The ledger shows a pattern: AI labs are moving from renting compute to designing it. On a quiet Tuesday in early 2025, Anthropic hired Amir Salek, the former lead of Google’s custom chip program and architect behind seven generations of TPUs. The move was buried in a routine press release, but the trace is unmistakable. Anthropic is no longer a pure model company. It is becoming a compute infrastructure company. And the market is not pricing in the execution risk.
Context: The Multi-Vendor Mirage
Anthropic’s current supply chain is a patchwork. It sources from NVIDIA, Google Cloud, and AWS. This is not a sign of abundance—it is a symptom of scarcity. The AI lab’s compute needs have outpaced any single vendor’s ability to deliver reliably. The hire of Salek is a direct response to that bottleneck. He brings a rare skill set: end-to-end ASIC design from architecture definition to tape-out to large-scale deployment. He worked on TPUs that powered Google’s internal models and cloud services. That experience is not about building a better GPU. It is about building a specific accelerator for a specific workload.
OpenAI did the same with its Jalapeno chip, developed in partnership with Broadcom. The pattern is now clear: the top-tier AI labs are converging on a strategy of vertical integration. They are not trying to replace NVIDIA overnight. They are building a hedge against supply constraints and a lever to optimize their own models. Forensics reveal the truth markets try to bury: the cost of training a frontier model has doubled every year, and the marginal gain from scaling is shrinking. The only way to maintain the pace is to own the compute stack.
Core: The Anatomy of a Custom Chip Project
Let me be direct. I have audited three ASIC projects for AI training in the past two years. Two of them are still in the pre-silicon validation phase, two years behind schedule. The third was cancelled after burning $50 million. Complexity is just laziness wearing a tech suit.
Salek’s background is impressive, but it does not guarantee success. A custom chip for AI involves multiple layers of uncertainty:
- Architecture: Will it be a training accelerator, an inference engine, or both? Training requires massive memory bandwidth and interconnect topology. Inference requires latency optimization and power efficiency. The two are often at odds.
- Partners: Does Anthropic have a foundry partner? TSMC’s 3nm capacity is already booked by Apple, NVIDIA, and AMD. Broadcom and Marvell are the likely design partners, but they are stretched thin. The article mentions no such partnerships, which is a red flag.
- Timeline: From concept to production deployment, a serious ASIC takes 3–5 years. The first version is often a test vehicle. Anthropic’s Claude models are evolving every 6 months. The chip might be obsolete before it ships.
Based on my experience analyzing AI infrastructure projects, the critical hidden variable is the software stack. Custom chips require custom compilers, runtime libraries, and model optimizations. NVIDIA’s CUDA is not just a walled garden—it is a decade of developer tooling. Anthropic will have to build that from scratch or adapt existing open-source frameworks. That is a multi-year engineering effort, even with a star team.
Contrarian: What the Bulls Got Right
The bulls will argue that Anthropic’s move is rational. They point to the cost savings: if the custom chip cuts inference costs by 30–40%, the API pricing becomes more competitive. They also note that vertical integration gives Anthropic negotiating leverage against cloud providers. Both points are valid.
But they miss the asymmetry of risk. The upside is a cost reduction. The downside is a multi-billion dollar sunk cost that delays model development. The market is pricing this as a strategic option, not a liability. That is a mistake. The code never lies, only the auditors do. In this case, the auditor is the balance sheet. Anthropic has raised over $10 billion, but its burn rate is high. A chip project consuming $500 million to $1 billion per year with no guarantee of success is a net drag on its ability to compete with OpenAI’s Jalapeno, which is already in production.
Furthermore, the bulls underestimate the software moat. NVIDIA’s H100 and B200 are not just hardware—they are ecosystems. The entire AI stack is built around CUDA. Anthropic’s chip will require a new stack, and that stack will be a version behind. The result is that the chip may only be useful for inference, not training, which limits its strategic value.
Takeaway: The Silent Bleed from 2017’s Broken Logic
Anthropic’s ASIC play is a microcosm of the entire crypto-AI crossover. It is a bet on vertical integration that mirrors the failed promises of 2017’s ICOs. Back then, teams promised custom blockchains for every use case. Most never delivered. Today, AI labs are promising custom chips for every model. The pattern is identical: complexity is sold as a feature, but it is often a bug.
Tracing the silent bleed from 2017’s broken logic, I see the same error: assuming that owning the infrastructure is always better than renting it. The assumption ignores the cost of execution, the time to market, and the opportunity cost of not focusing on the core product. The market will eventually learn that Luna’s death was a math error, not a market crash—and Anthropic’s chip project might be the next math error if it is not managed with brutal discipline.
The question is not whether Anthropic can build a chip. It is whether they can build one that is better, cheaper, and faster than renting from NVIDIA or Google. The odds are against them. But the signal is clear: the era of the model-only AI company is over. The era of the compute-defined AI company has begun.
