Many people assume AI companies compete only on whose model is smarter. But behind that is an equally important race: who can secure enough compute. Anthropic doesn’t build its own chips, and how it fights this battle decides how far it can go. To get to know the company first, start with what kind of company Anthropic is.
In one sentence: Anthropic “spreads” its compute risk across several clouds, at the cost of “binding” itself into several huge long-term contracts.
Many Clouds, Many Chips
Anthropic deliberately avoids putting all its eggs in one basket. It runs on three main compute tracks:
- Amazon AWS: AWS’s in-house Trainium chips. AWS is Anthropic’s earliest and deepest compute partner, and Amazon is also a major investor.
- Google Cloud: Google’s TPUs (Tensor Processing Units). Google is also a major investor.
- Microsoft Azure: NVIDIA GPU compute through Azure.
Anthropic also already uses AMD MI355X chips and plans up to 2GW of the MI450 series, with the first deployments expected in the first half of 2027. It also rents Colossus GPUs from SpaceX and, according to Bloomberg, has signed a compute contract with Lambda. Spreading across suppliers spreads risk, but it also means more long-term contracts to manage.
Huge Long-Term Contracts: How Big the Compute Commitments Are
| Partner | Compute commitment |
|---|---|
| Amazon AWS | More than $100 billion in technology spending over ten years; AWS provides up to 5GW of compute |
| Google/Broadcom | Multi-gigawatt TPU capacity, coming online in stages from 2027 |
| Microsoft Azure | About $30 billion of compute, up to 1GW of capacity |
| SpaceX | Renting GPU capacity on Colossus 1 and Colossus 2 |
| Lambda | About $35 billion (according to Bloomberg, September 2026) |
According to Reuters in September 2026, the prospectus lists about $518 billion in compute and infrastructure obligations over the next ten years, roughly 80% of which cannot be canceled or must be paid regardless of usage. The minimum spend with Google is $111.1 billion; if Anthropic doesn’t use it, it still pays the difference.
This is the flip side of a high valuation. Anthropic’s Series H post-money valuation is about $965 billion (May 2026), which assumes the market believes this compute will eventually turn into enough revenue. Its officially reported annualized revenue was more than $47 billion as of May 2026. If growth slows, the contracts turn from fuel for growth into a heavy fixed cost. For how to read the valuation, see Anthropic’s valuation and IPO.
The Supply Chain It Can’t Escape
Whether it’s AWS’s Trainium, Google’s TPU or NVIDIA’s GPU, trace them upstream and almost all are made on TSMC’s advanced processes, then go through CoWoS (advanced packaging), which packages the compute chip and memory together. Anthropic has spread out its cloud suppliers, but in the end it still shares the same supply chain with the entire AI industry.
Memory is another bottleneck. In its May 2026 Series H announcement, Anthropic named three big memory makers, Micron, Samsung and SK hynix, as strategic infrastructure partners. That shows HBM (high-bandwidth memory) has become a scarce resource it must lock in before it can expand. For how the whole chain works, see the AI hardware supply chain end to end.
Export Controls and Geopolitical Exposure
US export controls affect chip supply, and they can also restrict model services directly: Fable 5 was taken offline under a directive in June 2026 and came back on July 1. The TSMC supply chain carries its own geopolitical risk.
Another easily overlooked variable is electricity. These contracts are measured in gigawatts, which means enormous demand for power and data centers. Getting power and facilities in time is now as real a limit on how fast it can expand as chip capacity.
Penchan’s summary
Anthropic doesn’t design its own chips. It focuses on models and products and spreads supply risk across clouds and chips, which makes strategic sense. The cost is that its lifeline is in others’ hands: compute prices, capacity and policy winds aren’t fully in its control. More subtly, its biggest compute partners are often also its investors. This “funder and supplier” relationship is both a backstop and a dependency; see Anthropic’s investors and partners.
Compute isn’t just a cost. It connects valuation, competitiveness and geopolitical risk. Watching how its compute footprint changes often tells you more about the company’s long-term strength than any single model’s benchmark scores.