According to a report by The Information on Thursday, Anthropic is in talks with Samsung about custom AI chips and has initiated early-stage development work on its own AI chips. If these custom server chips eventually enter mass production, it will mark a significant step for the company behind Claude in advancing hardware autonomy.
This move is seen as Anthropic following in OpenAI's footsteps.
OpenAI has been progressing its custom AI chip project earlier, collaborating with chip design and manufacturing partners in an attempt to build a more independent and efficient computing infrastructure for products like ChatGPT. The actions of both companies point to the same trend: large model companies are shifting from pure algorithmic competition to integrated software-hardware competition.
The market impact first falls on three fronts: the bargaining environment for external GPU suppliers like Nvidia, opportunities for foundries like Samsung in AI chip orders, and the future financing and IPO pace of AI startups.
According to Barron's, Deutsche Bank analysts recently suggested that OpenAI and Anthropic should not delay their IPOs for too long, one reason being that developing their own chips and computing infrastructure requires massive long-term capital.
Developing Own Chips is Primarily a Matter of Computing Power Control
Currently, training and running large models require vast amounts of high-performance computing resources. The AI computing market heavily relies on Nvidia's GPU architecture, and tight supply-demand keeps model training and inference costs high. For model companies like OpenAI and Anthropic, chips are no longer just procurement items but core means of production.
The demand for Anthropic's Claude model has grown significantly in 2026. TradingKey reported that Anthropic executives previously disclosed the company's annualized revenue has exceeded $300 billion, compared to about $90 billion at the end of 2025. Business expansion drives rapid increases in computing power demand and also amplifies the impact of external chip supply uncertainties on company operations.
Anthropic still relies on various third-party chip solutions, including TPUs designed by Alphabet's Google and Amazon's self-developed chips. Reports indicate Anthropic has also entered long-term TPU supply agreements with Google and Broadcom, related to its previously announced $50 billion U.S. computing infrastructure investment plan.
This means that developing one's own chips does not equate to completely breaking away from external suppliers. A more realistic goal is to master core design capabilities, create technical alternatives, and enhance leverage in future business negotiations.
Cost is Just the Entry Point; Hardware-Software Co-design is Key
The most direct reason for developing one's own chips is to reduce costs. Through custom ASICs, AI companies can optimize computing processes around their own model architectures, reducing unnecessary modules in general-purpose chips, thereby improving energy efficiency. If Anthropic's chips are successfully taped out and deployed, reports suggest they could significantly lower API call costs and influence pricing structures in the enterprise AI application market.
But cost is not the only variable. Dylan Patel, founder of SemiAnalysis, emphasized in an interview that the greatest room for AI efficiency improvement doesn't come solely from faster chips, but from co-design across models, kernels, and silicon. He believes single-layer optimization might yield a 2x improvement, but cross-layer co-design could bring effects far greater than a simple multiplication.
This explains why OpenAI and Anthropic are moving towards deeper hardware involvement. Model architectures are not naturally suited to all chips. Dylan Patel stated that OpenAI models are more sparse-oriented, while Anthropic models are relatively more dense. They have significant differences in areas like matrix multiplication unit size, attention mechanism structure, and expert layer shapes, which naturally inclines the two companies towards different hardware directions. "In fact, given the direction OpenAI models are heading, using TPUs could be a bad decision for them; similarly, given the direction of Anthropic and Google models, using GPUs for training could be a bad decision for them," he said.
In other words, developing one's own chips isn't just about replacing Nvidia GPUs with proprietary ones. The real goal is to allow models, from their initial design, to fit the underlying hardware, thereby improving inference speed, energy consumption, throughput, and unit economics.
Not an Immediate Replacement for Nvidia, but Long-term Balancing
The process from R&D, tape-out, verification, to final mass production and deployment of self-developed AI chips typically takes 18 to 24 months. Even if Anthropic successfully reaches an agreement with Samsung, its self-developed chips are unlikely to substantially replace existing computing power supplies in the short term.
OpenAI is progressing earlier. TradingKey reported that OpenAI chose to collaborate with Broadcom and TSMC, planning to deploy its first inference chip in the second half of 2026. Compared to Anthropic, OpenAI is more proactive and closer to deployment on the custom chip path.
The direction of large model companies developing their own chips does point towards reducing dependence on suppliers like Nvidia. But this doesn't mean Nvidia's position will be rapidly weakened. Dylan Patel noted in the interview that Nvidia GPUs still hold advantages in generality, as many models and the open-source ecosystem are already optimized for GPUs. He also mentioned that the so-called CUDA moat isn't just CUDA itself, but the fact that a vast downstream ecosystem of models and software has been adapted for Nvidia's hardware form. If a model's expert structure, hidden dimensions, and communication patterns are inherently more suited for GPUs, migrating to other chips, even if advantageous, might not be straightforward.
Therefore, developing proprietary chips is more like establishing a second route. OpenAI and Anthropic will likely continue using GPU, TPU, Trainium, and other computing resources, while deploying self-developed ASICs for more specific, stable, and high-frequency workloads, especially inference scenarios.
Industry-wide "Computing Power Autonomy" Race Fully Underway
The shared logic behind OpenAI's and Anthropic's self-developed chips can be summarized in three points: reducing long-term computing costs, decreasing reliance on external supply, and improving model efficiency through hardware-software co-design.
Among these, the third point might be the most critical. As model companies scale, general-purpose computing power cannot fully meet the needs of differentiated architectures. Self-developed chips allow companies to place model design, system software, and underlying silicon within the same optimization framework.
But the direction is clear: competition among large models is extending from "whose model is stronger" to "who can better control computing power, capital, and the hardware stack." This is the real reason both OpenAI and Anthropic are moving towards developing their own chips.
Anthropic's exploration is not an isolated case. From Google's decade-long TPU series, to Amazon's Trainium series focused on training, to Meta's MTIA series for inference, and Microsoft's ongoing Maia series, leading tech companies have all deeply engaged in the self-developed chip race.
For Samsung, securing Anthropic's chip foundry order would provide a significant boost to its wafer foundry business's influence in the AI field. Samsung is currently fiercely competing with foundries like TSMC for advanced-node customers. Bringing in high-growth-potential AI clients like Anthropic would help expand its footprint in the AI semiconductor landscape.
This article is from WeChat public account: Wall Street News , author: Zhao Ying
