The ledger remembers what the market forgets. On February 27, 2025, Anthropic announced the hiring of Amir Salek, the former Google TPU leader who oversaw the first seven generations of Tensor Processing Units. The news didn't trigger a price spike in any token, nor did it dominate crypto Twitter. But for anyone who audits DeFi protocols for a living, this hire is a stress test of an entirely different kind—one that fractures the assumption that AI model companies will remain pure software players.
Salek's resume is not a list of academic papers. It is a product roadmap: chip architecture, compiler design, data center deployment, and the software stack that turns silicon into usable compute. He didn't just design TPUs; he shipped them. Google's TPU v1 through v7 are the only custom AI accelerators that have been validated at production scale outside of NVIDIA's CUDA ecosystem. By hiring the person who built that, Anthropic is signaling that it intends to move from buying compute to defining compute.
Context: The Multi-Source Dependency Trap
To understand the significance of this hire, we must first audit Anthropic's current compute supply chain. The company today sources chips from at least three vendors: NVIDIA (H100/B200 for training and inference), Google (TPU via Google Cloud), and Amazon (Trainium/Inferentia via AWS). This diversification is not a sign of strength—it is a hedging strategy. Each vendor locks Anthropic into a specific architecture, software stack, and pricing model. The cost of switching between them is not zero; it includes retraining models, rewriting kernels, and renegotiating contracts.
More critically, the market for AI chips is structurally constrained. NVIDIA's H100 supply remains tight, with lead times stretching beyond 12 months for large orders. Google's TPU is only available through its cloud, creating a single-vendor dependency if Anthropic deepens that relationship. Amazon's custom chips are still catching up in performance, especially for training large models. The result is that Anthropic, despite being a leading AI company, has limited control over its own cost structure and scalability.
OpenAI recognized this problem earlier. Its Jalapeno project, developed in partnership with Broadcom, is a custom inference chip expected to deploy in late 2025. Jalapeno is not a general-purpose GPU; it is an ASIC optimized for transformer inference, particularly for long-context and multi-turn reasoning workloads. The technical details are sparse, but leaked benchmarks suggest a 30-40% cost-per-token improvement over H100 inference when running GPT-4-class models. This is not a rumor—it is a competitive threat. If OpenAI can offer cheaper API calls, Anthropic’s Claude loses pricing parity.
Core: Deconstructing the Amir Salek Hire
Let me be precise about what this hire means at the code and protocol level. I have audited smart contracts for five years, and I have learned that the most dangerous vulnerabilities are not in the execution logic but in the implicit assumptions about the underlying infrastructure. The same principle applies to AI chips: the real risk is not in the architecture design but in the assumption that general-purpose GPUs will always be the optimal substrate.
Salek's experience spans the entire TPU stack: chip microarchitecture, systolic array design, TFLOPS/Watt optimization, XLA compiler, and the TensorFlow/PyTorch integration layers. He knows how to design a chip that is not just fast on paper but fast in practice, because the compiler and runtime are co-designed with the hardware. This is the opposite of NVIDIA's approach, where CUDA is the universal interface and the hardware must be general enough to support any model. A custom chip, by contrast, can sacrifice generality for efficiency on a specific workload.
What workload will Anthropic target? The most likely answer is inference, specifically Claude's inference pipeline. Claude is a large language model with a mixture-of-experts (MoE) architecture, long-context windows (200K tokens), and frequent tool-calling capabilities. Each of these features stresses different parts of the compute stack: - MoE requires sparse activation patterns, which can be efficiently handled by a chip that dynamically routes tokens to the right expert. - Long-context windows require large KV-cache memory and high-bandwidth memory access, which general-purpose GPUs often underutilize. - Tool-calling introduces latency-sensitive inference, where the chip must respond quickly to external API calls.
A custom inference chip could optimize for these specific patterns. For example, an ASIC with a dedicated KV-cache compression unit could reduce memory bandwidth by 50% while maintaining accuracy. A chip that hardcodes the MoE routing logic could reduce latency by 30% compared to a software implementation on H100. These are not speculative numbers; they are documented in academic papers and industry patents. The question is whether Anthropic can execute on this vision.
But let's stress-test the contrarian angle. Self-designed chips are capital-intensive, time-consuming, and risky. The average custom ASIC for AI takes 3-5 years from concept to production, with costs exceeding $100 million for design and tape-out. Even Google, with its engineering depth, took three generations of TPUs to reach competitive performance. Anthropic, as a company founded in 2021, has less than $10 billion in total funding—enough to build chips, but not enough to survive a failed project.
Moreover, the chip is only half the battle. The other half is the software stack: compilers, runtime libraries, network interconnect, and data center integration. NVIDIA's CUDA ecosystem is a decade old and deeply entrenched. Anthropic's chip will need its own CUDA-killer, or at least a seamless integration with existing frameworks. Salek's experience with XLA gives him a head start, but building a production-grade compiler for a new architecture is a multi-year effort.
Contrarian: The Blind Spots Nobody Is Talking About
Most analysis of this hire focuses on the strategic upside: cheaper inference, lower dependency on NVIDIA, greater control. But the market is missing three critical blind spots.
First, the chip project could cannibalize model development. Anthropic's primary competitive advantage is its AI safety research and model performance. If the company diverts engineering talent and capital into chip design, it risks slowing down the next generation of Claude. Google's TPU program never distracted from its core search business because Google could afford to run parallel teams. Anthropic, with a fraction of Google's headcount, cannot do the same. The stress test here is not whether the chip works, but whether the organization can maintain focus on both fronts.
Second, the timeline mismatch between chip and model. AI models are advancing at a rate of 6-12 months per generation. A custom chip that takes 3 years to design will be optimized for the model architecture of 2025, not 2028. By the time the chip ships, the model architecture may have shifted—perhaps to a different attention mechanism, a new sparsity pattern, or a mixture-of-multimodal inputs. The chip becomes a stranded asset. This is a standard risk in hardware-software co-design, but it is amplified in AI where the software (the model) is evolving faster than Moore's Law.
Third, the supply chain exposure does not disappear. Even if Anthropic designs its own chip, it still needs a foundry to manufacture it. TSMC is the only realistic option for advanced nodes (3nm, 2nm), and TSMC's capacity is already oversubscribed by Apple, NVIDIA, AMD, and Broadcom. Anthropic will be a small customer, meaning it will have limited bargaining power and long lead times. The chip's cost advantage could be eroded by foundry pricing and packaging costs. The ledgers of capital expenditure do not lie: the unit economics of a custom chip only work if the volume is high enough to amortize the NRE costs.
Takeaway: The Future Is a Hybrid Compute Stack
Anthropic's move is not a bet against NVIDIA. It is a bet that the future of AI compute will be a heterogeneous mix of general-purpose GPUs, custom ASICs, and cloud-specific accelerators. The most successful AI companies will be those that can orchestrate across these substrates, routing each workload to the most efficient chip. Amir Salek's hire is the first step toward building that orchestration layer.
But the real signal will come in the next 12-18 months: Watch for the release of an inference chip, a partnership with a foundry, or a significant expansion of the semiconductor team. If those signals appear, Anthropic will have validated its transition from a model company to an infrastructure platform. If they do not, the hire will remain a speculative placeholder.
Verification precedes value. The block height does not lie, and neither does the chip tape-out date. The market is waiting for proof, not promises.