Custom AI silicon is going from avant-garde choice to standard equipment for frontier AI companies.

On August 5, Anthropic officially confirmed it's building an internal custom chip team to design purpose-built AI silicon for the Claude model family. Job listings for chip design roles are already posted, with annual compensation ranging from $320,000 to $485,000 — top-tier pay in the semiconductor industry, on par with senior engineer compensation at NVIDIA and Apple.

The Information previously reported Anthropic was in talks with Samsung about potential chip manufacturing partnership. TechCrunch's reporting confirmed this, stating that Anthropic's goal isn't to build general-purpose GPUs but to pursue hardware-software co-design specifically for Claude's inference and training workloads, making models run faster, more power-efficient, and cheaper.

From Buying GPUs to Building Chips: AI Companies Move Upstream

Anthropic isn't the first AI company to build custom chips, but its entry marks a trend firmly established: frontier AI companies are systematically pushing upstream, transforming from "GPU buyers" into "chip designers."

Let's walk through the timeline: Google has used custom TPUs since 2015, pioneering this path; Meta has the MTIA accelerator, iterated across several generations; Amazon AWS has Trainium and Inferentia; Microsoft is developing the Maia chip; OpenAI launched the Broadcom-partnered Jalapeño inference chip in June; and now Anthropic joins the fray.

Why is every company building custom chips? Three fundamental reasons.

First, the efficiency ceiling of general-purpose GPUs. NVIDIA GPUs were designed for graphics rendering and general-purpose parallel computing, not specifically optimized for LLM inference and training. LLM computation patterns are actually quite regular — most time is spent on matrix multiplication, attention has fixed memory access patterns, KV caching has specific optimization opportunities. An ASIC purpose-built for Transformer architectures can deliver 3–5x performance at 1/3 to 1/5 power consumption on specific workloads. Google's TPUs already proved this; OpenAI's Jalapeño is rumored to deliver over 4x inference efficiency per watt versus H100.

Second, supply chain security and cost control. NVIDIA GPUs aren't just expensive (H100s run $30–40K each; B200s cost more) — they're also supply-constrained, with production bottlenecked by TSMC CoWoS packaging and order backlogs stretching over a year. If your entire business depends on a single supplier (NVIDIA), not only can't you bring costs down, supply chain security becomes a critical issue. If NVIDIA prioritizes other customers, cuts off supply to certain regions, or has a recall, your entire operation grinds to a halt. Building custom chips requires massive upfront investment (hundreds of millions to billions), but long-term it dramatically reduces compute costs and ensures supply chain autonomy.

Third, the moat of hardware-software co-design. When you design your own chips, train your own models, and build your own framework optimizations, you can achieve deep co-optimization that general-purpose GPUs can never match. For example, you can directly support FlashAttention at the chip instruction set level, implement KV cache compression in hardware, or specialize for your model's sparsity patterns. This kind of deep hardware-software binding creates performance advantages that buying general-purpose GPUs can never catch up to — Apple's M-series chips already proved this in consumer electronics.

Anthropic's Advantages and Challenges

Entering the chip race now, Anthropic has its advantages and also significant challenges.

On the advantage side: First, it doesn't have to start from zero — it can partner with design services firms like Broadcom (OpenAI's Jalapeño was also designed with Broadcom), leveraging mature design teams and IP without maintaining a multi-thousand-person chip design organization in-house. Second, it already has large-scale compute agreements with AWS, Google, NVIDIA, and AMD; custom silicon is a supplement, not a replacement, so there's no existential "all-in on custom chips" pressure, and it can iterate deliberately. Third, the Claude model family's architecture is transparent and well-published (Anthropic has released extensive technical papers); the team intimately understands its models' computational characteristics, giving clear direction for hardware-software co-optimization.

The challenges are also obvious. First, chip design cycles are extremely long — from architecture to tapeout to mass production typically takes 2–3 years. Starting now, real deployment might not happen until 2028–2029. Will model architectures have changed again by then? If the dominant architecture shifts away from Transformers in 2–3 years, chips designed today could be obsolete. Second, talent competition is ferocious — AI chip designers are the scarcest talent on the planet; NVIDIA, Apple, AMD, Qualcomm, OpenAI, and Google are all competing. Whether $485K can attract enough top talent remains an open question. Third, manufacturing depends on TSMC — Samsung's 4nm/3nm yields still lag behind TSMC, and manufacturing with Samsung could mean performance gaps versus TSMC-built competitors.

But none of these challenges will stop the trend. AI has reached a point where compute is no longer a "cost center" but a "core competency." A frontier AI company that doesn't have its own secret sauce at the compute level will be increasingly disadvantaged. After signing a $10B long-term compute deal with Volta and locking in power with SpaceX and TeraWulf, Anthropic is now designing custom chips — this is a complete full-stack compute strategy.

The AI arms race has burned from the algorithm layer up to the silicon layer.

See you tomorrow.