Anthropic hires Google TPU veteran Amir Salek to lead custom AI chip push

Anthropic has hired Amir Salek, a founding executive of Google’s custom Tensor Processing Unit program, to lead its push into in-house semiconductor development. Salek, who led Google’s TPU business until 2022 and oversaw delivery of the first seven generations of the AI-specific chips, will report to James Bradbury on Anthropic’s compute team. The move is the clearest signal yet that the AI lab intends to build a long-term, multi-generation chip program rather than a one-off solution.

The hire matters because control over silicon is rapidly becoming as strategic as model innovation in the frontier AI race. For years, the dominant constraint on training and serving large language models has been access to Nvidia’s GPUs. Now, the largest AI labs are systematically reducing that dependency by designing their own accelerators. Anthropic, which has built its reputation on frontier models like Claude, is signaling that it no longer wants to be a pure consumer of third-party hardware. For executives and founders watching the AI supply chain, this development underscores a broader shift: compute independence is becoming a prerequisite for competing at the highest level.

From Google’s TPU to Anthropic’s compute team

Amir Salek’s background is among the rarest in the semiconductor industry. At Google, he was a founding executive of the TPU program, which began as an internal project to accelerate neural network inference and evolved into a multi-generational family of chips powering everything from search ranking to Gemini training workloads. Leading the TPU business until 2022, Salek oversaw seven generations of the chip, navigating the difficult transition from a research curiosity to a hyperscale production system deployed across Google’s global data centers. The TPU’s journey from a 2015 inference-only accelerator to a training and serving workhorse for Google’s largest models required not just architectural innovation but deep coordination with compiler teams, datacenter operations, and machine learning researchers. That operational knowledge is exactly what Anthropic needs as it moves from buying compute to building it.

At Anthropic, Salek will work under James Bradbury, who leads the company’s compute team. Bradbury himself has deep infrastructure experience, having worked on machine learning systems at Google and later at AWS before joining Anthropic. The pairing suggests Anthropic is assembling a team with direct experience in both the architectural design and the operational scaling of custom silicon. That combination is critical: building a chip is one challenge, but deploying it reliably across thousands of servers for training runs that last months is another entirely. A chip that works in simulation but fails under sustained thermal load or memory bandwidth pressure is useless for frontier model training. Salek’s tenure at Google spanned exactly those deployment challenges, from early TPU pods to the liquid-cooled TPU v4 systems that power some of Google’s largest AI workloads.

The hire does not come with a public product roadmap. Anthropic has not disclosed a target architecture, a manufacturing partner, or a timeline for deployment. Industry norms suggest that a serious chip program takes three to five years from team formation to volume deployment. Salek’s arrival marks the beginning of that clock, not the end of it. Still, the decision to bring in a leader with his specific track record indicates that Anthropic is planning for multiple chip generations, not a single experimental design. A one-off accelerator would not justify recruiting someone who spent nearly a decade building a sustained silicon roadmap at Google. The implicit message is that Anthropic intends to compete on hardware over the long term, with successive chip generations optimized for its own model architectures and serving patterns.

A crowded field of AI chip builders

Anthropic is not alone in pursuing custom silicon. Google’s TPU program remains the most mature, now spanning multiple generations and powering a significant share of Google Cloud’s AI offerings. Amazon has built its Trainium and Inferentia lines to reduce dependence on Nvidia for AWS workloads, with Trainium2 now deployed in EC2 UltraClusters for large-scale training. Meta has developed its MTIA chips for recommendation and inference, targeting the massive serving workloads that run across its family of apps. Microsoft has introduced Maia accelerators for Azure AI services, initially focused on internal OpenAI workloads and Copilot inference. OpenAI, too, is exploring custom silicon through partnerships, according to multiple reports, though it has not publicly committed to an in-house chip team of Anthropic’s scale.

What distinguishes Anthropic’s move is the timing and the company’s position in the market. Anthropic is smaller than Google, Amazon, or Microsoft, and it does not operate a public cloud with millions of customers to amortize chip development costs. Its hardware bet is therefore more concentrated. Anthropic must justify the investment purely through the efficiency gains and cost reductions it can achieve for its own model training and inference. That is a riskier proposition, but it also means the company can optimize its chips specifically for its own model architectures and serving patterns, without needing to satisfy a broad range of external cloud workloads. A hyperscaler designing a chip must serve thousands of customer use cases; Anthropic can design for Claude and its future models alone. That focus could yield meaningful performance per dollar advantages, but it also raises the stakes if model architectures shift faster than silicon design cycles allow.

Anthropic is also pursuing what it describes as a “build and buy” strategy. The company has placed an initial order worth approximately $250 million with Fractile, a UK-based chip startup focused on AI inference. That dual approach — developing in-house silicon while also purchasing from emerging suppliers — mirrors strategies used by other labs in the early stages of hardware independence. It hedges against the risk that internal development stalls while still signaling a long-term commitment to reducing reliance on Nvidia. The Fractile order also gives Anthropic near-term access to inference-optimized hardware that could lower serving costs for Claude while Salek’s team works on a first-generation internal chip. For a company whose primary cost is compute, even marginal improvements in inference efficiency translate directly into better unit economics.

The economics and the uncertainty

The financial stakes are substantial. Designing a competitive AI accelerator requires hundreds of millions of dollars in upfront engineering, plus ongoing investment in software tooling, compiler support, and system integration. A single high-profile hire, even one with Salek’s credentials, does not guarantee success. Sustained investment, engineering scale, and foundry partnerships are critical. Anthropic has not disclosed which foundry it would use, though the most advanced AI chips today are manufactured by TSMC, with Samsung and Intel Foundry also competing for leading-edge customers. The choice of foundry will be a key indicator of the program’s ambition: leading-edge nodes offer the best performance but come with higher costs and longer lead times, while older nodes are cheaper but less competitive for frontier training workloads.

There is also the question of software. Nvidia’s dominance rests not only on its hardware but on CUDA, the software ecosystem that has accumulated over more than a decade and is deeply embedded in AI research and production workflows. Any custom chip program must build a comparable software stack or rely on emerging open standards. Google’s TPU succeeded in part because Google controlled its own machine learning frameworks, particularly TensorFlow and later JAX, and could optimize them for its silicon. Anthropic will need to make similar investments in compilers, kernels, and libraries if its chips are to be useful beyond a narrow set of internal workloads. Without a robust software layer, even a well-designed accelerator will struggle to attract internal adoption, let alone external developers. The fact that Anthropic has not announced a software strategy alongside the Salek hire suggests that this piece of the puzzle is still taking shape.

Adding to the pressure, Anthropic is reportedly moving toward a potential initial public offering. Public market investors will want clear delineation between long-term chip-program spending and other infrastructure costs. A multi-year silicon roadmap with uncertain returns could complicate the company’s financial narrative, especially if the chip program’s expenses are lumped together with day-to-day compute costs. At the same time, demonstrating progress on hardware independence could strengthen Anthropic’s long-term competitive position and reduce its exposure to Nvidia’s pricing power and supply constraints. The tension between near-term profitability and long-term hardware investment will be a central theme in any IPO prospectus. Salek’s hiring makes the hardware bet more credible, but it also makes the associated costs more visible.

What it means for the broader AI ecosystem

For international professionals across technology, finance, and policy, Anthropic’s hardware push is a leading indicator of where the AI industry is heading. The “AI chip wars” are intensifying, and control over compute is becoming a strategic asset on par with model quality. Companies that rely entirely on third-party GPUs face rising costs, potential supply bottlenecks, and limited ability to optimize hardware for their specific workloads. Those that build their own silicon gain flexibility but take on enormous capital and engineering risk. The threshold for entering this race is high: only a handful of organizations have the financial resources and technical depth to sustain a multi-year chip program. Anthropic’s decision to cross that threshold signals that it sees hardware independence as essential to its long-term survival, not merely a cost-saving exercise.

The trend also has implications for the broader semiconductor supply chain. As more AI labs bring chip design in-house, the market for commercial AI accelerators could fragment. Nvidia will likely remain dominant for the foreseeable future, but the share of AI compute running on custom silicon is growing. Foundries, memory suppliers, and advanced packaging companies stand to benefit from the increased design activity, while traditional server OEMs may see their role diminish as hyperscalers and AI labs specify their own hardware more directly. For international suppliers and partners, the shift creates both opportunity and risk: those who align early with custom silicon programs could capture long-term contracts, while those tied to general-purpose GPU supply chains may face slower growth.

Salek’s move from Google to Anthropic is one data point, but it is a significant one. It shows that the competition for talent in custom silicon is now as intense as the competition for AI researchers. The pool of executives who have led a hyperscale chip program from inception through multiple generations is extremely small, and their career moves are closely watched. The next few years will reveal whether Anthropic can translate this hire into a working chip program, and whether the broader industry’s push toward compute independence reshapes the economics of frontier AI. For now, the message is clear: the companies that will lead AI over the next decade are not just building better models — they are building the machines that run them.