Why Anthropic and rivals are racing to build custom AI chips

For most Australians who use Claude, ChatGPT or Gemini, the machine behind the chat window is an abstraction. You type, it answers. But the physical layer underneath — the silicon those models run on — has quietly become the most contested piece of the AI industry. Anthropic, the company behind Claude, has been steadily deepening its hardware partnerships and moving toward more customised compute, following a path already worn by Google with its TPUs and Amazon with its Trainium and Inferentia chips.

Why does this matter to anyone who isn't a chip designer? Because the reliability, price and safety of the AI tools we're all starting to depend on hinge, more than most people realise, on who owns the metal.

The compute squeeze is real — and users feel it

The clearest reminder that AI runs on physical infrastructure comes when it breaks. In late 2024, users noticed Claude going dark mid-conversation, prompting explainer coverage from outlets like Android Authority asking simply: Is Claude down for you? Outages of this kind are not just bad luck. They reflect the razor-thin margin between demand for large language model inference and the GPUs available to serve it.

Every AI provider is essentially renting scarce compute — mostly Nvidia H100s and H200s — at premium rates. When usage spikes, or when a hyperscaler reallocates capacity, downstream services stutter. For a company like Anthropic, whose entire commercial pitch is a dependable, safety-tuned assistant sold to enterprises and developers, an outage is more than an inconvenience. It's a churn event.

Custom silicon is the strategic answer. If Anthropic co-designs chips with a partner (its Amazon relationship is the obvious axis), it can lock in capacity, tune the hardware to Claude's specific model architecture, and stop competing with every other AI startup for the same GPU pool.

Claude is being asked to do heavier things

The other pressure driving custom hardware is that people are pushing these models far beyond chatbot territory. Consider the reporting in Tom's Hardware on an enthusiast using Claude Code to defeat RSA-2048 signature checks and unlock 55 hidden BIOS settings on a consumer motherboard. Whatever you think of that particular experiment, it's a striking demonstration of how deeply Claude is being embedded into low-level, iterative reasoning tasks: reverse engineering, code generation, cryptographic problem-solving.

Agentic workloads like these look nothing like a quick chat. They involve long context windows, many chained inferences, tool use and heavy back-and-forth. Each session can consume orders of magnitude more compute than a casual query. General-purpose GPUs handle it, but they're not optimised for the exact mix of memory bandwidth, interconnect and low-latency inference these agents demand.

Custom accelerators — designed around a specific transformer variant, a specific context length, a specific serving pattern — can crunch the same workload at a fraction of the energy and cost. That's the whole logic behind Google's TPU program, and it's why Anthropic quietly reserving huge blocks of purpose-built silicon is more consequential than any single model release.

Safety and security are compute problems too

There is another, less obvious reason AI labs want to own their hardware stack: safety enforcement is expensive. Anthropic recently disclosed that it had disrupted what it described as a state-linked cyber operation abusing Claude — Information Age covered the episode under the headline "Anthropic's AI escapes, hacks three companies". Detecting and stopping that kind of misuse means running additional classifiers, monitors and guardrail models alongside every user request.

All of that supervision is inference — and inference costs money. If you're paying spot-market GPU prices for both the primary model and the safety layer, the economics get ugly fast. Owning cheaper, purpose-built silicon is how you keep safety-heavy pipelines viable without passing costs to customers or watering down the guardrails.

The same argument applies to privacy. As ZDNET has laid out in its guide on keeping conversations with ChatGPT, Gemini, Copilot and Claude private, users are increasingly expected to manage what their AI provider retains, trains on and stores. Enterprise customers — hospitals, banks, government departments — want stronger guarantees: dedicated instances, isolated inference, sometimes on-premise deployment. Delivering that at scale is far easier when the provider controls the chip roadmap and can offer confidential-compute variants without begging Nvidia for allocation.

What it means for accessibility

The narrative around custom AI chips is usually framed as a battle of titans — Nvidia versus Google versus Amazon versus a scrappy pack of startups. That misses the more interesting question for ordinary users: what happens to access if Anthropic and its peers succeed?

There are two plausible futures.

Both futures are live. Which one arrives depends less on the technology than on regulatory posture — whether competition regulators treat vertically integrated AI stacks the way they've treated app-store lock-in — and on whether open-weight models running on commodity hardware can keep pace with frontier proprietary systems.

The takeaway for Australian readers

Australia is not going to build a domestic frontier AI lab any time soon, and it isn't going to fab leading-edge chips either. But we are heavy consumers of these tools, and our institutions — from Services Australia to the big four banks to the CSIRO — will increasingly buy AI capacity from whichever providers offer the best mix of reliability, price and sovereign-data guarantees.

That means the custom-chip story is not a spectator sport. When Anthropic locks in silicon capacity with Amazon, or when Google announces the next TPU generation, it directly shapes what Australian businesses can afford, what latency they can promise their own customers, and how much control they retain over sensitive data.

The chatbot in the browser tab is the tip of an iceberg. The interesting movement — the movement that will decide whether AI stays a subscription luxury or becomes genuinely ubiquitous infrastructure — is happening one layer down, in fabs and rack designs and interconnect protocols. Anthropic's growing appetite for custom compute is one of the clearest signals that the era of renting someone else's GPUs is coming to an end.

Related on Bleen

Sources