Skip to content
Deekoya AI

AI Hardware

Specs as published by each vendor on its official product page; prices only where the vendor lists one.

Data-centre GPUs & accelerators

Chip / systemMakerMemoryBandwidthPower / compute
NVIDIA H100 SXM ↗NVIDIA80 GB HBM33.35 TB/sup to 700 W
NVIDIA H200 SXM ↗NVIDIA141 GB HBM3e4.8 TB/sup to 700 W
NVIDIA B200 (HGX) ↗NVIDIA180–192 GB HBM3eup to 8 TB/s~1,000 W
NVIDIA GB200 NVL72 ↗NVIDIA72 Blackwell GPUs + 36 Grace CPUs per rack130 TB/s NVLinkrack-scale
AMD Instinct MI300X ↗AMD192 GB HBM35.3 TB/s750 W
AMD Instinct MI325X ↗AMD256 GB HBM3E6 TB/s1,000 W
Intel Gaudi 3 ↗Intel128 GB HBM2e3.7 TB/sup to 900 W (OAM)
Google Cloud TPU v5p ↗Google95 GB HBM per chip2,765 GB/spods of 8,960 chips
Google Cloud TPU v6e (Trillium) ↗Google32 GB HBM per chip1,640 GB/spods of 256 chips
AWS Trainium2 ↗AWS96 GiB HBM per chip2.9 TB/sTrn2 instances, 16 chips
Cerebras WSE-3 ↗Cerebras44 GB on-chip SRAM, 900,000 cores21 PB/s memory bandwidthCS-3 system
Groq LPU ↗Groq230 MB on-chip SRAM80 TB/s on-dieinference-only

Desktop & workstation GPUs

Chip / systemMakerMemoryBandwidthPower / compute
NVIDIA GeForce RTX 5090 ↗NVIDIA32 GB GDDR71,792 GB/s575 W
NVIDIA GeForce RTX 4090 ↗NVIDIA24 GB GDDR6X1,008 GB/s450 W
NVIDIA DGX Spark (GB10) ↗NVIDIA128 GB unified LPDDR5x273 GB/sup to 1 PFLOP FP4
NVIDIA Jetson AGX Orin 64GB ↗NVIDIA64 GB LPDDR5204.8 GB/s275 TOPS, 15–60 W
NVIDIA Jetson Orin Nano Super ↗NVIDIA8 GB LPDDR5102 GB/s67 TOPS, 7–25 W

NPUs & AI PC chips

Chip / systemMakerMemoryBandwidthPower / compute
Apple M4 (Neural Engine) ↗Appleup to 32 GB unified120 GB/s38 TOPS NPU
Qualcomm Snapdragon X Elite ↗Qualcommup to 64 GB LPDDR5x135 GB/s45 TOPS NPU (Copilot+ PC)
Intel Core Ultra 200V (Lunar Lake) ↗Intel16/32 GB on-packageLPDDR5X-8533up to 48 TOPS NPU, 120 platform TOPS
AMD Ryzen AI 300 ↗AMDLPDDR5X / DDR5—up to 50 TOPS XDNA 2 NPU

Chip-making supply chain

Chip / systemMakerMemoryBandwidthPower / compute
TSMC ↗Foundry (Taiwan)N3 / N2 process nodesCoWoS advanced packaging for AI GPUsmakes most leading AI accelerators
ASML ↗Lithography (Netherlands)EUV and High-NA EUV (EXE:5000) scannerssole supplier of EUV toolsrequired for leading-edge nodes
SK hynix / Samsung / Micron ↗HBM memoryHBM3E stacks for AI GPUs—Micron assembles/tests in Sanand, India

Data centres by country

Counts of data centres listed in the Cloudscene directory (colocation, cloud and enterprise facilities). Directory counts measure listed facilities, not capacity (MW). Statista's May 2026 snapshot of the same directory shows US 5,427, Germany 529, UK 523, China 449, Canada 337 and India 155. As of November 2025.

  1. 1United States5,427
  2. 2Germany529
  3. 3United Kingdom523
  4. 4China449
  5. 5Canada337
  6. 6France322
  7. 7Australia314
  8. 8Netherlands298
  9. 9Russia251
  10. 10Japan222
  11. 11Italy168
  12. 12India153
  13. 13Poland144
  14. 14Spain144
  15. 15Hong Kong122
  16. 16Switzerland121
  17. 17Singapore99
  18. 18Sweden95
  19. 19Indonesia88
  20. 20New Zealand83
  21. 21Belgium81
  22. 22Austria68
  23. 23Malaysia62
  24. 24Ukraine58
  25. 25Ireland55
  26. 26Denmark50
  27. 27Finland48
  28. 28Norway47
  29. 29South Korea43
  30. 30Thailand42

Bar length uses a square-root scale. Sources: Cargoson: Number of Data Centers by Country (Nov 2025, Cloudscene data) · Statista: Which countries have the most data centers (May 2026) · Cloudscene India market

By operator (facilities listed on Wikidata)

198 individually documented facilities from Wikidata (CC0). This is a partial list, not a market-share ranking.

India: fabs & data centres

1,789 MW operational, 147 data centres, 33 operators (H1 2026, Cushman & Wakefield); JLL: ~1.6 GW. Mumbai leads (890 MW). Source: Cushman & Wakefield. Full details on India AI → chips.

Latest AI hardware news