The Inference Report

April 21, 2026
From the Wire

The real leverage in AI infrastructure is shifting from model builders to cloud providers and chip manufacturers, and the deal structures reveal who's winning. Amazon's $5 billion investment in Anthropic paired with a $100 billion spending commitment is not a bet on Claude's capabilities, it's a lock-in mechanism dressed as partnership. Anthropic gets capital and the compute it desperately needs after this year's outages; Amazon gets contractual guarantee of revenue and a reason to prioritize Anthropic's workloads on AWS. This is circular capitalism with teeth. Meanwhile, GitHub is pausing new Copilot sign-ups because agentic workflows consume far more compute than the original pricing model anticipated. The infrastructure can't keep up with what the product is actually doing in the field. That gap between what vendors promised and what autonomous agents actually require is where the real business problem lives, not in model performance, but in cost structure and reliability.

Anthropic's Mythos model presents a different kind of leverage problem, one that cuts against the company's own interests. The model is restricted, meaning Anthropic tried to limit its use. Yet the NSA is using it anyway, and the startup faces the classic dilemma: restrict a powerful tool and lose influence over how it gets deployed, or open it and lose control of the narrative around safety. Separately, a Chinese startup called Colleague Skill is instructing workers to train AI agents to replace themselves, a preview of what happens when the technology is cheap enough and the labor market precarious enough. The friction isn't coming from regulation or corporate caution. It's coming from the ground up: workers in China are already pushing back, and in the U.S., tech workers in suburbs are likely to lead the backlash against AI-driven job displacement. That's where the real political economy of AI will be decided, not in frameworks or voluntary commitments.

The infrastructure constraints are real and measurable. RAM shortages will persist through 2027 because AI demand is consuming capacity faster than new fabs can come online. Fermi, the AI nuclear power startup backed by Rick Perry, just lost its CEO and CFO as Amazon pulled a $150 million investment, a sharp reminder that speculative infrastructure plays don't survive contact with actual economics. Apple's succession from Tim Cook to John Ternus matters here because Apple's relationship with TSMC gives it first access to sub-1 nanometer chips by 2029, which means the company that controls consumer hardware will also control the performance frontier. The constraint isn't capability anymore. It's silicon, power, and real estate. Companies that can't secure those inputs won't scale, no matter how good their models are.

Sloane Duvall