Anthropic confirmed for the first time that it has assembled an internal chip team, joining the ranks of AI labs that are designing custom processors rather than renting capacity from Nvidia.
The company, best known for its Claude models, said it is designing custom AI processors for its own workloads, with the goal of improving speed, efficiency and scalability. The confirmation, reported by Jiemian News, had been expected: SK Group Chairman Chey Tae-won said earlier this year that Anthropic had approached SK Hynix about supplying memory for an in-house chip project, and reports in early July described talks with Samsung Electronics about custom chip manufacturing.
The move puts Anthropic in a club that includes the biggest names in AI. OpenAI is working with Broadcom on custom silicon. Google has designed its own TPUs for more than a decade and now builds most of the hardware for its AI services. Meta has developed its own inference chips, and Microsoft has designed accelerators for its cloud. The trend has a clear logic: Nvidia’s accelerators are expensive, scarce and tailored to its own priorities, and the labs that depend on them are paying a premium for hardware that does not fit their needs exactly.
The economics are the driving force. Training and running frontier models consumes more computing power than any other software workload in history, and the largest labs now spend more on compute than on salaries. A custom chip designed for a specific model architecture can cut that cost substantially, and a lab that controls its silicon no longer has to queue for Nvidia’s allocation. The trade-off is that chip design is slow, expensive and unforgiving: a design mistake costs years and hundreds of millions of dollars, and the fastest route to a working chip still runs through TSMC’s fabs and the memory makers’ HBM lines.
The supply-chain angle is just as important. Anthropic’s overtures to SK Hynix and Samsung show that the company is thinking about memory as well as compute, and that it wants relationships with the suppliers who control the components that AI chips cannot do without. Securing HBM supply is as strategic as designing the processor itself, and a lab that locks in memory capacity for its own silicon has taken a step toward independence from the open market, where HBM has been in short supply for two years.
The shift in thinking, analysts said, has moved from renting to building: the question for the top labs is no longer how many GPUs they can rent, but what chips they want to build. Nvidia is the target of that shift, and its dominance, while intact, is being nibbled at from every direction. The custom-chip efforts are still small relative to Nvidia’s volumes, and the designs are years from production, but each one removes a customer from the queue and adds capacity to the ecosystem.
For Anthropic, the chip program is a long-term bet. The company is preparing for an IPO that could come as early as October, and a custom-silicon program is the kind of investment that public investors have historically rewarded, the promise of control over the cost structure of the most expensive part of the business. It is also the kind of investment that takes years to show returns and can fail outright. The company has said little about the team’s size, the design’s progress or the expected timeline, and until it does, the program is best read as a hedge: a bet that the era of buying every processor from one company is ending, and that the labs that build their own will lead the next phase.
The custom-chip trend is, in part, a response to Nvidia’s own success. Nvidia’s margins are the envy of the industry, and its allocation decisions, who gets GPUs and when, have effectively decided which AI companies can grow. The labs that can afford to design their own silicon are doing so as a form of insurance, and the ones that cannot are paying the price in queue times and premiums. Anthropic, with a valuation that has climbed toward $1 trillion and an IPO on the horizon, is one of the few companies with the resources to join the group.
There are reasons for caution. Custom silicon has a long history of failure, and even successful designs take years to reach production, by which time the model architectures they were built for may have changed. The memory side is equally demanding: HBM, the component that surrounds AI processors, is supplied by a handful of companies, and securing supply requires relationships that take time to build. Analysts said Anthropic’s approach, signaled by its conversations with SK Hynix and Samsung, suggests the company understands that the chip program is as much about supply chains as about design. What it will mean for Claude’s users is simple: faster models at lower cost, if the bet pays off.
