Everyone is buying from everyone now

If you’d predicted the AI industry’s 2026 supply chain map two years ago, you would have been wrong about almost every arrow.

Apple this week unveiled an Apple Intelligence architecture built on foundation models co-developed with Google, served through Private Cloud Compute that runs on Nvidia hardware inside Google Cloud. That is one company’s flagship platform pitch using three different competitors’ core IP at three different layers of the stack.

Meanwhile, a careful breakdown of xAI’s recent compute deals points out that xAI (now part of SpaceX) is renting roughly $1.25bn a month of capacity to Anthropic and another $920mn a month to Google, on cancellable terms, using infrastructure that was supposedly built to train xAI’s own Grok. The model lab is becoming a data center landlord for its direct competitors.

Two anecdotes, one pattern: the assumption that vertical integration is the AI moat is not aging well.

The old chart had each frontier lab owning its model, its serving stack, its chips, its data centers, and its product. The new chart has Apple buying Google models running on Nvidia chips in Google’s cloud, Anthropic buying SpaceX-owned compute, Google buying SpaceX-owned compute, and Google dual-sourcing its own TPU foundry from Intel. Everyone is buying from everyone.

This is good news and bad news, depending on where you sit.

Good news for buyers: heterogeneous, multi-source AI stacks are no longer a sign of immaturity. They are the production norm. Negotiating leverage is back in a way it wasn’t six months ago.

Good news for ISVs and integrators: a horizontal, multi-model, multi-cloud agent platform with consistent governance and observability across substrates is exactly the right shape of product for this market. Customers will not let themselves be locked in this time around.

Bad news for anyone whose strategy assumed a single hero stack. If your pitch deck still has a clean “we own everything from silicon to UX” slide, the rest of the market is quietly deciding not to operate that way.

Microsoft’s agent platform pitch from Build read as a long argument for openness at every layer: many models, flexible substrates, partner-friendly governance. A week ago that read like positioning. This week, looking at Apple’s and xAI’s moves, it reads like an accurate description of where the puck is going.

The vendors who pretend otherwise will sell short-term lock-in and miss the longer-term composability deal. The ones who lean into the messiness will win the platform.