AI accelerator specs

First availability: Sep 2025est — estimated / secondary-source figure

Interconnect: Infinity Fabric

Power: 1,400 W

Process: TSMC 3/6nm

CDNA 4, liquid-cooled. Values use AMD's dense BF16/OCP-FP8 and MXFP4 specifications; sparse headline rates are excluded.

Est. purchase price: ~$25,000 · Analyst estimate of average selling price · CNBC, citing HSBC, 2025-07-10 ↗

$8.60 / GPU-hour · Cheapest on-demand · Oracle OCI · observed 2026-10-14

Coverage: 1 provider · Oracle OCI

BM.GPU.MI355X.8; 8 × MI355X 288GB

Specifications reviewed 2026-10-15. Use explicit dense BF16/OCP-FP8 and MXFP4 rates, not the adjacent sparse headline values.

First availability: Aug 2025est — estimated / secondary-source figure

Interconnect: NVLink 5 · 1.8 TB/s

Power: 1,400 W

Process: TSMC 4NPest — estimated / secondary-source figure

One of the 72 Blackwell Ultra GPUs in a GB300 NVL72 rack; dense FP4 is 1.5× GB200's, with BF16 and FP8 unchanged. Launch month is an estimate: the first racks shipped in July 2025 and the first cloud instances followed in August.

No tracked rental quote; other billing models may exist.

Specifications reviewed 2026-10-24. Datasheet's individual-GPU GB300 NVL72 column: FP4 20 | 15 PF sparse | dense; FP8 10 and BF16 5 PF sparse, halved as the footnote directs to 5 and 2.5 PF; 279 GB HBM3E at 8 TB/s; TDP configurable up to 1,400 W. The rack page's totals divided by 72 agree. The process node comes from NVIDIA's Blackwell Ultra technical blog.

First availability: Feb 2025 · NVIDIA blog: CoreWeave GB200 NVL72 instances ↗

Interconnect: NVLink 5 · 1.8 TB/s

Power: 1,200 W

Process: TSMC 4NP

One of the 72 Blackwell GPUs in a GB200 NVL72 rack (two per Grace Blackwell superchip); rentals are priced per GPU. Launch month is an estimate: the first generally available cloud instances were announced in February 2025.

$16.00 / GPU-hour · Cheapest on-demand · Oracle OCI · observed 2026-10-14

Coverage: 1 provider · Oracle OCI

BM.GPU.GB200.4; 4-GPU GB200 NVL72 instance, 189GB per GPU

Specifications reviewed 2026-10-15. Datasheet's individual-GPU column for GB200 NVL72: its sparse FP4/FP8/BF16 figures (20/10/5 PF) halved, as its footnote directs, to 10/5/2.5 PF dense; the rack page's totals divide to the same. 186 GB HBM3E at 8 TB/s and a TDP configurable up to 1,200 W per GPU.

First availability: Nov 2025est — estimated / secondary-source figure

Interconnect: ICI · 9,216-chip pods

Power: Not published

Google Cloud TPU7x. Compute is per chip, comprising two TensorCores; memory capacity follows the documentation's GiB specification. GA since March 2026 and available on demand (quota required); Spot, Flex-start and calendar-mode reservations are allowlist-restricted.

No tracked rental quote; other billing models may exist.

Specifications reviewed 2026-10-15. Current TPU7x table (7,380 GB/s shown as 7.38 TB/s); GA since March 31, 2026. No FP4 rate or per-chip power is published.

First availability: Sep 2025est — estimated / secondary-source figure

Interconnect: NVLink 5 · 1.8 TB/s

Power: 1,100 W

Process: TSMC 4NP

HGX variant: dense FP4 is 108 PF per eight-GPU system divided by eight. Current datasheet specifies 270 GB and 1,100 W; GB300 NVL72 has different per-GPU specifications.

$7.40 / GPU-hour · Cheapest on-demand · Hyperstack · observed 2026-10-25; quotes checked through 2026-10-25

Coverage: 6 providers · AWS, Hyperstack, Nebius, Oracle OCI, Together AI, Verda

NVIDIA B300 288GB; maximum per GPU: 28 pCPUs and 240 GB RAM; actual allocation not specified; published per GPU-hour

Specifications reviewed 2026-10-15. Current HGX column specifies 270 GB, 7.7 TB/s and 1,100 W; older reference tables list 288 GB. Dense FP4 is 108 PF/system ÷ 8 (the per-GPU figure is rounded to 14 PF in the PDF).

First availability: Dec 2024est — estimated / secondary-source figure

Interconnect: NVLink 5 · 1.8 TB/s

Power: 1,000 W

Process: TSMC 4NP

Dual-die Blackwell; first NVIDIA part with FP4 tensor formats.

Est. purchase price: ~$35,000 · Midpoint of the range given for Blackwell · CNBC, quoting NVIDIA's CEO, 2024-03-19 ↗

$6.00 / GPU-hour · Cheapest on-demand · Hyperstack · observed 2026-10-25; quotes checked through 2026-10-25

Coverage: 7 providers · AWS, Hyperstack, Lambda, Nebius, Oracle OCI, Together AI, Verda

NVIDIA B200 192GB; maximum per GPU: 31 pCPUs and 256 GB RAM; actual allocation not specified; published per GPU-hour

Specifications reviewed 2026-10-15. HGX B200 column: 7.7 TB/s per GPU (62 TB/s per board) replaces 8 TB/s, which DGX B200 pages still list. Per-GPU rates are derived from the eight-GPU table.

First availability: Sep 2024est — estimated / secondary-source figure

Interconnect: 24×200 GbE RoCE

Power: 900 W

Process: TSMC 5nm

Priced aggressively; FP8 and BF16 share the same MME throughput.

Est. purchase price: ~$15,625 · Eight-accelerator kit list price ÷ 8, baseboard included · Intel Computex 2024 keynote (archived), 2024-06-04 ↗

No tracked rental quote; other billing models may exist.

Specifications reviewed 2026-10-15. Intel's July 2025 paper gives 1,678 TFLOPS BF16/FP8 MME at 900 W; its 2024 Hot Chips deck said 1,835.

First availability: Feb 2025est — estimated / secondary-source figure

Interconnect: Infinity Fabric

Power: 1,000 W

Process: TSMC 5/6nm

MI300X refresh with HBM3e.

No tracked rental quote; other billing models may exist.

Specifications reviewed 2026-10-15. Dense rates, 256 GB HBM3E and 6 TB/s confirmed.

First availability: Dec 2023 · AMD Advancing AI 2023 ↗

Interconnect: Infinity Fabric

Power: 750 W

Process: TSMC 5/6nm

CDNA 3 chiplet design; led on memory capacity at launch.

$2.99 / GPU-hour · Cheapest on-demand · Hot Aisle · observed 2026-10-25

Coverage: 2 providers · Hot Aisle, Oracle OCI

1x AMD MI300X 192GB VM (Small plan); 8 or 13 CPU cores and 224GB system RAM per VM; published per GPU-hour; billed by the minute

Specifications reviewed 2026-10-15. Dense rates confirmed; the datasheet's more precise 1,307.4/2,614.9 TFLOPS round to the table values.

First availability: Nov 2023 · Lambda Cloud Clusters with GH200 ↗

Interconnect: NVLink-C2C · 900 GB/s

Power: Not published

The Hopper GPU of a Grace Hopper superchip, in its 96 GB HBM3 version (the one rented by the hour). NVIDIA publishes power only for the whole module, CPU and memory included.

No tracked rental quote; other billing models may exist.

Specifications reviewed 2026-10-24. 96 GB HBM3 column: 990 BF16 and 1,979 FP8 TFLOPS dense, up to 4 TB/s, NVLink-C2C 900 GB/s. Power is published only for the module (450 to 1,000 W, memory, CPU and GPU together), so power stays empty.

First availability: Jun 2024est — estimated / secondary-source figure

Interconnect: NVLink 4 · 900 GB/s

Power: 700 W

Process: TSMC 4N

Same compute as H100; HBM3e lifts capacity 1.8× and bandwidth 1.4×.

$3.99 / GPU-hour · Cheapest on-demand · Hyperstack · observed 2026-10-25; quotes checked through 2026-10-25

Coverage: 8 providers · AWS, Azure, Hyperstack, Jarvislabs, Nebius, Oracle OCI, Together AI, Verda

NVIDIA H200 SXM 141GB; maximum per GPU: 22 pCPUs and 225 GB RAM; actual allocation not specified; published per GPU-hour

Specifications reviewed 2026-10-15. SXM column; sparse tensor figures divided by two, with existing whole-TFLOPS rounding retained.

First availability: Oct 2022 · NVIDIA Q3 FY2023 results ↗

Interconnect: NVLink 4 · 900 GB/s

Power: 700 W

Process: TSMC 4N

Defined the 2023–25 training buildout; the market’s reference chip.

Est. purchase price: ~$27,500 · Midpoint of an analyst range · CNBC, citing Raymond James, 2024-01-18 ↗

$3.20 / GPU-hour · Cheapest on-demand · Hyperstack · observed 2026-10-25; quotes checked through 2026-10-25

Coverage: 10 providers · AWS, Azure, Google Cloud, Hyperstack, Jarvislabs, Lambda, Nebius, Oracle OCI, Together AI, Verda

NVIDIA H100 SXM 80GB; maximum per GPU: 24 pCPUs and 240 GB RAM; actual allocation not specified; published per GPU-hour

Specifications reviewed 2026-10-15. SXM column; sparse tensor figures divided by two, with existing whole-TFLOPS rounding retained.

First availability: Oct 2024est — estimated / secondary-source figure

Interconnect: ICI · 256-chip pods

Power: Not published

4.7× the peak compute of v5e.

$2.70 / chip-hour · Cheapest on-demand · Google Cloud · observed 2026-10-14

Coverage: 1 provider · Google Cloud

Trillium / TPU v6e; per chip-hour

Specifications reviewed 2026-10-15. Per-chip architecture comparison; FP8 throughput is explicitly listed, not inferred from BF16.

First availability: Mar 2024est — estimated / secondary-source figure

Interconnect: NVLink bridge · 600 GB/s

Power: 400 W

Dual-slot PCIe H100 with 94 GB of HBM3, sold in NVLink-bridged pairs. Launch month is an estimate: the first general availability found is Azure's NC H100 v5 in March 2024.

No tracked rental quote; other billing models may exist.

Specifications reviewed 2026-10-24. NVL column: sparse 1,671 BF16 and 3,341 FP8 TFLOPS halved; 94 GB HBM3 at 3.9 TB/s (3,938 GB/s in the brief), 400 W maximum by default, NVLink bridge at 600 GB/s. Neither document states a process node.

First availability: Dec 2025 · Amazon: Trainium3 UltraServers now available ↗

Interconnect: NeuronLink v4 · 2.56 TB/s

Power: Not published

Process: 3nm

Per-chip NeuronCore-v4 specifications: MXFP8 and MXFP4 share the published throughput; dense BF16 is lower and must not be inferred by halving FP8.

No tracked rental quote; other billing models may exist.

Specifications reviewed 2026-10-15. 671 TFLOPS dense BF16; 2,517 TFLOPS MXFP8 and MXFP4. Dense BF16 cannot be inferred by halving FP8; AWS publishes no per-chip power.

First availability: Dec 2024est — estimated / secondary-source figure

Interconnect: NeuronLink v3 · 1.28 TB/s

Power: 500 West — estimated / secondary-source figure

AWS Neuron publishes dense BF16/FP16/TF32 and FP8 separately from sparse throughput. Board power remains an earlier estimate.

No tracked on-demand quote. Capacity Blocks: $2.24 / chip-hour · AWS (trn2) · observed 2026-10-14

trn2.48xlarge; 16 × Trainium2 chips

Specifications reviewed 2026-10-15. Dense compute and per-chip memory/interconnect; the 500 W board power remains a secondary estimate.

First availability: Dec 2023est — estimated / secondary-source figure

Interconnect: ICI 3D torus · 8,960-chip pods

Power: Not published

Announced alongside Gemini 1.0 (Dec 2023), which was trained on TPU v4 and v5e.

$4.20 / chip-hour · Cheapest on-demand · Google Cloud · observed 2026-10-14

Coverage: 1 provider · Google Cloud

TPU v5p; per chip-hour

Specifications reviewed 2026-10-15. Per-chip table; FP8 runs at the BF16 rate and capacity is specified in GiB.

First availability: Oct 2023est — estimated / secondary-source figure

Interconnect: PCIe 4.0

Power: 350 W

Ada Lovelace data-center card with 48 GB of GDDR6, not HBM, and no NVLink. Launch month is an estimate: NVIDIA said in August 2023 that it would be available from that fall.

$1.55 / GPU-hour · Cheapest on-demand · Nebius · observed 2026-10-20; quotes checked through 2026-10-25

Coverage: 4 providers · AWS, Nebius, Oracle OCI, Verda

L40S with Intel CPU; 8–40 vCPUs and 32–160GB RAM range

Specifications reviewed 2026-10-15. Dense tensor rates are the unstarred figures (BF16 362.05, FP8 733 TFLOPS); the starred sparse ones are excluded, and no FP4 format is listed. 48 GB GDDR6 at 864 GB/s, 350 W, PCIe Gen4 x16 without NVLink. The page states no process node.

First availability: Jan 2025 · NVIDIA: GeForce RTX 5090 and 5080 out now ↗

Interconnect: PCIe 5.0

Power: 575 W

Process: TSMC 4N

Consumer Blackwell with FP4 support. Tensor rates are NVIDIA's FP32-accumulate figures; the card's 2× FP16-accumulate rates are not comparable with data-center specifications.

Est. purchase price: ~$1,999 · Launch list price (MSRP) · NVIDIA GeForce RTX 50 Series launch, 2025-01-06 ↗

No tracked on-demand quote. Serverless / per-second: $1.09 / GPU-hour · Beam · observed 2026-10-20

RTX 5090 PCIe 32 GB serverless GPU task; one GPU; rate published per second, billed by the millisecond

Specifications reviewed 2026-10-15. Dense FP32-accumulate tensor rates: BF16 209.5 TFLOPS, FP8 419 TFLOPS, FP4 1,676 TFLOPS; process TSMC 4N. The whitepaper's 2× FP16-accumulate rates are not comparable with data-center specifications.

First availability: Aug 2023est — estimated / secondary-source figure

Interconnect: ICI 2D torus

Power: Not published

Efficiency-tier TPU; INT8 393 TOPS.

$1.20 / chip-hour · Cheapest on-demand · Google Cloud · observed 2026-10-14

Coverage: 1 provider · Google Cloud

TPU v5e; per chip-hour

Specifications reviewed 2026-10-15. 197 TFLOPS BF16 and 393 INT8 TOPS per chip; no FP8, FP4 or power figure is published. 0.82 TB/s follows Google's Trillium comparison; the v5e table's 800 GiBps would read 0.86.

First availability: Mar 2023 · NVIDIA GTC 2023 inference platforms launch ↗

Interconnect: PCIe 4.0

Power: 72 W

Process: TSMC 4N

Low-profile 72 W Ada inference card with 24 GB of GDDR6 and no NVLink.

No tracked rental quote; other billing models may exist.

Specifications reviewed 2026-10-24. Dense BF16 121 and FP8 242 TFLOPS, the unstarred halves of the sparse figures; 24 GB GDDR6 with ECC at 300 GB/s, 72 W, PCIe Gen4 x16; TSMC 4N. No FP4 format is listed (INT4 only).

First availability: Jan 2026 · Microsoft: Maia 200 ↗

Interconnect: Ethernet scale-up

Power: 750 W

Process: 3nm

Microsoft’s inference chip, live in Azure Jan 2026; serves GPT-5.x-class models at claimed 30% better perf/$. Its 1,268 TFLOPS BF16 peak runs on the vector unit (the tensor unit handles FP8/FP6/FP4), so no tensor BF16 rate is listed.

No tracked rental quote; other billing models may exist.

Specifications reviewed 2026-10-15. Peak-specification table: 5,072 TFLOPS FP8, 10,145 TFLOPS FP4, 750 W. BF16 (1,268 TFLOPS) is a vector-unit rate, so tf16 stays empty.

First availability: Jul 2025est — estimated / secondary-source figure

Interconnect: PCIe 5.0

Power: 600 W

Passively cooled data-center Blackwell card with 96 GB of GDDR7 and no NVLink. NVIDIA publishes only headline tensor rates, which its Workstation sibling footnotes as sparse, so no dense rate is shown. Launch month is an estimate from the first cloud instances.

No tracked rental quote; other billing models may exist.

Specifications reviewed 2026-10-24. 96 GB GDDR7 with ECC at 1,597 GB/s, up to 600 W (300 W minimum), PCIe 5.0 x16, NVLink not supported. NVIDIA's 1 / 2 / 4 PFLOPS FP16-BF16 / FP8 / FP4 headlines are not marked dense, and the Workstation Edition footnotes the same FP4 figure as sparse, so no dense rate is recorded.

First availability: Oct 2026est — estimated / secondary-source figure

Interconnect: Helios rack-scale (72 GPU)

Power: 2,500 W

Process: TSMC 2/3nm

MI400 Series flagship: CDNA 5 with HBM4. AMD's 20.1 PF FP8 and 40.3 PF MXFP4 are dense rates; it lists sparsity only for FP16/BF16/INT8. Customers: OpenAI 6 GW, Meta up to 6 GW (custom MI450-based GPU), Anthropic up to 2 GW. Launch month is an estimate.

No tracked rental quote; other billing models may exist.

Specifications reviewed 2026-10-15. Row is the MI455X, whose dense rates AMD publishes directly (5.0 PF BF16, 20.1 PF FP8, 40.3 PF MXFP4); 23.3 TB/s HBM4 bandwidth and 2,500 W maximum board power.

First availability: Aug 2026 · NVIDIA Q2 FY2027 earnings call ↗

Interconnect: NVLink 6 · 3 TB/s

Power: 1,800 W

NVIDIA's individual-GPU table specifies dense BF16, FP8/FP6 training and NVFP4 training directly. FP4 here uses its 35 PF dense training figure, not the separate 50 PF sparse inference claim. Bandwidth and NVLink follow the Oct 2026 NVL72 figures (19.2 TB/s, 3 TB/s); older HGX pages still say 22 and 3.6 TB/s. 1,800 W is NVIDIA's NVL4 benchmark power assumption, and launch timing is an estimate.

No tracked rental quote; other billing models may exist.

Specifications reviewed 2026-10-15. Rubin GPU column: 19.2 TB/s and NVLink 6 at 3 TB/s (HGX and preliminary DGX Rubin NVL8 pages still imply 22 and 3.6 TB/s). 1,800 W is the datasheet's NVL4 benchmark power; NVIDIA states no process node.

First availability: Oct 2022est — estimated / secondary-source figure

Interconnect: NVLink bridge · 600 GB/s

Power: 350 W

Process: TSMC 4N

The 350 W PCIe card on the H100 die, with HBM2e at 2 TB/s. NVIDIA's H100 page now lists only the SXM and NVL parts. Launch month is an estimate: NVIDIA began shipping H100 in October 2022 without naming the form factor.

No tracked rental quote; other billing models may exist.

Specifications reviewed 2026-10-24. Whitepaper Table 3 PCIe column: 756 BF16 and 1,513 FP8 TFLOPS with FP32 accumulate, half the sparse figures; 350 W, 4N. Bandwidth follows the product brief's 2,000 GB/s peak (the whitepaper says 2,039). No FP4 format is listed; legacy status is editorial, as NVIDIA's H100 page now lists only SXM and NVL.

First availability: Nov 2020est — estimated / secondary-source figure

Interconnect: NVLink 3 · 600 GB/s

Power: 400 W

Process: TSMC N7

Workhorse of the GPT-3→GPT-4 era.

Est. purchase price: ~$10,000 · Reported price in February 2023, not at launch · CNBC, 2023-02-23 ↗

$1.35 / GPU-hour · Cheapest on-demand · Hyperstack · observed 2026-10-25; quotes checked through 2026-10-25

Coverage: 7 providers · AWS, Google Cloud, Hyperstack, Jarvislabs, Lambda, Oracle OCI, Verda

NVIDIA A100 80GB; maximum per GPU: 28 pCPUs and 120 GB RAM; actual allocation not specified; published per GPU-hour

Specifications reviewed 2026-10-15. 80GB SXM column; process TSMC N7.

First availability: Oct 2022 · NVIDIA: GeForce RTX 4090 out now ↗

Interconnect: PCIe 4.0

Power: 450 W

Process: TSMC 4N

Consumer anchor with no NVLink, out of production since late 2024 by press reports (NVIDIA has published no end-of-life notice). Tensor rates are NVIDIA's FP32-accumulate figures; the card's 2× FP16-accumulate rates are not comparable with data-center specifications.

Est. purchase price: ~$1,599 · Launch list price (MSRP) · NVIDIA GeForce RTX 40 Series launch, 2022-09-20 ↗

No tracked on-demand quote. Serverless / per-second: $0.69 / GPU-hour · Beam · observed 2026-10-20

RTX 4090 PCIe 24 GB serverless GPU task; one GPU; rate published per second, billed by the millisecond

Specifications reviewed 2026-10-15. Dense FP32-accumulate tensor rates: BF16 165.2 TFLOPS, FP8 330.3 TFLOPS; no FP4 tensor format. The whitepaper's 2× FP16-accumulate rates are not comparable with data-center specifications.

First availability: Jun 2017est — estimated / secondary-source figure

Interconnect: NVLink 2 · 300 GB/s

Power: 300 W

Process: TSMC 12FFN

First tensor-core GPU; FP16 only (no BF16).

$0.23 / GPU-hour · Cheapest on-demand · Verda · observed 2026-10-25

Coverage: 4 providers · AWS, Lambda, Oracle OCI, Verda

1-GPU instance 1V100.6V; Tesla V100 16GB; 6 vCPUs and 23 GB RAM per instance; published per instance-hour, which is per GPU-hour

Specifications reviewed 2026-10-15. SXM2 column; FP16 tensor only (no BF16, FP8 or FP4).