AI chip market share 2026: Nvidia, custom silicon, training and inference

Nvidia still sells most of the world's AI chips by revenue, with an estimated 75 percent or so in 2026, but custom chips from Google, Amazon, Broadcom's customers and others are growing faster. Shares depend heavily on whether you count dollars or units, and no figure here comes from an audited industry census.

The question sounds simple and the honest answer is not. Nvidia reports its numbers in public filings. Most custom chips are made for one buyer and never sold on the open market, so their "revenue" is partly a guess. This post lays out what is reported, what is estimated, where sources disagree, and what the split between training chips and inference chips really means. All figures are as of 19 October 2026, and each one is labelled by how firm it is.

Key numbers at a glance

• Nvidia revenue, quarter ended 26 July 2026: $96.2 billion, up 106 percent from a year earlier (Nvidia results as republished by Storage Newsletter, 27 August 2026).
• Nvidia Data Center revenue, same quarter: $89.0 billion, up 117 percent year on year and 18 percent on the prior quarter (same source).
• Nvidia guidance for the next quarter: $108.0 billion, plus or minus 2 percent, with no Data Center compute revenue from China assumed (company outlook, 26 August 2026).
• Broadcom AI semiconductor revenue, quarter ended 2 August 2026: $16.7 billion, up 221 percent year on year (Broadcom release, filed with the SEC, 2 September 2026). Broadcom's next-quarter AI guidance: $21.7 billion.
• AMD Data Center segment revenue, quarter ended 27 June 2026: $6.718 billion, up 107 percent (AMD release, 4 August 2026). This segment includes server CPUs, not only AI GPUs.
• Amazon chips business: annual revenue run rate "over $25 billion" (Andy Jassy, Amazon earnings call, 30 July 2026, via a transcript summary). It includes Trainium and Graviton CPUs.
• Nvidia share of AI accelerator revenue: about 75 percent projected for 2026, down from a 2024 peak near 87 percent (Silicon Analysts estimate, published 21 February 2026).
• Custom chip shipment growth in 2026: about 44.6 percent, against 16.1 percent for merchant GPUs (attributed to TrendForce in a Tech Times article, 26 May 2026).
• Custom ASIC share of AI server shipments in 2026: 27.8 percent (reported as TrendForce, May 2026, in two secondary summaries).
• Unit forecast: about 16.3 million AI accelerators shipped in 2026, up 62 percent, with Nvidia plus AMD GPUs at about 58 percent, down from 68 percent in 2025 (JPMorgan, relayed in a social media post dated 19 August 2026).
• 2027 units: about 12.5 million custom chips against about 10.9 million GPUs (same JPMorgan relay).

First, what counts as an AI chip

An AI accelerator is a processor built to do the huge amount of multiplication that neural networks need. The best known are graphics processing units, or GPUs, which Nvidia and AMD sell to anyone. The other family is the application specific integrated circuit, or ASIC. An ASIC is designed for a narrow job and made in quantity for one customer or a few. Google's tensor processing units (TPUs), Amazon's Trainium and Inferentia, Microsoft's Maia, Meta's MTIA and OpenAI's new Jalapeño chip are all in this second group.

The difference matters for market share. Nvidia sells chips for cash and records the sale. When Google runs its own TPUs in its own data centers, no chip sale appears anywhere, and analysts have to estimate a value. Broadcom and Marvell design many of these chips for the cloud companies, and Broadcom does report its AI revenue, but that figure also includes networking parts. So "custom silicon share" is always built on a looser base than "Nvidia share".

The revenue view: Nvidia is far ahead, and still growing

Nvidia's own numbers are the firmest in this whole story. In the quarter ended 26 July 2026 it reported $96.2 billion of revenue and $89.0 billion from Data Center. It also reported a gross margin of 75.0 percent, meaning that of each dollar of sales, about 75 cents was left after the direct cost of making the product. That is a company statement, and it is the only margin figure in this post.

Multiplying one quarter by four is crude, but it gives a feel for scale. $89.0 billion in a quarter is roughly $356 billion a year at that pace. That is simple multiplication, not a forecast, and Nvidia's quarters are not flat.

Compare that with the others, with the same caveat. Broadcom's $16.7 billion AI quarter would be about $67 billion a year at that pace. AMD's Data Center quarter of $6.7 billion would be about $27 billion a year, and some of that is ordinary server processors. Amazon says its whole chips business runs at more than $25 billion a year.

One third party estimate puts the picture in share terms. Silicon Analysts, writing in February 2026, estimated Nvidia's share of AI accelerator revenue at 87 percent in 2024, 81 percent in 2025 and a projected 75 percent in 2026. The same page estimates custom silicon from Google, Amazon, Microsoft and Meta at about $8 billion in 2024 and $25 billion or more in 2026, which works out to roughly 12 percent of its total market estimate. The page does not publish its method, and its own numbers do not add up cleanly, so use it as a rough guide only.

A second summary, from a site called Alatirok, gives a split by company for 2026: AMD about 6 to 8 percent, Google TPU about 8 percent, AWS Trainium and Inferentia about 5 percent, Microsoft and Meta about 3 percent combined. It calls these its own synthesis of several analyst sources, and it says analyst ranges for Nvidia run from 70 to 86 percent depending on method. Treat the whole thing as an informed guess.

The unit view: custom chips are catching up faster

Dollars favour Nvidia because its chips sell at high prices. Count chips instead of dollars and the picture shifts. A custom chip can cost far less than a top Nvidia part, so a given dollar share hides a bigger share of units.

Two forecasts point the same way. TrendForce, as quoted by Tech Times in May 2026, expects custom ASIC shipments from cloud providers to grow about 44.6 percent in 2026, against 16.1 percent for merchant GPUs. A 27.8 percent share of AI server shipments for ASIC based systems in 2026 is also attributed to TrendForce by the Alatirok summary, though the Tech Times article does not name a source for that figure. The original TrendForce release was not available to check, so both figures rest on secondary reports.

JPMorgan's forecast, available only as relayed in a social media post, is more striking. It puts 2026 shipments at about 16.3 million accelerators, with Nvidia and AMD GPUs together at about 58 percent, down from 68 percent in 2025. For 2027 it expects about 12.5 million custom chips against about 10.9 million GPUs, which would mean custom chips pass GPUs in unit terms. These are forecasts, not counts, and the original research note is not public.

Put the two views together and the sensible reading is this. Nvidia has the large majority of the money. Custom silicon has a growing and possibly very large share of the chips. Both can be true at once.

Who is behind the custom chips, and how big are the bets

The custom chip story is now tied to named, very large contracts. All of the following are company statements unless marked otherwise.

• Google announced two eighth generation TPUs at its Cloud Next 2026 event: TPU 8t for training and TPU 8i for inference, both due to be generally available "later this year" according to Google's blog. See Google TPU 8t and 8i explained.
• Anthropic announced in April 2026 an agreement with Google and Broadcom for multiple gigawatts of next generation TPU capacity starting in 2027. A Broadcom filing put it at about 3.5 gigawatts, according to Data Center Knowledge, published 8 April 2026. A gigawatt is a measure of power draw, used as shorthand for the size of a data center build. See Anthropic's compute commitments.
• Amazon said on its first quarter 2026 call that it had more than $225 billion in revenue commitments for Trainium, with multi year, multi gigawatt commitments from Anthropic and OpenAI. Trainium3 began shipping at the start of 2026 and Trainium4 is about 18 months from broad availability, according to Jassy.
• OpenAI and Broadcom unveiled Jalapeño, an inference chip, on 24 June 2026. First servers are due by the end of 2026, according to SiliconANGLE. The 10 gigawatt Broadcom agreement and OpenAI's other deals are set out in OpenAI's compute commitments explained.
• Microsoft announced Maia 200, an inference chip, on 26 January 2026. It is running in a US Central region and serving models including OpenAI's GPT-5.2, according to Microsoft.
• Meta said on 11 March 2026 that it had deployed "hundreds of thousands" of its own MTIA chips, built with Broadcom, and set out four generations through 2027, with the inference focused MTIA 450 scheduled for mass deployment in early 2027. For Amazon, Microsoft, OpenAI and Meta side by side, see Trainium, Maia and Jalapeño compared.

Note what is missing from the list: these companies are also some of Nvidia's biggest customers. Google Cloud, Microsoft Azure and Oracle are named by Nvidia as running its new Vera Rubin racks. Google's own Next 2026 post says it will be among the first to offer Vera Rubin instances. AMD names Anthropic, Meta, Microsoft and OpenAI as Helios rack deployers, and says Anthropic plans up to 2 gigawatts of MI450 GPUs. Almost every large buyer is buying from several suppliers at once.

Training and inference: why the split matters

Training is the slow, expensive process of building a model. The system reads vast amounts of data and adjusts billions of internal numbers. It is done in large bursts, needs thousands of chips working in lockstep, and rewards raw compute and fast links between chips.

Inference is what happens when you ask the finished model a question. It runs for every user, all day, and the bill grows with usage. Inference tends to be limited by how fast the chip can move data in and out of memory, because the model has to be read again for each word it writes. That is why inference chips emphasise memory size, memory speed and low delay.

The chips are starting to split along that line. Google's TPU 8t is described as a training chip and TPU 8i as an inference chip, with 288 GB of high bandwidth memory and 384 MB of on-chip memory on the 8i (Google blog, April 2026). Microsoft calls Maia 200 an inference accelerator. OpenAI says Jalapeño is for inference, and TechCrunch's reading is that pre-training will probably stay on Nvidia hardware, which is the reporter's view and not an OpenAI statement. Nvidia's results say its Groq 3 LPX inference accelerator has entered full production.

The line is not clean. Amazon's Trainium is named for training, yet Jassy said on the Q1 call that Bedrock runs most of its inference on Trainium. Data Center Knowledge describes Google's Ironwood generation as designed for large scale inference, and IDC's Dave McCarthy said Anthropic is preparing for a wave of agentic inference, which means AI systems that take many steps on their own. SK hynix says it expects the spread of inference to keep raising demand for memory.

What this means for share: training chips are bought in fewer, larger lumps and favour Nvidia's software and networking. Inference is where most of the long run volume is expected, and it is where cost per answer matters most. That is where custom chips argue they win. It is an argument from the vendors, and the independent evidence is still thin.

What is firm and what is uncertain

Firm, because the companies reported it:

• Nvidia's quarterly revenue, Data Center revenue, gross margin and guidance.
• Broadcom's AI revenue and guidance.
• AMD's Data Center revenue and its named Helios deployers.
• Amazon's stated run rate and commitments, and Google's and Microsoft's chip announcements.

Uncertain:

• Every market share figure. Silicon Analysts, Alatirok, TrendForce, JPMorgan and IDC differ in method, and several are available only as relayed by other sites.
• Whether a share is by revenue or by units, which can differ by tens of points.
• Vendor performance claims such as price performance gains. These are company comparisons, not independent benchmarks.
• Private deals. Contract sizes and commitments are company statements or reports, not audited figures.
• Where sources conflict: Nvidia's 2026 share is shown as "about 80 percent" in one page title and "75 percent" in the same page's text.

None of this is investment advice, and nothing here predicts which supplier will gain or lose.

What to watch next

• Nvidia's next results, guided at $108 billion for the quarter, and whether China is added back.
• Whether TPU 8t and 8i reach general availability "later this year", as Google says.
• First Jalapeño servers by the end of 2026, and any public performance data.
• Broadcom's guided $21.7 billion AI quarter, and AMD's Helios ramp in the second half.
• Memory and packaging supply, which limits every chip on this list. See HBM memory and TSMC packaging.

What this means for a customer

If you buy AI services, you may never choose a chip. But the chip mix affects price, supply and where your data runs. A provider that uses its own chips can cut its costs, and it may or may not pass that on. Ask where your workloads run and which regions are served. Safeguard offers SBOM and dependency scanning, plus a free scan at gold.safeguard.sh/scan, for the software that runs on top of all this hardware.

Quick check

Sources

• Nvidia fiscal 2Q27 results (Storage Newsletter, 27 August 2026): https://www.storagenewsletter.com/2026/08/27/nvidia-fiscal-2q27-financial-results/
• Nvidia Q2 FY2027 coverage (The Energy Mag): https://theenergymag.com/news/market-news/nvidia-q2-revenue-hits-96-2-b-as-data-center-surges-117-and-vera-rubin-ramps
• Broadcom Q3 fiscal 2026 release (SEC filing): https://www.sec.gov/Archives/edgar/data/0001730168/000173016826000076/avgo-08022026x8kxex99.htm
• AMD second quarter 2026 results: https://ir.amd.com/news-events/press-releases/detail/1295/amd-reports-second-quarter-2026-financial-results
• Amazon Q1 2026 chips remarks (Amazon): https://www.aboutamazon.com/news/company-news/amazon-ceo-andy-jassy-amazon-chips-business-q1-2026-earnings
• Amazon Q2 2026 call summary (Webull): https://www.webull.com/news/15364638750483456
• Silicon Analysts, Nvidia accelerator share: https://siliconanalysts.com/analysis/nvidia-ai-accelerator-market-share-2024-2026
• Alatirok, AI chip market share 2026: https://alatirok.com/ai-chip-market-share-2026/
• Tech Times, custom chips outpace GPU growth: https://www.techtimes.com/articles/317225/20260526/custom-ai-chips-outpace-nvidia-gpu-growth-2026-asic-shipments-set-triple-gpu-rate.htm
• JPMorgan forecast as relayed on Moomoo: https://www.moomoo.com/community/feed/jpmorgan-ai-accelerator-shipments-will-surge-by-62-in-2026-117121297678342
• Google, eighth generation TPUs: https://blog.google/innovation-and-ai/infrastructure-and-cloud/google-cloud/eighth-generation-tpu-agentic-era/
• Google Cloud, AI infrastructure at Next 26: https://cloud.google.com/blog/products/compute/ai-infrastructure-at-next26
• Microsoft, Maia 200: https://blogs.microsoft.com/blog/2026/01/26/maia-200/
• Meta, Four MTIA chips in two years (11 March 2026): https://ai.meta.com/blog/meta-mtia-scale-ai-chips-for-billions/
• TechCrunch, OpenAI first custom chip: 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/
• Data Center Knowledge, Anthropic TPU deal: https://www.datacenterknowledge.com/data-center-chips/anthropic-secures-multi-gigawatt-tpu-deal-with-google-broadcom
• Blocks and Files, SK hynix Q2 2026: https://www.blocksandfiles.com/flash/2026/07/29/sk-hynix-announces-extraordinarily-high-revenues-but-misses-expectations/5280499

Last reviewed: 19 October 2026.