Anthropic Hires Silicon Experts to Build Custom AI Chips

Anthropic is making a decisive move into the hardware sector by hiring the engineer responsible for Google’s most successful custom chip program. This strategic recruitment indicates that the AI laboratory intends to build its own silicon to power future generations of its Claude large language models.

Strengthening Internal Hardware Expertise

Amir Salek spent nearly a decade at Google, where he founded and managed the Tensor Processing Unit (TPU) program. During his tenure, he oversaw the development and deployment of seven generations of custom silicon. These chips became the backbone of Google’s massive artificial intelligence infrastructure. After a brief period working in venture capital at Cerberus Capital Management, Salek is returning to the front lines of chip design. He joins Anthropic’s compute team and reports directly to James Bradbury, the current lead of that department.

The addition of Salek is not an isolated event but part of a broader recruitment drive. Earlier this year, Anthropic secured Clive Chan, a key figure from OpenAI’s custom chip initiative. Chan played a vital role in developing the Jalapeno inference accelerator, a product designed in collaboration with Broadcom. These high-profile hires demonstrate that Anthropic is aggressively building a world-class hardware team from scratch. Until recently, the company lacked any internal hardware design capabilities, relying entirely on partners for its processing power.

Anthropic confirmed in early August that it has established a dedicated in-house silicon team. The primary objective of this group is to co-design hardware alongside the Claude models. By tailoring the silicon to the specific requirements of its software, Anthropic expects to reduce inference costs by approximately 50 percent. This level of optimization is difficult to achieve when using general-purpose hardware. Preliminary reports suggest the company has already engaged in manufacturing discussions with Samsung to bring these future designs to life.

The decision to build internal hardware follows a path blazed by other tech giants. Meta, Microsoft, and Amazon have all invested billions into proprietary chips to avoid the high margins and supply constraints associated with third-party vendors. While Anthropic is entering the race later than its competitors, the caliber of its recent hires suggests it intends to close the gap quickly. The focus on inference efficiency is particularly important as more users interact with AI models daily, driving up operational expenses.

Navigating Complex Industry Partnerships

The hiring of Salek creates a fascinating dynamic between Anthropic and its primary infrastructure provider, Google. In late 2023, Anthropic signed a massive deal to purchase up to one million TPUs from Google. This multi-billion dollar agreement was expanded earlier this year to include even more capacity through 2027. This means Google is currently supplying the very chips that Anthropic aims to eventually replace with its own internal designs. It also means Google has lost a top hardware architect to a company that is both a major customer and a direct competitor in the AI space.

This complex relationship highlights the tension at the top of the AI industry. Companies must cooperate to build the necessary infrastructure while simultaneously competing for talent and market share. Anthropic maintains that its long-term strategy involves a mix of technologies. The company states that Nvidia GPUs, Google TPUs, and AWS Trainium chips will remain essential to its scaling efforts for the foreseeable future. However, the investment in a custom silicon team suggests that Anthropic wants to control its own destiny rather than remaining fully dependent on the roadmaps of other corporations.

For investors monitoring the semiconductor landscape, these developments carry significant weight. Every major AI lab that successfully launches a custom silicon program represents a potential decrease in long-term demand for Nvidia’s hardware. While the total market for AI chips is growing, the shift toward application-specific integrated circuits (ASICs) could slowly erode the dominance of general-purpose GPUs. Broadcom stands out as a potential winner in this shift, as it often provides the underlying intellectual property and design services for companies looking to build their own accelerators.

The timeline for these hardware projects is measured in years, not months. It typically takes a long period to design, verify, and manufacture a new chip at scale. Consequently, the immediate financial impact on companies like Alphabet or Nvidia remains minimal. The real shift is structural. Anthropic is transforming from a pure software research lab into a vertically integrated technology firm. This evolution mirrors the early days of the smartphone industry, where the most successful players eventually moved hardware design in-house to gain a competitive edge.

The Competition for Specialized Talent

The battle for dominance in artificial intelligence has moved beyond software engineers and data scientists. Today, the most intense bidding wars involve hardware architects and silicon engineers. These professionals possess the rare skills needed to translate complex mathematical models into physical circuitry. There are far fewer individuals capable of leading a global-scale chip program than there are machine learning researchers. This scarcity makes hires like Salek particularly impactful for a young company like Anthropic.

In other sectors of the technology industry, this aggressive poaching has led to legal disputes. For instance, some media companies have faced lawsuits after hiring executives away from competitors before their contracts expired. Anthropic avoided such complications with the Salek hire because he had already spent time away from Google in the private equity sector. This allowed the company to acquire elite talent without the risk of litigation over non-compete agreements or intellectual property theft. It serves as a blueprint for how startups can build sophisticated technical teams by identifying key veterans during career transitions.

Despite the impressive roster of talent, Anthropic faces significant hurdles. Building a chip is one of the most capital-intensive and technically difficult tasks in the world. The company must manage a global supply chain, navigate manufacturing bottlenecks at major foundries, and ensure its software stack remains compatible with the new hardware. Many well-funded startups have failed to bring a viable chip to market in the past. Anthropic’s late start means it must execute perfectly to catch up with the established programs at Amazon and Microsoft.

The ultimate success of this initiative will be measured by its impact on the cost of intelligence. If Anthropic can truly cut its inference costs in half, it will have a massive advantage in the pricing of its AI services. This would allow the company to offer more capable models at lower prices than competitors who remain tied to expensive off-the-shelf hardware. The next few years will reveal whether a company founded on AI safety and research can successfully transform into a hardware powerhouse capable of challenging the incumbents of Silicon Valley.