Custom AI Chips Gain Ground on Nvidia as Trainium3 and TPUs Ramp

The year custom silicon stopped being a side project

The four largest buyers of AI compute spent 2026 putting their own chips into production. In earlier years these programs looked like hedges. Now they come with named customers, gigawatt commitments and published prices. Amazon's Trainium3, Google's Ironwood TPU, Microsoft's Maia 200 and Meta's MTIA line are all aimed at the same target: the recurring cost of serving AI models to users. Nvidia has answered with Vera Rubin, which it says cuts the cost of producing each token by a factor of ten.

The reporting supports a fairly clear conclusion. Custom silicon is taking a real and growing share of inference, which is the high-volume, predictable work of running models that are already trained. Nvidia still dominates revenue and the most flexible training work. The contest is less about whose chip is fastest and more about who can generate a token most cheaply inside their own data centers.

Amazon: Trainium3 finds customers beyond Anthropic

AWS announced Trainium3 at re:Invent 2025 and made it generally available in December 2025. It is a TSMC 3nm part rated at 2.52 petaflops of MXFP8 compute, with 144GB of HBM3e memory.1 The flagship UltraServer links up to 144 chips for about 362 FP8 petaflops. AWS says that is 4.4 times the compute of the prior generation and four times the energy efficiency.910 AWS also says early testing showed customers saving up to 50% compared with GPU-based training.10

The customer list matters more than the spec sheet. TechCrunch reported that Trainium has won over Anthropic, OpenAI and Apple.3 Uber's April 2026 expansion with AWS moved ride-matching onto Graviton4 processors and began piloting model training on Trainium3.4 Anthropic plans to bring close to a gigawatt of Trainium capacity online by the end of 2026.67 Andy Jassy wrote in April that Trainium3 was "nearly fully-subscribed." On the July earnings call he said there was "a real chance" Amazon could sell Trainium outside AWS.7

The reported figures do not all agree, so some specifics should be treated with caution. One outlet lists the chip's power draw at about 500W.1 Another puts it at about 750W, and also cites internal pricing of roughly $1.80 per chip-hour against about $4.80 for on-demand Nvidia H200 capacity.4 A third report says Trainium3 launched in July 2026 with 128GB of memory, which conflicts with the December 2025 availability and 144GB figure reported by most other coverage.2 Software maturity is the main caveat. SemiAnalysis noted that the programming mode most ML researchers prefer would not be supported until mid-2026, which slows adoption beyond expert teams like Anthropic's.5

Google: TPUs become a product sold to other companies

Google's change in 2026 was commercial rather than technical. Ironwood, the seventh-generation TPU, became generally available at Cloud Next in April. Google calls it its first TPU built "for the age of inference." Each chip delivers 4.6 petaflops, and a 9,216-chip pod reaches 42.5 exaflops.47 Google also previewed an eighth generation split into two designs: a training chip developed with Broadcom and an inference chip developed with MediaTek, both planned on TSMC 2nm for late 2027.47

Anthropic is the anchor customer. According to SemiAnalysis, the first 400,000 Ironwood chips are being sold to Anthropic as finished racks by Broadcom, worth about $10 billion. The remaining 600,000 are rented through Google Cloud, a commitment SemiAnalysis estimates at around $42 billion.41 Accounts of the deal's size differ. Some describe it as expanding to about 3.5 gigawatts in 2027.4247 Coverage of a Bloomberg report describes it as 5 gigawatts over five years, worth up to $40 billion.44 Meta reportedly signed a multiyear, multibillion-dollar lease for TPU capacity in February, with an option to buy Ironwood hardware from 2027.45

One point is disputed in the coverage: whether Google now sells TPUs as hardware. One analysis says TPUs are still available only through Google Cloud.45 SemiAnalysis's description of Broadcom selling racks directly to Anthropic points the other way.41 The better reading is that Google has begun selling hardware, but only to a very small number of strategic customers, not on an open market.

Pricing is where Ironwood's story gets more specific. SemiAnalysis estimates that external customers can see total cost of ownership up to about 30% below Nvidia GB200 systems and about 41% below GB300.41 Google later published on-demand Ironwood pricing of $12 per hour in Iowa, falling to $5.40 per hour on a three-year commitment. Anthropic's negotiated rate is reported at roughly $1.60.50 The gap between list and contract pricing is large. In practice, the cheapest custom silicon is mostly available to buyers big enough to negotiate it.

The demand is also creating strain inside Google. Bloomberg reporting describes DeepMind researchers waiting for TPU time while capacity goes to paying customers.44

Microsoft: Maia 200 is built only for inference

Microsoft announced Maia 200 in January as a dedicated inference accelerator. It is a 3nm chip with more than 140 billion transistors, over 10 petaflops of FP4 and over 5 petaflops of FP8 compute, 216GB of HBM3e and a 750W power envelope.11 Microsoft says it delivers 30% better performance per dollar than the newest hardware already in its fleet. It also claims triple the FP4 performance of Trainium3 and higher FP8 performance than Google's TPU v7.11 The chip is serving OpenAI's GPT-5.2 and Microsoft 365 Copilot from Azure's US Central region, with Arizona next.14

Those head-to-head claims deserve some skepticism. Redmond Magazine noted that Microsoft did not publish full test configurations.14 One independent profile pointed out that the 30% performance-per-dollar baseline is not clearly defined.12 The design also has limits. Because the hardware is tuned for low-precision math, training work at higher precision falls back to slower processing units. Analysts read this as Microsoft keeping Nvidia for training and moving serving onto its own silicon.17 Maia is not sold outside Microsoft.12

Meta: a fast release cadence and many suppliers

In March, Meta laid out four MTIA generations: the 300, 400, 450 and 500. All are deployed or scheduled for 2026–2027. The program is expanding from ranking and recommendation systems into generative AI inference.21 MTIA 400 supports a 72-chip scale-up domain, and Meta describes it as competitive with leading commercial products. The 450 and 500 are planned for mass deployment in 2027.2125 In April, Broadcom extended its partnership with Meta through 2029, starting with more than a gigawatt of MTIA capacity.2427 Reuters reported that production of a chip code-named Iris would begin in September.24

Meta's situation shows the limits of the custom-chip trend. In the same quarter it detailed MTIA, it also signed for millions of Nvidia GPUs, agreed to a deal for up to 6 gigawatts of AMD accelerators, and rented Google TPUs.24 Coverage has also reported layoffs among former Rivos staff and a paused training chip.24 Meta calls this a "portfolio approach" that matches each workload to the chip with the best total cost of ownership.27 In practice, MTIA covers the workloads where Meta can design for one known job, and outside suppliers cover the rest.

Nvidia's answer: lower its own cost per token

Nvidia's response is Vera Rubin, which launched at CES and then at GTC. The company claims a 10x reduction in inference cost per token compared with Blackwell, and four times fewer GPUs needed to train large mixture-of-experts models.3332 One technical analysis notes that the 10x figure applies specifically to mixture-of-experts inference with long inputs. For conventional dense models it projects a smaller gain of two to three times.39 Rack-level pricing is estimated at $3.5–4 million, about a 25% premium over Blackwell.39

Nvidia's revenue position is still very strong. One analysis puts its fiscal second-quarter data center revenue at $89 billion and its accelerator market share at roughly 73–80%.45 Another estimate puts the share closer to 63%.4 Market forecasts have custom ASICs growing at about 44.6% a year, against roughly 16% for general-purpose GPUs.31 One analysis argues that Nvidia can lose share and still grow if cheaper tokens expand total demand fast enough.34

What the reporting adds up to

The sources largely agree on how the market is dividing. Inference is becoming the larger share of AI computing. Workloads are getting more predictable, and power is a tighter limit than chip supply. Those conditions favor chips built for a narrow set of tasks.3114 Every hyperscaler covered here still buys Nvidia hardware, and each pitches its own chip on cost per token, not on peak speed.

The main uncertainty is software. Ease of programming outside Nvidia's CUDA ecosystem is still the weak point for Trainium, TPUs and the rest.542 For now, the economic benefits go mostly to a small group of very large labs that can negotiate contract pricing. Smaller buyers will likely see the effect indirectly, through lower API and cloud prices. That pressure on Nvidia's margins may matter more over time than any individual chip announcement.