Key Takeaways
- Anthropic is developing custom AI chips targeting a 2028 deployment, with the goal of cutting its roughly $19 billion annual compute spend by up to 65% in total cost of ownership compared to general-purpose GPUs.
- AMD’s deal to deploy up to 2 gigawatts of Instinct MI450 Series GPUs, plus a potential $5 billion investment in Anthropic, shows the external partnerships Anthropic will run alongside any in-house silicon.
- Anthropic is reportedly pursuing a $6 billion acquisition of chip optimisation software startup Decart and offering salaries as high as $485,000 to recruit chip design engineers, signalling that the 2028 target depends on significant hiring and M&A execution. Anthropic’s compute bill runs to roughly $19 billion a year, and custom silicon is how the company plans to bring it down. The Anthropic chip programme, targeting deployment around 2028, mirrors moves already made by Google, Amazon and Meta, and comes as the company’s annualised revenue hits $65 billion, giving it the financial runway to fund the effort.
The Build-vs-Buy Calculus
For a frontier lab processing tokens at Anthropic’s scale, the arithmetic on compute costs is unforgiving. The company’s estimated $19 billion compute spend in 2026, against annualised revenue of $65 billion as of August 2026, puts hardware economics at the centre of its profitability problem. The key variables are long-term operational cost, how tightly hardware can be tuned to Claude’s model architectures, supply chain control, and the capital required to get there. Each of those pulls in a different direction depending on whether you buy from NVIDIA or build your own.
The External Hardware Stack
Anthropic’s current infrastructure spans several large partnerships. On the Google side, the company has secured roughly 3.5 gigawatts of next-generation Google and Broadcom TPU capacity expected online in 2027, building on a gigawatt of capacity from a Google Cloud agreement signed in October 2025. On the AWS side, close to one million Trainium2 chips were reportedly used to train and serve Claude through Project Rainier by end of 2025.
In July 2026, AMD added a third major thread: a deal committing up to 2 gigawatts of Instinct MI450 Series GPUs, with the first gigawatt deploying in the first half of 2027, plus a potential $5 billion investment in Anthropic. These partnerships spread capital expenditure and provide access to current hardware generations, but they also mean Anthropic is exposed to pricing set by manufacturers whose gross margins on AI accelerators have run above 70%. Supply chain bottlenecks and hardware roadmap decisions stay outside Anthropic’s control as long as the company relies entirely on third parties, a dependency that only grows as Claude scales. For more on how cloud providers are navigating GPU supply constraints the structural pressures are well-documented.
The Case for Custom Silicon
Custom chips let Anthropic co-design hardware and model together, tuning silicon specifically for Claude’s architectures rather than optimising Claude to fit hardware built for the general market. Tom’s Hardware reported in August 2026 that custom silicon can deliver a total cost of ownership up to 65% lower than general-purpose GPUs, a figure that, at Anthropic’s compute volumes, translates to billions of dollars annually.
The upfront cost is real. Designing an advanced AI chip runs to roughly $500 million, and when software stack development is included, total costs can exceed $1 billion. Anthropic’s Q2 2026 revenue of approximately $11.6 billion provides the headroom. To accelerate the programme, the company is reportedly exploring a $6 billion acquisition of Decart, a startup specialising in chip optimisation software, and is recruiting chip design engineers at salaries as high as $485,000. The 2028 target also runs into physical constraints: advanced packaging capacity and high-bandwidth memory are in short supply, and leading foundries including TSMC have capacity booked out to at least 2028. Anthropic’s custom chip work faces the same queue everyone else does, as covered in our look at how hyperscalers are managing their own silicon build-outs.
Trade-offs at Each Layer
The external GPU path offers capacity on demand and no chip design risk, but Anthropic pays NVIDIA’s margins and accepts hardware roadmaps it cannot influence. The custom silicon path offers cost control and model-hardware co-optimisation, but requires years of development, specialised talent, and successful navigation of a constrained foundry market.
Enterprise revenue, reported at over 70% of Anthropic’s Q2 2026 total, is particularly sensitive to inference cost and latency. Custom chips optimised for high-frequency inference workloads would directly improve margins on that business, but the benefits only materialise if the 2028 deployment holds, which depends on hiring execution, the Decart deal closing, and manufacturing slots coming through.
A Hybrid Strategy by Design
Anthropic has been explicit that this is a multi-chip strategy, not a supplier replacement. AWS, Google, NVIDIA and AMD will remain part of the stack even after in-house silicon arrives. The custom chips are intended to cover the most cost-sensitive, high-frequency workloads, primarily inference, while external accelerators handle training and specialised tasks where Anthropic’s own designs may not be optimal. That split reduces single-source risk and keeps access to the broader hardware ecosystem intact.
The 2028 window is tight. TSMC’s foundry calendar is heavily subscribed, and every major hyperscaler is running a parallel custom silicon programme. Whether Anthropic can deliver on schedule will determine how much of that projected 65% cost reduction actually reaches the income statement.
Originally published at https://autonainews.com/anthropic-builds-custom-chips-to-cut-its-19b-annual-compute-bill/
Top comments (0)