Topic
Compute
The frontier lab leaders all agree the AI race is a compute race, but they diverge sharply on how much raw hardware matters versus what you do with it. Sam Altman frames the buildout in almost civilizational terms: he told an Indian audience in February 2026 that if every person on Earth wanted even a thousand GPUs working for them, the world would need 8 trillion GPUs, which "just seems ridiculous." He called AI infrastructure "the most expensive, complex infrastructure project the world has ever collectively taken on," endorsed Jensen Huang's five-layer AI stack, and doubled down on his earlier claim that frontier models can't be built for $10 million ("it's even more true now"). The Stargate partnership with SoftBank put $1 billion into SB Energy, with OpenAI signing a 1.2 GW data center lease for the initial buildout. Greg Brockman, co-founding that effort, was blunt about OpenAI's philosophy: when critics said they were "just throwing bigger computers" at AI, his reaction was "yes, that's exactly what we need to do." Sometimes scaling is the innovation.
Anthropic tells a different story. Dario Amodei describes scaling laws as a chemical reaction (data, compute, and model size are the ingredients, intelligence is the product) but emphasizes that Anthropic has "always had a fraction of the compute" its competitors had and still produced the most performant models for several years. Daniela Amodei made the same case on CNBC in January 2026, arguing the company is proving "you don't need the most chips to win," while simultaneously banking on Amazon's Trainium2 chips and a billion-dollar Indiana data center as critical infrastructure. Their bet is to invest early in hardware access that pays off in years, while extracting more per FLOP than rivals. Jack Clark, in his Import AI newsletter, provides the most granular analysis: he tracked decentralized training networks peaking at the 6e22-6e23 FLOP range (roughly 1,000x less compute than Grok 4), noted that intelligence per watt improved 5.3x over two years thanks to MoE architectures and better silicon, and flagged a Google paper on space-based computing that observes the Sun emits 100 trillion times humanity's total electricity production. Altman, for his part, called the idea of space data centers "ridiculous" with current launch costs. At Google DeepMind, Demis Hassabis announced a $15 billion AI hub with gigawatt-scale compute and a submarine cable gateway, framing infrastructure as a means to an end rather than an end in itself. Mira Murati's Thinking Machines Lab took perhaps the most compute-efficient stance of anyone: their LoRA research showed fine-tuning can match full training at roughly two-thirds the FLOPs, and their Tinker product gives researchers granular control over training loops while offloading distributed compute to shared infrastructure. Ilya Sutskever, whose early conviction that big neural networks on big compute would work drove much of the field, put it most simply: a big neural network, a big computer cluster, a learning formula, and a lot of data. The surprise, he always said, was the simplicity itself.
People on this topic
Perspectives
Sutskever vs. the Scaling Consensus
Ilya Sutskever arguably proved scaling laws work better than anyone alive. He co-authored AlexNet, co-founded OpenAI, and oversaw the GPT series that turned neural scaling from a research curiosity into the dominant paradigm. Then he left and announced that the age of pure scaling is over. His argument: scaling sucked the air out of genuine research. Models trained on massive compute ace benchmarks but fail at generalization, like a student who memorized 10,000 competitive programming problems but can't architect real software. He thinks RL training produces meta-reward-hacking, where the researchers themselves (not just the models) are unconsciously overfitting to evaluation metrics. This puts him directly at odds with his former colleagues. Altman and Brockman are spending hundreds of billions on data centers and custom chips, betting that more compute is the binding constraint. Amodei still describes intelligence as a chemical reaction with known ingredients. Even Hassabis, who shares some of Sutskever's skepticism about pure scaling, is building gigawatt-scale infrastructure. Sutskever's counter-thesis is that AI needs something analogous to human emotions: a learned value function that provides fast, approximate feedback about whether a course of action is promising, rather than the sparse end-of-trajectory reward signals that current RL relies on. He cites a neurological case study of a stroke patient who lost emotional processing but retained full IQ, and became unable to make even trivial decisions. The implication: without an internal compass, raw intelligence is paralyzed. SSI is his bet that ideas, not compute budgets, are what's actually missing.