Hook
On a quiet Tuesday in Taipei, the crypto noise was mostly about ETH’s inability to break $2,400 and another failed DAO proposal. Then I saw the news: Amir Salek, the engineer who led seven generations of Google’s TPU, joins Anthropic to lead custom chip development.
This isn’t a headline about a new ASIC. It’s a narrative shift.
In the same way that a reentrancy vulnerability in TheDAO exposed a deeper trust crisis, Salek’s move reveals that the real battle in AI is no longer model architecture—it’s who controls the compute. And for those of us who track the intersection of code, culture, and capital, this is the kind of signal that demands a deeper read.
Context
Anthropic, the company behind Claude, has been a multi-source compute buyer—NVIDIA H100s, Google TPUs, AWS Trainium. That’s a smart hedging strategy for a startup that burned through $1.8B in 2024 alone. But the narrative of ‘rent everything’ is fragile. OpenAI already launched its own chip, internally codenamed “Jalapeno,” in partnership with Broadcom. Google has its TPU empire. Microsoft is rumored to be designing its own AI accelerator.
The pattern is clear: head AI labs are moving from “buying compute” to “defining compute.” This is not a new story—it’s the same narrative arc we saw in the 1980s when Ford started building its own steel mills, or in the 2010s when Apple transitioned from Intel to M-series chips. The logic: if your core product depends on a scarce resource, you eventually try to own that resource.
But here’s the nuance that most coverages miss. Anthropic isn’t trying to build a general-purpose GPU to compete with NVIDIA. That would be suicide. Instead, they are likely building a domain-specific accelerator (DSA) optimized for Claude’s inference patterns—long context, multi-modal, agentic workflows. The proof is in the hire: Salek’s TPU experience is about large-scale, purpose-built accelerators, not about competing in the CUDA ecosystem.
Where code meets culture, the real value emerges.
Core
The core insight is not about the chip itself. It’s about the narrative of vertical integration.
Let me explain through the lens of my own experience. In 2020, I audited a DeFi protocol that promised 10,000% APY through liquidity mining. The code was fine, but the narrative was flawed: the APY came from inflating the token supply, not from real yield. When the incentives stopped, the TVL evaporated. The same principle applies here. Anthropic’s chip strategy is a bet on long-term cost reduction, but the immediate narrative value is independence.
From the parsed data, the technical roadmap is clear: Salek will oversee the entire stack—from architecture definition to tape-out to data center deployment. The hiring signals that Anthropic is serious about reducing reliance on NVIDIA, Google Cloud, and AWS. But the real story is the hidden information: Anthropic is likely evaluating a “custom chip + custom data center” combo. That’s a massive capital expenditure, but it also creates a new moat.
Let’s break down the technical dimensions:
- Training vs. Inference: The chip will probably prioritize inference. Why? Because inference is where the recurring costs are. Claude’s long-context capabilities (up to 200k tokens) are memory-bandwidth hungry. A custom ASIC could optimize for HBM3e stacks and efficient attention mechanisms, cutting inference cost by 30-50%.
- Interconnect: Standard Ethernet is too slow for large model parallelism. Anthropic might design a custom interconnect (like NVLink or InfiniBand replacement) to reduce communication overhead. This is where Salek’s TPU experience shines—TPUs use a custom torus topology.
- Power Efficiency: With data center power constraints becoming a geopolitical issue, a custom chip that delivers 2x more tokens per watt could be a decisive advantage.
But here’s the contrarian view that most analysts ignore: this is a high-risk, long-payoff project. ASIC development cycles are 3-5 years. The budget is likely $500M+ just for the first tape-out. And if the chip doesn’t deliver the expected performance, Anthropic will have wasted precious time and capital that could have been used to improve Claude’s model quality.
Searching for truth in the noise of the network.
Contrarian
The popular narrative is that Anthropic’s chip move will make it more independent and competitive. I think the opposite: it could increase its dependency on a single point of failure—its own hardware.
Let me give you a crypto analogy. In 2022, many L1 blockchains built their own VMs (e.g., Solana’s Sealevel, Aptos’s MoveVM). The idea was to optimize for specific use cases. But the result was fragmentation: developers couldn’t easily port code, and liquidity was siloed. Similarly, if Anthropic builds a custom chip that only runs Claude efficiently, it becomes locked into its own hardware roadmap. If the chip fails to deliver, the entire model strategy is compromised.
Moreover, the move toward vertical integration in AI chips is anti-decentralization. It concentrates compute power in the hands of a few labs. For the blockchain world, where we value permissionless access, this is a concerning trend. Small AI startups will struggle to get competitive hardware, while Anthropic and OpenAI get even more entrenched.
The narrative is the asset; the code is the proof.
The second contrarian angle: this project might not be about reducing costs at all. It could be about negotiation leverage. By signaling that they can build their own chips, Anthropic can go to NVIDIA and say, “Give us better pricing, or we’ll build our own.” That’s a classic bargaining chip. The actual chip might never see production; the threat alone is enough to improve terms with existing suppliers.
Takeaway
So what is the next narrative? It’s not about Anthropic vs. OpenAI. It’s about the “compute stack” as a competitive moat. In the next 12-18 months, we will see a bifurcation: the top 5 AI labs will own their hardware, while everyone else will rent from cloud providers. This will create a new class of “compute scarcity” that could spill into crypto.
Imagine a world where AI compute becomes a tokenized resource—where you can stake tokens to access inference time on Anthropic’s custom chips. That’s not science fiction; it’s the logical extension of what we saw with Filecoin for storage and Render for GPU. The narrative of “compute as a commodity” is dying. The new narrative is “compute as a strategic asset.”
If Anthropic succeeds, it will not just be a model company. It will be a compute infrastructure company. And if it fails, it will be a cautionary tale about the perils of vertical integration.
For now, I’m watching the hiring signals. If Anthropic starts hiring HBM memory designers and data center cooling engineers, the narrative accelerates. If they start hiring salespeople for custom chips, that’s a different story. But as of today, the signal is loud and clear: the era of “rented compute” is ending. The era of “defined compute” is beginning.
And in that noise, I’m searching for the truth.
