In the quiet corridors of competitive advantage, where algorithms meet the brute physics of silicon, a single name can speak volumes. Late last week, the news rippled through my network not as a shock, but as a confirmation of a trajectory I had been tracking for months. Anthropic, the company that taught us to speak of AI safety with a moral urgency, has hired Amir Salek, the man who spent a decade shepherding Google’s TPU through its first seven generations. It was not the kind of move that gets a standing ovation on crypto twitter, but for those of us who curate the soul of infrastructure, it was the click of a lock turning. Anthropic is no longer just buying power; it is learning to forge it.
The context here is not just about one company’s supply chain. We are witnessing the final maturation of the AI industry. For years, the conventional wisdom held that the model was the moat. You build a better transformer, you win. But the 2024-2025 cycle has revealed a deeper truth: the model is a fraction of the cost, but the compute is the empire. OpenAI’s ‘Jalapeno’ project, a custom chip developed with Broadcom, is already out of the lab and moving into the data center. Google has had its TPU for years. AWS has Trainium and Inferentia. The only major player left relying on the kindness of strangers and the current availability of NVIDIA GPUs was Anthropic. This hire is the admission that the era of the ‘pure model company’ is a fairy tale.
Based on my audit of the signals, Salek’s presence is not about building a general-purpose GPU to rival NVIDIA. That would be a fool’s errand. Instead, the technical trajectory points to a targeted, bespoke accelerator. The move is a 'curation' of raw compute. Look at the mathematics: Anthropic’s Gemini, I mean, Claude models are notorious for their long-context windows and complex reasoning paths. These workloads are not the standard matrix multiplications that NVIDIA optimizes for; they involve heavy memory bandwidth and KV cache management. A generic GPU wastes energy on the pixel-crunching and geometry that an LLM never sees. A custom chip, however, can be sculpted to the exact silhouette of the Claude architecture. It can prioritize the sparse activation of Mixture-of-Experts models, accelerate the attention mechanisms, and manage the memory hierarchy for those 200k-token contexts.
This is the "curating the soul in a world of derivative clones" principle applied to hardware. The goal is not to sell chips but to achieve a specific economic outcome. If Anthropic can cut the unit cost per token by a significant margin—let’ say 30% to 40%—they alter the landscape of their API pricing. They can undercut competitors or simply buy back their own margin to fuel further research. This is a cost structure play, not a revenue play. The economics are clear: when you have the scale of a leading model, a 10% reduction in inference cost is the equivalent of a massive funding round that doesn’t require the paperwork.
However, here lies the contrarian, often unspoken, part of the ledger. The true bottleneck of the TPU isn’t the chip itself; it’s the software ecosystem. Google spent billions on the compiler, the XLA, and the on-chip networking that makes the TPU actually usable. In my experience working with early TPU deployments, the hardware was often 5 years ahead of the software, and it was the software that ate the returns. Anthropic’s challenge is not just designing a silicon die; it is building the entire stack around it—the custom kernels, the low-level libraries, the cluster networking that allows 100,000 chips to act as one coherent brain. They are entering a race where they are building the track and the car at the same time. This is where projects go to die.
Furthermore, let’s not forget the diplomatic tightrope. Anthropic has a deep partnership with AWS. Amazon has its own silicon. They are also big customers of Google Cloud. If Anthropic goes full custom, they are effectively telling their landlords, ‘I don’t need your most premium product anymore.’ This shifts the power dynamic. The cloud providers will still host the machines, but they will be less critical as a source of the chip itself. This could lead to a renegotiation of contracts, or worse, a cold war with those providers who see Anthropic as a whale, not a partner. The risk is that they get squeezed on rack space even as they try to decouple on silicon.
I am also, as a writer, driven by a cautious heart. I remember the Ethereum ASIC mania of the last cycle. The promise was that custom silicon would finally decentralize the hash. In reality, it centralize it into the hands of the few who could afford the supply chain. Anthropic is entering the same terrain. They are building a moat that very few can ever cross. This will further the gap between the top three labs and the rest of the world. It might be good for Anthropic’s treasury, but it makes the entire sector less equitable, less a public. The fear is that this is not about ‘alignment’ or safety, but about building an infrastructure that is so capital intensive that only the Titans survive.
The more likely outcome, however, is a hybrid. In the next 12 to 18 months, I expect to see a ‘mixed compute’ strategy. Anthropic will continue to buy NVIDIA’s H100s and B200s for those massive training runs. But the majority of the inference, the part that interacts with users, will run on the custom chip. This creates a stratified supply chain. The training becomes a golden goose of capital expenditure, while the inference becomes the engine of the margin. The first chip they deploy will not be a home run; it will be a single-use, high-efficiency token factory. If they can get the power efficiency right, they can place these accelerators in smaller data centers, bringing the compute closer to the data, reducing latency for the users in the Asia-Pacific regions.
So, what does this mean for the future? The hiring of Amir is not the event. The event will be the hiring of the compiler engineers, the networking architects, and the data center cooling specialists over the next six months. If I see the team grow from 20 to 200 people, the signal is confirmed. If I see a partnership with a foundry like TSMC or a design house like Marvell, the signal is confirmed. If I see the emergence of a new API pricing model that reflects a radically lower cost per token, the signal is confirmed. If I see nothing but silence, then we can chalk this up to a pre-emptive defensive move against OpenAI’s Jolapeno.
We are entering a time where the writer and the architect are the same. Anthropic is building the infrastructure for its own existence, not just the code. This is a beautiful, and frightening, act of vertical integration. I am sure they will succeed in building the chip. The question is, will the success of the chip allow them to build a company that can survive its own complexity? The answer will be written in the silicon, not in the whitepaper. I am watching the market for the evidence of the soul of the machine, and for the first time, I see it in the shape of a hiring, not in the training run. The future is not just about the weight of the weights, but the skeleton that holds them. Curating the soul in a world of derivative clones has just begun.
