According to three people familiar with the project who are deeply involved in the project, following the self-developed chip of competitor OpenAI, in order to take the initiative in the expensive computing system behind the large model, Claude developer Anthropic has started preparatory work for self-developed AI chips and has started negotiations with Samsung Electronics, intending to use Samsung as a potential chip foundry partner.
If this AI company officially advances its self-developed chip project, Anthropic can only be regarded as a latecomer in the field of self-developed server-level AI chips. Google and Amazon AWS have been working hard for many years and have successfully launched self-developed chips; Meta and Microsoft have also launched self-developed processors. OpenAI has joined hands with Broadcom to start self-developed chip design as early as 2024. Last month, the first product of the cooperation between the two parties, Jalapeño inference chip, was officially released. This chip can greatly improve the operating efficiency of large language models.
Three people familiar with the matter said that Anthropic is still in the planning stage: it has not determined the functional positioning and computing power specifications of this processor, nor has it finalized the deployment plan of the chip in servers and server clusters. Although the company has started exchanges with a number of chip design companies, it has not yet entered the detailed design, testing and mass production stages.
The research and development of AI processors is extremely difficult. Engineers need to take into account the five dimensions of computing speed, power consumption, memory, network transmission and heat dissipation. It is twice as difficult to achieve stable and large-scale mass production.
This self-research project reflects the general trend of the industry: AI companies represented by Anthropic are trying to firmly control the underlying infrastructure of large models, covering full-chain resources such as chips, cloud service contracts, power supply, and data centers. Ultra-large-scale AI models need to rely on massive processor clusters to run. At this scale, even a small improvement in computing efficiency can significantly reduce operating costs and release scarce computing resources. Self-developed AI chips can also give AI companies more bargaining power in the industry competition for processors, computer rooms, and power resources.
Although Anthropic has started recruiting chip engineers, the self-research project may eventually be shelved. At the beginning of this month, the company successfully recruited Clive Chen, a core member of OpenAI’s first-generation self-developed chip team.
In response to media interviews and questions, Anthropic responded that Amazon AWS's Trainium chips, Google's Tensor Processor TPUs and Nvidia GPUs are still the core hardware choices of the company's computing power expansion strategy, and did not disclose more details about the route of self-developed chips; Samsung declined to comment on this cooperation negotiation.
Samsung and Anthropic have long had capital ties. As the world's leading memory chip manufacturer, in May this year, Samsung joined forces with two other major storage giants, SK Hynix and Micron Technology, to participate in Anthropic's total financing of US$65 billion. At that time, the global supply of memory chips exceeded demand, and consumer electronics companies such as Apple were raising product prices. This strategic investment allowed Anthropic to bind the core memory chip suppliers needed for its own business expansion.
South Korea recently announced a 10-year industrial investment plan worth hundreds of billions, led by Samsung Group and SK Group (the parent companies of Samsung Electronics and SK Hynix respectively). The two companies have invested a total of US$518 billion to build four new memory chip factories in South Korea.
If the two parties finalize the foundry cooperation, it will become a blockbuster order with great industry influence for Samsung's wafer foundry business. Although Samsung is the world's leader in memory chips, it has been trying to expand its AI chip foundry business and narrow the gap with TSMC - the latter's advanced process production lines have always been the industry benchmark for the world's cutting-edge AI processor manufacturing. Currently, orders for AI chips are full and TSMC's production capacity is tight. Samsung has entered a window period to promote its 2-nanometer process to more customers. The media has previously reported that Google is considering using Samsung to manufacture some of its next-generation TPU tensor processors.
Two of the three people familiar with the matter revealed that Anthropic plans to use Samsung's 2nm process technology and advanced packaging technology. 2nm is an industry process designation, not a physical size. It represents one of the most advanced chip manufacturing technologies, allowing processors to achieve higher integration and lower power consumption. Advanced packaging technology can shorten the physical distance between the main processor and high-speed memory, greatly increasing the data transmission speed inside the chip.
For a long time, Anthropic's differentiated competitive strategy has been to diversify the purchase of server chips to avoid being highly dependent on Nvidia hardware like OpenAI and xAI. The company currently uses AI server chips from Amazon, Google, and NVIDIA, and is also negotiating to access self-developed chip solutions from Microsoft and British start-up Fractile.
Although the current inference chip track is booming in financing and various companies are gathering together for research and development, media calculations show that Nvidia's market share has risen instead of falling in recent years, reaching 74%. NVIDIA CEO Jensen Huang insists that the comprehensive performance of his own chips in inference scenarios is still ahead of all competing products.