TL;DR
· Anthropic is reported to be exploring a custom AI server chip but has not yet confirmed the design, tapeout, or production schedule.
· OpenAI has unveiled the Jalapeño inference chip and begun testing, aiming for deployment by the end of 2026.
· Samsung may benefit from the AI chip outsourcing trend, but Anthropic still relies on AWS, Google, and NVIDIA's computing power in the short term.

The discussion around Anthropic's in-house AI server chip is heating up, but it is not yet a finalized chip order line. The key focus of external attention is that Claude's inference costs, GPU supply, data center power, and rack capacity are becoming hard constraints for large model companies. OpenAI has disclosed the Jalapeño inference chip developed in collaboration with Broadcom, and Anthropic has also been reported to be evaluating dedicated chips more suitable for its own models. However, based on current public information, whether Samsung is involved in manufacturing, and whether the project has entered formal design, are still unconfirmed.

Anthropic is still in the early exploration stage, not on the eve of mass production

Anthropic is reported to be exploring a direction for a server chip that is more suitable for its AI model operation. Compared to general-purpose GPUs, a successful custom chip may reduce costs, improve efficiency, and reduce reliance on external chip supply for specific inference tasks.

The challenge of such chips lies not only in the performance of a single chip. Large model companies need to consider computing speed, memory bandwidth, interconnection networks, power consumption, heat dissipation, and cluster stability simultaneously. The real difficulty lies in enabling thousands of chips to collaborate stably in data centers and continually serve training or inference tasks.

The more prudent description at present is that Anthropic is still in the early evaluation and definition stage. There are no clear public answers to questions such as which AI tasks the chip will primarily handle, how performance and power consumption goals are set, how servers and clusters will be adapted, and whether external chip design companies need to be involved.

The company's external communication remains cautious. In April of this year, Anthropic announced an expanded partnership with Amazon, with plans to invest over $100 billion in AWS technology over the next decade, securing up to 5GW of capacity at most, and stating that they have utilized over 1 million Trainium2 chips to train and serve Claude. Anthropic also emphasizes a diverse hardware strategy, but AWS remains its primary training and cloud service provider.

This means that even as in-house chip exploration continues to advance, it will be difficult to replace existing suppliers in the short term. AWS Trainium, Google TPU, and NVIDIA GPU are still integral parts of Anthropic's scaled computing power system.

OpenAI Takes the Lead, Making Inference Cost Pressure More Direct

Anthropic is now being brought into the discussion of proprietary chips, with an important backdrop being OpenAI's prior reference.

An official announcement from Broadcom revealed that on June 24, 2026, OpenAI and Broadcom unveiled Jalapeño, positioned as an accelerator for large language model inference, also known as the Intelligence Processor. OpenAI and Broadcom stated that this chip took about 9 months from initial design to tapeout, with engineering samples already running in the lab and plans to begin deployment by the end of 2026.

It's important to distinguish two phases here. While Jalapeño has been announced and is in testing, it does not mean it has already been massively commercially deployed. It signifies that leading model companies are starting to incorporate inference costs into a deeper level of hardware control, rather than an immediate replacement of GPU demand.

Inference is the computational process where a model generates answers when users interact with products like ChatGPT and Claude. Compared to training, inference occurs more frequently, and as user scale increases, cost pressures will continue to rise. For large model companies, even if the cost of a single inference decreases by a small percentage, when placed in the context of massive requests and long-term data center expenditures, it could turn into significant savings.

Anthropic's pace is notably earlier. They have not disclosed chip specifications, performance metrics, partner lists, or mass production schedules. OpenAI's progress just shows the market a direction: the top model companies are no longer just buying GPUs but are also attempting to bring some of the compute infrastructure under their own control.

Samsung's Imagination Space Heats Up, But Orders Are Not Yet Set

Samsung is drawing market attention because it possesses advanced manufacturing capabilities and is also seeking more AI chip foundry opportunities. Following the news surrounding Anthropic's funding and infrastructure partnerships, the outside world naturally connects Samsung with potential AI accelerator manufacturing opportunities.

However, this point needs to be viewed with caution. Publicly available information confirms that companies such as Samsung, SK Hynix, and Micron have appeared in discussions with Anthropic regarding infrastructure partnerships. Micron announced a strategic agreement with Anthropic on June 22, 2026, which includes AI architecture design for memory and storage, supply agreements, internal adoption of Claude by Micron, and a strategic investment in Anthropic Series H.

These partnership signals cannot be directly equated to Samsung having secured an order for Anthropic's in-house chip. The claim that Anthropic has engaged with Samsung for manufacturing cooperation lacks sufficient publicly verifiable information. A more cautious assessment would be that if Anthropic's in-house chip project advances to the manufacturing stage, Samsung could become one of the market's potential participants of interest, but it cannot currently be definitively stated as a done deal.

For a chip project, from early evaluation to final production, it still needs to go through architecture finalization, design validation, manufacturing process selection, packaging testing, and supply chain coordination. As long as the chip design is not finalized, the foundry role is also challenging to firmly establish.

Recruitment to Increase Credibility, but Path Forward Still Uncertain

The talent movements have brought more attention to Anthropic's hardware clues. Reportedly, Clive Chan, an early member of the OpenAI custom chip team, has joined Anthropic. Public records show that he was involved in the early development of the OpenAI chip team and also has experience related to Tesla Dojo. Anthropic has also recently intensified its recruitment of chip engineers.

This indicates that the company is at least preparing for hardware capabilities. For a modeling company, a hardware team knowledgeable in models, inference workloads, and data center systems can help determine which tasks are suitable for custom chips and which should still rely on GPUs, TPUs, or cloud vendor chips.

However, talent acquisitions and recruitment expansion are still only early signals of investment. Whether the project can continue depends on whether the chip can gain sufficient advantages in cost, performance, power consumption, and deployment complexity. If the custom chip can only improve efficiency on paper but cannot run stably on a large scale, or if the manufacturing and software adaptation costs are too high, the company may still primarily rely on external chips.

This is also why NVIDIA is not easily replaceable in the short term. NVIDIA GPUs are still the mainstay of AI training and inference, with mature software ecosystems, and data center customers have built a large number of systems around its platform. In-house chips are more likely to initially share some workload in specific inference scenarios rather than fully replace GPUs.

For investors, the real-world impact of Anthropic's in-house chip discussions is more like a short-term supply chain game. Leading model companies hope to gain more computing power choices, and cloud vendors, Broadcom, Samsung, TSMC, memory manufacturers, and advanced packaging supply chains could all benefit from this trend. However, in Anthropic's case, clear facts are still limited: in-house exploration is still in the early stages, Samsung's role is unconfirmed, and Claude's scalable computing power still relies on AWS, Google, and NVIDIA.

Welcome to join the official BlockBeats community:

Telegram Subscription Group: https://t.me/theblockbeats

Telegram Discussion Group: https://t.me/BlockBeats_App

Official Twitter Account: https://twitter.com/BlockBeatsAsia