Google sold Anthropic access to as many as a million Tensor Processing Units in October 2025411. One order, promising more than a gigawatt of added compute through 2026. That sale turned a ten-year internal chip program into a business able to outfit a frontier lab. It also marked a pattern that now runs through every major AI company, one that reads as hedging until you notice how much money each of them still hands Nvidia in the same breath. Buy Nvidia. Order AMD. Build your own chip anyway.
Broadcom's Second Customer Becomes a Product Line
A custom chip took shape quietly from 2024, when Broadcom and OpenAI began co-designing hardware to cut OpenAI's reliance on Nvidia23. Demand had climbed past what one supplier could meet, and OpenAI wanted a design it controlled from the schematic out, rather than a place in somebody else's allocation queue. The project surfaced on June 24, 2026, under a name with personality: Jalapeño9. TSMC builds it on a 3-nanometer process, and one design handles both training and inference. OpenAI aimed to ship by the close of 2026. That puts finished silicon in production barely two and a half years after the work began.
Broadcom, led by president and chief executive Hock Tan, runs this as a repeat business rather than a favor. Its XPU division builds chips to a client's exact spec3. It has supplied Google's TPU line since that program began, co-builds Meta's MTIA chips, and counts ByteDance as a customer. OpenAI's arrival puts it inside a system Google and Meta helped build. Read the irony straight. OpenAI fights both for frontier models and hires the same chip house they use.
Google Turns a Decade of TPUs Into a Sales Pitch
Ironwood, Google's seventh TPU, launched at Google Cloud Next 25 on April 9, 20251. It is the first Google chip built for inference rather than training. One pod, 9,216 chips, delivers 42.5 exaflops. Each chip holds 192 gigabytes of fast memory, six times what Trillium held, moving data at 7.37 terabytes per second. Call that a 4.5-fold jump. Power efficiency doubled against the prior chip, and rose roughly thirtyfold against Google's first TPU from 2018.
Amin Vahdat, who runs machine learning, systems and cloud AI at Google, stated the purpose plainly. "Ironwood is our most powerful, capable and energy efficient TPU yet, designed to power thinking, inferential AI models at scale"1. Google framed the larger shift in its own announcement. "Ironwood represents a significant shift in the development of AI and the infrastructure that powers its progress. It's a move from responsive AI models that provide real-time information for people to interpret, to models that provide the proactive generation of insights and interpretation." Read what that sells. Chips for models that think longer, which is the work Google expects to bill for.
Anthropic's October 2025 order for up to a million of these chips tested the pitch at a scale few outside Google had reached4. A lab with every reason to shop around picked Google's hardware anyway. Its 2026 buying list shows how widely it shops: a roughly $45 billion Nscale deal, a Microsoft-brokered $30 billion Azure commitment, and gigawatts of AMD chips.
Amazon's Rainier and the Trainium Line
Amazon built Project Rainier around Trainium2, the chip now running Anthropic's Claude models in production5. Trainium3 followed, and Amazon disclosed deals for it with both Anthropic and OpenAI. Two labs that fight hard elsewhere run on the same Amazon chips. Trainium4 arrived at re:Invent in late 2025, holding a pace of roughly one better chip a year.
Emphasis splits them. Ironwood sells itself as an inference specialist1. Trainium sells on cost and supply, a chip Amazon makes at the volume its own cloud customers demand, outside Nvidia's queue5. Rainier hosting Claude gives Amazon something Google's TPU program lacks. A top frontier model runs on Amazon-designed hardware, in Amazon's own data centers. That is the proof smaller buyers ask for.
Meta Hedges With MTIA and Nvidia at Once
A plan for four new in-house chips arrived March 11, 2026, under Meta's Training and Inference Accelerator program610. It follows years of catching up. In 2022 the company still ran AI on general-purpose processors that rivals had already moved onto GPUs. Less than a month before the plan, Meta announced a long-term deal with Nvidia. Read the order as insurance rather than contradiction. Meta buys Nvidia for the volume its Llama successors and ranking systems need today, while MTIA takes the narrow work where a custom chip beats a GPU on cost.
Anthropic ordered up to 1 million Google TPUs in October 2025; Meta and OpenAI unveiled their own chips by June 2026
Dates of four custom-silicon moves this piece reports, from Ironwood's launch to Jalapeño's debut
Each point marks one move on a real time scale: the magenta point is Anthropic's TPU order, and the gray points are three custom chips or chip plans made public.
Sources: Google Cloud Blog [1]; Wikipedia [2]; Wikipedia [4]; Wikipedia [6]. Chart by Kelly Dennis.
The numbers behind this chart
| Date | Event |
|---|---|
| April 9, 2025 | Google launches Ironwood, its seventh TPU. The first Google chip built for inference |
| Oct. 2025 | Anthropic orders up to 1 million TPUs. More than a gigawatt of added compute through 2026 |
| March 11, 2026 | Meta plans four new MTIA chips. Weeks after Meta's long-term Nvidia deal |
| June 24, 2026 | OpenAI unveils its Broadcom chip, Jalapeño. Built by TSMC on a 3-nanometer process |
Every hyperscaler runs both. Meta's timing, an Nvidia deal and then a chip plan weeks apart, makes the hedge easy to see. Broadcom co-designs MTIA beside its OpenAI and Google work, so one design house now touches three of the industry's largest chip programs236. That pool of engineers and fab ties gives Broadcom outsized say over how fast each customer ships.
What Custom Silicon Does to the Cost of a Token
Custom silicon earns its cost through one thing above all. It strips a GPU vendor's margin out of every unit of compute a company uses itself. Where the numbers allow, it resells that edge to outside buyers at a thinner markup than Nvidia charges. Ironwood's memory speed and power efficiency turn straight into inference served per dollar of electricity1. That line grows as reasoning models multiply the tokens each query burns. Amazon's Trainium line chases the same end through volume rather than design focus5.
The tell is who buys whose chips. Anthropic runs on Google TPUs, Amazon Trainium and AMD GPUs at once, and pays Google as a customer4. A lab racing Gemini still buys Google hardware when the numbers work. Chip rivalry and model rivalry come apart cleanly. Cost per token now beats brand loyalty in deciding which silicon wins a job.
Hock Tan, Broadcom's president and chief executive, put a number on that shift on the company's third-quarter earnings call in September 2026. Broadcom builds the custom chips behind Google's TPU line, OpenAI's Jalapeño and Meta's MTIA alike, so its results track the business across every customer at once. "Our 6 XPU customers are accelerating the adoption of custom accelerators and AI semiconductor revenue more than tripled year-over-year to $16.7 billion," Tan told analysts on the call7. A hedge rarely triples in a year. Tan counted six customers on that same call, which places Anthropic's million-chip order inside a longer customer list.
Jensen Huang, Nvidia's founder and chief executive, has pressed the opposite case for longer than Broadcom has had six XPU customers to count. An analyst asked him on Nvidia's own earnings call in November 2025 whether years of dismissing ASICs as a threat still held up. "It's not about just putting a random ASIC into a data center," Huang answered, pointing to the off-take, the cloud reach and the software stack Nvidia sells alongside the chip itself8. Read the claim next to Tan's number. Huang's own customer list, every hyperscaler in this piece among them, still buys Nvidia GPUs by the gigawatt, so the chip rivalry he dismisses keeps funding itself either way.
By the numbers
- One million: the cap on TPUs Google agreed to supply Anthropic in October 2025, worth more than a gigawatt4.
- Peak compute of 42.5 exaflops arrives at full 9,216-chip pod scale, announced April 9, 20251.
- Six times: Ironwood's per-chip memory gain over Trillium, reaching 192 gigabytes of high-bandwidth memory1.
- June 24, 2026: the day OpenAI unveiled Jalapeño, its Broadcom-designed chip, after two years of joint work2.
- Four new in-house chips make up the MTIA roadmap Meta disclosed March 11, 20266.
- Trainium4 arrived at AWS re:Invent in late 2025, the fourth chip in Amazon's line5.
- A $45 billion Nscale agreement sits beside Anthropic's Google TPU and AWS Trainium commitments4.
What to watch
Jalapeño's yields, once TSMC ships volume in late 2026, test one question2. Does OpenAI's chip close a real gap against Nvidia pricing, or serve mainly as leverage in the next GPU contract? Google selling TPU capacity to other labs at similar scale would confirm the million-unit deal is strategy rather than one odd arrangement14. Trainium4 uptake outside Anthropic deserves a close look5. So does whether MTIA ships fast enough to matter before Meta's Nvidia deal locks in years of parallel demand6.
Sources
- Google Cloud, "Ironwood: The first Google TPU for the age of inference," Google Cloud Blog, April 9, 2025, https://blog.google/products/google-cloud/ironwood-tpu-age-of-inference/
- Wikipedia contributors, "OpenAI," Wikipedia, accessed Sept. 1, 2026, https://en.wikipedia.org/wiki/OpenAI
- Wikipedia contributors, "Broadcom," Wikipedia, accessed Sept. 1, 2026, https://en.wikipedia.org/wiki/Broadcom
- Wikipedia contributors, "Anthropic," Wikipedia, accessed Sept. 1, 2026, https://en.wikipedia.org/wiki/Anthropic
- Wikipedia contributors, "AWS Trainium," Wikipedia, accessed Sept. 1, 2026, https://en.wikipedia.org/wiki/AWS_Trainium
- Wikipedia contributors, "Meta Platforms," Wikipedia, accessed Sept. 1, 2026, https://en.wikipedia.org/wiki/Meta_Platforms
- Motley Fool Transcribing, "Broadcom (AVGO) Q3 2026 Earnings Call Transcript," The Motley Fool, Sept. 2, 2026, https://www.fool.com/earnings/call-transcripts/2026/09/09/broadcom-avgo-q3-2026-earnings-call-transcript/
- Motley Fool Transcribing, "Nvidia (NVDA) Q3 2026 Earnings Call Transcript," The Motley Fool, Nov. 19, 2025, https://www.fool.com/earnings/call-transcripts/2025/11/19/nvidia-nvda-q3-2026-earnings-call-transcript/
- Russell Brandom, "OpenAI unveils its first custom chip, built by Broadcom," TechCrunch, June 24, 2026, https://techcrunch.com/2026/06/24/openai-unveils-its-first-custom-chip-built-by-broadcom/
- Meta, "Expanding Meta's Custom Silicon to Power Our AI Workloads," Meta Newsroom, March 11, 2026, https://about.fb.com/news/2026/03/expanding-metas-custom-silicon-to-power-our-ai-workloads/
- Maham Fatima, "Google Seals Multi-Billion Dollar Deal to Supply Anthropic with 1M TPUs in Major AI Hardware Commitment," Yahoo Finance (via Insider Monkey), Oct. 29, 2025, https://finance.yahoo.com/news/google-seals-multi-billion-dollar-152458472.html