Megawatts, not compute
LinkedIn, 23 October 2026 (activity timestamp 18:40 CEST).
Status labels sit next to the section they apply to. Tender clauses: CONFIRMED (two documents read in full). Deadline move to 3 Dec: CONFIRMED. Lot 1 move to 4 Feb 2027: SINGLE-SOURCE (EuroHPC page). Power deals: CONFIRMED each. Nebius and CoreWeave figures: CONFIRMED. Queue figures: SINGLE-SOURCE per country. OVHcloud break-even: SINGLE-SOURCE. GB300 rack cost: SINGLE-TRANSACTION plus two analyst estimates. Utilisation: ASSUMPTION.
The author is a venture partner at DEPO Ventures, a Czech seed fund, and holds no position in any company named on this page.
Core claim
Text as posted. Corrections are in the log at the bottom, not in this text.
Want hyperscalers in the EU? Stop buying compute, start selling megawatts! AI data center is great business. It pays for itself in about two years. So why EU officials scream for AI and cannot get anything? Why Europe's tender for seven "AI gigafactories" moved its deadline twice this month, to 3 December and, for the four smaller sites, to February 2027. Because the EU and the host state offer to each site that they buy computing time worth about 40 million euro a year for five years, capped at 17 percent of the hardware cost. Looks like a good offer? Not at all - on a 900 million euro site that is about 5 percent of capacity, peanuts. And who buys the other 95 percent? - Not European companies: they buy AI through Azure, AWS and Google, so their demand lands in those books. - Not European labs: the largest, Mistral, targets 200 megawatts by the end of 2027, about one and a half of these sites. The buyers of 100-megawatt blocks are Microsoft and Anthropic (surprised yes?). The tender itself exempts hyperscaler anchor customers from its own sovereignty rules. So let's see where the motivations sits... if you are EU datacenter business with a hyperscaler contract you do not need the tender. Without one, the tender does not save you - you don't have enough customers. What the hyperscalers are short of is not money. It is MEGAwatts availability and speed. NOT ENERGY PRICE. Microsoft, Amazon, Meta and Google each signed 17 to 25 year nuclear power contracts in the US; in Europe, one, Google at a Finnish plant, with 13 billion euro of investment announced the same day. I modeled the numbers for one site. If i double the electricity price in model, the return falls (only) three points. But if i add waiting one year for the grid connection with the chips bought - it falls eighteen! The scarce input is not cheap power. It is megawatts on a date. But let's check how it will go, I am right if fewer than seven gigafactory contracts are signed by 31 December 2027, and the first site to pass technical validation has a non-EU anchor customer for more than half its capacity. If all seven sign and the first one runs on European demand - I am wrong. Position (not political), and I can be wrong: Europe is negotiating a 5 percent anchor for providers... while the tenants it wants sign 20-year power contracts elsewhere. The policy that brings them is not a compute purchase. It is: - a reactor life extension - a really fast (not "EU fast") grid connection with a date on it - a government willing to sign for the megawatts. ANd In my opinion this is why the governments should exist - to bring this kind of strategic stuff, not buy/sell the things they have no clue about.
Panel
Text description of this panel
Horizontal bar chart on one basis, megawatts of electricity under a signed long-term purchase contract. Title: Megawatts booked for AI. Europe's tender books zero. Five rows with dark bars: Amazon, Susquehanna, Pennsylvania, 17 years to 2042, up to 1,920 MW; Meta, Clinton, Illinois, 20 years from 2027, 1,121 MW; Microsoft, Three Mile Island, Pennsylvania, 20 years, restart 2027, 835 MW; Google, Duane Arnold, Iowa, 25 years from 2029, about 615 MW; "Also Google, but EU", Loviisa, Finland, half the plant 2030 to 2049, 507 MW. Sixth row in red, below a hairline: EU AI gigafactory tender, seven sites, buys computing hours, not power, about 5 percent of one site, five years, 0 MW of power. Footer gives the deal dates and parties, the tender reference and the clause that building, utilities and operations are covered by the consortium, and the derivation of the 5 percent (EUR 40m a year against a 54,000 H100-equivalent site, author's model, inputs published).
1. What the post claims and what settles each claim
| Claim in the post | Settled by | Status |
|---|---|---|
| A full AI data centre pays back in about two years | Nebius Q2 2026 call, 19 Aug 2026: "a yield of $20 million to $25 million per megawatt", "less than 2-year payback period" | CONFIRMED (transcript, two carriers) |
| Deadline moved twice this month, to 3 Dec and, for the four smaller sites, 4 Feb 2027 | EuroHPC call page (modified 22 Oct 2026); first move public by 11 Oct (zakazka.cz) | 3 Dec CONFIRMED; 4 Feb SINGLE-SOURCE |
| EU and host state buy computing time worth about 40 million euro a year for five years, capped at 17 percent of the hardware cost | Tender Specifications §4 and §3.d | CONFIRMED |
| About 5 percent of a 900 million euro site | Model: 54,000 H100-eq site, EUR 895m capex, public EUR 40m a year at bid price EUR 1.80 per H100-eq-hour = 4.7 percent of capacity | DERIVED |
| European companies buy AI through Azure, AWS and Google | Mechanism statement; enterprise GPU utilisation near 5 percent (one study, one carrier) | SINGLE-SOURCE, printed as mechanism |
| Mistral targets 200 MW by end-2027 | DCD, 30 Mar 2026, company statement | CONFIRMED |
| Microsoft and Anthropic are the buyers of 100 MW blocks | Nscale S-1 (Sep 2026); IREN and Microsoft; Nebius with Microsoft and Meta | CONFIRMED |
| Tender exempts hyperscaler anchor customers from its sovereignty rules | Spec §3.g(a) and §3.b | CONFIRMED (verbatim) |
| Four US nuclear contracts, 17 to 25 years; one in Europe with 13 billion euro | Section 3 table | CONFIRMED each |
| Electricity doubled: minus three points; one-year grid wait with chips bought: minus eighteen | Model, section 4 | DERIVED |
2. The tender in its own words
CONFIRMED (Tender Specifications V1 and model framework contract)
- Phase 1 firm / phase 2 indicative (§4).
- 17 percent of IT CAPEX, and the IT CAPEX definition (§4, Annex 4).
- Consortium covers capex and opex (§3.e).
- Public role "limited to act as a strategic anchor customer" (§3.e).
- Anchor-customer exemption (§3.g(a)): anchor customers "may maintain operational autonomy ... and such specific services are not subject to the overarching AIGF sovereignty and data residency conditions".
- Price fixed; revisions only for future specific contracts (FWC I.23.2).
- No compensation on termination (FWC II.18.4).
- 25,000 / 40,000 H100-eq phase 1 minimums and 120 / 150 MW phase 2 (§3.d, §3.g).
Arithmetic the spec does not print: full EU draw needs IT CAPEX of EUR 588m (Lot 1) or 1,176m (Lot 2), which at today's EUR 10,900 per H100-equivalent is 54,000 or 108,000 units, 22 or 45 MW of IT; the phase 2 EUR 400m needs EUR 2.35bn of IT CAPEX, about 216,000 units against a 75,000 minimum.
Full conditions extract: aigf-tender-conditions-extract-2026-10-23.md.
3. Megawatts booked: the five power contracts
CONFIRMED each (party announcement plus trade press)
Basis: megawatts of electricity under a signed long-term purchase contract, as announced by the parties. The EU row is zero because the tender procures computing time and no electricity, grid capacity or site ("shall be covered by the AIGF Consortium").
| Buyer | Plant, place | MW | Term | Announced | Sources |
|---|---|---|---|---|---|
| Amazon | Susquehanna, Pennsylvania (Talen) | up to 1,920 | 17 years, to 2042; ramp 840 to 1,200 MW by 2029, full by 2032 | Jun 2025 | POWER 12 Jun 2025; nuclear tracker |
| Meta | Clinton, Illinois (Constellation) | 1,121 | 20 years from Jun 2027 | 3 Jun 2025 | Constellation PR; tracker |
| Microsoft | Crane, Three Mile Island 1, Pennsylvania (Constellation) | 835 | 20 years; restart 2027 | Sep 2024 | WNN 26 Jun 2025; tracker 1 Oct 2026 |
| Duane Arnold, Iowa (NextEra) | about 615 | 25 years; back online early 2029 | 28 Oct 2025 | WNN; tracker | |
| Loviisa, Finland (Fortum) | 50 percent of 1,014 = 507 | 2030 to 2049, tickets from 2028 | 9 Sep 2026 | ANS 11 Sep 2026; WNN 9 Sep; Google PR (EUR 13bn in Finland over two years) | |
| EU AI gigafactory tender | seven sites | 0 MW of power | buys computing hours, five years | 30 Jul 2026 | Tender Specifications |
4. The model: inputs, results, what it says
Downloads: aigf_model.py, results.csv, results.txt, sensitivity.csv. Reproduced by running the script in this folder.
4.0 Inputs (base case)
| Input | Base value | Source and status |
|---|---|---|
| Reference hardware | GB300 NVL72, 72 GPUs, 4.55 H100-eq per GPU | spec performance index |
| All-in IT cost per rack | USD 4.0m | IREN/Microsoft transaction; Loop Capital and Morgan Stanley estimates (TWO-SOURCE) |
| Power per rack | 135 kW | vendor class; assumption |
| EUR/USD | 1.12 | ECB 22 Oct 2026 |
| Facility | EUR 11m per MW IT | Turner and Townsend 2025; JLL 2026; Pantera (TWO-SOURCE) |
| PUE | 1.15 | direct-to-chip 1.10 to 1.20 |
| Electricity | EUR 165/MWh (Czechia) | Eurostat nrg_pc_205, 2026-S1, excl. VAT |
| Market price | EUR 2.00 per H100-eq-hour year one, falling 20 percent a year | Thunder Compute; SemiAnalysis |
| Utilisation | 70 percent | ASSUMPTION (no operator discloses) |
| Hurdle | 10 percent unlevered | neocloud capital structure |
4.1 Phase 1 sites, base case
Public share: 4.7 percent of capacity, 8.6 percent of five-year revenue, 22.4 percent of capex nominal, 15.4 percent of capex in present value at 10 percent. IRR is the same across sizes because the model is linear in scale.
| Case | GPUs | IT MW | Total capex | Public/yr | Y1 rev | IRR | NPV@10% |
|---|---|---|---|---|---|---|---|
| Lot 1 min 25k | 5,495 | 10.3 | 414 | 18.5 | 311 | 37.5% | 349 |
| Lot 1 score cap 50k | 10,989 | 20.6 | 828 | 37.1 | 621 | 37.5% | 698 |
| Lot 1 full draw 54k | 11,868 | 22.3 | 895 | 40.0 | 671 | 37.5% | 754 |
| Lot 2 min 40k | 8,791 | 16.5 | 663 | 29.7 | 497 | 37.5% | 559 |
| Lot 2 score cap 70k | 15,385 | 28.8 | 1,160 | 51.9 | 870 | 37.5% | 977 |
| Lot 2 full draw 108k | 23,736 | 44.5 | 1,789 | 80.0 | 1,342 | 37.5% | 1,508 |
EUR millions. Source: results.csv from aigf_model.py.
4.2 Scenarios, Lot 1 full draw
| Scenario | IRR | Public share of revenue |
|---|---|---|
| Base | 37.5% | 8.6% |
| Ramp 40/60/70 | 26.3% | 10.1% |
| Nebius mid-term contracts flat | 30.9% | 8.8% |
| Conservative | 8.8% | 11.8% |
| Bear | minus 15.1% | 22.1% |
| Empty (20% util) | minus 3.8% | 24.8% |
| CoreWeave FY2025 realised | minus 9.6% | 33.2% |
4.3 Sensitivity: utilisation against annual price decline
IRR, Lot 1 full draw, otherwise base. From sensitivity.csv.
| util \ decline | 0% | 10% | 20% | 30% | 40% |
|---|---|---|---|---|---|
| 50% | 35% | 28% | 22% | 16% | 10% |
| 60% | 43% | 36% | 30% | 24% | 18% |
| 70% | 51% | 44% | 38% | 31% | 24% |
| 80% | 58% | 52% | 45% | 38% | 31% |
| 90% | 65% | 59% | 52% | 45% | 38% |
4.4 Power price against grid delay
Base NPV at 10 percent: EUR 754m on EUR 895m of capex (results.csv). Electricity doubled (EUR 330/MWh): IRR 37.5 to 34.3, NPV loss EUR 102m (reproduced from aigf_model.py). One year waiting for the connection with hardware bought and the market price 20 percent lower at start: IRR 19.0, NPV loss EUR 436m. Two years: IRR 10.0, NPV loss EUR 753m. Bruegel PB 18/2026: a one-year delay above 5.5 percent of lifecycle value, doubled energy 4.5 percent.
4.5 Break-evens, Lot 1 full draw, IRR = 10 percent
- Utilisation 35 percent at the base price path.
- Year-one price EUR 1.06 per H100-eq-hour at 70 percent utilisation.
- Annual price decline 64 percent.
- Rack cost USD 9.1m.
- Market revenue flat for five years EUR 11.5m per MW per year (USD 12.8m), between CoreWeave's realised USD 6m (FY2025) and Nebius's contracted USD 20 to 25m.
4.6 Per gigawatt, for scale
7,400 NVL72-class racks, about 2.4 million H100-equivalents, IT capex EUR 26.5bn plus facility EUR 11bn, EUR 37bn in all (Epoch AI: USD 38bn per GW; Pantera: USD 41bn). Revenue at Nebius's contracted rate about EUR 19.6bn a year; electricity 8.1 TWh a year at PUE 1.15 and 80 percent draw. Europe's seven sites at full scale (about 0.95 GW of IT) would need on the order of EUR 20bn a year of compute sales; CoreWeave booked USD 5.1bn worldwide in 2025.
4.7 Limits
- Base case is a snapshot of the rental market's best quarter.
- Utilisation undisclosed by every operator.
- Rack cost: one transaction plus two estimates.
- H100-equivalence is the spec's index.
- No tax, working capital, sales cost, Layer 3 service staffing or sovereignty compliance cost in the model.
- OVHcloud's break-even (Reuters 23 Sep 2026) implies a cost base about twice the model's; printed, not reconciled.
5. Who can fill 95 percent: the demand chain
POSITION
Three European buyers that cannot fill it alone: enterprises via hyperscaler clouds; labs (Mistral's 200 MW); public sector (the AI Factories already exist). Contracts that can: Nscale (Microsoft up to USD 43.8bn, Anthropic up to USD 44.6bn, weighted contract life 5.7 years, prepayment 23 percent); IREN and Microsoft USD 9.7bn for 200 MW; Nebius with Microsoft and Meta (Nebius as anchor of the Rotterdam project); PPC in Greece "in confidential talks with hyperscalers"; Telekom willing only with guaranteed demand. Reading: with a hyperscaler contract a site does not need the tender; without one the tender does not fill it; the tender therefore selects for sites that could not get a contract.
6. The queue: where the megawatts are not
SINGLE-SOURCE per country (regulator or operator)
- Ireland: data centres 23 percent of metered electricity in 2025 (CSO, 7 Jul 2026); greater Dublin "fully constrained", 2 GW contracted, 6 GW more sought (EirGrid, May 2026).
- Netherlands: 15 GW on the waiting list, data centres about a third (Netbeheer Nederland via NL Times, 16 Oct 2026).
- Germany: Brandenburg 22 GW of data-centre enquiries since 2022, 100 MW connected (Clean Energy Wire, 25 Jul 2025); four TSOs, 147 GW of transmission requests in H1 2026, mostly batteries (23 Oct 2026).
- Europe: ENTSO-E, May 2026, data-centre demand 87 TWh (2024) to at least 134 TWh (2030).
- IEA: FLAP-D connection queues "average seven to ten years".
7. What a government could sell instead: the auction idea
POSITION (author's proposal, not evidence)
Precedents: spectrum auctions carry build-out dates and penalties; gas pipelines and LNG terminals are sized by open seasons; JAO auctions cross-border transmission capacity; the UK reordered its connection queue in 2025 to first ready, first connected; the tender's own liquidated-damages clause (0.1 percent per week, capped at 5 percent) shows the mechanism exists, pointed the other way.
8. Call, dated
Call: fewer than seven AI gigafactory framework contracts are signed by 31 December 2027, and the first site to pass technical validation has a non-EU anchor customer for more than half its phase 1 capacity. Made 23 Oct 2026. Published: LinkedIn post. Scored by: EuroHPC award notices and the consortium's own announcement at technical validation. Falsifier: all seven signed and the first to run on European demand. Note: the project's earlier call 3 (12 Nov) moved with its source to 3 Dec 2026 (Lot 2) and 4 Feb 2027 (Lot 1); state that on /calls/ when built.
Related
Outcome
[PASTE impressions, reached, reactions, comments, saves, sends at 72h and 7d]
Corrections log
23 Oct 2026. Empty at publication.
Sources
Read date for items below: 23 Oct 2026 unless stated. Grouped; full URLs in the tender extract and economics note in this folder.
- Tender: EuroHPC call page; Tender Specifications V1 (PDF); model framework contract.
- State of play: Europa Press via Demócrata 22 Oct; zakazka.cz 11 Oct; Euronews 16 Oct (budget).
- Nebius: Motley Fool transcript 19 Aug 2026.
- Nscale: S-1 summary (Mostly Metrics, 18 Sep 2026). IREN: PR 3 Nov 2025. Mistral: DCD 30 Mar 2026.
- OVHcloud: Reuters 23 Sep 2026 via Global Banking and Finance and Techzine.
- Power deals: POWER, Constellation, WNN, ANS, Google PR (section 3).
- Queues: CSO, EirGrid, NL Times, Clean Energy Wire, ENTSO-E, IEA.
- Prices and costs: Eurostat nrg_pc_205; SemiAnalysis; Thunder Compute; Turner and Townsend 2025; JLL; Epoch AI; Pantera; CoreWeave and Lambda financing PRs. Bruegel PB 18/2026.
- Model files in this folder: aigf_model.py, results.csv, sensitivity.csv, results.txt.