The AI Stack Has Consolidated Around a Handful of Suppliers
When Apple rents compute from Google to run Siri on Nvidia silicon, the question of who controls the layer everyone depends on stops being theoretical.
Consider the arrangement announced this week. Apple — the most vertically integrated large technology company, with its own silicon and operating systems — will run its overhauled Siri on Google's fleet of Nvidia Blackwell B200 chips.
If Apple is renting, the compute layer has consolidated.
How the layers stack
- Silicon — dominated by a very small number of designers, with Nvidia holding the strongest position in AI accelerators.
- Fabrication — leading-edge manufacturing concentrated in a handful of facilities globally.
- Data centres — the capital required to build at scale limits the field to a few hyperscalers.
- Frontier models — a small number of laboratories, most with substantial investment from the companies operating the layers below.
Concentration at every level, with the same names recurring across layers.
Why it happened
The consolidation is not primarily a story of anticompetitive conduct. It follows from the economics:
- Capital intensity — data centres and fabrication plants cost sums few organisations can raise.
- Learning curves — advantage compounds with volume in ways new entrants cannot shortcut.
- Software ecosystems — accumulated tooling around a platform raises switching costs enormously.
These are the classic conditions for concentration, and they have produced it.
What follows
Salesforce booked a $2.6 billion gain on strategic investments this quarter, including its stake in AI developer Anthropic — an illustration of how thoroughly the enterprise software layer is now financially entangled with the model layer.
The pattern is one of interlocking dependency: model developers depend on chip suppliers, application companies depend on model developers, and the same firms hold equity across the chain.
Where the pressure goes
The consolidation creates its own opposition. Nvidia's partnership with MediaTek pushes toward edge computing, where different constraints apply and different suppliers may compete. China is pursuing an "AI Plus" strategy of diffusion across existing industries rather than frontier competition, projecting AI-related industries above 10 trillion yuan by 2030.
Both are attempts to compete somewhere other than the layer that has already consolidated — which is generally what happens when a layer consolidates.