The plans are early. Nothing is finalized. Anthropic hasn't committed to a chip design, hasn't assembled a dedicated hardware team, and may still decide to keep buying compute from Nvidia and its cloud partners. But the fact these conversations are happening at all tells you something about where the AI hardware industry is heading.
Quick Rundown
Anthropic is in early-stage talks with Samsung to manufacture a custom AI accelerator chip, targeting Samsung's 2-nanometer fabrication process. The company currently leans on Google TPUs, Amazon Trainium/Inferentia, and Nvidia GPUs for training and inference. Nothing is committed yet , no design, no dedicated chip team.
The financial context matters here. Anthropic's run-rate revenue crossed $30 billion in early 2026, up from roughly $9 billion at the end of 2025. A chip program could cost around $500 million to execute, which is meaningful but not crazy at that scale. The company is also reportedly finalizing a $20 billion funding round.
So Why Bother?
Anthropic already has access to some of the best chips on the planet. Google's TPUs, Amazon's Trainium and Inferentia, Nvidia's H100s and H200s. That's a genuinely strong lineup. So what's the problem?
Money and control. Pretty much always comes down to those two things.
Compute is the single largest cost in running a frontier AI lab. Training a cutting-edge model like Claude costs hundreds of millions in raw compute. Inference , serving responses to millions of users every single day . Adds up fast too. Even a 20 or 30 percent improvement in chip efficiency at Anthropic's scale translates to savings that are genuinely hard to wave away.
Then there's the control issue. Right now, Anthropic's hardware roadmap is essentially dictated by whatever Google, Amazon, and Nvidia decide to build. If Google optimizes its next TPU for a different workload profile, or if Nvidia's supply gets constrained, Anthropic feels it downstream. Your own chip means you get to define the memory bandwidth, the precision formats, the interconnect architecture . All the stuff that determines how efficiently your specific models actually run.
When your entire business is built around one family of models, that kind of hardware-software co-design stops looking like a luxury.
The Samsung Angle
According to The Information, Anthropic has begun early-stage work on a custom AI chip and held talks with Samsung as a potential manufacturing partner. The chip would reportedly target Samsung's 2-nanometer fabrication process, one of the most advanced nodes currently in or near production.
Timing is interesting. This reporting landed just days after OpenAI announced its own custom chip, codenamed "Jalapeño," developed with Broadcom. The AI chip race is clearly picking up speed.
Why Samsung and not TSMC? Fair question. TSMC manufactures chips for Apple, Nvidia, Qualcomm, basically the whole industry. But TSMC's leading-edge nodes are notoriously oversubscribed . Getting a slot at 2nm or 3nm is competitive and expensive. Samsung has been aggressively courting new customers and has more accessible capacity at advanced nodes. For a company designing its first chip, working with a partner that's hungry for your business, even if yield rates aren't quite TSMC-level, might just be the smarter play.
Reuters also reported in April 2026 that Anthropic was weighing whether to build chips at all. The company may still decide to only buy AI chips rather than design any. Genuinely exploratory territory.
What Anthropic Actually Runs On Today
Google TPUs handle a significant chunk of Anthropic's compute. Google has invested heavily in Anthropic, and TPUs are well-suited to the transformer architectures powering Claude , purpose-built for the matrix multiplications that dominate neural network work, refined since 2016.
Amazon Trainium and Inferentia cover the AWS side. Anthropic has a major partnership with AWS, and as part of deal uses Amazon's custom chips: Trainium for training, Inferentia for inference.
Nvidia GPUs fill in the gaps, particularly where the CUDA software ecosystem makes development faster or where specific model components run more efficiently on general-purpose hardware.
A custom Anthropic chip probably wouldn't replace all of this. More likely it handles a specific slice , probably inference, tightly optimized for Claude's architecture . While training keeps running on TPUs and GPUs.
Everyone's Doing This
Anthropic isn't doing anything unusual. If anything, it's following a well-worn path.
Google started designing TPUs internally in 2013, deployed them in data centers by 2015, and is now on the seventh generation. Amazon launched Inferentia in 2019 and Trainium in 2021. Meta developed its MTIA chip, which powers recommendation models and is increasingly used for generative AI inference. And OpenAI just announced "Jalapeño" with Broadcom . Its first foray into custom silicon after years of total dependence on Nvidia.
The pattern is consistent. Once an AI company reaches a certain revenue and compute scale, the economics of custom silicon start clicking. You're spending so much on compute that even a $500 million upfront investment in chip design pays off within a few years of efficiency gains.
But It's Actually Hard
Building a chip is genuinely difficult, and Anthropic would be taking on a lot.
Talent is scarce. Chip design requires RTL engineers, physical design specialists, verification engineers, compiler writers . A skill set in high demand and short supply. Companies like Apple, Nvidia, and Qualcomm have spent decades building these teams.
The timeline is long. From initial architecture design to first silicon typically takes 18 to 24 months, minimum. Then comes validation, software stack development, production ramp. A chip Anthropic starts designing today realistically won't be in production until 2028.
Software is often harder than hardware. This is where many custom chip programs quietly fail . The hardware works, but the software ecosystem isn't ready, and teams end up falling back on Nvidia's CUDA anyway. A custom chip is useless without the compilers, kernels, and tooling researchers need to actually run models on it.
Manufacturing risk is real too. Working with a new partner on a leading-edge node means early production runs with potentially low yields. First chips off the line often have high defect rates. Expensive.
None of these are dealbreakers. Google, Amazon, and Meta have all navigated them. But they explain why Anthropic's plans are still early-stage . And why the company might ultimately decide the juice isn't worth the squeeze, at least not yet.
What About Nvidia?
Short term? Not much changes. Anthropic is nowhere near shipping a custom chip, and even if it does, it would likely cover a subset of workloads rather than replacing Nvidia across the board. Every company that has built custom silicon, Google, Amazon, Meta still buys Nvidia GPUs in massive quantities. The ecosystems are complementary.
Longer term, the trend matters. As more AI companies build chips optimized for their specific models, the addressable market for general-purpose GPU compute grows more slowly. Nvidia knows this, which is why it has invested so heavily in software CUDA, cuDNN, the broader developer ecosystem to keep switching costs high. If your entire codebase is written in CUDA, walking away from Nvidia is painful even if a competing chip offers better raw performance.
But the software lock-in argument gets weaker every year. Frameworks like JAX, PyTorch 2.0, and emerging compiler technologies like MLIR are making it easier to target multiple hardware backends. Anthropic already runs models on TPUs, non-CUDA hardware, so it has more experience here than most labs.
The Bigger Picture
There's a strategic dimension beyond cost savings. AI labs controlling their own hardware can iterate faster, optimize more aggressively, and reduce dependence on suppliers who may have competing interests.
For Anthropic specifically a company positioning itself as a safety-focused lab having more control over the full stack is probably a good thing. Custom hardware means building in the memory architectures, precision formats, and compute patterns safety research actually needs, rather than working around constraints designed for someone else's use cases.
Geographic diversification matters too. The concentration of advanced chip manufacturing in Taiwan and ongoing US-China semiconductor tensions are real concerns for any company whose business depends on a steady supply of leading-edge chips. Samsung's fabs in South Korea offer some buffer against that.
Where This Leaves Us
Anthropic exploring custom AI chips is a logical next step for a company at its scale. The economics make sense.But precedent is well-established. And the Samsung partnership, if it materializes, would give Anthropic a credible path to first silicon within a few years.
What's interesting isn't just that Anthropic is doing this. It's the entire AI industry is converging on the same conclusion: if you want to be serious at the frontier, you eventually need to control your own hardware. The era of AI labs being pure software companies, entirely dependent on third-party chips, is quietly ending.
Whether Anthropic's chip program actually ships remains to be seen. The company could still decide to keep buying compute instead of designing it. But the conversation is clearly underway. And in this industry, that's usually how these things start.
Sources
- CNBC / Reuters — "Anthropic weighs building its own AI chips". Https.//www.cnbc.com/2026/04/10/anthropic-weighs-building-its-own-ai-chips-reuters.html
- 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/
- Techzine — "Anthropic is considering developing its own AI chip with Samsung". Https.//www.techzine.eu/news/infrastructure/142662/anthropic-is-considering-developing-its-own-ai-chip-with-samsung/
- UPI — "Anthropic eyes South Korea's Samsung for custom AI chip" (July 3, 2026). Https.//www.upi.com/Top_News/World-News/2026/07/03/Anthropic-Samsung-Electronics/7811783128641/
- Taipei Times — "Anthropic is weighing building its own artificial intelligence chips" (April 11, 2026). Https.//www.taipeitimes.com/News/biz/archives/2026/04/11/2003855386
- SiliconANGLE — "Anthropic reportedly in talks with Samsung to manufacture custom AI chip" (July 2, 2026). Https.//siliconangle.com/2026/07/02/anthropic-reportedly-talks-samsung-manufacture-custom-ai-chip/
- Yahoo Finance / Investing.com — "Anthropic explores Samsung 2nm chip partnership" (July 2026): https://finance.yahoo.com/technology/ai/articles/anthropic-explores-samsung-2nm-chip-144844786.html