The AI Roadmap Is a Power Purchase Agreement

An AI roadmap used to be a model card and a launch date.

Now it is a utility filing.

OpenAI’s infrastructure spending plans have reportedly ballooned to $750 billion, including a $20 billion data center campus in Georgia expected to draw at least 3.2 gigawatts of power. That is not “we will scale later.” That is a small nation’s worth of commitment hiding inside a product roadmap.

The chip side is just as concrete. AMD is investing up to $5 billion in Anthropic, alongside a plan for Anthropic to deploy up to 2 gigawatts of AMD Instinct GPUs using its Helios rack-scale system. And Etched just hit a $10.3 billion valuation by arguing that inference needs purpose-built systems, not just more general-purpose GPU capacity.

Put differently: the AI roadmap is moving down the stack.

For partners, this matters because customers still talk about AI like it is software procurement. It is not only that anymore. It is capacity planning, energy exposure, chip availability, data-center geography, depreciation cycles, and workload forecasts. A regulated-industry customer asking “can this agent scale?” is also asking “what happens to cost, latency, region, and vendor risk when everyone else wants the same compute?”

The answer cannot be a shrug and a bigger token budget.

The better answer is workload-aware architecture: what runs now, what can wait, what can route elsewhere, what gets cached, what gets smaller, and what deserves premium inference.

AI strategy has left the slide deck.

It is in the power contract now.