The race among AI model developers to create their own silicon is intensifying. Following rival OpenAI's launch of its first in-house inference chip named Jalapeño on June 24, AI startup Anthropic is now reported to have initiated early-stage development of its own AI chip and is in talks with Samsung Electronics Co., Ltd. regarding potential manufacturing collaboration.
According to three informed sources, Anthropic has commenced planning for its proprietary AI chip project, though it remains at the conceptual stage. The company has not yet finalized the processor's functional focus, computing specifications, or its deployment scheme within servers. While preliminary discussions have been held with several chip design firms, the project has not progressed to detailed design, testing, or mass production phases.
Choosing Samsung's 2nm Process: A Window Opens Amid TSMC's Capacity Constraints
Sources indicate Anthropic plans to utilize Samsung's 2-nanometer process technology and advanced packaging. The 2nm process enables higher transistor density and energy efficiency, while advanced packaging allows for tighter integration of computing chips with high-speed memory, boosting data transfer efficiency.
For Samsung, this potential order is significant. With robust demand for AI chips and persistent tightness in TSMC's advanced capacity, Samsung is encountering a strategic window to promote its high-end manufacturing capabilities to more clients. Previous reports suggest Alphabet is also evaluating outsourcing future TPU production to Samsung.
Analysis suggests Anthropic's choice of Samsung may be closely tied to capacity timelines. TSMC's 2nm capacity is reportedly booked until 2028-2029, while Samsung's Taylor fab is expected to achieve mass production by 2027 and may offer integrated solutions bundling HBM4 memory with packaging.
Samsung and Anthropic already share capital ties. In May this year, Samsung, alongside SK Hynix and Micron Technology, participated in Anthropic's $65 billion Series H funding round. This strategic investment effectively pre-secured key memory chip suppliers for Anthropic's business expansion.
OpenAI's Jalapeño: Designed in 9 Months, Aiming to Halve Inference Costs
Anthropic's chip development plan follows closely behind its primary competitor, OpenAI. On June 24, OpenAI and Broadcom jointly unveiled their first custom AI inference chip, Jalapeño. This ASIC, designed specifically for Large Language Model inference, reportedly took just nine months from initial design to tape-out.
Broadcom CEO Hock Tan stated that Jalapeño's performance is comparable to NVIDIA's Blackwell chips and Google's TPUs. In terms of cost, the chip is projected to reduce inference expenses by approximately 50%. Deployment is slated to begin in late 2026, with the chips and server systems intended for OpenAI's internal use only, not for external sale.
OpenAI has established a multi-generation chip roadmap. Jalapeño represents the first step in a multi-generation computing platform co-developed with Broadcom, with the first chips expected to enter commercial use at Microsoft and other partners by year-end. Broadcom indicated the platform will begin deployment in gigawatt-scale data centers from late 2026.
Anthropic, meanwhile, has taken a step forward in talent acquisition for chip development. Earlier this month, the company recruited Clive Chan, a core member of OpenAI's original in-house chip team. This move coincides closely with the initiation of Anthropic's chip project, signaling a systematic effort to build chip design capabilities.
Anthropic responded by stating that Amazon AWS's Trainium chips, Google TPUs, and NVIDIA GPUs remain central to its compute expansion strategy, declining to disclose further details on its custom chip roadmap. Samsung declined to comment.
Anthropic has long pursued a diversified chip procurement strategy, utilizing Amazon Trainium, Google TPU, and NVIDIA GPUs, and is reportedly in talks with Microsoft and UK startup Fractile to introduce their chip solutions.
Industry Trend: A Paradigm Shift from Using to Building Chips
Anthropic's chip plan reflects a significant trend shift in the AI industry, where model developers are extending competition from model capabilities to underlying hardware infrastructure. Alphabet, Amazon, Microsoft, and Meta have all embarked on developing custom chips.
The primary drivers for in-house chip development are cost control and supply chain autonomy. In large-scale model training and inference scenarios, even minor improvements in chip efficiency can significantly reduce operational costs and free up scarce compute resources. Amidst sustained growth in AI compute demand and continued tight supply of advanced chips, custom silicon is seen as a crucial means to enhance cost control and bargaining power within the supply chain.
Notably, participants in custom chip development are no longer limited to cloud giants. OpenAI's entry on June 24, followed by Anthropic, signals AI model companies are comprehensively entering the chip design arena.
From Model Race to Full-Stack Race
If Anthropic ultimately partners with Samsung, it would mark a significant escalation in the AI industry's chip-making trend. For Anthropic, developing its own chips signifies extending its reach from the model layer to the entire chain encompassing silicon, data centers, power, and cloud infrastructure, aiming to reduce reliance on external supply chains. In the context of surging AI compute demand, mastering chip design equates to gaining leverage over compute cost negotiations.
For Samsung, securing an order from Anthropic would represent important progress for its foundry business in attracting heavyweight AI clients. With TSMC's advanced capacity persistently in short supply, Samsung is presented with a strategic window to narrow the gap with its rival.
For the industry, the collective move by AI model developers into chipmaking signifies that the AI race has evolved from a "model competition" to a "full-stack competition." An all-encompassing arms race, spanning algorithms, chips, models, and data centers, is reshaping the power dynamics of the entire AI industry chain. As industry observers note, AI companies are attempting to expand competition from the models themselves to the entire infrastructure ecosystem.
However, developing AI server chips presents extremely high barriers. The design process requires balancing multiple metrics like computational performance, power consumption, memory, network bandwidth, and thermal management, while also solving subsequent challenges related to large-scale, stable mass production. Anthropic's custom chip project remains in its early stages and could potentially be shelved.
Comments