Track global GPU distribution, infrastructure investments, and compute capacity across the world's leading AI organizations. Real-time intelligence on who controls the hardware powering the AI revolution.
H100-equivalent GPU distribution across major AI players. Figures represent deployed capacity including H100, H200, B100, B200, and equivalent accelerators.
| Organization | Est. GPU Count | Primary Hardware | 2025 Blackwell Orders | Region |
|---|---|---|---|---|
|
G
Google
DeepMind, Gemini
|
1,000K – 1,500K | TPU v5, H100, B200 | ~400K | 🇺🇸 USA |
|
O
OpenAI
GPT, Stargate
|
~1,000K+ | H100 (Azure), GB200 | 400K+ | 🇺🇸 USA |
|
M
Microsoft
Azure AI
|
750K – 900K | H100, H200, GB200 | 700K – 1,400K | 🇺🇸 USA |
|
Meta
Llama, FAIR
|
~600K | H100, MI300X | 350K+ | 🇺🇸 USA |
|
A
Amazon
AWS, Trainium
|
250K – 400K | Trainium, H100, H200 | ~360K | 🇺🇸 USA |
|
X
xAI
Grok, Colossus
|
~200K | H100, H200 | 300K+ (H2 2025) | 🇺🇸 USA |
|
B
ByteDance
TikTok, Doubao
|
~230K | H800, H20, Ascend | Export Restricted | 🇨🇳 China |
|
T
Tencent
Hunyuan
|
~230K | H800, H20, Ascend | Export Restricted | 🇨🇳 China |
2025 marks the largest technology infrastructure investment cycle in history. Hyperscalers are deploying capital at unprecedented scale to secure AI compute dominance.
$500B AI Infrastructure Initiative (2025–2029)
AI is driving unprecedented energy demand. Data centers consumed 415 TWh globally in 2024—projected to reach 945 TWh by 2030 (3% of global electricity).
| Model | TDP | Purchase Price | Cloud $/hr | Status |
|---|---|---|---|---|
| H100 SXM | 700W | $25,000 – $40,000 | $2.10 – $5.00 | Available |
| H200 | 700W | $30,000 – $40,000 | $3.72 – $10.60 | Available |
| B100 | 700W | $30,000 – $35,000 | Limited | Shipping |
| B200 SXM | 1,000W | $30,000 – $50,000 | $6.25 – $9.00+ | Shipping |
| GB200 Superchip | 2,700W | $60,000 – $70,000 | N/A | Shipping |
| GB200 NVL72 Rack | 120 kW | ~$3,000,000 | N/A | Liquid Cooled |
Calculate annual electricity costs for your GPU cluster. Factor in utilization rates, regional electricity pricing, PUE, and GPU power consumption.
| Region | $/kWh | 100K GPU Cost/yr |
|---|---|---|
| Texas (US) | $0.090 | $252M |
| Virginia (US) | $0.106 | $297M |
| Nordic (EU) | $0.095 | $266M |
| Germany | $0.233 | $652M |
| China (Inland) | $0.070 | $196M |
| Solar PPA (US) | $0.057 | $160M |
Training compute for frontier AI models has grown ~4–5× per year since 2010. Over 30 models have now exceeded the 10²⁵ FLOP threshold (GPT-4 scale).
| Model | Organization | Training Compute | Est. Cost | Release |
|---|---|---|---|---|
| Grok-3 (Est.) | xAI | ~2.85×10²⁷ FLOPs | Billions | 2025 |
| Gemini Ultra | ~5×10²⁵ FLOPs | $30–191M | Dec 2023 | |
| Llama 3.1 405B | Meta | 3.8×10²⁵ FLOPs | ~$100M+ | Jul 2024 |
| GPT-4 | OpenAI | ~2×10²⁵ FLOPs | $100M+ | Mar 2023 |
| Claude 3.7 Sonnet | Anthropic | <10²⁶ FLOPs | Tens of millions | 2025 |
Critical dynamics shaping the AI infrastructure landscape in 2025 and beyond.
NVIDIA maintains 80–92% market share in AI accelerators. Blackwell is the fastest-ramping product in company history with $11B+ in Q4 FY2025 revenue.
Data center power demand projected to reach 945 TWh by 2030 (3% of global electricity). Grid investment of $720B needed through 2030.
US maintains ~3× cost advantage in training. China faces 4-year capability gap despite Huawei Ascend 910C reaching 60% of H100 performance.
Combined hyperscaler CapEx of $315–380B in 2025—50% more than entire US semiconductor industry revenue. Stargate alone represents $500B over 4 years.
US controls 54% of global hyperscale capacity (up from 51% in 2023). Northern Virginia alone hosts 15% of global capacity.
Human-generated text (~300T tokens) may be exhausted between 2026–2032. Synthetic data and test-time compute scaling emerging as alternatives.