Sales by segment relative to datacenters online this quarter.
Q2 Real-World Alignment: Power Footprint vs. NVIDIA Sales
| Customer Segment (NVIDIA Ledger) | Actual Q2 Power Additions (GW Deployed) | NVIDIA Q2 Compute Sales (GW Sold) | Net Physical Power Gap (The Deficit) |
|---|---|---|---|
| 1\. Hyperscalers (Self-Builds: MSFT, GOOG, AWS) | 0.90 GW | 1.43 GW | +0.53 GW (530 MW) |
| 2\. Tier-2/3 AI Clouds & Labs (Colocation: CoreWeave, xAI) | 0.45 GW | 0.97 GW | +0.52 GW (520 MW) |
| 3\. Sovereign AI Networks (State-Backed/Off-Grid) | 0.10 GW | 0.37 GW | +0.27 GW (these are often secret) |
| 4\. Fortune 500 Enterprise (On-Premises/Legacy Cages) | 0.15 GW | 0.18 GW | +0.03 GW (30 MW) |
| Total Global Footprint (Ex-China) | 1.60 GW | 2.95 GW | +1.35 GW (1,350 MW Overhang) |
Since the start of 2025, Nvidia has always sold way more AI datacenter GPUs than reported online datacenters. But it is reaching almost double oversold. The most vulnerable overbuyers are those tier2 “neoclouds” with well over double the purchases compared to deployments. These firms generally have debt costs over 10%, and lose money every quarter. Coreweave, especially locks in customers at massive structural losses that make its bankruptcy completely certain.
A big issue for Nvidia customers is all of the price hikes that have occurred this year. well over 50% including a 15% hike this week. They are retroactive to all undelivered sales.
2027 sales forecast to match, the now only 6.3gw of datacenter additions requires even more price increases.
Finally, Rubin is complete crap in training time/$ both relative to blackwell, but especially Huawei. Both Rubin and Blackwell are worse than H200 in bandwitdh/$ for inference throughput. Rubin 80% worse than blackell at inference. Huawei well under 50% cost/performance, largely as a result of being the greatest tech engineering force on the planet. Much of their cost advantage comes from not needing buildings. Outdoor sited containerized pods.
Even when considering Rubin’s new expensive performance network subsystem, compared to 8096 sized Huawei superpod at tflops equivalence for training a 1t parameter dense llm (about the same for 1t total experts active in a extremely large MOE) model
The 90-Day Amortized TCO Matrix
| Financial & Operational Metrics | NVIDIA GB300 Layout | NVIDIA Rubin Layout | Huawei 8,096 SuperPod |
|---|---|---|---|
| Fully Burdened Capital Layout | $260.5 Million (High RE) | $309.5 Million (Premium HW) | $150.0 Million (no building costs) |
| Required Training Days to Equalize Work | 90 Days | 83 Days | 128 Days |
| Localized 5-Year Energy Cost Rate | $80.00 / MWh (US Utility Grid) | $80.00 / MWh (US Utility Grid) | $22.00 / MWh (Stranded Chinese Hydro) |
| Total Power Cost Spent During the Run | $1.24 Million | $1.14 Million | $0.86 Million |
| Total Burdened Infrastructure Layout | $261.74 Million | $310.64 Million | $150.86 Million |
| True Cost per Training Day | $2.90 Million / Day | $3.74 Million / Day | $1.17 Million / Day |
Much larger clusters than this do have a networking advantage for Nvidia, but nowhere close enough to overcome the 3x cost, and superpod has 6x the vram per tflops, and so can train much larger MOE models than nvidia unless more hardware is thrown at it.
For inference, NPU/TPU/ASiCs from Nvidia’s main customers are far ahead, which means those relying on Nvidia for inference are at a massive cost disadvantage, when not only are hyperscaller chips more efficient, they don’t pay for Nvidia’s 74-75% gross margins.
The big takeaway/opportunity for countries facing trade or other aggression wars by the US, that are close to the US, are stupid to use US tech stack for training or inference, and in fact have great opportunity to sell inference at low latency to US customers, and break US AI bubble by deploying Huawei quicker, and making partnerships with Chinese labs for sovereign AI, to share public/customer’s open training data sets.