Anthropic has just quietly recruited a chip engineer from OpenAI while planning to cooperate with Samsung
In early June 2026, before any news about Anthropic negotiating with Samsung was announced, the company behind Claude quietly recruited Clive Chan to its team. Chan is not a random name: he is the second engineer to join OpenAI's custom chip team, having spent two and a half years there helping design an inference chip (an inference chip - a chip that runs the model after it has been trained, different from a chip used for training) in collaboration with Broadcom, which OpenAI just announced as "Jalapeño" on June 24. Chan left with exactly the kind of knowledge an AI company needs if it wants to design its own chips from the software layer down. The fact that he joined Anthropic, which happened before the news of negotiations with Samsung was revealed, in my opinion is the clearest signal showing how serious Anthropic is.
Negotiations are only at a very early stage
According to The Information, who first reported this news, Anthropic began preliminary discussions with Samsung Electronics in early July 2026 about producing a customized AI chip. Three sources familiar with the matter told The Information that Anthropic is still in the process of determining the specifications, power requirements and server cluster layout for the chip. No physical prototype has been built yet, nor is there a specific production timeline. In other words, this is the type of news that if you read it at a glance, it's easy to think "there's nothing to discuss", but I think because it comes right after the hiring of Clive Chan, the story becomes much more worth following.
Why Samsung, and why now?
Samsung's attraction to Anthropic comes from two different directions: existing financial relationships, and production capacity that no other partner of Anthropic has.
On the production side, discussions are currently revolving around Samsung's 2nm process, officially named SF2, and the company's advanced packaging (the technique of combining many different semiconductor components into one chip package). The 2nm process uses a transistor structure that surrounds the entire channel (Gate-All-Around, or GAA), instead of the FinFET structure that has dominated the past decade. Basically, GAA allows the transistor gate to cover the entire channel from all four sides instead of just three sides like FinFET, helping to control the current more tightly, thereby achieving about 15% higher performance at the same power consumption, or significant power savings at the same performance level. Samsung will begin mass production of this process from the end of 2025. High-performance AI chips rarely consist of just a single piece of silicon, but often combine many logic, memory and connection components into one package using 2.5D and 3D layering techniques, and this is the strength of Samsung's packaging segment.
Why the target is the inference chip, not the training chip
This is the technical detail I find most important in the whole story. Anthropic's talks are not aimed at replacing chips used to train models, where Nvidia GPUs still dominate and there are no reliable alternatives at scale. The real target is the inference chip, which is the processing that happens continuously every time Claude responds to millions of users every day. These two types of chips have completely different design priorities. The training chip needs to maintain extremely high throughput on huge volumes of data, so a flexible GPU-style architecture is still the best choice. The inference chip needs to respond quickly, have low latency, and have a low cost per query, which means eliminating general-purpose computing costs to focus all the silicon on exactly one type of calculation: transformer model calculations. The Jalapeño chip that OpenAI announced in partnership with Broadcom on June 24 is designed specifically for this purpose, and Broadcom CEO Hock Tan once told Bloomberg that initial tests show a cost savings of about 50% compared to using conventional GPUs for inference tasks. If that number holds at real production scale, it explains why the economics of custom inference chips are so attractive to leading AI companies. Anthropic, which powers Claude for millions of individual and business users, faces exactly this inference cost structure, and a chip specifically optimized for Claude's model architecture and serving style could deliver similar savings.
Geopolitical factors hidden in choosing Samsung
There's a structural reason that makes Samsung an attractive partner, beyond cost and technical prowess. A chip designed in-house and manufactured at Samsung's factories in Korea, or at the large-scale factory in Taylor, Texas expected to begin 2nm production from 2027, will be outside the influence of US-China semiconductor tension hot spots. Samsung's factories operate under the legal jurisdiction of South Korea and the US, without Chinese-style data sharing constraints or the technology transfer concerns that still haunt some other partners. For an AI company that processes enterprise data at the scale of Anthropic, this legal clarity carries real weight, not just a side detail.
The opportunity comes from where OpenAI just stumbled
One detail that makes this story more interesting: Samsung had previously developed a custom AI chip for OpenAI itself, a neural inference processor based on the ARM architecture, before those negotiations stalled in early June 2026 because of what Korean media described as "strategic differences" between the two sides. OpenAI CEO Sam Altman later canceled a visit to Seoul that was expected to boost the relationship. If Samsung redirects the technical capacity and 2nm capacity originally devoted to the OpenAI project to Anthropic, the competitive impact will compound in both directions: Anthropic gets a foundry partner with the most recent experience designing AI inference chips, and Samsung gets a major customer at the right time when its foundry business is trying to close the gap with TSMC.
The barrier that Samsung cannot avoid: standard chip ratio
Samsung's 2nm ambition has a well-documented limit: the ratio of qualified chips per wafer (yield). This low ratio means high costs per chip and unpredictable supply. The first generation of the SF2 process is expected to reach only 50-60% through most of 2025, well below the 70-80% level that analysts consider the economically viable threshold for large-scale production. Meanwhile, TSMC's competitive N2 process is said to have reached a rate of 65-80% and entered mass production with large customers such as Apple and Nvidia. The SF2P version, the second-generation performance optimization in Samsung's 2nm roadmap, is said to have approached 70% by early 2026, but whether that number will be stable at large production scale is yet to be tested. For a company like Anthropic, which cannot accept erratic chip sources at unpredictable costs, this is a factor that needs to be seriously considered, not a technical detail that can be ignored. Anthropic has not confirmed anything beyond a brief answer that Amazon Web Services' Trainium chip, Google's Tensor processors and Nvidia's GPUs will remain central to the company's computing strategy. That answer is not at all inconsistent with parallel development of a custom chip: what Anthropic is using today is one thing, and what they could be running in three to five years, if conversations with Samsung progress from concept to engineering to real silicon, is another story. For me, what's worth watching is not whether this chip will be born or not, but the fact that more and more leading AI laboratories, from OpenAI to now Anthropic, are coming to the same conclusion: if you want to control the cost of serving a model at a scale of millions of users, sooner or later you will have to design your own hardware.