megawatts of new generation and uprates. The wiki’s “collapse in 2027/28” falsifier did not fire. PJM’s own estimate is that without the collar the 2028/29 auction would have cleared $554.72, taking total cost from $16.4 billion to nearly $30 billion. Constellation’s 2028/29 book is 15,700 megawatts of nuclear against 3,175 megawatts of fossil and other, all at $325. The company banks the collar. The scarcity, if you believe the uncapped shadow price, is about 70 percent higher. SemiAnalysis argues part of the original spike is a modeling artifact — gas plants rated as if winter never lifts their output, Storm Elliott failures still in the study after hundreds of units winterized — and puts the ratepayer overcharge at about $11.6 billion across two auctions. The wiki leaves that open. What it will not do is treat the error as a cut to Constellation’s check: a requirement correction of a few gigawatts has to eat through the entire $230 gap before the capped price a generator is paid moves at all.
FERC met the end-June RM26-4 clock on 18 June with six Federal Power Act section 206 show-cause orders. Commissioner LaCerte’s line: “rather than a proposed final rule at this time.” The emergency-backstop docket is still a plan waiting on a signature. Comments on ER26-3380 closed Friday, August 21. As of September 24 the Commission had still not ruled. A Federal Register combined notice dated 21 September listed other PJM tariff items — not an order on the Reliability Backstop Procurement. PJM’s July 31 transmittal still asks the Commission to act by 29 September so a window can open 30 September. The bid window, if accepted, is still September 30 to October 21. PJM’s window — a $555-per-megawatt-day cap on 6.8 gigawatts — remains contingent on that signature. Silence is not a ruling, and it is not a Constellation re-rate.
PJM capacity, two delivery years
Data centers were 63% of the load-growth increment. Numbers from the PJM-to-nuclear-premium chain.
The book that followed is mostly brownfield. Meta has assembled about 7.8 gigawatts of nuclear, including Vistra megawatts that start late 2026. Google is paying NextEra to restart Iowa’s Duane Arnold, shut since 2020, for 2029 delivery. AWS is on the meter at Talen’s Susquehanna and at Vistra’s Comanche Peak. Cumulative hyperscaler nuclear commitments were 9.8 gigawatts across thirteen projects when last tallied. Small modular reactors remain a 2030s product. Cameco’s Tim Gitzel said data centers are “bringing dead reactors to life.” That is the near-term trade: existing plants, not new designs.
Meta’s Hyperion campus in Richland Parish is the conventional grid-tied version of the same wait: 5 gigawatts of IT announced in July, originally December 2024, not described as energized. Colossus skipped that clock by reusing a factory and sitting behind the meter. Those are different units. The wiki does not treat Hyperion’s still-pre-online months as a falsifier of the behind-the-meter path, and the new campus tape does not name Constellation or add a PPA.
Michael Cembalest, who does the energy math at JPMorgan, leaves a hole in that story without closing it. The West, he says, builds nuclear at four to five times the cost of China, Korea, and India. He names Flamanville, Olkiluoto, Hinkley, Vogtle. The demand can be right and the listed beneficiaries can still disappoint if new Western steel cannot be poured at a returns-positive cost. Restarts are what the hyperscalers have actually signed. Greenfield is the wish.
Sarah Guo put the same bind as safety and price. If you want nuclear cheap enough to be baseload, she said, you first have to convince people it is safe. Working backwards from a hall of GPUs — cooling, power, training, inference — she said the stack looks a great deal like energy independence. Conviction has put money into nuclear energy and into alternative chip architectures. That is a framing, not a new power-purchase agreement, and it does not re-rate Constellation or Cameco.
When the rack outruns the wire
Bridge power buys time at the building. Inside the hall, a different physics problem appears as GPU racks approach six hundred kilowatts. At fifty-four-volt AC the current is on the order of eleven thousand amperes; at eight hundred volts DC it falls to about seven hundred fifty. Resistive losses scale with current squared — the wiki’s cited pair puts the reduction near two hundred twenty-fold. Eight-hundred-volt DC is mandated by physics at that density, not preferred by engineers.
The transition runs in phases through the end of the decade. Nvidia, Schneider, and Vertiv independently confirm that fifty-four volts is impractical at a few hundred kilowatts and that eight-hundred-volt DC is the Kyber-class architecture. The official path is a hybrid sidecar in the second half of 2026 — Nvidia’s MGX-compatible rack, then named boxes from Vertiv, Delta, and Schneider — a row-level design in 2027, and solid-state transformers later. That is not SemiAnalysis’s four labeled phases. Sequential equipment TAM in the cited industry work peaks near eleven billion dollars for power racks in 2028 and near thirteen billion for SST in 2030. Independent Nvidia and OEM pages do not replicate those dollars, or the three-point-six to four-point-eight-million-dollar content per megawatt, or a four-hundred-to-five-hundred-thousand-dollar sidecar ASP. The fetched SemiAnalysis body even says SST TAM near thirty-two billion, not thirteen. Forward curves, not realized share. NEC 2029 is the next NFPA work cycle, not a ban until then. Eaton, ABB, Schneider, and Vertiv sit on the later incumbent list. Do not re-rate Eaton or Vertiv on the TAM table.
Oilfield modular, data-center urgent
Permitting and interconnect queues are not the only bind. Hyperscale construction also rewards whoever can deploy modular, remote, harsh-environment infrastructure fast — a competency oilfield services spent decades building. SLB’s CEO Olivier Le Peuch reoriented the portfolio toward “higher-return, technology-driven, and less cyclical growth”; data-center solutions grew forty-five percent year on year in Q1 2026 with an NVIDIA partnership as validation, then eighty percent year on year in Q2 with management raising the target to exceed two billion dollars annualized run-rate by end-2027 on new hyperscaler customers, plural.
That is a non-traditional picks-and-shovels leg — not Caterpillar on a skid, not PWR pulling wire, not the grid-transformer supercycle in the aging-infrastructure article. SLB sells integrated modular data-center infrastructure from an oilfield toolkit. Middle East force majeure still masks the segment in reported oilfield revenue; the re-rate narrative depends on the DC line compounding faster than Hormuz headwinds fade. NVIDIA partnership terms remain undisclosed — referral versus anchor matters for how durable the credential is.