The subsidy phase of AI is over. According to Moonshots with Peter Diamandis, labs are no longer trading cash for compute; they are trading equity, signing $73 billion in combined deals with Amazon and Google simply to secure the electricity and chips needed to survive.
On The AI Daily Brief, Nathaniel Whittemore describes a "vertical wall of demand" hitting a fixed global compute supply. GitHub and Microsoft are abandoning flat-rate pricing because every token is sold out. On No Priors, Baseten CEO Tuhin Srivastava provides the on-the-ground view: his clusters run at mid-90s utilization, and to get a significant allotment of chips now requires three-to-five-year commitments with 30% paid upfront.
"The compute market faces a severe supply crunch with very little slack compute, forcing Baseten to run large clusters at mid-90s utilization across 18 clouds globally."
- Tuhin Srivastava, No Priors
The AI bottleneck has shifted from code to the physical world. The Intelligence from The Economist notes that while software updates ship in weeks, building a semiconductor fab or securing electrical transformers takes years. On All-In, Chamath Palihapitiya argues Microsoft, Google, and Amazon are signing power purchase agreements at twice the spot rate, morphing from asset-light software giants into debt-heavy industrial utilities.
This capital intensity is crushing free cash flow for the hyperscalers - Amazon's is down 97% - but it creates a strategic weapon. David Sacks argues on All-In that Sam Altman’s early commitment to $600 billion in data center spend gave OpenAI the capacity to keep serving tokens while a compute-gated Anthropic struggles. The race is now about owning the token factories.
"Even if the consumer business is soft, OpenAI is winning the enterprise market by simply being the only shop that can reliably serve tokens."
- David Sacks, All-In
The frontier model wars are secondary to the inference layer where real business is built. Srivastava of Baseten reveals over 95% of the tokens his platform serves are from custom models, not vanilla open-source weights. Companies like the medical scribe Abridge survive by sitting inside specific workflows, capturing user data that becomes a proprietary training signal impossible for horizontal labs to replicate.
A consensus emerges across the podcasts: pure software moats are collapsing. As Ben Horowitz argued on The a16z Show, AI lets capital bridge technical gaps, shifting competitive advantage to physical constraints - compute, power, and the organizational design to wield them. The game is no longer about the best model, but who can secure the grid to run it.
