Anthropic is co-designing custom AI inference ASICs with Samsung as manufacturing partner, targeting inference-specific optimization to reduce compute costs and Nvidia GPU dependency.

Anthropic has announced it's building an in-house chip development team to co-design custom ASIC processors for handling AI inferencing workloads. As described to Business Insider, Anthropic is hiring engineers to design the chips with an unspecified partner and is proceeding at pace, with job listings indicating potential hires would need to work to a demanding schedule to deliver the chip design.

This is the latest major AI company to announce custom hardware development. As global economic pressures squeeze chip supply and models increasingly require optimization to improve efficiency and profitability, designing custom silicon for proprietary data centers has become strategically important. Anthropic now joins Google, Meta, Microsoft, Amazon, and OpenAI in building their own inference chips.

Training advanced AI models requires Nvidia GPUs—even Chinese developers with limited access still use them, smuggled or otherwise. However, running inference workloads for agentic and generative AI models allows for much wider hardware options. While Nvidia GPUs are effective, they're expensive—custom silicon can achieve up to 65% lower total cost of ownership—and power-intensive. More efficient alternatives exist: Chinese labs use domestic hardware, and many Western developers have built or are building their own solutions.

Google has been developing Tensor Processing Unit (TPU) chips for 12 years, working with Broadcom on each generation. Amazon produces Trainium and Inferentia chips. Meta recently announced new MTIA designs for deployment through 2027. Microsoft developed its Maia line. Tesla pivoted to AI5 and AI6 designs after years developing Dojo.

Anthropic told Business Insider it is co-designing chips to allow Claude to run faster and more efficiently at the scale its customers require. The company has experienced explosive growth over the past year, expanding significantly in consumer markets and securing major government contracts. Agentic AI also demands far more tokens than traditional single-prompt interactions, driving the need for better inference infrastructure.

Anthropic has not disclosed its design and development partner. While Broadcom and Marvell dominate the ASIC co-design market with approximately 95% combined share, The Information reported that Anthropic was in talks with Samsung for manufacturing.

Designing, packaging, and manufacturing custom ASICs for AI inference is neither cheap nor straightforward. Beyond developing the hardware, companies must build supporting software stacks and optimize models to leverage the hardware's advantages. This complexity deters most companies from pursuing custom silicon, but for major AI firms with sufficient capital and scaling capacity, the investment yields significant returns.

While AI developers benefit, the broader landscape shows clear winners and losers. Nvidia has been among the few companies profiting substantially from the AI boom, while Meta, Google, Microsoft, and OpenAI lose enormous sums on AI initiatives. The custom ASIC co-design market generates major profits: TSMC, Broadcom, and Marvell are the clear winners.

Broadcom has been Google's co-design partner for years and recently was tapped to build OpenAI's inference chips. It also partners with Meta on MTIA and holds contracts for custom ASICs with ByteDance and Fujitsu. Its strong interconnect and networking hardware enable complete solutions for customers. Broadcom reports a $73 billion order backlog and expects to generate over $100 billion in annual AI chip revenue by end of 2027.

Marvell holds significant contracts with Amazon for Trainium chips and Microsoft for Maia, expected to generate over $11 billion in 2026 from these co-design engagements.

If Samsung becomes Anthropic's partner, it would be a relatively small player in this space but brings substantial manufacturing and chip design expertise, plus access to critical memory supplies in short global supply.

The largest winner appears to be TSMC. The Taiwanese company produces most of the world's cutting-edge silicon and is involved in nearly all the chips discussed here. Advanced packaging with HBM integration requires TSMC's CoWoS technologies. TSMC also handles significant packaging for Nvidia and AMD GPUs and produces many underlying wafers.

Anthropic's entry into the custom ASIC race is unsurprising and reinforces an emerging trajectory: major hyperscaler AI companies will increasingly handle inference with custom hardware. This approach is more efficient, allows better control over features and specifications, and scales more readily when optimized for internal models. For Chinese companies, it also sidesteps international trade blockades and tariffs.

While companies will likely never rely entirely on proprietary chips—AI demand is too vast and Nvidia too dominant in supply chain control—each new custom chip installed is one fewer Nvidia GPU needed for inference. This may gradually reduce Nvidia's stranglehold on the market. Training, however, will likely remain Nvidia's strongest position for the foreseeable future.