TL;DR
· Anthropic has been reported to be exploring in-house AI server chip development, but its design, tape-out, or mass-production plans have not yet been confirmed.
· OpenAI has disclosed the Jalapeño inference chip and begun testing, with plans to deploy it by the end of 2026.
· Samsung could benefit from the outsourcing trend for AI chips, but in the short term Anthropic still relies on compute capacity from AWS, Google, and Nvidia.

Discussions at Anthropic around developing its own AI server chip are heating up, but this is not yet a line of chip orders that has actually landed. The key outside focus is that the reasoning cost behind Claude, GPU supply, data center power, and rack capacity are becoming hard constraints for large-model companies. OpenAI has already disclosed its Jalapeño inference chip in collaboration with Broadcom, and Anthropic has also been reported to be evaluating dedicated chips that are better suited to its own models. However, based on the publicly available information so far, it has not been confirmed whether Samsung is involved in manufacturing or whether the project has entered formal design.

Anthropic is still exploring early on—not on the eve of mass production

The direction Anthropic has been reported to explore is a server chip better suited to the way its own AI models operate. Compared with general-purpose GPUs, if the custom chip design succeeds, it could reduce costs, improve energy efficiency in specific inference tasks, and lessen reliance on external chip supply.

The difficulty of this kind of chip isn’t just the performance of a single chip. Large model companies need to handle, at the same time, compute speed, memory bandwidth, interconnect networks, power consumption, thermal management, and cluster stability. What’s truly hard is making thousands of chips work together stably in a data center and continue to serve training or inference tasks.

At present, the safest phrasing is that Anthropic is still in the early evaluation and definition stage. Which AI tasks the chip will mainly handle, how its performance and power-consumption targets will be set, how servers and clusters will be adapted, and whether an external chip design company is needed—none of these questions have clear publicly confirmed answers yet.

The company is also cautious about its external messaging. In April this year, Anthropic announced an expansion of its partnership with Amazon. In the next decade, it said it would invest more than $100 billion in AWS technology, with up to 5GW of capacity locked in, and it claimed to have used over 1 million Trainium2 chips to train and serve Claude. Anthropic also emphasized a diversified hardware strategy, but AWS remains its primary provider for training and cloud services.

This means that even if in-house chip exploration continues to move forward, it is still unlikely to replace existing suppliers in the short term. AWS Trainium, Google TPU, and Nvidia GPUs remain key components of Anthropic’s scaled compute infrastructure.

OpenAI goes one step ahead—more direct pressure on inference costs

An important background for placing Anthropic into the in-house chip discussion at this point is that OpenAI has already set a reference precedent.

Broadcom’s official announcements show that OpenAI and Broadcom released Jalapeño on June 24, 2026. It is positioned as an accelerator for large language model inference, and is also known as an Intelligence Processor. OpenAI and Broadcom state that the chip will take about nine months from initial design to manufacturing tape-out. Engineering samples have been running in labs, with plans to begin deployment by the end of 2026.

Here, two stages need to be distinguished. Jalapeño has already been released and entered testing, but that does not mean it is already in large-scale commercial use. It represents the start of top model companies bringing inference cost under deeper hardware control, rather than meaning GPU demand will be immediately replaced.

Inference is the computing process where the model generates answers after users ask questions to products like ChatGPT and Claude. Compared with training, inference happens more frequently. As the number of users grows, cost pressure will continue to rise. For large model companies, even if the cost of each inference only drops by a small percentage, spread across massive requests and long-term data-center spending, it could turn into substantial savings.

Anthropic’s pace is clearly earlier. It hasn’t published chip specifications, disclosed performance metrics, listed partners, or provided a mass-production timeline. OpenAI’s progress only shows the market one direction: the very top model companies are no longer just buying GPUs—they’re also trying to bring part of their compute infrastructure under their own control.

Samsung’s imagination is heating up, but no orders have been finalized

Samsung is getting market attention because it has advanced manufacturing capabilities and is also seeking more opportunities to act as a foundry for AI chip partners. After news emerged about Anthropic’s fundraising and infrastructure cooperation, it’s natural that people connected Samsung with potential AI accelerator manufacturing opportunities.

But this point needs to be treated with a cooling-off mindset. What public information can confirm is that companies such as Samsung, SK Hynix, and Micron have appeared in discussions about Anthropic’s infrastructure partners. Micron announced on June 22, 2026 that it has reached a strategic agreement with Anthropic, including memory and storage AI architecture design, supply agreements, Micron’s internal adoption of Claude, and a strategic investment in Anthropic Series H.

These partnership signals cannot be directly equated with Samsung having already secured an order for Anthropic’s in-house chip. Claims that Anthropic has been in contact with Samsung regarding manufacturing cooperation are not supported by enough publicly verifiable information. The more prudent assessment is that if Anthropic’s in-house chip project progresses to the manufacturing stage, Samsung could become one of the potential players the market watches. But for now, it cannot be written as a definite deal.

For chip projects, from early evaluation to final mass production, there are many steps in between: architecture confirmation, design validation, selecting manufacturing processes, packaging and testing, and coordinating the supply chain. As long as the chip design hasn’t been finalized, the foundry role is also hard to truly lock in.

Hiring aggressively increases credibility, but the roadmap is still not set

Talent moves have made Anthropic’s hardware hints draw more attention. According to reports, Clive Chan, an early member of OpenAI’s custom chip team, has joined Anthropic. Public information indicates he participated in early efforts to build OpenAI’s chip team and also has experience related to Tesla Dojo. Anthropic has also been strengthening its recruitment of chip engineers recently.

This suggests the company is at least preparing for hardware capabilities. For a model company, having a hardware team that understands models, inference workloads, and data-center systems can help decide which tasks are suitable to hand over to custom chips, and which still need to rely on GPU, TPU, or chips from cloud providers.

But the involvement of talent and expanded hiring are still only early investment signals. Whether the project can continue depends on whether the chip can deliver sufficient advantages across cost, performance, power consumption, and deployment complexity. If a custom chip can only improve efficiency on paper but cannot run at scale reliably, or if manufacturing and software adaptation costs are too high, the company may still continue to rely mainly on external chips.

This is also why Nvidia is not easy to replace in the short term. Nvidia GPUs are still the mainstay for AI training and inference, the software ecosystem is mature, and data-center customers have already built many systems around its platform. In-house chips are more likely to take on part of the workload in specific inference scenarios first, rather than fully replacing GPUs.

For investors, the real-world impact of Anthropic’s in-house chip discussions is, in the short term, more like a supply-chain standoff. Leading model companies want more options and control over compute capacity. Cloud providers, Broadcom, Samsung, TSMC, memory makers, and advanced packaging supply chains could all benefit from this trend. But in Anthropic’s case, the concrete facts remain limited: its in-house exploration is still at an early stage, Samsung’s role hasn’t been confirmed, and Claude’s scaled compute still relies on AWS, Google, and Nvidia.