Hook
Anthropic just hired the man who built Google's TPU from scratch. Amir Salek, former VP of TPU/Cloud AI at Google, is now heading their chip strategy. This is not a rumor. It's a confirmed hire. And it's the clearest signal yet that Anthropic is moving from 'buying compute' to 'defining compute.'
The market hasn't priced this in. Most headlines will frame it as 'Anthropic enters chip race.' But the real story is about compute dependency. The kind of dependency that kills margin, slows iteration, and hands control to NVIDIA's supply chain.
I've seen this playbook before. In 2020, when Uniswap v2 launched, the real alpha was in understanding the liquidity mechanics—not the code. Similarly, Salek's move is not about the chip itself. It's about the compute liquidity that Anthropic is trying to control.
Context
Let's rewind. The AI chip race is already a bloodbath. OpenAI has Jalapeno—a custom inference chip co-developed with Broadcom, targeting 2026 deployment. Google has TPU, now in its seventh generation. AWS has Trainium and Inferentia. Even Meta has started in-house silicon for recommendation systems.
Anthropic has been the outlier. They've been sourcing chips from NVIDIA, Google Cloud, and AWS—a multi-vendor strategy that screams 'we don't have a long-term plan.' But that's changing. Salek's background is not just chip architecture; it's the entire stack: from silicon design to compiler, to data center deployment, to the software layer that makes the hardware sing.
His resume: 18 years at Google, led the TPU team through seven generations, oversaw the deployment of TPUs across Google's data centers for both training and inference. He knows how to take a chip from a whiteboard to a rack of servers serving millions of requests.
This is not a 'let's explore options' hire. This is a 'we're building a team' hire.
Core
Here's what we know. Anthropic is currently spending a fortune on NVIDIA H100s and B200s. They also use Google's TPUs for some training workloads and AWS's Trainium for inference. The cost structure is insane. Each Claude query burns through GPU cycles. The margins at scale are razor-thin.
Custom silicon is the only way to escape the NVIDIA tax. But the path is not about building a general-purpose GPU. That would be suicide. The real play is a custom ASIC—an accelerator designed specifically for Claude's model architecture.
What does that look like? Claude is a mixture-of-experts (MoE) model with heavy reliance on long-context attention and KV cache. The inference bottleneck is not compute—it's memory bandwidth and cache efficiency. A custom chip could hardwire the MoE routing logic, reduce the overhead of sparse activation, and integrate a massive on-chip SRAM for KV cache.
Compare this to OpenAI's Jalapeno. Jalapeno is rumored to be a inference-only chip with a focus on transformer decoding. It's designed to reduce the cost per token by 5-10x over H100. That's the target—not raw FLOPS, but cost per token.
Anthropic's chip will likely aim for the same goal. But with one critical difference: they can co-design the chip with the model. Claude's architecture is evolving. If they can lock in the next version of Claude to exploit specific hardware features (e.g., sparse attention, variable precision, dedicated tensor cores for expert routing), they can achieve a level of efficiency that NVIDIA's general-purpose GPUs can't match.
This is the holy grail of AI infrastructure: vertical integration from model to silicon.
Speed beats analysis when the graph is vertical. Salek's hiring is the first step. But the real work starts now. The chip team needs to grow from a handful of people to hundreds. They need to lock in a foundry partner (TSMC, Samsung, or Intel?). They need to design the chip, tape it out, debug it, and get it into production. That's a 3-5 year timeline, minimum.
But here's the kicker: Anthropic doesn't have to build a full chip right away. The first signal to watch is whether they announce a 'co-design' partnership with a cloud provider. For example, they could work with AWS to design a custom version of Trainium optimized for Claude. That would be a faster path to value than building from scratch.
I don't read whitepapers; I read order books. So what does the order book say? Anthropic is still buying NVIDIA chips. They just signed a multi-billion dollar deal with Google Cloud for TPU access. But they are also hiring chip architects. That's the classic 'hedge' strategy: keep buying off-the-shelf while building the in-house alternative.
The true test will be the next 18 months. If Anthropic announces a first-generation chip target for 2027, they are serious. If they only talk about 'exploring' chip design, it's a talent grab.
Contrarian
Everyone is framing this as Anthropic taking on NVIDIA. That's the wrong narrative. The real story is about Anthropic's dependency problem. They are currently at the mercy of three players: NVIDIA for supply, Google for cloud, and Amazon for alternative compute. That's a fragile position. A single supply chain disruption—like the 2022 GPU shortage—could cripple their ability to train and serve models.
Custom silicon is a hedge, not a silver bullet. It's about reducing dependency, not about winning a chip war.
But here's the contrarian angle: self-designed chips are a massive capital drain. The cost of a single tape-out at 3nm is over $50 million. The team cost is another $100 million per year. If the chip doesn't deliver a significant cost advantage over H100s or B200s, it's a net negative. The ROI equation is brutal.
The best news is the news that moves the price. And the price of Anthropic's chips is not just money—it's opportunity cost. Every dollar spent on chip development is a dollar not spent on model research, safety, or go-to-market. If the chip project delays Claude 5 by six months, that's a strategic loss that no hardware efficiency can compensate.
OpenAI's Jalapeno is a cautionary tale: it's been in development for over two years and still not in production. The timeline slips. The costs balloon. The model architecture changes, and the chip becomes obsolete before it's deployed.
Anthropic's move is defensive, not offensive. They are trying to avoid the 'vendor lock-in' trap that has killed many AI companies before. But the cure might be worse than the disease.
Takeaway
What to watch? Three signals: hiring pipeline (chip architects, compiler engineers, data center specialists), foundry partnerships (TSMC or Broadcom?), and first chip target (inference or training?). If Anthropic announces a chip within 12 months, it's likely a rebranded custom ASIC from a partner. If they go silent, they are still in the 'exploration' phase.
Either way, the next 18 months will define whether Anthropic becomes a true AI infrastructure platform or remains a model company renting compute from others. The cheetah is sprinting—but the finish line is years away.