The Real Question: Why Doesn't Anthropic Depend on NVIDIA?

The AI hardware market looks like a simple capital game — whoever secures the most NVIDIA H100/B200 chips wins. But Anthropic, the company behind the Claude model family, is playing a different game entirely. They run workloads across Google's TPU v5e, Amazon's Trainium2, and NVIDIA GPUs — deliberately avoiding dependence on any single vendor. Here's why that matters.

1. Inference Economics: Why General-Purpose GPUs Are Overkill

NVIDIA GPUs are general-purpose powerhouses backed by CUDA's massive ecosystem. But once a model is trained, the economics shift to inference — serving billions of API responses per day. At that scale, paying for general-purpose flexibility you don't need becomes expensive waste.

  • TPU v5e's sweet spot: According to Google, the TPU v5e delivers up to 2.5× the inference performance per dollar compared to its predecessor. While this is a vendor-reported figure and real-world gains vary by workload, the directional advantage of purpose-built ASICs for tensor operations is well established — lower electricity costs, lower latency, better cost efficiency at scale.
  • Trainium2's parallel bet: AWS's Trainium2, developed by Annapurna Labs, takes a similar approach from the Amazon side — a custom chip optimized for both training and inference, claiming 4× the performance of its predecessor. Anthropic's multi-billion-dollar deal with AWS means Trainium2 is just as central to their stack as Google's TPUs.
  • Network architecture matters: At LLM scale, inter-chip communication bandwidth often matters more than single-chip performance. Both Google's TPU pods (using Optical Circuit Switches) and AWS's custom networking fabric are designed to minimize this bottleneck at a level that commodity InfiniBand clusters struggle to match.

2. Software as the Unlock: Hardware Abstraction

The biggest barrier to leaving NVIDIA is the software stack. The entire AI industry is locked into CUDA, and switching chips means rewriting and re-optimizing code from scratch.

  • JAX and XLA as likely enablers: Based on Anthropic's engineering job postings and technical blog posts, the company appears to make heavy use of Google's JAX framework and XLA (Accelerated Linear Algebra) compiler — tools that abstract away hardware-specific details and compile the same code to run on GPUs, TPUs, and custom accelerators. While Anthropic hasn't published a detailed infrastructure paper, these hints point to a deliberate strategy of hardware portability.
  • Multi-architecture as a moat: The ability to seamlessly train and serve the same massive model across NVIDIA GPUs, AWS Trainium, and Google TPUs is extraordinarily difficult to build. Most AI companies are locked into a single vendor's toolchain. This cross-platform execution capability — not any single model — may be Anthropic's most durable competitive advantage.

3. The Capital Structure: Compute Credits, Not Just Cash

The financial architecture behind Anthropic's chip strategy is as important as the technical one.

  • Google's deal: Google has invested up to $2 billion in Anthropic. A significant portion of the early investment came as Google Cloud compute credits rather than cash, though later rounds have included more direct capital. This structure gives Anthropic access to massive TPU capacity without the upfront hardware procurement that would drain a startup's cash reserves.
  • AWS's deal: Amazon has committed up to $8 billion, structured similarly — a combination of cash and AWS credits. This is the larger of the two deals and ensures Trainium2 gets serious production workloads from a frontier AI lab.
  • The hedge: With two major cloud providers competing for Anthropic's workloads, the company has effectively neutralized the biggest operational risk in the industry: "We can't ship because the chips aren't here yet." If NVIDIA GPUs are delayed, shift to TPU. If Google has capacity issues, lean on AWS. This optionality is something only a handful of companies in the world have achieved.
  • Pricing power: Playing three vendors against each other gives Anthropic extraordinary leverage in negotiations. "Raise your rates and we'll shift workloads elsewhere" is a credible threat when you actually have the software stack to back it up.

Bottom Line

Anthropic's multi-chip strategy sends a clear signal: the AI race is no longer determined by how many NVIDIA GPUs you can stockpile.

By investing in hardware abstraction at the software layer — compilers, frameworks, and architecture design — they've turned what most companies treat as vendor lock-in into a strategic advantage. In an industry where NVIDIA dominance feels inevitable, Anthropic's approach stands as the most sophisticated example of building real optionality into AI infrastructure.

Whether this translates to a lasting business advantage depends on execution. But the engineering bet is clear: own the software layer, and the hardware becomes a commodity you can swap.