Material Insights · Understanding the Material World

AI accelerator chip: Use case

Use cases

Each use, its share of demand and volume, how critical the product is to it and why, and the price paid

Use caseVolumeUnitCriticalityReasonYearSource
Training large AI models in data centres3.76million NVIDIA data-centre GPUs shipped in 2023 (training and inference together)EssentialTraining a large model needs thousands of accelerators working together; general-purpose processors are far too slow.2023[45]
Serving trained models to users (inference) in cloud data centres– – ImportantSmall models can run on ordinary processors, but large ones need accelerators; operational time for inference is assumed at 40% against 80% for training.2024[22]
Public AI compute for research and start-ups, India38,000GPUs onboarded on the IndiaAI compute portalImportantThe state pools GPUs to rent out at an affordable rate, because buying chips is beyond most start-ups.2026[23]

as of 2026-11-16. No source gives demand shares by use. TechInsights counts only NVIDIA units and the 2023 figure includes all uses [45].

Industries that buy it

The industries and kinds of firm that buy it, for which use case, and how critical it is to them

Used by 4
  • Hyperscale cloud providers Training and serving models in cloud data centres · criticality essential NVIDIA names Amazon Web Services, Google, Meta, Microsoft, Oracle and others among the organisations expected to adopt Blackwell [14]; Google also builds its own TPUs [39].
  • Foundation model developers Training and running frontier models · criticality essential Anthropic has announced plans to use up to a million Google TPUs [39]; OpenAI is among Blackwell adopters [14].
  • Hyperscale clouds and GPU cloud operators Renting GPU time to firms that cannot buy chips · criticality important GPU cloud operators buy the same chips and resell their time; no share found.
  • Government IT and digital public infrastructure providers Pooled public AI compute (India) · criticality important The IndiaAI Mission has onboarded over 38,000 GPUs on a common compute portal [23].

as of 2026-11-16

What it competes with

The products and technologies that compete with it in each use case

Competes with
  • Processor chip Inference and on-device AI General-purpose processors, and the neural-processing blocks built into them, can run small models, at lower speed and efficiency. No source opened measures how far they can replace accelerators.
  • Nm Neuromorphic computing Low-power inference at the edge An emerging approach that copies the brain's way of computing. No source opened compares its cost or speed with accelerators.

as of 2026-11-16

How it is used

Lifetime, use per year and utilisation, by variant

MeasureValueUnitVariantPlaceYearSource
Lifespan assumed for AI machines in a Google life-cycle study6yearsCustom acceleratorWorld2025[3]
Average lifespan assumed for the A100 GPU in a cradle-to-grave study3yearsData-centre GPUWorld2025[17]
Operational time of servers doing training (LBNL model assumption)80% of the yearData-centre GPUUnited States of America2024[22]
Operational time of servers doing inference (LBNL model assumption)40% of the yearData-centre GPUUnited States of America2024[22]
Average power of an 8-H100 node in saturated training, as a share of rated power78% (about)Data-centre GPUUnited States of America2024[22]
Average power usage effectiveness (PUE) of US data centres1.4ratio of facility to IT energyData-centre GPUUnited States of America2023[22]
Share of average power that Google's TPU machines draw when idle60%Custom acceleratorWorld2025[3]
Ratio of TDP to actual average power for TPUs (TDP overstates real use by this factor)2 to 6timesCustom acceleratorWorld2025[3]

as of 2026-11-16. Hours in a year are 8,760. Estimates of utilisation and lifespan differ by study; the energy figure on the Footprint tab uses the LBNL operational time for training and the LBNL PUE.

Prices

Unit prices by market or use, with year and source

MarketValueUnitYearSource
NVIDIA H100, low end of analysts' estimates25,000USD per chip2024[46]
NVIDIA H100, high end of analysts' estimates40,000USD per chip2024[46]
NVIDIA Blackwell, low end of the chief executive's statement30,000USD per chip2024[46]
NVIDIA Blackwell, high end of the chief executive's statement40,000USD per chip2024[46]
Cost to make the GPUs of one DGX H100, per GPU (USD 26,600 divided by eight, calculated; not a price)3325USD per GPU2023[47]

as of 2026-11-16. Unit values from trade data are calculated by the site and are not researched here.

Sources

  1. Life-cycle emissions of AI hardware: a cradle-to-grave approach and generational trends Schneider et al. (Google), arXiv preprint 2502.01671, 2025
  2. NVIDIA Blackwell Platform Arrives to Power a New Era of Computing NVIDIA investor relations, 2024
  3. More than Carbon: Cradle-to-Grave environmental impacts of GenAI training on the Nvidia A100 GPU (arXiv:2509.00093v3) arXiv preprint, 2025
  4. 2024 United States Data Center Energy Usage Report (LBNL-2001637) Lawrence Berkeley National Laboratory, 2024
  5. IndiaAI Mission expands AI ecosystem with affordable compute and startup support Digital India, Government of India, 2026
  6. Google's Ironwood TPUs represent a bigger threat than Nvidia would have you believe The Register, 2025
  7. Nvidia GPU shipments totaled 3.76m in 2023, equating to a 98% market share: report (TechInsights) Data Center Dynamics
  8. Nvidia's latest AI chip will cost more than $30,000, CEO says CNBC via NBC San Diego, 2024
  9. AI Server System BOM & Rack Cost: Historical Data & Chart (newest print March 2026) Silicon Analysts, 2026