Europa 2031 · four briefs · July–August 2026

Where Europe can still build: compute, power, silicon, robots

Four condensed briefs from the Europa 2031 project, gathered on one page. Each keeps its full content; the working documents and explorers in the project folder carry the evidence.

Discussion starter · 21 Aug 2026AI Gigafactories tender: where we stand Venture assessment · Aug 2026A European sovereign inference company Findings brief · Jul 2026What it takes to turn an electron into an AI token Investor lens · 21 Aug 2026Why humanoids will not scale to tens of millions a year soon

Europa 2031 · discussion starter

AI Gigafactories tender: where we stand

EuroHPC call EUROHPC/2026/OP/0008 · Lot 1 and Lot 2 · deadline 12 November 2026, 16:00 Luxembourg. Prepared from the official tender bundle, 21 August 2026.

What it is

Not a grant. The EU plus 18 member states pre-buy five years of managed AI compute (L2 platform + L3 model-as-a-service) from up to seven privately financed gigafactories. The consortium builds, finances, owns and operates. Public money arrives only after go-live (around late 2028), covers at most ~34% of IT CAPEX (17% EU + matching state), and nothing of the building, power, site or OPEX. The call itself prices one gigafactory at €4–5bn.

The two lots

Lot 1 · medium: 25k H100-equivalents at go-live, 75k later. €100M EU + €100M state. Hosting: FR, FI (+DK/EE/LV/SE sites), DK, CZ, PL.
Lot 2 · large: 40k at go-live, 100k later. €200M EU + €200M state. Hosting: DE, ES, IT, PT, EL.

Same rules otherwise; only scale, money and the procured service mix differ.

Five facts that decide our position

The Netherlands is not a hosting country. We can coordinate, but the facility must stand in one of the countries above, and each country effectively funds one project.
We can't pass selection alone. The bidding group must show >€200M average turnover, a 5 MW build-and-operate track record, and €5M of AI-service revenue in the last three years. Partners can supply this, but they must do that part of the work.
The coordinator must be EU-controlled through every ownership layer for the full 78-month contract, and any change of control needs public approval. That constrains our cap table and exit.
The evaluation scores what is already signed: land options, grid interconnection studies for 120–150 MW, financing term sheets, GPU vendor LOIs. Plans score "acceptable"; contracts score "excellent".
The real cut-offs are earlier than 12 November.

~1 Oct · state amounts firm up 4 Nov · questions close before 12 Nov · SPV must exist 12 Nov 16:00 · submission

Scoring in brief

Stage 1, quality: Technical 40%, Impact 20%, Financial 40%. Need 60/100 overall and a floor on every one of ~15 sub-criteria; one miss and the bid is out.
Stage 2, ranking: 50% stage-1 score, 25% scale and service depth, 25% price relative to the cheapest bid, normalised per chip (newer silicon looks cheaper).
Financial viability counts as much as all technical criteria combined and is audited by independent financial experts. The benchmark sub-criterion has the steepest floor (10/15).
Roughly a third of the final score sits in service depth (L3: hosted models, APIs, RAG, agents, guardrails, AI-Act auditing), EU-run sovereign operations and a European software stack within four years. That is the part closest to what we want to build. No clear L3 at day one is explicitly penalised.

The decision to make first

Which role: (a) coordinator of a new consortium in someone else's country, (b) service-layer and inference-stack partner inside a consortium led by an established operator, or (c) identified subcontractor to two or three consortia to build track record for the next round. The documents allow (a), reward (b), and make (c) the cheapest way into the room.

Suggested go/no-go: 1 October. By then we either have a consortium term sheet with an operator, a balance-sheet partner and a site in a hosting country, or we switch to role (c).

Full analysis: AIGF-Tender-Analysis-Lot1-Lot2.docx in the Europa 2031 folder. Sources: Tender Specifications V1, Annexes 1–7, Model Framework Contract, FAQ and public Q&A to 18 Aug 2026.

Europa 2031 · Venture assessment · August 2026

A European sovereign inference company

Owned EU datacenters for AI inference with an OpenRouter-style layer serving open-weight models. The short version.

Defensible idea, wrong shape. It is two businesses: a capital-heavy token factory and a neutral router. Owning the capacity kills the router's neutrality. The real venture is a power-and-capital company that sells tokens.

You cannot bid for an AI Gigafactory alone; the consortium role is the first decision.

Pros

Demand is real: tokens up 14x a year, open-weight taking share, inference now the dominant workload.
Law now protects the market: CADA Level 3/4 excludes US-controlled providers; EuroHPC and 18 states pre-buy capacity.
Cheap Nordic and Iberian power, heat-reuse revenue, and European banks already funding this shape (Mistral, Nscale, Nebius).

Cons

Late and crowded: Nebius, Nscale, Mistral Compute, incumbent EU clouds and eight EU routers already exist. Your own analysis ranked this Tier 3.
Sovereignty earns no price premium at the API tier; only procurement-gated contracts (about 10% of sensitive public spend) pay for it.
Power connection, utilization, GPU pricing and depreciation schedule decide the economics, not electricity price. A 2028–29 delivery lands in the likely glut.
Nvidia/TSMC/CUDA infra is not truly sovereign; a venture exit likely means a US buyer, breaking the sovereignty promise.
Team fit is weak: this is a power, finance and enterprise-sales business, not a hardware-engineering one.

Decide together

Which consortium, and in what role.
First customer type: procurement-driven or self-serve.
Whether the router also serves non-EU and closed-weight models.
Target CADA level, and what it implies for cap table and exit.
Chip strategy: Nvidia only, plus AMD, or a European-silicon lane.
Whether energy-per-token instrumentation is the better entry for your team.

Condensed from the full memo (Sovereign-Inference-Venture-Assessment.docx) in the Europa 2031 folder. Sources: Exponential View, State of the AI Economy (June 2026); EuroHPC AI Gigafactories call; Electron to Token v1.4.

Europa 2031 · Findings brief · July 2026

What it takes to turn an electron into an AI token

Eighteen steps researched and sourced. The essentials only — the working document carries the evidence.

Fuels → Generation → Grid → On-site power → Shell → Power conversion → Cooling → Litho → Foundry → HBM → Packaging → Chip design → Cluster & network → Data → Training → Model → Inference → Delivery

The chain

AI demand moves in months; its foundations move in decades. That clock mismatch is nearly every bottleneck — and where the durable businesses live.

Energy

The constraint is the calendar: turbines sold out to 2030, grid connections 5+ years — so 130+ GW is being built behind the meter, mostly gas.
Flexibility is the most underpriced asset: it buys faster grid connection and the $50–100/MWh gap between non-firm renewables and firm power.

Silicon

TSMC's allocation effectively plans global AI compute — 2nm, packaging and 2027 memory are all sold out.
Value is migrating into the package (power, cooling, memory, optics) — and its three key disciplines are Dutch: ASML, Besi, PhotonDelta.

The token

Energy per token spans 100× (~0.15 Wh for an efficient answer, ~15 Wh for a reasoning marathon).
Inference is now two-thirds of all AI compute; prices fell ~50× while volumes exploded.

Patterns & risks

Every shortage (turbines, transformers, memory, packaging) overshoots: today's scarcity is likely 2029's glut — bet on transitions, not shortages.
Regulation keeps creating markets on schedule: curtailment rules, heat-reuse law, energy disclosure.
Two unhedged risks: Taiwan, and an algorithmic breakthrough that voids the capacity math.

For the venture

Top opportunities: test & metrology for the advanced package, photonic assembly automation, the flexibility/firmness stack.
For a Dutch hardware team, photonic assembly ranks first; the 800 VDC window closes ~2027.

Condensed from “From Electron to Token” v1.4 — 18 sourced deep dives, 14 opportunity clusters, 8 dated predictions (first resolves H2 2026). Working document and interactive explorer: Europa 2031 project folder.

Europa 2031 · investor lens · 21 August 2026

Why humanoids will not scale to tens of millions a year soon, and where that leaves a European investor

Humanoids ship in the tens of thousands, not millions: about 19,000 units in the first half of 2026, 97% of them Chinese, most going to labs rather than jobs. Western volume claims run 10–1,000× above verified output.

Four problems that capital does not solve

Gears and screws: Each robot needs roughly 28 ground precision reducers or roller screws; ten million robots need 30–70× today's world capacity. The limit is micron-grade grinding on machines from about six companies, run by scarce process engineers. The fix is a process, or a transmission design, that scales without scaling Swiss-German machine tools 30×.
Magnets: Hot joints need dysprosium or terbium in their magnets. China makes 94% of them and licenses Dy/Tb exports; its suspension of broader controls expires on 10 November 2026. The fix is Dy-free magnet grades or actuators that run cool enough not to need them.
Heat: Harmonic reducers are 65–75% efficient and joints reach 120 °C, so robots do demos rather than shifts and run two to four hours. The fix is roughly doubling continuous torque density through better windings, in-joint cooling and transmissions. More battery only adds mass, torque demand and heat.
Reliability: About 95% task success today against a 99.9% industrial need. Gains scale with data, log-linearly: twenty times the data buys about 2.4× the completion rate. No vendor publishes uptime or teleoperation-intervention figures.

Hardest hardware, in order: hands, actuator thermals, C3-grade roller screws, Dy-lean magnets, tactile skin. Batteries, chips, castings and training compute are not the constraint.

The European investor angle

European OEMs (Neura at $7B, Humanoid at $1.35B) are fighting Figure ($39B), Unitree ($53–66B) and an entirely American foundation-model layer on capital and volume, which is Europe's weakest axis.

The asymmetric bets sit upstream, inside the first and third problems above. The gear- and thread-grinder makers (Reishauer, Klingelnberg, Kapp-Niles) that every reducer on earth passes through, and that no analyst covers. Schaeffler, the only Western one-stop supplier of actuators, gears, roller screws and bearings, still priced as a bearings company. Maxon, Faulhaber and Renishaw, if cheap parts keep failing at one to two million cycles. Siemens and Hexagon digital twins, used by every OEM regardless of who wins.

EU certification rules (Machinery Regulation from January 2027, AI Act embedded-AI deadline August 2028) are a lever over market access, but they will also slow European deployment more than any component shortage.

Not yet considered

Demand proof may be the real gate: only about 10% of shipped units do paid work, and nobody has shown repeat deployments with positive unit economics. Wheels may beat legs and erase half the chokepoints. Households mean privacy law, liability, insurance products that do not exist in the West, and roughly 100,000 teleoperators at a 1% intervention rate on ten million units.

Condensed from the Humanoid Scaling Explorer (Europa 2031 project). Figures are latest public reports as of 21 August 2026; ratings are analytical judgements, not investment advice.