The week's most telling pattern isn't about AI capability, it's about who controls the infrastructure that capability runs on, and what happens when that control becomes scarce. Anthropic's annualized revenue jumped to $65 billion in two months, but that number only matters if the company can secure the compute to deliver it. Nvidia is now guaranteeing its chips power OpenAI's data centers through a $1.5 billion investment in SoftBank's infrastructure arm, and separately pledging $100 billion backing for another OpenAI facility in Ohio. This is not venture capital. This is vertical integration masquerading as partnership. Groq raised $350 million at a $3.5 billion valuation to pivot away from making AI chips toward running Nvidia-powered data centers, a tacit admission that the chip wars are over and the real money is in owning the buildings where inference happens. The pattern is clear: frontier model companies no longer compete on silicon. They compete on access to Nvidia silicon, which means they compete on relationships with the companies that can secure land, power, and capital at hyperscale.
Meanwhile, the cost structure of actually using these models is inverting in real time. Token prices are collapsing, good for the model makers' margins, but Gartner predicts inference costs for agentic workflows will increase more than fivefold over the next two years. Builders are using more tokens per task as models get more complex, and agents require orchestration, tool calls, and multi-step reasoning that balloons the token count. Microsoft delayed Exchange Server's first cumulative update because its AI-assisted development process surfaced so many flaws that compatibility testing became the bottleneck. GitHub's eight-hour outage knocked out Actions, APIs, and Copilot for an entire user base dependent on AI-accelerated development. The infrastructure that was supposed to compress timelines is now a single point of failure. Amazon is destroying rare books to train its models, Relay shut down and its team joined Google's Chrome division to build AI directly into the browser, and MongoDB is embedding itself as the database layer for coding agents. Every move locks users deeper into a specific stack.
The real competition isn't between models anymore. It's between control of the hardware layer, the data layer, and the distribution layer. Anthropic can announce $65 billion in annualized revenue, but without guaranteed access to Nvidia chips and power infrastructure, that revenue is theoretical. Groq stopped trying to build chips and started renting Nvidia capacity. Google absorbed Relay's team to own the browser interface where agents will live. Amazon is burning rare texts because the data advantage matters more than the optics. What looked like a race to build better models is actually a race to own the pipes, the power, and the default surface where users interact with them. The model makers are running fastest toward whoever controls the infrastructure that makes their models useful.
Sloane Duvall