Crusoe says each new gigawatt is harder and Oklo will not sell power until 2028

On September 17 Chase Lochmiller sat for an interview the same morning Crusoe announced a $3.9 billion Series F at a $30.9 billion valuation, co-led by Atreides Management, Mubadala Capital and Valor Equity Partners. Then he said the thing you do not expect from a man who has just been handed that much money to build data centers. The scarce input is not chips. It is power, and each additional slice of it costs more than the last. Over the following 24 hours two more founders in our corpus described the same wall and offered three incompatible ways around it. Only one of those three is in production today, and it is the one that needs no concrete.

The gigawatt is getting more expensive, not cheaper

Crusoe builds the Stargate campus in Abilene, Texas. Lochmiller has more reason than almost anyone to say the buildout is going fine. He mostly did, and then he slipped in the caveat.

the electrons are flowing and they're turning into tokens and, you know, helping accelerate this new era of AI abundance. ... each incremental gigawatt is getting increasingly more challenging.

Chase Lochmiller, September 17, 2026, source

He also gave a progress count that runs ahead of the public trackers. Buildings three through six at Abilene are online, he said, with seven and eight landing before year end, which would put the campus at full eight building scale. Epoch AI's site-by-site tracker estimates four of eight operational at roughly 0.3 GW against a projected 1.2 GW. Somebody is counting differently. I would take the builder's number over the tracker's on what is energized, but the gap is worth holding onto, because every forecast downstream of Abilene assumes a fill rate.

Crusoe's newest product is deliberately small

The interesting move is what Crusoe is selling next. Not another gigawatt campus. A box.

You don't need a gigawatt-scale computer. You need lots of smaller clusters of compute to rapidly serve inference infrastructure and really dramatically reduce that time to token.

Chase Lochmiller, September 17, 2026, source

That is Crusoe Spark, and the argument for it is a utility argument rather than a compute one. Getting a gigawatt from the grid in one location is very hard right now. Getting 10 to 100 megawatts from a pocket of spare capacity at some node the utility already has is not. Lochmiller went further and claimed that plugging inference boxes into those pockets should push costs down for local ratepayers, because you are pushing more megawatts across transmission infrastructure that is already built. I would want to see a utility confirm that before repeating it. But the direction is clear enough. Training wants one enormous coherent cluster. Inference does not, and inference is where the token volume is going.

Oklo built a reactor in eleven months and it does not make electricity

Jacob DeWitte recorded his interview sitting next to the thing itself, and the headline number is real. Oklo's Groves reactor in Lockhart, Texas went from empty ground to first criticality in under a year, announced August 6, 2026. DeWitte calls it the fastest privately sited and privately built reactor in history and I have not found anything that contradicts him.

Groves makes isotopes. It does not make power. DeWitte is candid about it, and about what that means for revenue.

power generation, we expect to start selling in 2028. We do have prepayments from customers

Jacob DeWitte, published September 18, 2026, source

Isotope revenue, he said, arrives in early 2027. Electricity is 2028. That is the whole gap between the nuclear story the market is buying and the nuclear product that exists.

Why the hyperscalers come to him

DeWitte's actual insight is commercial, not nuclear. The industry standard was to design a reactor to roughly 80 percent completion, charge a nine figure fee for the drawings, and hand the customer the job of permitting, siting, building and operating it. Oklo sells the electricity instead, on twenty year contracts, and keeps the plant.

But Meta doesn't want to build a power plant.

Jacob DeWitte, published September 18, 2026, source

Six words, and they explain the order book. He is blunt about the demand side too: AI "accelerated all of our wildest dreams by 20 years or more," because Meta, Google, Amazon, OpenAI, Anthropic and SpaceX all desperately need electricity. A decade ago, he said, the people running data centers told him nuclear was too big for what they needed.

One date does not line up

Lochmiller said Crusoe's first reactor powered AI factory goes live by mid 2027. Our transcript renders his partner's name as Oklo. Crusoe's own newsroom says the partner is Aalo Atomics, announced July 30, 2026, with a 2027 proof of concept at Idaho National Laboratory pairing a 10 MWe Aalo-X reactor with a Crusoe Spark unit, and 50 MWe commercial plants targeted for end of 2029. Two similar sounding company names, one of them spoken fast. Either way the number to keep is 10 megawatts in 2027 as a demonstration, and Oklo's first electricity sales in 2028. Nuclear is not relieving the gigawatt squeeze this decade at a scale that matters to Abilene.

The third answer needs no permit

Stefano Ermon is the Stanford professor whose lab produced the score based work that became modern diffusion models. He now runs Inception, which sells diffusion language models. His claim is that the industry is burning the GPUs it already owns.

autoregressive models are still sequential the computation is one left to right one token at a time. You cannot generate the 10th token until you've generated everything that comes before it. That kind of workload does not map well to GPUs. That kind of workload is extremely memory bound. You're spending most of your time moving around weights across the memory hierarchy and you're doing very little arithmetic

Stefano Ermon, September 18, 2026, source

Every vendor claims its thing is the efficient thing, so I went looking for the customer evidence. Ermon gave one, and it cuts the right way.

they were previously using serving their LLMs on Cerebras. So they were using custom chips to get to the kind of speed that they need to deliver the best experience to their customers. And then they switched over to our diffusion based LLMs because they can essentially get the same speed as what you would get if you were to run an autoregressive model on custom hardware.

Stefano Ermon, September 18, 2026, source

A voice agent company gave up specialized silicon and went back to Nvidia parts because a different model architecture bought back the latency. If that holds, it is the cheapest of the three answers by a wide margin, because it consumes no land, no interconnect queue and no fuel. Ermon is honest about the ceiling: he puts the latency sensitive share of workloads at 20 to 30 percent, and says Inception's Mercury models match the speed tier of frontier offerings rather than the frontier itself. Inception has raised $50 million against Crusoe's $3.9 billion. That ratio is the whole story of how this industry allocates.

So which one shows up first

Software, then small power, then nuclear, and the spacing is years. Ermon's models are serving production traffic now. Crusoe Spark units can land wherever a utility has 10 to 100 spare megawatts, which is a 2026 and 2027 activity. A 10 MWe reactor demonstration arrives in 2027 and Oklo's first power revenue in 2028, against a single campus in Abilene that will want 1.2 GW on its own.

That ordering has an uncomfortable implication for the capex debate I wrote about after Hock Tan reaffirmed his 2028 number. If the binding constraint is megawatts and megawatts are slow, then the marginal dollar of compute spend earns its return through efficiency, not through more steel. Lochmiller, the person with the most steel, is the one who said each new gigawatt is harder. He is also the one now selling a smaller box. Read those two together.