NVIDIA held a 55.2% slice of the installed Windows video-card base in the 30 days ending 18 September 2026, with AMD at 20.0%, Intel at 19.2%, and other vendors at 5.7% [S1]. On the AI accelerator side, NVIDIA's revenue share is estimated at 75-81% for 2026, AMD at 5-7%, Intel near 1%, and hyperscaler custom silicon (Google TPU, AWS Trainium, Microsoft Maia, Meta MTIA) at 12-18% [S5][S6].
Three different "shares" are circulating in 2026: installed-base benchmarks, discrete add-in card sales, and AI accelerator revenue. Each tells a different story about the same three vendors, so the headline number you quote depends entirely on which pool you measure. For a process engineer sizing server hardware or workstation graphics, the distinction matters when you are paying for PLCs on the factory floor and GPUs in the rack.
Installed-base benchmark: NVIDIA 55.2%, AMD 20.0%, Intel 19.2%
PassMark's PerformanceTest aggregate of Windows x86 and ARM submissions, updated 18 September 2026, shows NVIDIA at 55.2% of video-card units in the prior 30 days, AMD at 20.0%, Intel at 19.2%, and all other vendors combined at 5.7% [S1]. The dataset reflects cards in active use rather than cards purchased, which is why Intel's share is structurally higher here than in discrete retail-sales trackers: integrated and Arc GPUs sit in millions of working office machines that never get benchmarked at retail.
The same series also publishes quarterly history back to Q3 2005, and the 30-day window is the freshest cut available. AMD's 20.0% reading is the strongest installed-base figure the vendor has posted in years, driven by Ryzen-era system builds that pair Radeon graphics with AMD CPUs; Intel's 19.2% mostly reflects Arc discrete cards plus the iGPU population in 12th- through 15th-generation Core laptops and desktops. NVIDIA's 55.2% is held up by the GeForce RTX 40- and 50-series installed base across gaming rigs and workstation towers [S1].
Discrete gaming GPU sales: NVIDIA near 94-95%, AMD near 5%
For retail add-in board sales, the picture is much more concentrated. A report indicated NVIDIA dominated the gaming GPU market with 95% share while AMD Radeon sales fell to a historical low of 5% [S2]. TechPowerUp's September 2025 coverage of the prior JPR quarter had NVIDIA at 94% and AMD at roughly 6% [S7].
The gap between the 55.2% installed-base figure and the 94-95% discrete-sales figure is real, not a measurement error. Installed-base data carries forward every NVIDIA card shipped in the last decade, including older GeForce GTX 10-series and RTX 20-series boards that linger in working PCs. Discrete-sales data only counts new retail boxes, where NVIDIA's RTX 50-series has dominated since the late-2024 launch. AMD's RX 9000-series launch in early 2025 did not materially shift the discrete split in either quarter measured [S2][S7]. For system integrators pulling fresh GPUs into industrial image-processing or flow-meter vision cells, the retail-sales number is the one that affects lead time and price.
AI accelerator revenue: NVIDIA 75-81%, AMD 5-7%, Intel 1%
NVIDIA's fiscal 2026 data-center revenue reached $193.7 billion, with the quarter ended 26 April 2026 alone delivering $75.2 billion in data-center sales and $14.8 billion in data-center networking, up 199% year-on-year [S5]. On the basis of that revenue, Silicon Analysts pegs NVIDIA's 2026 AI accelerator share at roughly 75%, while IDC puts it closer to 81% [S5][S6]. The 2024 peak sat near 87%.
AMD's data-center segment recorded $5.8 billion in Q1 2026, up 57% year-on-year, with the Instinct MI350 and MI355X lines generating an estimated $7-8 billion across full-year 2025 [S5]. That maps to 5-7% accelerator share. AMD's two largest 2026 AI commitments are the reported 6-gigawatt deals with Meta and OpenAI, both built around the MI350 and forthcoming MI450 families [S5]. Intel's AI GPU share sits near 1% while Gaudi 3 ramps and the Crescent Island accelerator is not sampling until the second half of 2026 [S5]. Custom ASICs from Google, Amazon, Microsoft, and Meta capture 12-18% combined, and TrendForce reports their shipment growth at 44.6% in 2026 versus 16.1% for merchant GPUs [S5].
Headline numbers by segment, side by side
The three measurement pools give three different rankings, and a buyer has to pick the one that matches the use case.
For installed-base share on Windows PCs as of 18 September 2026, NVIDIA leads at 55.2%, AMD is second at 20.0%, Intel third at 19.2%, and other vendors collectively at 5.7% [S1]. For discrete add-in card retail sales in Q1 2026, NVIDIA is at 95% and AMD at 5% [S2]. For 2026 AI accelerator revenue, NVIDIA is at 75-81%, AMD at 5-7%, Intel at about 1%, and hyperscaler custom silicon at 12-18% [S5][S6]. A buyer pulling GPUs for a new inference cluster is paying attention to the third column; a system builder quoting a workstation for industrial-valve qualification software lives in the second column; an IT manager sizing a refresh of mixed office and CAD seats lives in the first.
Why the AI gap to AMD keeps closing slowly
NVIDIA's share drop from 87% in 2024 to 75-81% in 2026 looks small in percentage points but is large in absolute dollars, because the total market grew from under $80 billion in 2024 to over $200 billion in 2026 [S5]. CUDA is still the moat: more than 5 million developers build on it, which keeps training workloads locked to NVIDIA hardware even where AMD's MI350X matches raw FP8 throughput per [silicon analysts](silicon analysts) [S5]. The MI355X competes on memory bandwidth and HBM3E capacity, but porting cost and the absence of a CUDA-equivalent compiler stack keep most hyperscalers buying NVIDIA for the bulk of training and using AMD as a second source for inference.
Custom silicon is the faster-growing threat. Google TPU, AWS Trainium, Microsoft Maia, and Meta MTIA grew 44.6% in 2026 versus 16.1% for merchant GPUs [S5]. These are not counted in AMD's or NVIDIA's share, but they cap the addressable market for both. Intel's path back into the AI top three runs through the Crescent Island launch in H2 2026 and Gaudi 3 volume; until those land, the "Intel 1%" line item will stay flat [S5].
Where the data gets fuzzy and what to watch next
No audited public table exists for AI accelerator market share. Revenue share flatters NVIDIA because its chips carry premium pricing, while a TPU or Trainium part is an internal cost line for Google or Amazon rather than a market transaction, so they often get excluded from headline counts [S5]. JPR's discrete-GPU numbers are the closest thing to an audited benchmark, but they are paid-subscription and lag by one quarter, which is why March 2026's 95% figure is the most recent public cut [S2]. PassMark's installed-base series is free and updated daily, but it skews toward enthusiast and gaming users who run benchmarks [S1].
Two signals to track through the rest of 2026: AMD's MI450 volume ramp, which the OpenAI and Meta deals are supposed to anchor, and Intel's first Crescent Island revenue quarter, expected in the Q4 2026 earnings cycle. Either will show up first in NVIDIA's quarterly data-center growth rate, which has run above 90% year-on-year for four straight quarters and is the cleanest leading indicator of where the merchant AI GPU share is actually moving [S5].
This topic is covered further in Quarry Excavator Specs: Power, Bucket, and Dig Reach by Size Class.