Table of Contents
The move shifts from rumor to a dedicated in-house silicon effort
In sum – what we know:
- Dedicated silicon team – Anthropic is assembling a full-stack hardware and software team to design custom AI processors specifically for Claude.
- Inference-first focus – The new chips target query serving rather than model training, leveraging predictable workloads to maximize efficiency and cost savings.
- Multi-chip strategy – Anthropic will maintain partnerships with AWS, Google, and Nvidia while exploring Samsung for advanced 2nm manufacturing of its custom ASICs.
Anthropic is officially building an in-house team to design custom AI chips for its Claude models. — something that was already rumored, but has now been confirmed. The move marks a decisive shift from the exploratory phase Reuters first reported back in April, when sources stressed the plans were preliminary and Anthropic might never commit to a design at all. Now, there’s a dedicated silicon team being assembled, with hiring underway across the hardware and software stack.
It does, of course, make sense. Anthropic serves billions of tokens daily against a revenue run-rate cited around $47 billion, and at that scale, shaving even a fraction off the cost of every query becomes economically compelling. The company says the goal is to create chips that let Claude run faster and more efficiently at the scale customers need.
Hardware strategy and co-design
The planned hardware is a set of custom AI processors tailored specifically for Claude’s workloads rather than general-purpose GPUs. The emphasis is inference-first. These chips are meant to serve queries, not train models, and training will likely continue to lean on GPUs and other accelerators for the foreseeable future. That’s a sensible split. Inference is where Anthropic can guarantee utilization and predictable workloads, which is exactly where a custom ASIC pays off.
The more interesting piece is the design philosophy. Anthropic describes the effort as “co-design” — building the chip and Claude together so each shapes the other, rather than treating hardware as a fixed external constraint the model has to work around. It’s the same logic behind Google’s TPU program and OpenAI’s collaboration with Broadcom. When you know exactly what model you’re serving, you can strip out everything a general-purpose GPU carries for flexibility’s sake and spend that silicon budget on what actually matters.
Against Nvidia and AMD GPUs, the expected wins are the usual ones for purpose-built silicon — better performance-per-watt, lower latency, and reduced cost per FLOP. Anthropic markets itself around “constitutional AI” and safety, and custom hardware could eventually incorporate features that enforce or optimize safety constraints at the hardware level.
Multi-silicon strategy and partnerships
To be clear, this isn’t Anthropic walking away from its existing suppliers — at least not yet. The company has stressed it will maintain a “multi-chip” strategy, and its current hardware dependencies are enormous. Amazon previously announced that Anthropic would use up to one million of its custom Trainium2 chips under — and Amazon has invested billions in the company, so that relationship isn’t going anywhere. Google TPUs and Nvidia and AMD GPUs stay in the mix too, running alongside whatever in-house silicon eventually materializes.
On the manufacturing side, things are still very early, too. Multiple reports say Anthropic has held preliminary discussions with Samsung Electronics about using its advanced 2-nanometer process and packaging technology. Those talks are exploratory — there’s no formal manufacturing agreement and no taped-out design as of early August. Fabrication will almost certainly be outsourced to a foundry like Samsung rather than done internally, since Anthropic has neither the capital nor the reason to build fabs. But at this stage, the company is still defining specifications, power requirements, and server cluster configurations. The foundational architectural decisions haven’t been made yet.
Everyone is making chips now
Anthropic’s move lands in an increasingly crowded race. OpenAI unveiled its own custom AI chip called “Jalapeño,”developed with Broadcom, a few months ago, and it mirrors what Google, Amazon, and Meta have been doing for years— designing custom accelerators to reduce dependence on Nvidia’s ecosystem and tailor hardware to their own models. The proximate cause is the same across the board. Advanced AI chips remain in short supply, prices keep climbing, and anyone serving inference at scale is watching hardware costs eat into margins.
If Anthropic pulls this off, the upside is obvious. Drastically lower per-query costs would let the company price its enterprise products more aggressively or simply keep the margin. Custom silicon also buys strategic leverage — it’s a lot easier to negotiate with cloud providers and chip vendors when you have a credible alternative in your back pocket.
That said, designing the chips, of course, won’t be cheap. Designing a cutting-edge ASIC on a 2nm node is enormously complex and capital intensive. There’s also a delicate balancing act with partners — AWS, Google Cloud, Nvidia — who are simultaneously suppliers, investors, and in some cases competitors. Pursuing a chip that could eventually reduce dependence on those relationships, without straining them in the meantime, is a genuine strategic tension.
And the scope may stay narrower than the ambition. Whether these chips ever expand into training workloads is unclear, since training demands the high memory bandwidth and interconnect scale where Nvidia’s ecosystem still dominates. Some observers expect both Anthropic and OpenAI to ship first-generation ASICs within roughly a year, but that timeline is speculative.
