Research cutoff · August 30, 2026

NVIDIA GPUs by country, through 2027

A country-by-country audit of physical NVIDIA data-center GPUs, dated deployments, modeled capacity and announced ceilings.

26 countries · UN 2027 population denominator · links preserved
#11
Armenia in timing-consistent 2027 scenario
#2
Armenia by scenario GPUs per capita
#5
Armenia by 2027 B200-equivalent scenario
1 : 41
Armenia residents per scenario GPU
How to read these estimates

Public evidence does not provide precise national GPU inventories.

This report estimates how many physical NVIDIA data-center GPUs may be located in each country by the end of 2027. It combines company and government announcements, regulatory filings, reported orders, dated construction plans and—where explicitly labeled—capacity modeling. Sources sometimes report performance-equivalent rather than physical GPUs, global fleets without country allocation, overlapping customer and operator capacity, or multi-year programs whose later phases extend beyond 2027. Each country row therefore separates firm evidence, a comparable 2027 scenario and longer-term ceilings. The figures are transparent evidence-based estimates, not audited inventories or guaranteed forecasts.

Main dataset

Country evidence table

Click any column heading to sort. The default “2027 scenario” is the broad comparison number from the audit; it mixes firm counts, modeled capacity and program ceilings, all explicitly labeled.

Total rank ↕Country ↕2027 scenario GPUs ↕GPUs / million ↕Per-capita rank ↕2027 population ↕Evidence & sources
Population-normalized

The small-country advantage

Population is the UN World Population Prospects 2024 medium projection for July 1, 2027. Per-capita values inherit every uncertainty in the GPU numerator.

Generation-normalized scenario

B200-equivalent compute

A separate, model-sensitive view of theoretical dense BF16 Tensor Core throughput. Physical counts above remain the primary inventory view.

One B200 equivalent = one B200 GPU's published dense BF16 Tensor Core peak. From NVIDIA's HGX specifications ↗, B300 counts as 1.00 under this deliberately narrow metric and preliminary Rubin as 1.78; GB200/GB300 ↗ counts as 1.11; H100/H200/GH200 ↗ as 0.44; A100 ↗ as 0.14; H20 as about 0.066; and V100 as 0.056. The ranges vary the undisclosed model mix while holding each country's physical 2027 point estimate fixed. They do not measure memory capacity, NVLink topology, utilization, software, power, FP4 inference or real workload speed. Download country scenarios ↧ · Download conversion factors ↧
B200e rank ↕Country ↕Physical GPUs ↕B200 equivalents ↕B200e / GPU ↕Reported or inferred GPU typesModel note
Beyond NVIDIA

Custom AI accelerators

Actual TPU and Trainium quantities, kept separate from the NVIDIA rankings because complete country allocation is not public.

6.59M
Modeled global TPU shipments through Q2 2026
2.13M
Global TPU B200-equivalents through Q2 2026
1.4M
Trainium2 chips landed globally by end-2025
These are quantities, not merely conversion factors. Epoch's August 27, 2026 downloadable shipment model ↗ estimates 6.59M Google TPU chips shipped globally through Q2 2026, with a 4.38M–9.11M uncertainty range. Applying published dense-BF16 specifications gives about 2.13M B200e globally, with a 1.44M–2.89M range. Delivered chips may not yet be installed, and the dataset does not locate them by country.
Accelerator generationGlobal units through Q2 2026B200e / chipGlobal B200eEvidence treatment
TPU v4 + v4i293,6870.116 weighted34,054Legacy group; compute uses Epoch's mapping.
TPU v5e2,357,6190.088206,423Google BF16 specification.
TPU v5p665,3000.204135,721Google BF16 specification.
TPU v6e2,599,1600.4081,060,457Google publishes 918 TFLOPS BF16 per chip.
TPU v7677,2621.025694,419Google publishes 2,307 TFLOPS BF16 per chip.
All Google TPUs6,593,028
4,382,010–9,108,422
0.323 weighted2,131,074
1,441,204–2,892,520
Global modeled shipments; not a U.S. inventory.
AWS Trainium21,400,000 global≈0.30≈420,000 globalCompany-disclosed units; B200e uses Epoch's compute mapping.
Illustrative U.S. TPU sensitivity

3.30M–4.62M TPU chips

If 50%–70% of Google's modeled global TPU stock is U.S.-located—the same broad location sensitivity used for the NVIDIA ownership cross-check—it would equal about 1.07M–1.49M B200e. This is an explicit assumption, not a disclosed Google country allocation, and is not used in either ranking above.

Illustrative U.S. all-accelerator sensitivity: adding the 50%–70% TPU location case and 500K–1M documented/path Trainium2 to the 7.4M central NVIDIA estimate gives roughly 8.62M–9.19M B200e and 12.30M–14.12M physical accelerator packages. This is not promoted to a country ranking because Google does not publish a complete U.S. TPU inventory and the latest open TPU series ends at Q2 2026 rather than December 2027. Download the custom-accelerator data ↧
Conversion scope: Google publishes per-chip TPU BF16 specifications ↗. Trainium2 ≈ 0.30 uses Epoch's shipment/compute model ↗ because AWS's public 20.8-PFLOPS figure for a 16-chip Trn2 instance does not provide an equally clean dense-BF16 specification. Inferentia, Maia and MTIA remain unnormalized where comparable public precision and fleet data are insufficient.
Rules of comparison

Methodology and caveats

What is counted

  • Physical NVIDIA data-center accelerators located or credibly planned in-country.
  • Installed, received, ordered, scheduled, modeled and ceiling figures are kept separate in each row.
  • The sortable scenario number is a comparison device—not an audited installed inventory.
  • GPUs per million = scenario GPUs ÷ projected 2027 population × 1,000,000.

What is not silently counted

  • H100-equivalent performance as if it were physical chips.
  • Global hyperscaler fleets assigned entirely to headquarters countries; the US model uses an explicit 50–70% location assumption; the 74.5% performance statistic is only a cross-check.
  • Customer capacity contracts added again on top of the operator's hardware.
  • MW converted to GPUs without being labeled as modeling.

How B200 equivalents are calculated

  • Published peak dense BF16/FP16 Tensor Core throughput is divided by B200's 2.25 PFLOPS per GPU.
  • Known model counts are converted directly; undisclosed balances use a documented low, central and high generation mix.
  • The normalization range changes only the assumed type mix. It retains the same physical scenario count, so readers can compare the two tables without hidden inventory changes.
  • Conversion factors and every country assumption are downloadable in the data files.

Limits of a single compute unit

  • B300's extra memory and 2× attention performance are invisible in dense BF16, so B300 = B200 here.
  • Rubin specifications are preliminary and subject to change.
  • Rack interconnect, memory bandwidth, FP4/FP8 speed, availability and utilization can dominate real workloads.
  • “B200-equivalent” is therefore a transparent theoretical-compute scenario—not a benchmark or procurement valuation.
Evidence archive

All collected sources and links

Deduplicated from the original source ledger, every country comment and the expanded audit report.