- Anthropic has hired Amir Salek, a founder of the custom chip program at Alphabet’s Google, to begin work on its own AI semiconductors.
- The move signals an intent to reduce reliance on third-party accelerators from Nvidia, Google, and Amazon.
- Bloomberg reported the hire on August 21, 2026; Anthropic has not published a chip roadmap.
- Designing custom silicon is a multi-year, multi-billion-dollar undertaking that few AI labs have attempted directly.
What Happened
Anthropic has hired Amir Salek, described by Bloomberg as “a founder of the custom chip program at Alphabet Inc.’s Google,” as the company lays the groundwork for a push into making its own semiconductors. The hire, reported on August 21, 2026, places a veteran of one of the industry’s most successful in-house accelerator efforts at the center of Anthropic’s hardware ambitions.
Anthropic has not published a chip architecture or timeline, and framed the effort as early groundwork rather than a shipping product. The signal is the seniority of the hire: Salek helped start the program behind Google’s Tensor Processing Units, the accelerators that let Google train and serve large models without buying Nvidia hardware.
Why It Matters
Frontier model training is gated by access to AI accelerators, and Anthropic has publicly leaned on Google’s Tensor Processing Units and Amazon’s Trainium chips alongside Nvidia GPUs. Its expanded partnership with Amazon, which included an $8 billion total investment and the large Trainium cluster Amazon calls Project Rainier, already tied Anthropic to custom silicon it does not control. Designing its own chips would extend that logic one step further, from renting bespoke hardware to owning the design.
The move mirrors the path OpenAI has taken with Broadcom on a bespoke inference chip, and echoes the vertical integration that Google (TPU), Amazon (Trainium and Inferentia), Microsoft (Maia), and Meta (MTIA) each pursued with in-house data-center accelerators. Controlling the design lets a lab tune hardware to its own model architectures and, over time, lower the cost per token of both training and inference — the single largest line item in a frontier lab’s budget.
Technical Details
Custom AI chips are application-specific integrated circuits (ASICs) optimized for the dense matrix multiplication that dominates transformer workloads, trading the flexibility of a general-purpose GPU for higher throughput per watt on a narrower set of operations. Google’s TPU program, where Salek worked, is the clearest example of that trade paying off at scale across many chip generations. Labs typically target inference silicon first, because serving a fixed model is a more stable design target than training the next one.
A first-generation design typically takes two to three years from architecture to volume production and requires a foundry partner such as TSMC, plus committed advanced-packaging and high-bandwidth-memory capacity that is itself in short supply. That timeline means any Anthropic-designed chip is years, not months, away, and would not relieve near-term compute constraints.
Who’s Affected
Nvidia, still the default supplier of training hardware, faces the prospect of another large customer designing around it over time, though its GPUs remain the near-term standard. Google and Amazon remain both suppliers and cloud partners to Anthropic, complicating the relationship. For Anthropic’s engineering organization, the hire signals a durable hardware team rather than a one-off, and for enterprise customers of Claude it points to a longer-term effort to control compute cost and supply.
What’s Next
Anthropic has not disclosed a chip architecture, timeline, or foundry partner. The near-term signals to watch are additional silicon hires, a stated foundry relationship, and whether Anthropic’s future compute deals shift from buying third-party accelerators toward co-designing them.
