Anthropic Eyes Samsung to Build Its First AI Chip
Anthropic is in early talks with Samsung to build a custom AI chip. Here's what it signals about the race to break free from Nvidia's hardware dominance.
Key takeaways
- Custom AI chips are purpose-built accelerators designed around a specific lab's model architecture. They trade the general-purpose flexibility of a GPU for dramatically lower inference costs on targeted workloads.
- The shift toward in-house silicon is not a rejection of Nvidia, but a strategic hedge.
- The gap between design talks and a chip running in a data center typically spans two to three years and hundreds of millions in development costs.
- For Samsung Foundry, landing a marquee AI client would represent a meaningful challenge to TSMC's dominance of leading-edge chip manufacturing.
Anthropic has entered preliminary discussions with Samsung Electronics to manufacture its first custom AI chip. This move places the Claude developer alongside OpenAI, Google, Meta, and Amazon in a quiet but accelerating push to reduce the industry's dependence on Nvidia hardware.
The talks are still early-stage, with no commitments made on either side. But the direction of travel is clear.
Why Anthropic Needs Its Own Chip
The story behind the Samsung discussions is about control. Anthropic currently runs its Claude models on hardware it rents:
- Chips from Amazon's Trainium line
- Google's Tensor Processing Units
- Nvidia's GPUs
Every query processed, every API call served, every Claude Code session completed runs on silicon owned and priced by someone else. At a run-rate revenue of $47 billion and a user base that has grown 75% since January, the economics of renting compute are beginning to work against Anthropic.
Custom silicon changes that arithmetic. A chip designed specifically around Claude's inference workloads, the process of generating model responses in real time, eliminates the overhead that general-purpose GPUs carry for tasks beyond AI.
According to reporting by The Information, Anthropic is specifically evaluating Samsung's 2-nanometer manufacturing process and its advanced chip packaging facilities.
The company recently hired Clive Chan, the second engineer ever to join OpenAI's dedicated custom chip team, and the person who spent two and a half years building the architecture that became Jalapeño, as a signal that this effort has moved from exploration to active development.
The OpenAI Jalapeño Announcement Changed the Timeline
On June 24, 2026, OpenAI and Broadcom unveiled Jalapeño – a custom inference accelerator designed from scratch for large language model workloads, co-developed in just nine months. OpenAI called it the fastest ASIC development cycle ever achieved in high-performance semiconductors.
Jalapeño is designed for inference. That distinction matters. Training a frontier model requires months of computation on massive GPU clusters where Nvidia's hardware remains difficult to displace. Inference, however, happens billions of times per day across every consumer product an AI lab ships. It is where compute costs accumulate fastest, and where a purpose-built chip offers the clearest return.
OpenAI's announcement was a competitive signal the rest of the industry could not ignore. Anthropic's Samsung discussions surfacing eight days later was not a coincidence.
Why Samsung – and What It Gets From This Deal
Samsung's role in this story has a second dimension. Earlier this year, the Korean chipmaker had been developing a custom AI chip for OpenAI – an ARM-based inference neural processing unit. Those talks stalled in early June 2026, with OpenAI CEO Sam Altman cancelling a planned visit to Seoul.
If Samsung redirects the engineering work and 2nm manufacturing capacity it had pointed at the OpenAI project toward Anthropic instead, both sides gain something concrete.
- Anthropic would acquire a foundry partner with directly applicable experience in AI inference chip design.
- Samsung would secure a marquee client at a critical moment in its effort to close the gap with TSMC – Taiwan Semiconductor Manufacturing, which remains the dominant producer of leading-edge AI processors.
Samsung's position in this negotiation is also strengthened by an existing financial relationship. The company participated in Anthropic's $65 billion Series H funding round in May as a "strategic infrastructure partner," alongside SK Hynix and Micron. Samsung is the only one of those three memory and storage investors that also operates its own foundry business, making it the natural candidate for a manufacturing conversation.
Anthropic has confirmed the discussions are not exclusive. The company is also evaluating chips from Microsoft and UK-based startup Fractile, indicating a competitive process rather than a committed partnership.
>> Read more: What You Really Own When You Buy an “Anthropic Token”
Every Major Lab Is Now in the Custom Silicon Race
Anthropic and OpenAI are the most recent entrants in a pattern that has been building for years.
- Google has operated its own Tensor Processing Units since 2016 and currently deploys the latest generation at data center scale.
- Amazon offers its Trainium line to cloud customers.
- Meta has shipped multiple generations of its in-house accelerators for both training and inference.
- Microsoft launched its Azure Maia 200 chip on TSMC's 3-nanometer process in January 2026, already powering some of OpenAI's GPT-5.2 workloads.
- ByteDance is reportedly in active negotiations with Qualcomm for custom ASIC designs.
Nvidia, despite the competitive pressure, has not lost ground in absolute terms. Its share of the global AI chip market currently stands at approximately 74% – higher than before the custom silicon race intensified – because overall demand for AI training infrastructure has grown faster than alternatives have matured.
No frontier lab is abandoning Nvidia to build a custom chip. They are building around it.
What This Means for the Broader Market
Owning the chip design also means owning the leverage. No supply contract gives a lab the same control over its inference stack as proprietary silicon does, and that distinction is part of why capital has rotated from crypto into AI and semiconductor stocks throughout 2026.
The custom chip race also has two market-level implications that extend beyond AI company strategy.
- For Samsung Foundry, a confirmed Anthropic relationship would validate its 2nm process as commercially competitive – a meaningful development for an operation that has struggled with yields and customer acquisition relative to TSMC. Google is separately in discussions with Samsung about contributing to future Tensor Processing Unit production. If Samsung secures both, it changes the foundry competitive landscape.
- For the semiconductor sector broadly, the buildout signals that AI infrastructure spending is entering a new phase, one where the value capture shifts from chip volumes to chip design capability. Broadcom, which designs chips for OpenAI and Meta, has been one of the clearest beneficiaries of this transition. The companies that design are now at the center of the most durable AI infrastructure contracts.
Sources
- The Information – Anthropic in Talks With Samsung to Manufacture Custom AI Chip https://www.theinformation.com/articles/anthropic-talks-samsung-manufacture-custom-ai-chip
- TechCrunch – Anthropic is discussing a new custom chip with Samsung https://techcrunch.com/2026/07/02/anthropic-is-discussing-a-new-custom-chip-with-samsung/
- OpenAI – OpenAI and Broadcom unveil LLM-optimized inference chip (Jalapeño) https://openai.com/index/openai-broadcom-jalapeno-inference-chip/
- CNBC – OpenAI and Broadcom reveal Jalapeño, first AI chip in partnership https://www.cnbc.com/2026/06/24/openai-and-broadcom-reveal-jalapeno-first-ai-chip-in-partnership.html
- CNBC – Anthropic, SpaceX announce compute deal https://www.cnbc.com/2026/05/06/anthropic-spacex-data-center-capacity.html
FAQs
A GPU is a general-purpose parallel processor, originally designed for graphics and later adapted for AI. A custom AI chip, typically called an ASIC or accelerator, is designed from scratch for a specific type of workload. It sacrifices versatility for efficiency: it does fewer things, but does them significantly faster and cheaper than a GPU built for everything.
