Compute

Start with the one sentence nobody at the frontier disputes: intelligence is a chemical reaction, and compute is the fuel. Dario Amodei said it almost literally on Nikhil Kamath's podcast in February 2026, describing the scaling laws as a recipe where "those ingredients are data, compute, the size of the AI model... and what you get out is intelligence. Intelligence is the product of a chemical reaction" (Nikhil Kamath, Feb 2026). Daniela Amodei put the same idea in a balance sheet: the industry "runs on a capability equals compute" assumption (CNBC, Jan 2026). Greg Brockman put it in a press release. Shane Legg drew it as an exponential graph in 2010 and called it "almost terrifying." On the underlying physics, the nine people in this archive are unanimous. The disagreements are not about whether compute matters. They are about who gets to own it, whether the spending is a sane bet or a bubble, and whether efficiency might quietly route around the whole arms race.

The believers, and the one who needs you to believe with him

The clearest statement of compute-maximalism comes from Sam Altman, and it is striking how openly he frames it as an act of will rather than a forecast. His pitch in India was a thought experiment: "how many GPUs would you like working for you all the time?" People answer a thousand. Multiply by eight billion humans and "we have no way to deliver 8 trillion GPUs anytime soon... that just seems ridiculous" (The Indian Express, Feb 2026). The point isn't the number. The point is the ambition it licenses: building compute will be "the most expensive, complex infrastructure project the world has ever collectively taken on."

Brockman is the man actually pouring the concrete. His one appearance in this archive is the SB Energy announcement: OpenAI and SoftBank each dropping $500 million into an energy developer, a 1.2 GW data center in Milam County, all under the $500 billion Stargate banner (OpenAI Blog, Jan 2026). His line is pure operational confidence: "a fast, reliable way to scale compute through large, highly optimized AI data centers." No hedging, no bubble talk. Compute is a logistics problem, and OpenAI intends to out-logistics everyone.

Demis Hassabis sings the same tune in a different key. At Google's India summit he and Sundar Pichai announced a $15 billion AI hub in Vizag housing "gigawatt scale compute" plus new subsea cable routes (VERTEX, Feb 2026). But Hassabis keeps pivoting from raw horsepower to diffusion: AlphaFold reaching "over 3 million researchers... over 200,000 in India." His implied argument is that the interesting frontier isn't the size of the cluster, it's how widely you can spread what the cluster produces.

The believer who's nervous about it

Then there's Dario Amodei, who agrees with all the physics and is visibly uneasy about the politics. He recounts the founding conviction (scaling laws spotted "in 2019 with GPT-2") with no regret. But asked whether he's now the most relevant person in the world, he reaches for the layered stack to deflect: "there's the folks who make chips. There's the folks even earlier who make semiconductor manufacturing equipment. There's the folks who make models like us" (Nikhil Kamath, Feb 2026). And then the admission that separates him from Altman and Brockman: "I'm at least somewhat uncomfortable with the amount of concentration of power that's happening here." Same scaling laws, opposite mood. For Amodei the compute build-out isn't just expensive, it's a centralization machine, and he frames Anthropic's governance (the long-term benefit trust) and his support for "sensible regulation... even though I think it holds us back commercially" as deliberate friction against "the natural grain of this technology."

His sister sharpens the economic version of the same anxiety. Daniela Amodei's CNBC profile is the closest thing here to the bubble debate, and she names it directly: the industry's faith is "this sort of exponential graph that... continues until it doesn't." She notes Anthropic "pretty consistently had the most performant models for several years" while holding "a fraction of" the compute its rivals had, then turns around and signs onto the same logic everyone else does (the Trainium2 partnership with Amazon, a slice of the ~$1.4 trillion in AI commitments). The whole sector is a single wager: "win by outbuilding everyone." Her hedge is honest. If revenue keeps "climbing fast enough to justify the build out," the bet pays. If there's a "timing error, little bit," the most expensive infrastructure project in history becomes the most expensive mistake.

Jack Clark and the politics of who owns the meter

The most intellectually serious treatment of compute in this archive isn't from a CEO. It's Jack Clark, writing his Import AI newsletter through late 2025, treating compute as a question of power rather than performance. His recurring obsession is that "the frontier of AI is determined by basically 5 companies, maybe 10 in coming years" (Import AI 439, Jan 2026), all American, possibly soon Chinese, with zero frontier training runs from "academic, government, independent, or non-tech-industry actors." That's the real story compute tells: a handful of firms holding the only meters that matter.

His hope is decentralized training, which he calls "a political technology that will alter the politics of compute at the frontier." The numbers are brutal but trending: decentralized runs are roughly 1000x behind a model like Grok 4, yet growing "20x/year" against the frontier's "5x a year." And he keeps one eye on the ceiling everyone else is racing toward: Google's Project Suncatcher, putting TPUs in orbit to tap the Sun's "3.86 × 10^26 W," "more than 100 trillion times humanity's total electricity production" (Import AI 434, Nov 2025). Compute, taken to its logical end, becomes "turning energy into thoughts" as the main job of the entire civilization.

This is where Altman breaks ranks with his own maximalism, and it's the sharpest concrete fault line in the file. Asked about space data centers, the man building Stargate calls them "ridiculous" right now: "if you just do the very rough math of launch costs relative to the cost of power we can do on Earth, to say nothing of how you're going to fix a broken GPU in space" (The Indian Express, Feb 2026). Jack Clark thinks orbit is where the energy is. Sam Altman thinks it's a punchline. Same religion, different eschatology.

Mira Murati's quiet heresy

And then, off to the side, the position that threatens to make the whole argument moot. Mira Murati's Thinking Machines spent late 2025 publishing research arguing you might not need the biggest cluster at all. "LoRA Without Regret" (Thinking Machines, Sep 2025) showed that cheap low-rank fine-tuning "fully matches the learning performance" of full fine-tuning for reinforcement learning, "even with ranks as low as 1," because RL "absorbs ~1000 times less information per token" than supervised learning. Their tool Tinker is explicitly pitched at putting "frontier training into everyone's hands" rather than "centralizing power in a handful of labs" (AIM Network, Oct 2025). Jack Clark's intelligence-per-watt thread (Import AI 435) measures the same undercurrent: a 5.3x efficiency gain in two years, open models answering 88.7% of queries locally.

If Murati and the efficiency curve are right, the gigawatt arms race is partly a status game. If Altman and Brockman are right, efficiency just gets eaten by demand and you still need the 8 trillion GPUs. Nobody in this archive has settled it, which is exactly why they're all still building.

The tell is that even the skeptics keep signing the checks. Daniela Amodei can say the exponential "continues until it doesn't" and still commit to outbuilding everyone. Dario can be "uncomfortable with the concentration of power" and still chase the frontier. The fear isn't that compute won't deliver intelligence. They're all sure it will. The fear is being the one lab that flinched the year the graph kept going.