The news hit my feed at 6:47 AM Berlin time: Anthropic just poached Amir Salek, the man who shepherded seven generations of Google's Tensor Processing Units into production. My first reaction wasn't "Wow, they're building a chip." It was "Wow, they're building a new kind of trust architecture."
Let me explain. I've spent the last six years watching crypto projects chase the holy grail of decentralized compute — from Filecoin's storage proofs to render networks that promised to democratize GPU access. But the real story was always hiding in plain sight: the AI labs that consume the most compute are the ones rewriting the rules of hardware. And now, Anthropic is stepping into that arena.
Context: The Multi-Vendor Trap and the Race for Independence
Anthropic currently sources chips from NVIDIA, Google Cloud, and Amazon. It's a smart hedge — don't put all your H100s in one basket. But as any DeFi native knows, diversification isn't the same as sovereignty. When you rely on external suppliers for your core compute, you're renting the rails, not owning them. Every time NVIDIA delays a new architecture or Google reallocates TPU capacity to its own Gemini models, Anthropic's training pipeline gets squeezed. The 2022 chip shortage taught us that liquidity isn't just about capital; it's about the ability to scale your compute on your own terms.
Salek's background is the key signal. He didn't just design chips; he scaled them across Google's entire data center ecosystem. He understands the full stack: from the transistor-level architecture of the TPU to the cooling infrastructure that keeps 10,000 chips running in sync. That's the kind of institutional knowledge that can't be bought on a whiteboard. It's the difference between a toy chip and a production-grade accelerator that powers the world's most advanced models.
OpenAI already struck first with its Jalapeno chip, developed in partnership with Broadcom. Now Anthropic is following suit. But here's the nuance: this isn't about building a GPU competitor. It's about building a custom silicon that is deeply coupled with Claude's model architecture. Think of it as vertical integration for the AI age — model + inference engine + chip = a moat that's hard to replicate.
Core: The Decentralized Compute Stack Meets Institutional Reality
Let me break this down through the lens of my own experience. In 2020, during DeFi summer, I audited over 150 Uniswap V2 liquidity pools. I saw firsthand how a small vulnerability in slippage calculation could cascade into a $2 million loss. The lesson was clear: the difference between a robust system and a fragile one often lies in the layers you can't see. The same applies to AI compute.
Anthropic's move is about controlling the invisible layers — the interconnect topology, the memory bandwidth, the power efficiency. These aren't sexy. They don't make headlines. But they determine whether Claude can process a 200,000 token context window without hallucinating, or whether a multi-agent system can coordinate in real-time without hitting a latency wall.
From a values perspective, this is both exciting and troubling. On one hand, custom silicon could enable more efficient, less energy-intensive AI — a win for the environment and for accessibility. On the other hand, it concentrates power in the hands of those who can afford to build their own chips. The barrier to entry for training cutting-edge models just got higher. We didn't build a future; we built a mirror of the centralized power we sought to escape.
But let's get technical. Salek's expertise in ASIC/DSA design means Anthropic can optimize for specific workloads: training dense transformers, long-context inference, multimodal reasoning. The company's current multi-vendor strategy suggests they're not abandoning NVIDIA overnight. Instead, think of the custom chip as a high-performance lane on a multi-lane highway. For the most expensive, most latency-sensitive tasks, they'll use their own silicon. For everything else, they'll keep renting from the cloud giants.
This is where the contrarian angle bites.
Contrarian: The Myth of the Chip Savior
Everyone wants to believe that a custom chip will solve Anthropic's scaling woes. But the reality is messier. Building a competitive AI accelerator takes 3-5 years and billions of dollars. The first tape-out almost always underperforms. Google's first TPU was a inference-only beast that couldn't train a model. OpenAI's Jalapeno is still in early stages. And even if Anthropic's chip works, it needs its own software stack, driver support, and integration with existing frameworks like PyTorch and JAX. That's a massive engineering lift.
Here's the blind spot most analysts miss: the real bottleneck isn't raw compute — it's the communication fabric. When you train a 70-billion-parameter model, you're not just multiplying matrices; you're moving terabytes of data between thousands of chips. The interconnect is the hidden choke point. NVIDIA's NVLink and InfiniBand are proprietary and deeply entrenched. Building a custom alternative that achieves comparable bandwidth and latency is a moonshot. Salek has done it at Google, but that was at a company with infinite resources and a decade of experience.
Moreover, the self-chip project could become a distraction. Anthropic's core competency is model alignment and safety research. Diverting engineering talent to hardware design risks slowing down Claude's next-generation capabilities. In the crypto world, we saw this with projects that tried to build everything — consensus, storage, compute — and ended up with nothing. Focus is a feature.
Takeaway: The Frontier Is the Stack
Mining for truth in the noise of AI chip mania, I see a clear signal: the future of AI is not just models; it's the entire stack from silicon to soul. Anthropic's hiring of Salek is a declaration that they intend to own their compute destiny. But the road is long, and the risks are real.
If I were a developer or investor tracking this space, I'd watch for three signals: (1) whether Anthropic announces a chip partner (Broadcom, Marvell, or even AMD), (2) the first public benchmark comparing their custom silicon to NVIDIA B200 in inference latency, and (3) any change in their API pricing — a cost reduction that could reshape the competitive landscape.
Until then, remember: open source is not a license; it's a state of mind. And true decentralization requires not just code, but the infrastructure to run it without permission. Anthropic is building that infrastructure. Whether it will be a garden or a fortress remains to be seen.
