Stage 7 of 9
GPU-hours compute
Where the top-down company models live. The price of a GPU-hour is the hinge between the physical chain and the dollar chain.
- Unit
- GPU-hour
- Sells
- rented accelerator time under multi-year contracts (neoclouds) or used internally (hyperscalers)
- Buys from
- Data centre shell and cooling · GPUs, servers, racks and networking
- Lead time
- 1 year
- Status
- skeleton
Primer
What this stage is
Accelerator time, sold by the hour. Neoclouds rent GPU-hours under multi-year contracts; hyperscalers rent them and also use them for their own models. This is the stage where the physical chain becomes a dollar chain, and where the company models in this repository live.
How it works
The price of a GPU-hour has two markets. On-demand list prices are public and fall quickly as a generation ages. Contract prices are private, longer, and stickier, and most neocloud revenue is contracted before the GPUs are switched on. Utilisation, the share of installed hours that earn, is the number operators rarely disclose and the one that decides margins. Customers prepay part of a contract, and the GPUs are financed with loans secured on the contract itself.
The unit economics are simple to state: the hour must earn more than its energy, its facility, its running costs and the repayment of the GPU with interest, over the GPU's useful life.
What to watch
- On-demand prices by generation and provider, and their rate of decline.
- Contract terms: length, prepayment, take-or-pay.
- Capital spending guidance from the hyperscalers and what it implies in GPUs.
- Utilisation where disclosed; backlog and its conversion to revenue.
- The cost of debt secured on GPU contracts.
How it connects
A GPU-hour bought by a lab becomes tokens. Its price, against tokens per hour and price per token, decides how the dollar splits between the owner of the GPU and the owner of the model.
Figures
| Metric | Value | Unit | As of | Scope | Confidence | Source | Note |
|---|---|---|---|---|---|---|---|
| meta_capex_guidance_2026_low | 130 | USD bn (2026 capital expenditures incl. principal payments on finance leases, low end of range) | 2026-07-29 | Meta | high | Meta Platforms, Form 8-K Exhibit 99.1, 'Meta Reports Second Quarter 2026 Results' (29 Jul 2026) | Low end of Meta's full-year 2026 capex range of $130-145 billion (narrowed from $125-145 billion); Q2 2026 capex incl. finance-lease principal was $31.08 billion. |
| meta_capex_guidance_2026_high | 145 | USD bn (2026 capital expenditures incl. principal payments on finance leases, high end of range) | 2026-07-29 | Meta | high | Meta Platforms, Form 8-K Exhibit 99.1, 'Meta Reports Second Quarter 2026 Results' (29 Jul 2026) | High end of the same $130-145 billion range. |
| amazon_capex_guidance_2026 | 220 | USD bn (expected 2026 capital expenditures, stated by CEO on Q2 2026 call) | 2026-07-30 | Amazon | medium | Fortune, 'Andy Jassy said Amazon will spend $220 billion this year' (30 Jul 2026), reporting the Q2 2026 earnings call | Verbal guidance from the earnings call, raised from $200 billion because of memory prices; it does not appear in the 8-K press release or the 10-Q, which show $98.4 billion of purchases of property and equipment in the first half of 2026 and $169.0 billion trailing twelve months. Secondary source restating the CEO, hence medium. |
| alphabet_capex_guidance_2026_low | 195 | USD bn (full-year 2026 capital expenditures guidance, low end of range) | 2026-07-22 | Alphabet | medium | Investing.com transcript of Alphabet Q2 2026 earnings call (22 Jul 2026), CFO Anat Ashkenazi | Verbal guidance on the call (the 10-Q gives no range); raised from $180-190 billion because capacity is being delivered faster to meet demand. First-half 2026 capex was $80.6 billion per the 10-Q cash-flow statement. |
| alphabet_capex_guidance_2026_high | 205 | USD bn (full-year 2026 capital expenditures guidance, high end of range) | 2026-07-22 | Alphabet | medium | Investing.com transcript of Alphabet Q2 2026 earnings call (22 Jul 2026), CFO Anat Ashkenazi | High end of the same $195-205 billion range. |
| microsoft_capex_fy2026_incl_finance_leases | 175 | USD bn (fiscal year ended 30 Jun 2026 capital expenditures incl. finance leases, approximately) | 2026-07-29 | Microsoft | high | Microsoft Investor Relations, fiscal 2026 fourth-quarter earnings call transcript (29 Jul 2026), CFO Amy Hood | Microsoft's fiscal year runs to June, so this is the closest thing to a '2026 capex' figure; the 8-K cash-flow statement shows $115.9 billion of additions to property and equipment before finance leases. Hood added: 'We expect FY27 capital expenditures will grow year-over-year given demand signals across our portfolio.' |
| microsoft_capex_q1_fy2027_guidance | 50 | USD bn (expected capital expenditures incl. finance leases, quarter ending 30 Sep 2026, 'over') | 2026-07-29 | Microsoft | high | Microsoft Investor Relations, fiscal 2026 fourth-quarter earnings call transcript (29 Jul 2026), CFO Amy Hood | Microsoft guides capex one quarter at a time; Q4 FY26 capex was $41 billion plus $5.6 billion of finance leases, about two thirds of it short-lived assets (CPUs and GPUs). Hood said FY27 capex would grow on 'demand signals across our portfolio' without giving a full-year number. |
| oracle_capex_guidance_fy2027_low | 90 | USD bn (fiscal 2027 capital expenditures, low end of range; net cash capex not more than $70 bn) | 2026-09-10 | Oracle | medium | The Motley Fool transcript of Oracle Q1 fiscal 2027 earnings call (10 Sep 2026), CFO Hilary Maxson | Guidance given verbally on the call, not in the 10-Q; gross capex $90-95 billion for the fiscal year ending May 2027, with net cash outlay capped at $70 billion because customers or partners fund part of the build. Q1 FY27 capex alone was $28.5 billion (10-Q cash-flow statement). |
| oracle_capex_guidance_fy2027_high | 95 | USD bn (fiscal 2027 capital expenditures, high end of range) | 2026-09-10 | Oracle | medium | The Motley Fool transcript of Oracle Q1 fiscal 2027 earnings call (10 Sep 2026), CFO Hilary Maxson | High end of the same $90-95 billion range. |
| oracle_remaining_performance_obligations | 664 | USD bn (total RPO) | 2026-08-31 | Oracle | high | Oracle Corporation, Form 10-Q for the quarter ended August 31, 2026 (Q1 FY2027) | Contracted revenue not yet recognised; only about 13% is expected to convert to revenue in the next twelve months, so most of it depends on data-centre capacity still to be built. |
| microsoft_commercial_rpo | 678 | USD bn (commercial remaining performance obligation) | 2026-06-30 | Microsoft | high | Microsoft Corporation, Form 8-K Exhibit 99.1, fiscal 2026 fourth-quarter earnings release (29 Jul 2026) | Commercial contracted revenue not yet recognised at fiscal year-end; up 84% year on year and includes the incremental $250 billion Azure commitment from OpenAI announced October 2025. |
| alphabet_remaining_performance_obligations | 519.5 | USD bn (remaining performance obligations, 'revenue backlog', of which $513.9 bn Google Cloud) | 2026-06-30 | Alphabet | high | Alphabet Inc., Form 10-Q for the quarter ended June 30, 2026 | Almost all of Alphabet's backlog is Google Cloud; with Oracle's $664 bn and Microsoft's $678 bn commercial RPO, the three largest disclosed cloud backlogs total about $1.86 trillion. |
| coreweave_revenue_backlog | 104 | USD bn (revenue backlog) | 2026-06-30 | CoreWeave | high | CoreWeave, Form 8-K Exhibit 99.1, 'CoreWeave Reports Strong Second Quarter 2026 Results' (11 Aug 2026) | Contracted revenue to be delivered against 1.5 GW of active power and about 3.7 GW of contracted power at the same date; the release does not split the backlog by customer. |
| h100_on_demand_median_price | 3.25 | USD per GPU-hour (on-demand, median of publicly listed prices) | 2026-09 | global (75 providers) | medium | AIMultiple, 'Cloud GPU Rental Price Index' (updated 22 Sep 2026) | Median of publicly listed on-demand prices, so it overstates what large committed buyers pay; Silicon Data's neocloud-only H100 index was $2.72 on 26 Sep 2026 and Lambda's public 8-GPU H100 SXM on-demand price was $3.99. |
| h100_on_demand_median_price_getdeploying | 3.39 | USD per GPU-hour (on-demand, median across providers) | 2026-09-27 | global (43 providers) | medium | GetDeploying, 'H100 Cloud Pricing: Compare 53+ Providers (2026)' (27 Sep 2026) | Second tracker as a cross-check on AIMultiple's $3.25; the page says the median rose about 12% over the past 12 months, the only year-ago comparison found, and lists a range from $1.69 (Lium) to $11.07 (Google Cloud). |
| h100_neocloud_index_price | 2.72 | USD per GPU-hour (Silicon Data H100 Neo-Cloud index, standardised like-for-like rate) | 2026-09-26 | global (neocloud providers) | medium | Silicon Data, 'H100 Rental Price Index' (26 Sep 2026) | Index of observed rental rates standardised for term length, cluster scale and interconnect, so lower than the raw on-demand list-price median; the page shows only a 7-day change (+2.6%), no year-ago value. The quote is assembled from adjacent page elements. |
| h200_on_demand_median_price | 4.4 | USD per GPU-hour (on-demand, median of publicly listed prices) | 2026-09 | global (75 providers) | medium | AIMultiple, 'Cloud GPU Rental Price Index' (updated 22 Sep 2026) | Listings ranged from $2.09 per GPU-hour (Beam) to $13.78 (Microsoft Azure) in September 2026, so the median hides a wide spread between neoclouds and hyperscalers. |
| b200_on_demand_median_price | 6.52 | USD per GPU-hour (on-demand, median of publicly listed prices) | 2026-09 | global (75 providers) | medium | AIMultiple, 'Cloud GPU Rental Price Index' (updated 22 Sep 2026) | B300 median was $7.87 in the same month; Lambda's public 8-GPU B200 on-demand price was $6.69, close to the median. |
| lambda_h100_sxm_on_demand_list_price | 3.99 | USD per GPU-hour (public on-demand list price, 8x H100 SXM instance) | 2026-09-27 | Lambda | high | Lambda, 'GPU Cloud pricing' page (viewed 27 Sep 2026) | A single published price list from a large neocloud; the same page lists the 8x B200 SXM6 instance at $6.69 per GPU-hour and says reserved (contract) capacity is priced on request, so contract prices are not public. Quote is a table row. |
| oracle_gpu_utilization | 97.9 | percent (GPU utilisation, OCI fleet, quarter ended 31 Aug 2026) | 2026-08-31 | Oracle | medium | The Motley Fool transcript of Oracle Q1 fiscal 2027 earnings call (10 Sep 2026), CEO Clay Magouyrk | The only large provider to state a fleet utilisation number; the definition (time-based vs. capacity-weighted) is not given, and the transcript text has odd capitalisation. |
| oracle_datacenter_capacity_delivered_quarter | 850 | MW (additional data-centre capacity delivered in the quarter ended 31 Aug 2026) | 2026-08-31 | Oracle | high | Oracle Corporation, Form 8-K Exhibit 99.1, 'Oracle Announces Fiscal 2027 First Quarter Financial Results' (10 Sep 2026) | A quarterly delivery figure, i.e. the pace at which backlog turns into rentable capacity; on the call management said Q1 delivery was almost three times all of Q4. |
| nebius_contracted_power_target_2026 | 5 | GW (contracted power expected at end of 2026) | 2026-08-12 | Nebius | high | Nebius Group N.V., Form 6-K Exhibit 99.2, Letter to shareholders Q2 2026 (12 Aug 2026) | Contracted power (sites with signed power), not operating GPUs; the letter says Nebius plans to deploy more than 1 GW per year starting in 2027 and could sell its entire 2027 capacity today. |
| nebius_customer_commitments | 40 | USD bn (customer commitments / committed backlog) | 2026-08-12 | Nebius | high | Nebius Group N.V., Form 6-K Exhibit 99.2, Letter to shareholders Q2 2026 (12 Aug 2026) | Nebius' term for contracted future revenue, comparable to CoreWeave's $104 billion 'revenue backlog'; the letter says four recent deals averaged more than $1 billion of total contract value each and that 70% of deals closed included customer prepayments. |
| openai_nvidia_agreement_capacity | 10 | GW (at least; NVIDIA systems for OpenAI under a letter of intent) | 2025-09-22 | OpenAI / NVIDIA | high | NVIDIA Newsroom, 'OpenAI and NVIDIA Announce Strategic Partnership to Deploy 10 Gigawatts of NVIDIA Systems' (22 Sep 2025) | Letter of intent, not a definitive contract; NVIDIA intends to invest up to $100 billion in OpenAI progressively as each gigawatt is deployed, with the first gigawatt on Vera Rubin in the second half of 2026. No dollar value or end date for the systems purchases was stated. |
| openai_amd_agreement_capacity | 6 | GW (AMD Instinct GPUs across multiple generations) | 2025-10-06 | OpenAI / AMD | high | AMD, Form 8-K Exhibit 99.1, 'AMD and OpenAI Announce Strategic Partnership to Deploy 6 Gigawatts of AMD GPUs' (6 Oct 2025) | First 1 GW of MI450 GPUs to begin in the second half of 2026; AMD expects 'tens of billions of dollars' of revenue and issued OpenAI a warrant for up to 160 million AMD shares vesting on milestones. No total dollar value or end date stated. |
| openai_broadcom_agreement_capacity | 10 | GW (OpenAI-designed custom AI accelerators built with Broadcom) | 2025-10-13 | OpenAI / Broadcom | high | Broadcom / OpenAI joint press release via GlobeNewswire, 'OpenAI and Broadcom announce strategic collaboration to deploy 10 gigawatts of OpenAI-designed AI accelerators' (13 Oct 2025) | Deployment targeted to start in the second half of 2026 and complete by end of 2029; no dollar value was stated by either party. |
| openai_microsoft_azure_commitment | 250 | USD bn (incremental Azure services contracted by OpenAI) | 2025-10-28 | OpenAI / Microsoft | high | Microsoft Official Blog, 'The next chapter of the Microsoft-OpenAI partnership' (28 Oct 2025) | No duration was stated for the $250B purchase; the same post extends Microsoft's IP rights for models and products through 2032. Sits inside Microsoft's $678 billion commercial RPO. |
| openai_coreweave_contract_value | 22.4 | USD bn (total contract value, three agreements March-September 2025) | 2025-09-25 | OpenAI / CoreWeave | high | CoreWeave investor relations, 'CoreWeave Expands Agreement with OpenAI by up to $6.5B' (25 Sep 2025) | Sum of an initial up-to-$11.9 billion agreement (March 2025), a $4 billion expansion (May 2025) and a $6.5 billion expansion (September 2025); duration not stated in the release. |
| crusoe_contracted_capacity | 4.9 | GW (contracted AI infrastructure, data centres plus Crusoe Cloud) | 2026-06-09 | Crusoe | high | Crusoe newsroom, 'Crusoe's Contracted AI Infrastructure Capacity Approaches 5 Gigawatts Across Data Centers and Cloud' (9 Jun 2026) | Contracted, not operating: mixes data-centre leases (the 1.2 GW Abilene campus built for Oracle and a 900 MW Abilene campus for Microsoft) with Crusoe Cloud GPU capacity; Crusoe says its total pipeline exceeds 40 GW. |
| applied_digital_contracted_it_load | 1,410 | MW of contracted critical IT load (executed long-term leases) | 2026-05-31 | Applied Digital | high | Applied Digital Corporation, Form 10-K for the fiscal year ended May 31, 2026 | Approximately 1,410 MW across five campuses under roughly 15-year leases, including 400 MW to CoreWeave at Polaris Forge 1 (about $11.0 billion) and 200 MW to an unnamed investment-grade hyperscaler at Polaris Forge 2. |
| applied_digital_contracted_lease_revenue | 36.2 | USD bn (total contracted revenue over initial 15-year base lease terms) | 2026-05-31 | Applied Digital | high | Applied Digital Corporation, Form 10-K for the fiscal year ended May 31, 2026 | Arithmetic: $36.2 billion / 1,410 MW / 15 years = about $1.71 million per MW per year of lease revenue, a floor for what a landlord earns versus what a GPU renter earns on the same megawatt. |
| anthropic_fluidstack_us_datacenter_investment | 50 | USD bn (Anthropic investment in US data centres built with Fluidstack) | 2025-11-12 | Fluidstack / Anthropic | high | Fluidstack newsroom, 'Fluidstack selected by Anthropic to deliver custom data centers in New York and Texas' (12 Nov 2025) | Dollar figure only; neither Fluidstack nor Anthropic stated the gigawatts, and the money is Anthropic's capital rather than a Fluidstack backlog in the CoreWeave sense. |
| together_ai_humain_datacenter_capacity | 250 | MW (new AI data centre in Saudi Arabia under Together AI-HUMAIN partnership) | 2026-08-31 | Together AI / HUMAIN | medium | Business Wire press release syndicated on Yahoo Finance, 'Together AI and HUMAIN Form Strategic Partnership to Accelerate AI Infrastructure and Cloud Services in Saudi Arabia' (31 Aug 2026) | The release also says the partnership is expected to yield over $5 billion in gross annualised revenue in its first year; arithmetic: $5 billion / 250 MW = $20 million per MW-year, matching the $20-25 million per MW Nebius cited for its recent deals. |
Conversions
| From → To | Factor | Unit | As of | Confidence | Source | Note |
|---|---|---|---|---|---|---|
| GPUs, servers, racks and networking → GPU-hours | 56,000 | USD per GPU deployed in an 8-GPU HGX B200 server (system-level price, all components) | 2026-07-16 | low | Mercatus AI blog, 'NVIDIA B200 Server Price in 2026: What an 8-GPU HGX System Costs' (16 Jul 2026) | Arithmetic check: a typical $450,000 server / 8 GPUs = $56,250 per GPU, versus a bare B200 card list price of $30,000-40,000, so the server wrapper (CPUs, memory, NICs, storage, chassis, integrator margin) adds roughly 40-90% on top of the GPU; the $2.8-3.4 million GB200 NVL72 estimate implies $38,900-47,200 per GPU at rack level. Reseller compilation, not a filing. |
| GPU-hours → Data centre shell and cooling | 1.67 | kW of rack IT load per GPU (GB200 NVL72: 120 kW rack / 72 GPUs) | 2024-10-15 | high | NVIDIA Technical Blog, 'NVIDIA Contributes NVIDIA GB200 NVL72 Designs to Open Compute Project' (15 Oct 2024) | Arithmetic: 120 kW / 72 GPUs = 1.667 kW per GPU at the rack (GPU, Grace CPU, NVLink switches, in-rack cooling). Excludes scale-out network, storage and facility overhead (PUE), which the datacenter stage adds. |
| GPU-hours → Data centre shell and cooling | 1.27 | kW of system power per GPU (DGX H100: 10.2 kW max / 8 GPUs) | 2026-09-27 | high | NVIDIA DGX H100/H200 User Guide, 'Introduction to the NVIDIA DGX H100/H200 Systems', hardware specifications | Arithmetic: 10.2 kW / 8 GPUs = 1.275 kW per GPU for an air-cooled H100 server at maximum system power ('10.2 kW max.' on the same spec table; H100 SXM TDP is up to 700 W, so the server roughly doubles the GPU's own draw). Excludes network switches and facility overhead. |
| Data centre shell and cooling → GPU-hours | 5,256,000 | GPU-hours per MW-year of rack IT load (GB200 density, 100% uptime) | 2024-10-15 | medium | NVIDIA Technical Blog, 'NVIDIA Contributes NVIDIA GB200 NVL72 Designs to Open Compute Project' (15 Oct 2024); arithmetic by this note | Arithmetic: 1,000 kW / 1.667 kW per GPU = 600 GPUs per MW of rack IT load; 600 x 8,760 hours = 5,256,000 GPU-hours per MW-year if every GPU is rentable every hour. Sellable hours are lower by utilisation (Oracle reported 97.9%) and by facility overhead: at PUE 1.2 the figure per MW of facility power is 4,380,000. At H100 density (1.275 kW per GPU) it is 784 GPUs and 6,870,588 GPU-hours per MW-year. |
| GPU-hours → Data centre shell and cooling | 34.27 | USD m of revenue per MW-year of rack IT load at $6.52 per GPU-hour (B200 on-demand median, GB200 density, 100% utilisation) | 2026-09 | medium | AIMultiple, 'Cloud GPU Rental Price Index' (Sep 2026) for the price; NVIDIA GB200 NVL72 blog for the density; arithmetic by this note | Arithmetic: 5,256,000 GPU-hours per MW-year x $6.52 = $34.27 million per MW-year, a ceiling using on-demand list prices at full utilisation. At H100 density and the $3.25 H100 median: 6,870,588 x $3.25 = $22.33 million. Contract-level cross-checks are lower: Nebius cites $20-25 million per MW of annual contract value and Together AI/HUMAIN expect over $5 billion on 250 MW (= $20 million per MW-year). |
| GPU-hours → Data centre shell and cooling | 20 | USD m of annual contract value per MW (low end of Nebius' $20-25 million range on recent large deals) | 2026-08-12 | high | Nebius Group N.V., Form 6-K Exhibit 99.2, Letter to shareholders Q2 2026 (12 Aug 2026) | A company-stated contract yield per megawatt for multi-year GPU capacity deals, the best public anchor for 'revenue per MW-year on contract'. Implied contract price: $20-25 million / 5,256,000 GPU-hours (GB200 density) = $3.81-4.76 per GPU-hour, below the $6.52 B200 on-demand median. High end 25. |
| GPU-hours → Data centre shell and cooling | 1.71 | USD m of lease revenue per MW-year paid by a GPU cloud to its data-centre landlord (Applied Digital portfolio) | 2026-05-31 | medium | Applied Digital Corporation, Form 10-K for the fiscal year ended May 31, 2026; arithmetic by this note | Arithmetic: $36.2 billion / 1,410 MW / 15 years = $1.71 million per MW-year. The landlord's take is under a tenth of the $20-25 million per MW-year a GPU cloud books on the same megawatt, which shows where the margin (and the GPU capex) sits. |
| GPU-hours → Tokens | 10,000 | output tokens per second per GPU (Llama 3.3 70B, NVIDIA B200, NVFP4, at 50 tokens/s per user, InferenceMAX 1K/1K) | 2025-10-13 | medium | NVIDIA Technical Blog, 'NVIDIA Blackwell Leads on New SemiAnalysis InferenceMAX Benchmarks' (13 Oct 2025) | Vendor-run result on SemiAnalysis's open InferenceMAX suite for a 70B dense model at moderate interactivity: 10,000 tokens/s per GPU = 36 million tokens per GPU-hour (10,000 x 3,600); NVIDIA says this is more than 4x the H200, implying roughly 2,500 tokens/s per H200. Throughput falls steeply as per-user speed rises (for DeepSeek-R1 on B200 SemiAnalysis shows about 4,800 tokens/s/GPU at 42 tokens/s per user but about 500 at 100), so any single figure must carry its interactivity target. |
| GPU-hours → Tokens | 3,750 | output tokens per second per GPU (DeepSeek-R1 671B, FP4, 8x Blackwell B200 DGX, max-throughput mode) | 2025-03-18 | medium | NVIDIA Technical Blog, 'NVIDIA Blackwell Delivers World-Record DeepSeek-R1 Inference Performance' (18 Mar 2025) | Arithmetic: 30,000 tokens/s per 8-GPU DGX B200 / 8 GPUs = 3,750 tokens/s per GPU, or 13.5 million tokens per GPU-hour, for a 671B-parameter mixture-of-experts reasoning model at low interactivity; the same system serves over 250 tokens/s to one user when tuned for latency instead. Vendor benchmark at FP4, so an upper bound for production serving. |
| GPU-hours → Tokens | 0.19 | USD of rented GPU time per 1M output tokens (B200 at Lambda on-demand $6.69/GPU-hr, Llama 3.3 70B at 10,000 tokens/s per GPU) | 2026-09-27 | medium | Lambda, 'GPU cloud pricing: rent NVIDIA H100, H200, and B200' (on-demand instances table, as fetched 27 Sep 2026) | Arithmetic: 10,000 tokens/s x 3,600 s = 36,000,000 tokens per GPU-hour; $6.69 / 36 = $0.186 per 1M output tokens for GPU rental alone, before margin, idle time, input-token processing and the lower utilisation of real traffic. Against list prices of $3.96 (DeepSeek), $6 (Grok 4.7), $12 (Gemini 3.1 Pro) and $50 (gpt-6-astra, Claude Fable 5.1) per 1M output tokens, the raw GPU cost is a small fraction of the price of a frontier closed model but a large fraction of an open-weight one. Quote is the table text; the header column is PRICE/GPU/HR. |
| GPU-hours → Tokens | 0.1 | USD per 1M output tokens (DeepSeek-R1 on GB200 NVL72 at about 75 tokens/s per user; NVIDIA's own cost model) | 2025-10-13 | medium | NVIDIA Technical Blog, 'NVIDIA Blackwell Leads on New SemiAnalysis InferenceMAX Benchmarks' (13 Oct 2025) | NVIDIA's stated cost per million tokens using InferenceMAX throughput and its own GPU-hour cost assumptions (not disclosed on the page): $1.56 on H200 versus just over $0.10 on GB200 NVL72 for DeepSeek-R1 at about 75 tokens/s per user, a 15x gap; for gpt-oss-120b at 100 tokens/s per user it fell from $0.11 to $0.02 in two months of software updates. Shows how fast serving cost moves with hardware generation and software, independent of list prices. |
| Tokens → GPU-hours | 6 | FLOP of training compute per parameter per training token (the 6ND rule) | 2026-09-27 | high | Epoch AI, 'Data on AI models: documentation, estimation methods' (page as fetched 27 Sep 2026) | Standard heuristic for dense transformers: training FLOP = 6 x parameters x tokens (2 FLOP per parameter per token for the forward pass, about twice that for the backward pass). Example: Llama 3 405B x 15.6 trillion tokens x 6 = 3.79e25 FLOP, matching Meta's reported 3.8e25. For mixture-of-experts models use active rather than total parameters (Qwen3.8-Max: 95B active of 2.4T total). |
| GPU-hours → Tokens | 30,840,000 | H100 GPU-hours to train one frontier open model (Llama 3.1 405B, pretraining plus fine-tuning; Meta disclosure) | 2024-07-23 | high | Meta, 'Llama 3.1 Model Card' (llama-models repository, July 2024), training-time table | Meta's table gives 30.84M GPU-hours for the 405B model, 7.0M for 70B and 1.46M for 8B, 39.3M in total, on H100-80GB at 700 W TDP; at Lambda's on-demand $3.99 per H100-hour that is about $123 million of rented GPU time for the 405B run (30.84e6 x 3.99), versus Epoch's $50 million amortized-hardware estimate for the same run. Epoch estimates Grok 4 (2025) at 246 million H100-hours, eight times more. |
| GPU-hours → Tokens | 1,230,000,000,000,000,000 | FLOP of training compute realised per H100 GPU-hour (Llama 3 405B: 3.8e25 FLOP over 30.84M GPU-hours) | 2024-07-31 | high | Meta, 'The Llama 3 Herd of Models' (arXiv 2407.21783, 31 Jul 2024), section 1 and section 3.3 | Arithmetic: 3.8e25 FLOP / 30.84e6 GPU-hours = 1.23e18 FLOP per GPU-hour = 3.4e14 FLOP/s sustained, about 35% of an H100's 9.89e14 FLOP/s dense BF16 peak (model FLOP utilisation), on up to 16,000 H100s. Epoch's Grok 4 inputs (5.0e26 FLOP over 246M H100-hours) imply 2.0e18 per GPU-hour. Use this to turn a GPU-hour budget into a training-compute budget or vice versa. |
Players
| Company | Ticker | Role | Listed | Note |
|---|---|---|---|---|
| Microsoft | NASDAQ:MSFT | hyperscaler (Azure); OpenAI's primary cloud | yes | Commercial RPO of $678 billion at June 30, 2026 (up 84%) with Azure revenue growing 43%; buys overflow capacity from CoreWeave, Nebius, IREN and Lambda. |
| Alphabet | NASDAQ:GOOGL | hyperscaler (Google Cloud, TPUs) | yes | Google Cloud revenue backlog of $513.9 billion at June 30, 2026 and 2026 capex guided to $195-205 billion on the Q2 call; sells TPU capacity to Anthropic via Fluidstack sites. |
| Amazon | NASDAQ:AMZN | hyperscaler (AWS, Trainium) | yes | AWS sales grew 37% to $42.2 billion in Q2 2026; CEO guided 2026 capex to about $220 billion and said Amazon will not have enough capacity to meet demand in 2026 or 2027. |
| Meta | NASDAQ:META | hyperscale self-builder (no external GPU sales) | yes | 2026 capex guided to $130-145 billion; consumes compute for its own models and ads, so it competes for GPUs and power without selling GPU-hours. |
| Oracle | NYSE:ORCL | hyperscaler (OCI); largest OpenAI/Stargate landlord-operator | yes | RPO of $664 billion at August 31, 2026 with only about 13% converting in twelve months; delivered 850 MW in one quarter and reported 97.9% GPU utilisation. |
| CoreWeave | NASDAQ:CRWV | largest independent neocloud | yes | Revenue backlog about $104 billion at June 30, 2026 against 1.5 GW active and about 3.7 GW contracted power; OpenAI contracts total about $22.4 billion. |
| Nebius Group | NASDAQ:NBIS | neocloud (Europe and US) | yes | Expects 5 GW of contracted power by end-2026, cites $40 billion of customer commitments and $20-25 million per MW of annual contract value on recent deals; Q2 2026 revenue $582.3 million, up 454%. |
| Crusoe | private | neocloud and AI data-centre developer (Stargate Abilene) | no | 4.9 GW contracted across data centres and Crusoe Cloud as of June 2026, including the 1.2 GW Abilene campus for Oracle/OpenAI and a 900 MW Abilene campus for Microsoft. |
| Lambda | private | neocloud with public price list | no | Multibillion-dollar multi-year agreement with Microsoft (Nov 2025) for tens of thousands of NVIDIA GPUs incl. GB300 NVL72; value not disclosed. Publishes on-demand prices ($3.99 H100 SXM, $6.69 B200 per GPU-hour). |
| IREN | NASDAQ:IREN | bitcoin miner turned AI cloud (owns power and sites) | yes | $9.7 billion five-year GB300 contract with Microsoft (20% prepaid) on 200 MW of IT load at Childress, Texas, about $1.94 billion a year; targets $4 billion contracted ARR for 2026 capacity. |
| Applied Digital | NASDAQ:APLD | AI data-centre landlord to neoclouds and hyperscalers | yes | 1,410 MW of contracted IT load and $36.2 billion of 15-year lease revenue at May 31, 2026, including 400 MW leased to CoreWeave; shows the landlord layer's economics (about $1.7 million per MW-year). |
| Fluidstack | private | neocloud and build partner for labs (Anthropic, Google TPUs) | no | Selected by Anthropic for a $50 billion US data-centre build in Texas and New York with sites coming online through 2026; first TPU capacity reported live at Lake Mariner, NY in July 2026. |
| Together AI | private | inference-focused neocloud (open models) | no | Partnership with HUMAIN (Aug 2026) for a new 250 MW data centre in Saudi Arabia expected to yield over $5 billion of gross annualised revenue in its first year; raised $800 million at $8.3 billion in July 2026. |
| xAI | private | lab self-building compute (Colossus, Memphis) | no | Colossus campus expanding to nearly 2 GW with more than 555,000 NVIDIA GPUs (about $18 billion) per press reports; a large buyer of GPUs and power that sells no GPU-hours. |
Signals
| Date | Kind | Actor | Counterparty | Claim | Value | Confidence | Source |
|---|---|---|---|---|---|---|---|
| 2026-08-11 | contract | CoreWeave | customers | More than $25 billion of net new customer commitments added in early Q3 2026, on top of the $104 billion backlog at June 30 | 25 USD bn | confirmed | CoreWeave Q2 2026 earnings release (8-K exhibit 99.1, 2026-08-11) |
| 2026-03-31 | financing | CoreWeave | lenders | DDTL 4.0 facility with $8.5 billion of borrowing capacity at SOFR plus 2.25%, non-recourse to the parent, entered March 2026 | 8.5 USD bn | confirmed | CoreWeave 10-Q for the quarter ended June 30, 2026 (note 10) |
| 2025-09-08 | contract | Nebius | Microsoft | Five-year agreement to supply Microsoft with dedicated GPU capacity in tranches from a new data centre in Vineland, New Jersey; total contract value about $17.4 billion through 2031, with options that would increase it | 17.4 USD bn | confirmed | Nebius Form 6-K, 2025-09-08 |
| 2025-03-31 | contract | CoreWeave | OpenAI | Master services agreement under which OpenAI committed payments of up to approximately $11.9 billion | 11.9 USD bn | confirmed | CoreWeave final prospectus (424B4, 2025-03-31), pp. 33 and 158 |