six ways to build a sovereign business.

Switch the variant buttons to load each one's economics. The default Blend is what the strategy implies: win the student on UX for free, let the institution pay a platform fee because sovereignty unlocks procurement, meter the heavy AI usage so no user is unprofitable, sell the compliance/training package on top, and let grants fund the open sovereign layer while the commercial layer finds its feet. Self-host and OEM/consortium are where it goes once there's volume — the long-run moat and the up-market climb, not the place to start.

• The deepest risk isn't in this model — it's that nothing here is proprietary. The agent is an open-source fork (opencode), the intelligence is open-weights (GLM, of Zhipu / Beijing origin), the GPUs are rented (Scaleway). "Sovereign" is a data-residency configuration — real, and procurement-relevant — but replicable by any EU-native competitor, Mistral or Scaleway included, and grant funders back open base tech precisely because it's copyable. The moat is a head start in go-to-market and integration depth (SURFconext, GitLab, the managed workspace), not technology. If that head start doesn't harden into switching-cost lock-in before a commodity clone appears, the usage markup trends to zero. Related, unspoken: a sharp procurement officer asking "whose model is this?" hears "Beijing" — a live objection for a sovereignty pitch.
• You're charging a sovereignty premium where sovereignty isn't felt. The strategy's own beachhead is non-confidential student data with "no compliance wall" — the one segment where sovereignty doesn't bite. The segment that genuinely needs it (sensitive research) is years and a compliance-hardening project away. So early willingness-to-pay rests on UX and sanction, not on the sovereignty story — and the model's flat per-customer pricing makes the up-market value-climb (research seats worth €400–600, not €120) financially invisible. Treat the climb as upside the projection doesn't yet price.
• The softest knob is token cost per active user. A genuine agentic user can burn 100M+ tokens/month; at GLM-5.2's output rate that's real money. The whole model lives or dies on routing most traffic to cheap models + prompt-caching keeping effective COGS bounded — which is exactly why the Base case now defaults to €45, not the hopeful €30. Instrument this in the pilot before trusting any net number here.
• Customer counts are a sales hope until the pilot converts. "First paid customer in 2027" assumes the free pilot produces an institutional buyer. It might not. Drag "first paid customers" to 0: the honest read of the cash line is how big a hole grants + founder time must cover if paid revenue slips a year.
• Grants are time, not money. Competitive, slow, and frame-specific (the open infra, not the product). The model weights them to 2027 on purpose — apply from pilot strength. A plan that needs the grant to clear break-even is fragile; the grant should buy runway to reach commercial break-even, not be it.

What survived the critique and is baked in: revenue is shown as true reseller top-line and net margin is net÷revenue (no flattering 70% number); usage is always metered at a markup (no flat-rate heavy-user trap); services carry a delivery cost (~60% margin, not free); a hire is forced at 3 and 8 customers; the ISO-27001 spend lands at the first institutional sale (certs precede the buyer); grants taper to zero by 2029 (no treadmill); and the headline is cumulative cash, because revenue you can't bank is vanity for a bootstrapped team.

Deliberately legible over precise. Excludes VAT, working-capital timing, FX on non-sovereign routes, and equity raises (the point of this plan is to need none). The softest input is token-cost-per-active-user; instrument it in the pilot. Indicative, June 2026 — pair with the single-year unit-economics view.