Vendors & EquipmentAmazon AwsCustom SiliconAI AcceleratorsAsics

Amazon's custom chips top $25bn as ASIC gold rush accelerates

Amazon's custom chip business passed $25bn in annual revenue run rate in July 2026 as hyperscalers and startups race into ASICs, but fab capacity and costs constrain the field.

6 min read

ASIC ambitionVendors & Equipment
ASIC ambitionbluehouseskis / pixabay

Why it matters

  • Amazon reported its custom chip business surpassed $25 billion in annual revenue run rate in July 2026.
  • Startups Etched and Olix raised $700m and $300m in 2026, valued at $21bn and $3.3bn respectively.
  • A UK government report put leading-edge chip design costs at $50-100m and fab construction at about $20bn.
  • TSMC holds roughly 92 percent of global advanced chip production capacity; new capacity arrives near decade's end.
  • Amazon claims 40 percent better cost-performance and five times more tokens per watt versus its previous chip generation.

The story

Amazon's custom chip business passed a $25 billion annual revenue run rate in July 2026, the clearest evidence yet that application-specific silicon has moved from sideline experiment to core infrastructure economics for the largest cloud operators.

The milestone caps a frantic stretch of activity. Between June and August 2026 alone, OpenAI unveiled its 'Jalapeño' Intelligence Processor developed with Broadcom; Anthropic confirmed it is assembling an in-house chip design team; reports emerged that Google is building a new AI chip to run its Gemini model more efficiently; and chip startups Etched and Olix raised $700 million and $300 million respectively, reaching valuations of $21 billion and $3.3 billion.

The pattern traces back to October 2025, when Amin Vahdat, VP and GM of AI and infrastructure at Google, told an a16z Runtime panel in Utah that certain computations run on Google's Tensor Processing Units are "somewhere between 10 and 100 times more efficient per watt than a CPU." He added: "We know that there are other computations that if you built even more specialized systems for … for example, maybe serving agentic workloads, that would benefit from an even more specialized architecture."

Why is everyone building custom silicon now?

Custom silicon is hardware designed in-house to replace generic third-party chips. ASICs — chips hardwired for one distinct task — sit under that umbrella, distinct from general-purpose processors such as standard GPUs and CPUs. All ASICs are custom silicon; not all custom chips are strictly ASICs.

Since OpenAI pushed generative AI into mainstream use in 2022, hardware demand has risen exponentially. As workloads pivot from training to inference — the stage where a trained model answers user queries — companies are pouring money into silicon built specifically for inference.

Ryan Shrout, president and general manager of analyst firm Signal65, part of the Futurum Group, says the startup landscape has always been strong, but capital costs historically barred entry. That has changed. "There's just too much money involved now for [custom chips] to be something that isn't chased by basically everyone that has a good idea," he said.

What does Amazon get out of it?

Amazon entered the field through its 2015 acquisition of chip designer Annapurna Labs. Its portfolio now comprises Trainium for AI training and inference and Graviton for general cloud computing and agentic AI, both built on the Nitro System for security and networking.

In April 2026, Amazon CEO Andy Jassy said two of AWS's biggest customers had asked whether they could buy every single Graviton instance capacity available that year.

Mark Carroll, director of engineering at Amazon, frames the rationale as customer-driven. "The primary reason that we got into [the chip business] … is we're designing specifically to solve customer problems," he said, citing early Nitro offload silicon built to cut customer latency pain.

The commercial payoff is concrete:

  • Around 40 percent better cost-performance across multiple generations, by stripping out "extra features and hardware"
  • Five times more tokens per watt versus the previous generation
  • Power reductions generation over generation, designed around observed customer workload patterns
  • Faster time to market than relying wholly on generic GPUs, plus flexibility to optimize memory interfaces, memory usage and accelerator interconnects

Amazon controls the entire stack. "We do the silicon, we do the boards, we do the chassis, we do the systems, we do the firmware interim drivers, and we do the software on top," Carroll said, which yields feedback and changeability at every level.

What stands in the way?

Money is the largest barrier. A 2025 report from the UK government's Council for Science and Technology put the cost of designing a leading-edge chip at roughly $50-100 million, manufacturing at about $100 million, and a fabrication plant at about $20 billion.

Winners of the funding race then face talent sourcing, volatile supply chains and scarce manufacturing capacity. Most chip companies are fabless and depend on third-party manufacturers. Taiwan's TSMC holds roughly 92 percent of global advanced chip production capacity and is investing hundreds of billions of dollars across the US, Germany, Japan and Taiwan — but the bulk of that new capacity is unlikely to be available much before the end of the decade.

A global memory chip shortage has sharply raised prices over the past 12 months, with SK Hynix and Samsung warning the crunch could persist into 2027 and beyond. Startup Olix says its chips integrate SRAM architecture with photonics, forgoing HBM and insulating its hardware from those constraints.

For startups, prototyping remains feasible; scale does not. "If you need to build 1,000 of something, you could go do that," Shrout said. "The issue will be if a startup now needs to produce three million chips. There are no wafer starts for three million chips sitting around, and Nvidia or AMD or anybody else is not going to give up their capacity."

Timing compounds the problem. Nvidia and AMD have committed to yearly release cadences that smaller players cannot match. Vahdat noted in 2025 that even the world's best design teams take two and a half years to move specialized architecture from concept to production "at the speed of light" — and very few get everything right. "How do you predict the future two and a half years out?" he asked. "We have to shrink that cycle."

Carroll concedes that developing chips 18 months to two years out is "obviously stressful," though he calls it "about as bleeding edge as you can get in the technical industry." To mitigate the pressure, Amazon talks with customers — "whether it's Anthropic or others" — about their roadmaps and designs toward them.

Does this threaten Nvidia?

Shrout thinks not, at least not yet. He describes the current market as a "Wild West scenario" in which nobody is certain what inference or compute requirements agentic AI will demand — a uncertainty demonstrated by the resurgence of CPUs, which nobody predicted three years ago.

The general-purpose incumbents are hedging regardless. Shrout points to Nvidia's Groq pseudo-acquisition and AMD's acquisition of Taalas as signs that major general-purpose compute builders "have started to see the writing on the wall, that there is going to be some benefits to specialization of compute." But specialization requires well-defined workloads, and "we're still in the pretty early days of that, at least from an AI perspective."

His bottom line: general-purpose GPU compute from Nvidia, AMD, Intel and others will coexist with custom silicon for some time, and anyone claiming to see clearly how this resolves "is probably fooling you." The open threshold — at which point the benefits of building your own hardware outweigh the challenges — remains unknown.

For his part, Carroll reports dealing with "seven or eight different new workloads a month" and describes innovation in the sector as "full-throttle." With TSMC's expanded capacity not arriving much before decade's end and memory suppliers warning of shortages into 2027, the companies that secure fab access and define their workloads earliest will determine how far the ASIC wave runs.

Also reported

Share this article:

« PreviousNext »

More from James Calloway

James Calloway

Show full bio

Staff writer covering companies, deals and strategy in Carrier business, 5G and 6G, fiber and broadband at Telecom Gazette.

498 articles