When a Google TPU veteran walks into an AI lab, the market hears a different story than the one being told. This week, Anthropic announced the hiring of Amir Salek, the former lead of Google's custom chip program and architect of the first seven TPU generations. The immediate narrative: Anthropic is building its own silicon, following OpenAI's Jalapeno project, and finally escaping the shackles of NVIDIA and cloud dependence. The subtext, however, is far more complex—a story of desperation, misaligned incentives, and the hollow promises that arise when alchemy replaces engineering.
I've seen this narrative before. In 2017, during the ICO boom, I analyzed 42 whitepapers for the Buenos Aires Crypto Circle. Every project that hired a former Goldman Sachs trader or a retired Google engineer promised to reinvent finance, supply chains, or identity. The market bought the story, but the code rarely delivered. Today, Anthropic hires a chip legend, and the narrative repeats: 'We will build our own GPU, control our own destiny, and beat the bear market.' But alchemy fails when the intent is hollow. And the intent here is not technological superiority—it's survival.
Context: The Bear Market Squeeze
Anthropic currently sources its compute from NVIDIA, Google Cloud, and AWS. In a bull market, this multi-vendor strategy ensures flexibility. In a bear market, it becomes a liability. The demand for AI chips outstrips supply, and the waiting lists for H100s stretch into months. Anthropic's API revenue, driven by Claude, needs to scale without margin erosion. The math is simple: every query on a rented GPU eats into the gross margin. The only way to survive is to either cut costs or increase control. Self-designed chips promise both—but only if the promises are backed by execution.
Salek's background is impeccable. He oversaw the architecture definition, tape-out, deployment, and integration of seven generations of TPUs, each pushing the boundaries of die size, memory bandwidth, and thermal design. His experience spans the full ASIC lifecycle, from RTL to data center power delivery. But there is a critical difference between Google's TPU program and Anthropic's chip ambitions: Google had a captive use case, a massive internal revenue stream from search and ads, and years of iterative refinement. Anthropic, in contrast, is a startup burning cash in a bear market, with a single product line and no guarantee of future revenue.
Core: The Narrative Mechanism and the Bear Market Lens
The market interprets this hiring as a signal of strength: 'Anthropic is becoming an infrastructure company.' But the bear market lens reveals a different reading. Over the past seven days, the market cap of AI-focused tokens has dropped 12%, and the funding for AI startups has tightened. When capital is scarce, companies double down on narratives that attract investment. The 'self-designed chip' narrative is a powerful one—it implies technological moat, long-term vision, and independence from the NVIDIA tax. Investors love it. But the ethnographic reality is that chip design is a capital-intensive, time-consuming, and failure-prone endeavor. The average ASIC development cycle is 3-5 years, with costs ranging from $50 million to $500 million depending on the node and complexity. For a company that has already raised over $7 billion, this is not impossible, but it is a significant diversion of resources from the core mission: building better models.
I learned this lesson during the 2020 DeFi Summer. I wrote 'The Yield Farming Fable,' a series of newsletters explaining liquidity mining to non-technical users. The projects that promised to build their own blockchains or layer-2 solutions often failed, not because the technology was impossible, but because the execution was fragmented. They spread their talent thin, chasing infrastructure instead of product. Anthropic's move is a classic case of 'infrastructure envy'—the desire to own the entire stack, even when the stack is not the bottleneck. The real bottleneck for Claude is not GPU availability—it's the quality of training data, the efficiency of the model architecture, and the alignment of the reward function. Custom chips will not solve these problems.
The Technical Architecture: A Deeper Dive
From my MS in Blockchain Engineering, I understand the distinction between a general-purpose processor and a domain-specific accelerator. Salek's expertise is in ASICs designed for tensor operations, specifically for training and inference of large models. The TPU was optimized for Google's internal workloads—Transformer-based NLP models, ranking algorithms, and image recognition. Anthropic's Claude models are also Transformer-based, but with a strong emphasis on long-context reasoning and multi-modal inputs. The optimal chip for Claude would need to support extremely large memory bandwidth (for the 100k+ token context windows), high-bandwidth interconnect (for model parallelism across thousands of chips), and energy-efficient compute (to keep the power bill under control). These are not trivial constraints. They require custom HBM, advanced packaging, and a network topology that rivals the InfiniBand solutions used by NVIDIA.
But here is the hidden truth: chip design is not just about the silicon. It's about the software stack. Google's TPU success is driven by XLA, a compiler that maps TensorFlow operations to TPU instructions. Without a similar compiler for Claude, Anthropic's custom chip will be a paperweight. The company has not announced any software development plans, and the talent pool for ML compiler engineers is even smaller than for chip architects. Alchemy fails when the intent is hollow. If Anthropic's goal is to build a chip that runs Claude three times faster than an H100, it needs not just Salek, but an entire team of compiler engineers, system designers, and data center architects. The hiring of one person, no matter how brilliant, does not constitute a program.
Commercialization and the Bear Market Reality
In a bear market, survival matters more than gains. The reader wants to know if their assets—whether tokens, API credits, or company equity—are safe. Anthropic's chip move introduces a new risk: capital expenditure. The company will need to either raise more money or cut spending elsewhere. Given the current venture climate, raising another $2 billion for a chip project is not guaranteed. The alternative is to partner with a foundry like TSMC or a design house like Broadcom, which would dilute ownership and control. The commercialization analysis is clear: self-designed chips will not become a revenue stream for at least three years, if ever. In the meantime, Anthropic must continue renting GPUs from NVIDIA and Google, paying the same high prices as everyone else. The chip project is a hedge, not a solution.
But there is a second-order effect that the market is missing. If Anthropic succeeds in building a chip that is 20% cheaper per token than NVIDIA's offerings, it could undercut OpenAI on pricing, especially for enterprise customers. This is a plausible narrative. However, the bear market lens asks: what happens if the chip project fails? The opportunity cost is enormous. While Anthropic's best engineers are designing silicon, OpenAI is training GPT-5, and Google is iterating on Gemini. The model gap widens. The chip project becomes a distraction, and the company loses its competitive edge. This is the classic innovator's dilemma: the incumbents (NVIDIA, Google) have the resources to build custom chips without sacrificing model development. The challengers (Anthropic, OpenAI) must choose between the two, and the market is betting they can do both. History suggests otherwise.
Contrarian Angle: The Weakness Signal
The contrarian reading of this narrative is that Anthropic's hiring is a sign of weakness, not strength. The company is being forced to build its own chips because it cannot secure enough GPU supply from its cloud partners. This is a supply chain failure, not a technological leap. The market reads it as 'vertical integration,' but the reality is 'vertical desperation.' The same pattern occurred in the 2017 ICO boom: projects that promised to build their own blockchains often did so because they couldn't get a smart contract deployed on Ethereum's congested network. The result was a proliferation of buggy, insecure chains that never gained traction. Anthropic's chip project may be similarly misguided.
Furthermore, the comparison to OpenAI's Jalapeno chip is misleading. OpenAI has a partnership with Broadcom, a company with decades of ASIC design experience. Anthropic, as of now, has no announced chip partner. The distillation of the narrative is that Anthropic is trying to replicate Google's TPU success without Google's resources. This is a dangerous alchemy. The intent is to control the stack, but the hollow intent is to paper over the supply chain cracks. The market will eventually see through this.
Takeaway: The Next Narrative
When the hype fades, the next narrative will be about the 'chip gap' between AI labs. The companies that survive are those that can execute on their infrastructure promises, not just announce them. Anthropic's hiring of Amir Salek is a bet on the future, but it is a bet that comes with a built-in escape clause: if the chip fails, the narrative can pivot to 'we learned from our mistakes.' The real question is whether Claude can continue to improve without custom silicon. My experience in the 2022 bear market taught me that the most resilient protocols are those that focus on their core product, not on building the entire stack. Celestia's modular approach succeeded because it embraced specialization, not vertical integration. Anthropic would do well to remember that alchemy fails when the intent is hollow. The intent must be genuine, the execution must be relentless, and the resources must be aligned. Otherwise, the chip becomes a monument to hubris, not a tool for survival.