Anthropic, OpenAI, and Google: Moving Towards Their Own AI Chips

Anthropic, OpenAI, and Google: Towards Their Own AI Chips

A Confirmed "Silicon Team" at Anthropic, Without Specifications or Timeline

On Wednesday, August 5, 2026, Anthropic confirmed the creation of an internal team tasked with designing its own chips for Claude. The announcement comes after six weeks of rapid agreements between AI labs and their hardware suppliers, as the usage intensity of Claude Code regularly saturates the company's infrastructure. However, designing its own chips does not mean freeing itself from Nvidia.

The information was published by Business Insider and later confirmed to TechCrunch. A spokesperson for Anthropic stated that the company will now design its hardware and models together, so that Claude operates faster and more efficiently "at the scale that [its] clients need." Nevertheless, nothing will change in the short term. The spokesperson confirmed the continuation of a "multi-chip approach" where hardware from AWS, Google, Nvidia, and AMD remains central.

The job listing reveals the level of ambition. Anthropic is hiring for its "custom silicon team" profiles covering the entire design chain, from front-end design to packaging, with a salary range of $320,000 to $485,000 annually. "This is a position for someone who has already delivered silicon (silicon chips, editor's note), who has a realistic relationship with timelines, and who is comfortable making decisions without a large organization behind them," the listing specifies.

However, the project is still in its infancy. According to The Information, which revealed on July 2 discussions with Samsung based on three sources, Anthropic is still trying to define what its chip should do and how it will fit into a server. The lab is examining the 2-nanometer process from the Korean company, also discussing chips with Microsoft and the British startup Fractile, while having recruited in June Clive Chan, one of the early members of OpenAI's chip team. There is no guarantee that this project will come to fruition.

Inference: What Makes Custom Chips Profitable

If all these players are starting to design chips, it’s because the nature of the computational load has changed. Making a model respond, request after request, is a repetitive and predictable task. It can therefore be fixed in a dedicated circuit much more easily than training, which remains exploratory.

Training, inference, GPU, and ASIC... What are they?

• Training: creates the model by feeding it data. This is a heavy and one-time operation that requires considerable raw power.

• Inference: refers to the operation of the model once trained, when it produces a response. It repeats with each request, and its cumulative cost rises with adoption.

• GPU: a general-purpose chip capable of both tasks, which explains Nvidia's position.

• ASIC: a circuit designed for a single task. It loses versatility but gains performance per watt.

OpenAI took the plunge before Anthropic. On June 24, the company unveiled Jalapeño, a processor designed with Broadcom specifically for inference. "Jalapeño was designed from the ground up for LLM inference, based on detailed insights from our close collaboration with OpenAI researchers," explained Richard Ho, head of OpenAI's hardware program.

This specialization does not make these chips competitors to GPUs. Journalist Jérôme Marin already pointed this out in our columns regarding the OpenAI-Oracle contract. "The fact that you develop your own chip does not mean competing with Nvidia. Nvidia chips are essential for training. But OpenAI, with Broadcom, wants to design chips dedicated to specific needs, like inference," he explained. Ten months later, Jalapeño validates this perspective.

Nvidia has still responded in this arena. At the end of December, the group spent $20 billion for the intellectual property of Groq and the recruitment of its engineers. The startup's ultra-low latency architecture has become the LPU Groq 3, expected in the third quarter. Publicly, Nvidia continues to defend the versatility of its GPUs: "NVIDIA offers more performance, versatility, and fungibility than ASICs, which are designed for specific frameworks or functions," its X newsroom account stated in November.

Wafers, Memory, and Electricity: The Three Bottlenecks No One Can Avoid

This debate over architectures masks a constraint that no one can evade. Whether a player designs its own chip or buys one from Nvidia, they source from the same place. Broadcom is the best witness to this, as the American group co-designs accelerators for six major clients, including Google, Meta, OpenAI, and Anthropic. When asked on June 3 about the $30 billion in orders recorded in a single quarter, against $10.8 billion actually delivered, its president and CEO Hock Tan described what these clients must align before receiving anything. "It's not just about requesting wafers, getting the chips or memory, ensuring that HBM is available or that DRAM is available. You also need to have the necessary electricity," he detailed before analysts.

Wafer, DRAM, HBM, and TPU... What are they?

• Wafer: a silicon disk on which a foundry engraves several hundred chips at once. It is the unit of account in the industry, and a factory's capacity is measured in wafers per month.

• DRAM: dynamic random-access memory, used in PC memory sticks as well as servers.

• HBM: stacked DRAM then glued to the processor, to deliver data at very high speeds.

• TPU: Google's in-house accelerator, designed with Broadcom. It is not sold but rented via Google Cloud.

However, these bottlenecks are closing in on very few players. Only three manufacturers produce HBM – SK hynix, Samsung, and Micron – all of which entered Anthropic's capital in May. And the assembly of these memory stacks with the processor, what the industry calls advanced packaging, remains primarily a specialty of TSMC. It is because the demand for accelerators has saturated the Taiwanese foundry's lines that Samsung found an opening to approach new clients, notes The Information. Consequently, supply commitments are multiplying. On July 22, AMD committed to providing 2 gigawatts of Instinct MI450 GPUs to Anthropic, with an investment potentially reaching $5 billion in the lab. Two days later, Nvidia secured its own memory supply from SK hynix, in a partnership valued at over $500 billion.

If Anthropic is on this list, it is not for a chip of its design. Hock Tan corrected an analyst on this point during the same conference. "The deal we made with Anthropic is that we are using our TPU chips that we develop to provide computing capacity to Anthropic," he recalled, referring to the 3.5 gigawatts of TPU capacity signed in April with Google and Broadcom. Nvidia, for its part, has seen its market share grow in recent years to reach 74%, according to estimates from The Information.

Effects That Can Be Measured Even in RAM Prices

This competition has a concrete consequence for hardware buyers. Memory manufacturers have redirected their capacities toward AI servers, which are much more profitable than traditional DRAM. In the first quarter of 2026 alone, the contractual prices of the latter surged by ...