As Anthropic races toward an initial public offering that could value it at $2 trillion, the company is pivoting its strategic focus from securing raw computing muscle to actually manufacturing the silicon that powers its explosive growth, seeking to control the foundational infrastructure beneath its AI ambitions.
On Friday, the developer of Claude announced it has hired Amir Salek, the founder of Google's custom chip division, to establish and lead an in-house AI silicon design team.
In tandem, the company is aggressively expanding its external procurement network: it has signed an initial approximately $250 million chip agreement with British startup Fractile, deepened partnerships with Google and Broadcom to lock in roughly 3.5 gigawatts of next-generation TPU capacity for 2027, and is planning a joint custom hardware design effort with Samsung.
This flurry of moves sends a clear signal: leading AI model developers are rapidly accelerating their efforts to break free from absolute dependence on single-chip suppliers like NVIDIA. By pursuing a dual-track strategy of internal development combined with diversified external sourcing, Anthropic aims not only to alleviate compute supply bottlenecks but also to achieve structural improvements in inference costs—and the control over compute is fast becoming the defining variable in how much value the market assigns to frontier AI firms.
Securing the final piece of the in-house chip puzzle
The addition of Amir Salek marks a crucial step in Anthropic's march toward independent semiconductor production. Salek spent eight years at NVIDIA, where he founded and led the system-on-chip design team, working on GPU and Tegra processors. In 2013, Google recruited him to build its custom silicon unit from scratch, where he oversaw the development of the first seven generations of TPUs—a product line that remains the backbone of Google's AI infrastructure today.
After departing Google in 2022, Salek served as a senior managing director at Cerberus Capital Management, focusing on semiconductor and AI investments. In his new role at Anthropic, he will report to James Bradbury and will be responsible for assembling and directing the custom AI chip design team. Beyond internal efforts, Anthropic also plans to collaborate with Samsung on co-designing hardware and models to further enhance runtime speed and efficiency.
This approach mirrors broader industry momentum. Rival OpenAI has co-developed the Jalapeno chip with Broadcom, specifically engineered for inference workloads; Google DeepMind has long leaned on its proprietary TPUs; and Meta is advancing its MTIA accelerators. By designing custom silicon, Anthropic can tailor hardware to the precise architecture of its models, establishing a competitive moat at the silicon level while diminishing reliance on off-the-shelf general-purpose chips.
Building a multi-sourced compute network with tens of billions in commitments
With in-house chips still on the horizon, Anthropic is showing remarkable aggression in expanding its external compute capacity. On the procurement front, beyond the initial $250 million order from Fractile, the company is dramatically scaling up collaborations with established giants—expanding its work with Google and Broadcom to secure roughly 3.5GW of next-generation TPU capacity for 2027.
That figure is roughly equivalent to the electricity consumption of a mid-sized city, and it does not include the 1GW slated for delivery in 2026 under the October Google Cloud agreement. Broadcom plays a pivotal role here: it is responsible for developing and supplying custom TPUs and has committed to providing network components for Google's next-gen AI racks through 2031.
On the data center front, Anthropic's contracted capacity has reached tens of billions of dollars, spanning traditional cloud providers, major chipmakers, and emerging compute suppliers, including capacity agreements with Riot Platforms and Volta Infra Holdings. Notably, market estimates suggest that more than half of its compute has historically been hosted on Google Cloud TPUs. As Google's own total compute growth decelerates, Anthropic is under mounting pressure to source chips externally.
The economics of compute autonomy: solidifying the physical foundation for a $2 trillion IPO
Anthropic's aggressive investment in chips and computing power maps directly onto its rapidly improving financial trajectory. As of the end of July, the company's annualized revenue had surpassed $65 billion, with second-quarter preliminary revenue exceeding $11.5 billion—up more than 14-fold year over year. Gross margins on inference infrastructure have surged from 38% to over 70%, and adjusted operating profit has turned positive for the first time.
Chief Financial Officer Krishna Rao has described this as the company's "most significant investment in computing to date," designed to accommodate the exponential growth of its customer base. For Anthropic, preparing to enter the public markets at a valuation of $2 trillion or more, owning the pricing power over compute and securing supply chain autonomy is not merely a matter of cost—it is the linchpin for convincing investors that the steep growth curve can be sustained through the next phase of the AI race.
Comments