๐ Take the State of AI Report Survey!
STATE OF AI REPORT COMPUTE INDEX.
The State of AI Report Compute Index tracks the size of public, private and national HPC clusters, as well as the utilisation of various AI chips in AI research papers. As a key substrate upon which AI models are trained and run, the size of compute clusters and the popularity of specific chips helps us take a temperature check on the rate of progress of AI systems.
Cited chip usage in AI papers Last update: June 2026
Notes: This figure presents the number of open source AI papers that cite the use of specific AI chips according to analysis by Zeta Alpha. The "Big 6 Startups" category includes Habana, Graphcore, Cerebras, SambaNova Systemes, Cambricon and Groq. Figures for 2026 include counts as of 1 June 2026 and H2 predictions adjusted for volume.
Cited NVIDIA usage in AI papers Last update: June 2026
Notes: This figure presents the number of open source AI papers that cite the use of specific NVIDIA chips according to analysis by Zeta Alpha. Figures for 2026 include counts as of 1 June 2026 and H2 predictions adjusted for volume.
Cited AI chip startup usage in AI papers Last update: June 2026
Notes: This figure presents the number of open source AI papers that cite the use of specific AI chip startups according to analysis by Zeta Alpha. Figures for 2026 include counts as of 1 June 2026 and H2 predictions adjusted for volume.
Hopper GPU count (H100 / H200 / GH200) Data cutoff: 1 July 2026
Notes: Deployed NVIDIA Hopper GPUs (H100, H200, and GH200 Grace-Hopper superchips, counted as one Hopper GPU each) by operator, as of 1 July 2026. Public-cloud, private and national-HPC fleets combined. xAI Colossus 1 is owned by xAI/SpaceX and leased to Anthropic from May 2026; the Tesla and Meta figures include H100-equivalent and aggregate clusters. National-HPC counts are exact; large company figures are best available estimates. Sources: operator disclosures, EuroHPC, Top500 (June 2026), SemiAnalysis, Data Center Dynamics.
A100 GPU count Data cutoff: 1 July 2026
Notes: Deployed NVIDIA A100 (Ampere) GPUs by operator, as of 1 July 2026. Ampere is a legacy generation with no new builds; several fleets are being retired. Sources: operator and national-HPC disclosures, Top500.
Blackwell GPU count (B200 / B300) Data cutoff: 1 July 2026
Notes: NVIDIA Blackwell standalone GPUs (B200 and B300 "Blackwell Ultra"), split by Deployed / Installing / Announced, as of 1 July 2026. B100 did not ship at scale; most Blackwell volume ships as GB200/GB300 NVL72 (next chart). Public-cloud GA instance counts (Google, AWS, Azure, Lambda, others) are not disclosed. Sources: operator disclosures, NVIDIA, Data Center Dynamics.
Grace-Blackwell GPU count (GB200 / GB300) Data cutoff: 1 July 2026
Notes: NVIDIA Grace-Blackwell superchips (GB200 and GB300; one NVL72 rack = 72 GPUs), split by Deployed / Installing / Announced, as of 1 July 2026. The Announced segment is committed pipeline, not installed capacity. Counted by owner/operator; tenants (e.g. OpenAI on Oracle/Crusoe) are not added separately. China-bound and export-restricted parts are excluded. Sources: NVIDIA, operator disclosures, SemiAnalysis, Data Center Dynamics, EuroHPC.
Frontier-lab contracted compute (GW) Data cutoff: 1 July 2026
Notes: Frontier-lab contracted compute by silicon vendor, in gigawatts, for GW-disclosed deals only, as of 1 July 2026. OpenAI's non-NVIDIA share now reflects its AMD (6 GW, MI450) and Broadcom (10 GW, custom "Jalapeรฑo" ASIC) commitments; Anthropic's AWS Trainium and Google TPU commitments are only partially GW-disclosed, so its non-NVIDIA share remains conservative. Sources: company announcements, NVIDIA, AMD, Broadcom, CNBC.
Research topic signals by AI chip Last update: June 2025
Notes: The dataset covers 6,356 papers published between January 1 and June 1, 2025 and analysed using Zeta Alpha. Topic labels were assigned using GPT-4o-mini, and we filtered the results to include only chip-topic pairs with at least three papers. To focus the analysis on meaningful differentiation, we excluded topics whose percentage difference fell between -1 and +1 compared to the corpus-wide baseline. The results reveal clear skews between specific chips and research areas.
Authored on the interwebs by:
This work is licensed under a Creative Commons Attribution 4.0 International License.