Trainium, Maia and Jalapeño: the cloud and lab custom AI chips compared

Amazon's Trainium, Microsoft's Maia 200 and OpenAI's Jalapeño are custom AI chips each company designs for its own workloads, with partners handling manufacturing. Trainium is in its third shipping generation and is the only one of the three with a published commitments figure, which Amazon puts at more than $225 billion. Maia 200 and Jalapeño are newer and focus on inference, and most of their performance claims are unverified.

Nvidia sells the chips most people know. But the biggest buyers of AI chips are also designing their own, to lower cost, gain control over supply, and tune hardware to their models. This post compares what Amazon, Microsoft and OpenAI have actually published about their chips, in plain language, with dates and caveats, and adds a short note on Meta's MTIA line. Google's chips have their own explainer: Google TPU 8t and 8i explained. All figures are as of 19 October 2026.

Key numbers at a glance

• Trainium revenue commitments: "over $225 billion" (Andy Jassy, Amazon first quarter call, April 2026, via Amazon's own page).
• Amazon chips business run rate: over $20 billion at the first quarter, over $25 billion at the second quarter call on 30 July 2026, growing at triple digit rates (Jassy). The business covers Trainium and Graviton processors.
• Project Rainier: nearly half a million Trainium2 chips across multiple US data centers, with Anthropic expected to run on more than 1 million Trainium2 chips by the end of 2025 (Amazon page, published when the cluster went live).
• Trainium3: announced 2 December 2025, 3 nanometer process, 144 chips per UltraServer, up to 1 million chips in a deployment, over 4 times the speed and memory of Trainium2 (AWS claims, TechCrunch report).
• Trainium3 status: began shipping at the start of 2026, 30 to 40 percent more price performant than Trainium2, nearly fully subscribed (Jassy, April 2026).
• Trainium4: about 18 months from broad availability, much of it already reserved (Jassy, April 2026).
• Maia 200: announced 26 January 2026, TSMC 3 nanometer, more than 140 billion transistors, 216 GB of HBM3e at 7 TB per second, 272 MB of on-chip SRAM, 750 watt power envelope (Microsoft).
• Maia 200 compute: over 10 petaFLOPS at FP4 and over 5 petaFLOPS at FP8 (Microsoft).
• Jalapeño: OpenAI's first custom chip, shown 24 June 2026, built with Broadcom, inference only, first servers due by the end of 2026 (TechCrunch and SiliconANGLE).
• Trainium3 chip: 144 GB of HBM3e at 4.9 TB per second; a 144 chip UltraServer offers up to 362 petaFLOPS at MXFP8 and 20.7 TB of HBM3e (AWS Trainium product page, as of 19 October 2026).
• Meta MTIA: four generations (300, 400, 450 and 500) set out on 11 March 2026, with "hundreds of thousands" of MTIA chips already in production (Meta).
• Amazon capital spending: about $220 billion expected for 2026, raised from about $200 billion, which Jassy attributed to higher memory prices (30 July 2026 call).

Why build your own chip at all

There are three reasons the companies give or that follow from their statements.

Cost. If you run a very large amount of AI work, even a modest saving per answer adds up. Amazon says Trainium2 delivers about 30 percent better price performance than comparable GPUs, and that Trainium3 is another 30 to 40 percent better than Trainium2. Those are Jassy's claims on an earnings call, not independent measurements.

Supply. When one maker sells most of the chips, the buyers queue behind each other. A second source of chips gives a company more control over how fast it can add capacity.

Fit. A company that knows exactly which models it runs can design for them. Greg Brockman, OpenAI's president, said at the Jalapeño launch, "We have a deep understanding of the workload," according to TechCrunch.

A custom chip does not remove the dependence on others. These companies still buy from chip foundries, memory makers and packaging plants, and many still buy Nvidia chips as well. For those limits, see HBM memory and TSMC packaging.

Amazon: Trainium and the numbers behind it

Trainium is Amazon's line of AI chips. Its name suggests training, but Amazon uses it for both jobs. Jassy said Amazon's Bedrock service, which gives customers access to many AI models, runs most of its inference on Trainium, and that Bedrock has more than 125,000 customers.

Trainium2 and Project Rainier. Project Rainier is the large cluster Amazon built for Anthropic. Amazon says it is fully operational, uses nearly half a million Trainium2 chips, and that Claude is expected to run on more than 1 million Trainium2 chips for training and inference by the end of 2025. Amazon describes the building block as an UltraServer, made of four physical servers with 16 chips each, so 64 chips per UltraServer, linked by Amazon's own NeuronLink cables inside and EFA network adapters between them. Amazon also says the cluster gives Anthropic more than five times the compute used to train its previous models. No public source confirms that the 1 million chip figure was reached.

Trainium3. Announced at AWS re:Invent on 2 December 2025, it uses a 3 nanometer manufacturing process. That number describes the size of the smallest features on the chip, and a smaller process usually means more transistors and better efficiency. AWS claims over 4 times the speed and memory of Trainium2, 40 percent better energy efficiency, 144 chips per UltraServer, and up to 1 million chips in a deployment. TechCrunch names Anthropic, Karakuri, SplashMusic and Decart as early users reporting lower inference costs, per Amazon. On the April call Jassy said Trainium3 began shipping at the start of 2026 and is nearly fully subscribed.

AWS's Trainium product page, as of 19 October 2026, gives more detail on the chip itself. Each Trainium3 has eight NeuronCores, the processing blocks that do the math, plus 144 GB of HBM3e memory at 4.9 TB per second, which AWS says is 1.7 times Trainium2's memory bandwidth. AWS also claims up to twice Trainium2's compute at MXFP8, a compact 8 bit number format. A full UltraServer of 144 chips is listed at up to 362 petaFLOPS at MXFP8, 20.7 TB of HBM3e and 706 TB per second of combined memory bandwidth, with a switch AWS calls NeuronSwitch that doubles the bandwidth between chips. The same page names Anthropic, Databricks, Decart, OpenAI, Ricoh, SplashMusic and Uber as customers. These are AWS figures, and the page gives peak numbers, not results on a named model.

Trainium4. It is still on the roadmap. Jassy said it is about 18 months from broad availability, and TechCrunch reported it will support Nvidia's NVLink Fusion interconnect, which lets Trainium systems work alongside Nvidia GPUs in the same racks.

The commercial side. Jassy said Trainium has more than $225 billion in revenue commitments, including multi year, multi gigawatt commitments from Anthropic and OpenAI, with Uber and Pinterest also adopting it. A commitment is a customer promise to buy, not revenue already earned. On the July call he said there is "a real chance" Amazon will sell Trainium chips for customers' own data centers. All of these are company statements.

Inferentia. Amazon's earlier line of inference chips is called Inferentia, and it shows how long Amazon has been at this. AWS's Inferentia page describes two generations. The first powers EC2 Inf1 instances with up to 16 chips each, and each chip carries 8 GB of ordinary DDR4 memory. Inferentia2 moved to 32 GB of HBM per chip and up to 190 teraFLOPS at FP16, and AWS says it delivers up to 4 times the throughput and up to 10 times lower latency than the first generation. AWS does not say on that page how Inferentia's role is changing now that Bedrock runs most of its inference on Trainium, so the two lines may overlap for some time.

Microsoft: Maia 200

Microsoft announced Maia 200 on 26 January 2026, in a post by Scott Guthrie. Its published specs are:

• TSMC 3 nanometer process and more than 140 billion transistors.
• Over 10 petaFLOPS at FP4 and over 5 petaFLOPS at FP8. FP4 and FP8 are low precision number formats that make inference faster and cheaper, with a small loss of accuracy that models are trained to tolerate.
• 216 GB of HBM3e memory at 7 TB per second, and 272 MB of on-chip SRAM.
• A 750 watt power envelope per chip.
• 2.8 TB per second of two way scale up bandwidth per chip, over standard Ethernet with a custom transport layer, with predictable communication across up to 6,144 chips.
• Four chips per tray, linked directly, and second generation closed loop liquid cooling.

Microsoft says it is running now in its US Central region near Des Moines, Iowa, with US West 3 near Phoenix next. It serves multiple models, including OpenAI's GPT-5.2, for Microsoft Foundry and Microsoft 365 Copilot, and it is used by Microsoft's Superintelligence team for synthetic data and reinforcement learning. A software kit is in preview.

The claims are Microsoft's: three times the FP4 performance of Amazon's third generation Trainium, FP8 performance above Google's seventh generation TPU, and 30 percent better performance per dollar than the latest hardware in Microsoft's own fleet. The blog calls it "the most performant, first party silicon from any hyperscaler." The text gives no baselines or test conditions for those comparisons, so they are claims, not benchmarks. Note also that Microsoft does not say what share of its AI capacity runs on Maia.

OpenAI: Jalapeño

OpenAI introduced Jalapeño on 24 June 2026. It is the first chip from OpenAI's Broadcom partnership, which was announced in October 2025. For the wider deal, see OpenAI's compute commitments.

What was published is limited. OpenAI calls it an inference chip and the first step in "a multi generation compute platform." It says early results show much better performance per watt than current leading alternatives, and that the chip is still being tested. SiliconANGLE reports that OpenAI's blog says the architecture reduces data movement, and that servers will use custom racks with Broadcom networking, including Tomahawk switches, with a Toronto company, Celestia, helping design the systems. First Jalapeño servers are due online by the end of 2026.

What was not published: memory type and size, process node, compute throughput, power, and deployment scale. TechCrunch says pre-training will likely remain on Nvidia hardware, which is the reporter's view. OpenAI said its own AI models helped with the chip's development.

Meta and others

Meta's chip line is called MTIA, short for Meta Training and Inference Accelerator. In a post dated 11 March 2026, Meta said it had "deployed hundreds of thousands of MTIA chips in production" and set out four generations, all developed "in close partnership with Broadcom":

• MTIA 300, already in production for ranking and recommendation training, the systems that decide what appears in feeds and ads.
• MTIA 400, which Meta says has 400 percent higher FP8 compute and 51 percent more HBM bandwidth than the 300. Meta said lab testing was finished and the chip was on the path to data center deployment.
• MTIA 450, aimed at generative AI inference, with twice the HBM bandwidth of the 400, scheduled for mass deployment in early 2027.
• MTIA 500, also for inference, with 50 percent more HBM bandwidth than the 450 and up to 80 percent more HBM capacity, scheduled for mass deployment in 2027.

Meta says it can now ship a new chip roughly every six months and that the 400, 450 and 500 share the same chassis, rack and network. The absolute power and memory figures sit in an image table on Meta's page and are not given here. All of these are Meta's own statements, and the inference parts are still schedules, not deployments.

Other makers are left out because they have not published comparable details.

How to compare them fairly

Put them side by side and the problem is clear: each company reports different things in different ways.

• Amazon talks about commitments, run rate and price performance.
• Microsoft talks about peak compute, memory and a comparison against rivals.
• OpenAI talks about performance per watt, with no figures yet.
• Meta talks about generation to generation gains, in percent.

So there is no neutral scoreboard. Peak compute at FP4 tells you the ceiling on a lab test. Real cost per answer depends on the model, the software, how busy the chips are, and the price of power and memory. A fair comparison needs the same model run on each chip under the same load, and no public source provides that as of 19 October 2026.

Also watch the number format. FP4 figures run higher than FP8 figures for the same chip, so a headline in FP4 cannot be compared with one in FP8. AWS quotes Trainium3 at MXFP8, Microsoft quotes Maia 200 at FP4 and FP8, and Meta quotes MTIA 450 and 500 at MX4, so even the units differ.

Software is the other half of any comparison. A chip is only useful if common AI frameworks run on it without heavy rewriting. AWS lists PyTorch, Hugging Face Transformers and vLLM support through its Neuron software kit, and lets engineers write low level code through a tool it calls the Neuron Kernel Interface. Meta says MTIA is built natively on PyTorch, vLLM and Triton. Microsoft says a Maia software kit is in preview. OpenAI has not said how outside developers would program Jalapeño, which makes sense for a chip that, as described so far, serves OpenAI's own models. For a buyer, how mature the software is often matters more than the peak number on the spec sheet.

What is firm and what is uncertain

Firm:

• The existence, names and dates of the announcements.
• Microsoft's published Maia 200 specs, such as memory sizes, which come from the maker.
• Amazon's stated commitments and run rates, which are company statements.

Uncertain:

• All performance and price comparisons between chips.
• Whether Rainier's chip count target was met.
• Jalapeño's specs and scale, and when meaningful volume arrives.
• Meta's MTIA 450 and 500 schedules, which are plans, not deployments.
• How much of each company's workload really runs on its own chips versus Nvidia or AMD hardware.
• Commitments are promises. They are not audited revenue.

What to watch next

• Trainium4 timing, and whether Amazon sells chips outside AWS.
• Maia 200 spread beyond the first two regions, and a successor.
• First Jalapeño servers before the end of 2026, and public performance data.
• MTIA 450 mass deployment, which Meta has scheduled for early 2027.
• Memory prices. Jassy blamed higher memory costs for raising Amazon's capital spending guide by about $20 billion.

Safeguard offers SBOM and dependency scanning and a free scan at gold.safeguard.sh/scan, which works regardless of which chip your workloads run on. For the market share picture, see AI chip market share 2026.

Quick check

Sources

• Amazon, Project Rainier: https://aboutamazon.com/news/aws/aws-project-rainier-ai-trainium-chips-compute-cluster/
• TechCrunch, Trainium3 and Trainium4 roadmap (2 December 2025): https://techcrunch.com/2025/12/02/amazon-releases-an-impressive-new-ai-chip-and-teases-a-nvidia-friendly-roadmap
• Amazon, Jassy on the chips business, Q1 2026: https://www.aboutamazon.com/news/company-news/amazon-ceo-andy-jassy-amazon-chips-business-q1-2026-earnings
• AWS, Trainium product page: https://aws.amazon.com/ai/machine-learning/trainium/
• AWS, Inferentia product page: https://aws.amazon.com/ai/machine-learning/inferentia/
• Meta, Four MTIA chips in two years (11 March 2026): https://ai.meta.com/blog/meta-mtia-scale-ai-chips-for-billions/
• Amazon Q2 2026 call summary (Webull): https://www.webull.com/news/15364638750483456
• Microsoft, Maia 200 (26 January 2026): https://blogs.microsoft.com/blog/2026/01/26/maia-200/
• TechCrunch, OpenAI's first custom chip (24 June 2026): https://techcrunch.com/2026/06/24/openai-unveils-its-first-custom-chip-built-by-broadcom/
• SiliconANGLE, Jalapeño: https://siliconangle.com/2026/06/24/openai-broadcom-debut-custom-jalapeno-chip-llm-inference/

Last reviewed: 19 October 2026.