Anthropic confirmed on August 5 that it is building an in-house silicon team with a single mandate: co-design Claude and custom chips together. Job postings went live for a Silicon Engineer (salary: $320,000–$485,000, requirement: “must have shipped silicon”) and a Technical Program Manager. Samsung is reportedly in early talks to manufacture the chip on its 2nm process. The stated goal is to cut per-token inference costs by roughly 50%. This is not exploratory R&D — and it follows OpenAI by about two months.
Why Co-Design Is Different From Just Buying Better Hardware #
The standard response to compute demand is to buy more GPUs. Anthropic is betting on something harder and more valuable: designing the chip and the model together so each shapes the other’s architecture.
When Claude runs on Nvidia GPUs today, every operation passes through CUDA — Nvidia’s proprietary compute layer. That abstraction carries a real cost: extra memory transfers, translation overhead, and higher power draw per token. A chip designed specifically around Claude’s inference workload eliminates that abstraction layer. Operations run in the chip’s native instruction set. Fewer memory transfers means lower latency per token and higher throughput per watt — and lower cost per API call as a result.
Apple proved this principle with the M-series. Apple M3 Ultra delivers up to 23× better energy efficiency — measured in tokens per joule — compared to an Nvidia RTX 5090, not because it is faster in raw compute, but because unified memory architecture removes the PCIe bottleneck entirely. Google’s TPU program has run the same playbook since 2013. By the eighth TPU generation, Google split training and inference into separate chips because the two workloads want opposite hardware profiles. That specialization is only possible when you own the silicon.
OpenAI Is Already at Tape-Out. Anthropic Is Still Hiring. #
OpenAI unveiled its Jalapeño inference chip on June 24 — built with Broadcom, manufactured on TSMC’s 3nm process, claiming roughly 50% lower inference cost versus Nvidia GPUs. The development cycle was nine months, which Broadcom calls the fastest ASIC development timeline at this scale. Even so, Jalapeño will not reach full-scale production until the first half of 2028.
Anthropic is at team-formation stage. OpenAI is at tape-out-and-deployment-planning stage. In silicon development, two months on the calendar can translate to two-plus years in production timing. That gap is real, and it means OpenAI may offer materially cheaper inference before Anthropic’s custom chip reaches any customer.
This does not mean Anthropic made the wrong call — it means they made a late call. The competitive pressure from Jalapeño likely accelerated the announcement. Companies at Anthropic’s scale cannot afford to cede the infrastructure layer permanently.
The Timeline Is Longer Than the Headlines Suggest #
Custom chip development runs 3–5 years from team formation to production silicon at scale. Samsung’s 2nm process (the SF2 node, using Gate-All-Around nanosheet transistors) is still ramping and adds timing complexity. OpenAI’s nine-month tape-out was widely described as exceptional — and that project still will not fully serve production traffic until 2028.
Microsoft’s Maia chip offers a cautionary note. Maia 100 was designed primarily for image processing and never powered production AI services at scale. Maia 200 was delayed roughly six months due to design changes and team turnover. Custom silicon is hard, and the failure modes are expensive.
Realistic timeline for Anthropic’s chip to affect what developers pay: 2029–2030, if the program stays on track.
Nvidia Is Not Going Away #
Anthropic is keeping its multi-chip strategy. AWS, Google, Nvidia, and AMD hardware stays central to its infrastructure. Custom silicon will target inference — serving already-trained models in response to live queries — not training, which still requires Nvidia’s scale. This mirrors how OpenAI operates: Jalapeño for inference, Nvidia for training, and more Nvidia agreements signed simultaneously with the chip announcement.
What Developers Should Do Right Now #
Keep building on what exists. Current Claude API pricing reflects Nvidia GPU costs. Custom silicon that materially lowers inference costs is a 2029–2030 story at the earliest.
The $485,000 salary ceiling and the requirement that candidates personally “shipped silicon” tells you this is a serious commitment. Anthropic hired Clive Chan from OpenAI’s silicon team — the group that built Jalapeño. The program is real. But real programs still take years.
The directional signal for developers: Anthropic intends to compete on infrastructure cost, not just model quality. If they execute, Claude API pricing in the 2030 timeframe should be materially lower and inference significantly faster than today. That is worth knowing. It is not worth changing your architecture over today.