GPUDeck · Public GPU production census

Which GPU models remain in production

A longitudinal register of GPU models observed on the public fast network. It records presence and worker count, not ownership, capacity, utilisation, price or performance.

Latest census sample: 23 Aug 2026. Page regenerated 2026-08-24 11:45 UTC from the dated archive.

250 days · 39 GPU models · 3,219 model-day observations · 9 models present now · first observed 17 Dec 2025

Lifelines

A dated register of the GPU models observed on the fast network. Thickness is observed worker count on a shared scale. A band that stops is a model no longer observed; LAST OBSERVED marks its final dated appearance in this window. Persistent absence is consistent with operators withdrawing that class, but the census does not establish why.

Equipment-class production lifelines Daily worker counts by GPU model. Lines reaching the right edge remain present; severed lines print their exact last-observed date. GPU MODEL OBSERVED STATE 2025-12 2026-01 2026-02 2026-03 2026-04 2026-05 2026-06 2026-07 2026-08 RTX 4090 NOW 48 | PEAK 56 RTX 5090 NOW 30 | PEAK 40 RTX PRO 6000 Blackwell Server Edition NOW 6 | PEAK 6 RTX PRO 5000 Blackwell NOW 6 | PEAK 6 RTX 4080 NOW 2 | PEAK 2 RTX 4080 SUPER NOW 1 | PEAK 3 RTX 5070 Ti NOW 1 | PEAK 3 RTX PRO 4500 Blackwell NOW 1 | PEAK 2 RTX 3080 NOW 1 | PEAK 1 RTX 3090 LAST OBSERVED 2026-07-19 | PEAK 42 H100 80GB HBM3 LAST OBSERVED 2026-07-29 | PEAK 21 B200 LAST OBSERVED 2026-03-08 | PEAK 12 L40S LAST OBSERVED 2026-08-18 | PEAK 11 RTX PRO 6000 Blackwell Workstation Edition LAST OBSERVED 2026-08-21 | PEAK 8 RTX 3090 Ti LAST OBSERVED 2026-04-17 | PEAK 6 RTX 5080 LAST OBSERVED 2026-08-03 | PEAK 5
Roster rule: present now, or an all-window peak of at least 5 workers. Lane count and height are generated from that rule.Warm bone reaches the present; brass marks a model holding below its opening count. Rust marks a zero-now cutoff, never a claim that a model cannot return.

16 models plotted. 23 further classes peaked below 5 and are retained in the complete census and the table below.

View all 39 GPU models as a table
ClassFirst observedLast observedNowPeakIn figure
RTX 40902025-12-172026-08-234856yes
RTX 50902025-12-172026-08-233040yes
RTX PRO 6000 Blackwell Server Edition2025-12-172026-08-2366yes
RTX PRO 5000 Blackwell2026-05-292026-08-2366yes
RTX 40802025-12-172026-08-2322yes
RTX 4080 SUPER2025-12-172026-08-2313yes
RTX 5070 Ti2025-12-172026-08-2313yes
RTX PRO 4500 Blackwell2026-04-152026-08-2312yes
RTX 30802026-05-162026-08-2311yes
RTX 30902025-12-172026-07-19042yes
H100 80GB HBM32025-12-172026-07-29021yes
B2002025-12-172026-03-08012yes
L40S2025-12-172026-08-18011yes
RTX PRO 6000 Blackwell Workstation Edition2025-12-172026-08-2108yes
RTX 3090 Ti2025-12-172026-04-1706yes
RTX 50802025-12-172026-08-0305yes
L4 (22GB)2026-03-022026-03-1104no
H100 NVL2025-12-172026-05-1904no
RTX 6000 Ada Generation2025-12-172026-07-2904no
Unknown GPU2026-03-122026-07-0504no
RTX 4090 D2026-01-062026-08-2203no
A100 80GB PCIe2025-12-252025-12-2602no
RTX PRO 6000 Blackwell Max-Q Workstation Edition2026-03-282026-03-2802no
RTX 5060 Ti2025-12-312026-07-2802no
B300 SXM6 AC2025-12-172025-12-1702no
L42026-02-242026-03-1001no
RTX 4090 (24GB)2026-03-052026-03-1101no
H2002026-01-072026-03-3001no
H100 PCIe2026-01-022026-07-1701no
RTX 5090 D2026-03-232026-08-1601no
RTX 4070 Ti SUPER2026-04-132026-07-3001no
RTX A50002025-12-172025-12-2001no
RTX A45002026-04-262026-04-2601no
RTX A60002025-12-172025-12-2901no
A800 80GB PCIe2025-12-252026-01-2301no
A100-SXM4-40GB2026-06-032026-07-1901no
RTX A40002026-03-062026-07-2201no
Graphics Device2026-07-262026-07-2701no
L402025-12-252025-12-2501no

Observed workers are concentrated: the leading GPU model holds 48 of 96 observed workers (50%). 7 established classes no longer reach the present.

Data, scope and citation

The census is a separate open dataset. It is not a performance ranking and never weights the goodput ratios.

Scope. One public, unauthenticated fast-network census, sampled once per dated day from 2025-12-17 through 2026-08-23. Counts show observed workers, not capacity, utilisation, price, ownership or the reason a model appeared or disappeared.

Method. Lifelines include every GPU model present now or peaking at 5 workers or more. Band thickness uses one shared scale. The complete table and downloads retain every observed model, including those below the figure threshold.

Separate measurement. Workload-specific production goodput lives at /market. The census is context only and is not combined with it.

Citation. Reuse is licensed under CC BY 4.0.