---
type: idea
title:
created: 2026-07-04
status: seed
origin_sources: []
related_auto: []
used_in: []
---

## Claim

## Why it matters

## Evidence / examples

## Counterpoints

## Related

## Draft hooks
- Possible headline:
- Possible angle:
- What this is really about:


Rough estimates: 
- ~$27MM/MW for deployed AI compute infrastructure, including GPUs, servers, networking, and some data-center/leasehold equipment; not pure GPU chip cost and not total data center build cost (from Coreweave IR)
- That is roughly $27BN / GW 
- Datacenter provider incurs capex worth $10-13MM per MW which is $10-13BN per GW
- Operating gosts are $0.075 per kWh, 85% utilization, with 1.4 PUE

- $50BN estimate for all-in cost on Sarah Frair's interview on all-in 

- ~$35B/GW for IT equipment and ~$25B/GW for power/cooling/infrastructure according to BG2

Esimates on non-GPU capex

| Source                             | What It Supports                                                                                             | Per GW         |
| ---------------------------------- | ------------------------------------------------------------------------------------------------------------ | -------------- |
| CoreWeave clipping                 | Shell, cooling, electrical equipment borne by data-center provider: `$10M-$13M/MW`                           | `$10B-$13B/GW` |
| JLL 2026 outlook                   | Global shell/core construction avg: `$10.7M/MW` in 2025, forecast `$11.3M/MW` in 2026                        | `~$11B/GW`     |
| Applied Digital project financings | AI factory construction financings imply roughly `$9M-$15M/MW`, depending project and included reserves/fees | `$9B-$15B/GW`  |
| Epoch AI model                     | 1GW AI data center = `$38B` upfront; servers are 60%, so non-server/facility bucket is ~40%                  | `~$15B/GW`     |
| BG2 transcript                     | Switchgear, generators, transformers, shell, getting power; land/shell/power/cooling                         | `$20B-$25B/GW` |

The Applied Digital evidence is useful because it is near-primary and AI-specific:

- `$2.35B` notes for `100MW + 150MW` facilities at Polaris Forge 1 = rough `$9.4M/MW`, though proceeds also cover repayment/reserves/transaction costs: [Applied Digital](https://ir.applieddigital.com/news-events/press-releases/detail/134/applied-digital-announces-proposed-offering-of-2-35)
- `$1.59B` notes for `150MW` ELN-04 = rough `$10.6M/MW`, also including bridge repayment/reserves/expenses: [Applied Digital](https://ir.applieddigital.com/news-events/press-releases/detail/155/applied-digital-announces-proposed-offering-of-1-59)
- Polaris Forge 2 described as a `$3B`, `200MW` campus = `$15M/MW`: [Applied Digital Q3 FY26](https://ir.applieddigital.com/news-events/press-releases/detail/148/applied-digital-reports-fiscal-third-quarter-2026-results)


Estimates on GPU Capex

| Source                      | Estimate                                            | Scope                                                          | Quality                          |
| --------------------------- | --------------------------------------------------- | -------------------------------------------------------------- | -------------------------------- |
| CoreWeave/Rittenhouse model | `$27B/GW`                                           | GPUs + auxiliary equipment like networking                     | Analyst-derived                  |
| CoreWeave filings-derived   | `~$24B-$27B/GW`                                     | Technology equipment; higher if adding DC equipment/leaseholds | Primary inputs, derived math     |
| NVIDIA earnings transcript  | Hopper: `$20B-$25B/GW`; Grace Blackwell: `~$30B/GW` | NVIDIA content per GW                                          | Best direct anchor               |
| Epoch AI model              | `~$25B/GW` implied server capex                     | GB200 NVL72 server systems                                     | Independent research model       |
| BG2 transcript              | `$35B/GW`                                           | GPUs + silicon for training/inference                          | Podcast/investor model, high-end |
| American Compute            | `$250k-$400k` per 8-GPU H100 SXM server             | Unit server cost                                               | Secondary sanity check           |

The **strongest source** is NVIDIA. On its Q3 FY2026 earnings call, Jensen says Hopper was roughly **`$20B-$25B/GW`**, while **Grace Blackwell is around `$30B/GW` plus or minus**: [NVIDIA transcript](https://s201.q4cdn.com/141608511/files/doc_financials/2026/q3/NVDA-Q3-2026-Earnings-Call-19-November-2025-5_00-PM-ET.pdf).

The CoreWeave/Rittenhouse model uses **`$27MM/MW` = `$27B/GW`** for “GPUs including auxiliary equipment such as networking etc.” [local clipping line 81 (line 81)](/Users/darshshah/Documents/Darsh's Vault/Darsh's Vault/clippings/26-07-03 substack.com CoreWeave Unit Economics, Margin Profile, and Leverage Analysis.md:81). The author says this is based on CoreWeave Sep. 2025 balance sheet and MW disclosures [line 182 (line 182)](/Users/darshshah/Documents/Darsh's Vault/Darsh's Vault/clippings/26-07-03 substack.com CoreWeave Unit Economics, Margin Profile, and Leverage Analysis.md:182).

That derived math is roughly:

`Technology equipment only: $14.629B / 590MW = ~$24.8MM/MW`

`Technology equipment + data center equipment/leasehold improvements: ($14.629B + $1.139B) / 590MW = ~$26.7MM/MW`

Inputs: CoreWeave Sep. 2025 10-Q PP&E table: [SEC filing](https://www.sec.gov/Archives/edgar/data/1769628/000176962825000062/crwv-20250930.htm), CoreWeave Q3 active power of `~590MW`: [CoreWeave IR](https://investors.coreweave.com/news/news-details/2025/CoreWeave-Reports-Strong-Third-Quarter-2025-Results/), and Dec. 2024 `>360MW`: [S-1](https://www.sec.gov/Archives/edgar/data/1769628/000119312525044231/d899798ds1.htm).