unit economics · single year
the simple business model.
A handful of institutional buyers — TU Delft–sized universities and large companies. We resell sovereign AI coding access at a margin, and add training & customization on top. No GPUs to own, no 24/7 fleet — a lean team and a few good deals. For the time axis — ramp, runway, grants — see the 5-year projection.
resale + margintraining & customizationinstitutional dealslean team
the deals
A small number of institutional customers.
Universities + large companies. Each is one sales relationship, not a self-serve signup.
1005k10k
Share of licensed seats that actually use it monthly. Enterprise tools average ~50%.
platform — resale + margin
You buy tokens wholesale, sell access at a markup. Margin is built in.
≈ €20 / seat / month. Copilot Business ≈ $19/mo, Enterprise $39/mo.
Your wholesale cost, after routing ~70% to a cheap model + ~80% prompt-caching. Heavy users capped by fair-use so this stays bounded.
services — training & customization
High-margin labor. Often the real reason they buy.
Onboarding, training, custom skills/integrations, the DPO/compliance package. One-time + recurring.
your cost base
Keep it lean — this is the whole point.
The team that runs the gateway + delivers services. ~2 people to start (NL loaded ≈ €145k each).
net margin
—
—
total revenue / year—
↳ platform (seats)—
↳ services—
token cost / year—
gross profit—
gross margin—
operating cost / year—
net profit / year—
net per customer—
assumptions & sources
why it's shaped this way
| Resale, not self-host est | Below ~1,000 active seats, reselling a per-token EU gateway beats owning a GPU node — no fixed iron to fill. Self-hosting only wins later, at volume — the inference-economics tool models exactly where that crossover falls. |
| Markup removes heavy-user risk firm | Flat per-seat pricing for agentic coding is being squeezed — Copilot moved to usage-based billing in 2026. Charging cost + margin (with fair-use caps) keeps every user profitable. GitHub |
| Token cost / active user est | Typical agentic user ≈ 135M tok/mo; after ~70% routing to a cheap model + ~80% prompt-cache, effective wholesale lands ≈ €15–40/mo. morphllm, RouteLLM |
| Seat pricing comps firm | Copilot $19/$39, Cursor $20/$40, Tabnine (self-host) $39. Sovereignty + compliance supports the upper half. GitHub, Tabnine |
| Seat utilization firm | ~47–54% of enterprise SaaS seats are actively used. Zylo |
| Margin class firm | This is a managed-services / AI-app hybrid: net 15–25% at scale, not SaaS 80%. Services lift the blend. Gradient (MSP) |
| Team cost est | NL senior MLOps/SRE ≈ €145k loaded. Two people run the gateway + deliver services to start. PayScale |
how the math works
| Platform revenue | customers × seats × price/seat/yr |
| Services revenue | customers × services/customer/yr |
| Token cost | customers × seats × active% × token-cost/active-user/mo × 12 |
| Net | (platform + services) − token cost − operating cost |
Deliberately simple — one year, one snapshot. Excludes one-time setup (ISO 27001 ~€55k, build), sales ramp, and assumes services are delivered by the same team inside operating cost. The softest knob is token-cost-per-active-user — it depends on routing, caching and how heavy your users are; instrument it in a real pilot. For the multi-year view with grants, hiring and runway, see the 5-year projection. Indicative, June 2026.