Crossing the EU Line
Part 10 left Maestro-2 at 96% of the EU AI Act's compute line. A model is not trained in one stage, though, and by the time the full plan was added up the total was 3.7% over. This part is about a choice most start-ups would rather not face: shave the plan to stay under the line, or cross it on purpose and take on the obligations of a frontier provider.
eu_line.py and comply_plan.py were run (real output below). The 10^25 threshold and the obligations come from the AI Act's text (Articles 51–55); how each training stage is counted should be checked against the Commission's guidelines with counsel. Costs and timings are scenario estimates. Disclosure: Claude, made by Anthropic, helps write this site.
1. Month 36: three point seven percent
Hana Sato, the company's counsel, had a habit of reading spreadsheets from the bottom up. That was why she found it.
The pre-training line said 9.60e24. Everyone had seen that number; it was the one in the board deck, ninety-six percent of the line. Under it, in rows nobody had summed, were the long-context stage, the reinforcement-learning stage in the new environments, and a small allowance for customer skill fine-tunes. Hana added them up and wrote the total in red: 1.04e25. You are a systemic-risk model.
Gopala read it twice. Then he asked the only question that mattered: Is it cheaper to be under the line, or to be good?
2. The map: stages add up, and the line triggers duties
🕸️ From a compute ledger to obligations
Training stages in purple add up to the total (red). Crossing the line triggers four duties (green and blue).
- The point of the map.
- A model is trained in stages; the line is about the total.
- Each duty above the line has an owner and a deadline.
3. Adding up the ledger
"""eu_line.py — will Maestro-2 cross the EU AI Act compute line?
The AI Act presumes a general-purpose model has systemic-risk capabilities when
its cumulative training compute exceeds 10^25 FLOPs (Article 51(2)). Which later
stages count, and how exactly, should be checked against the Commission's
guidelines with counsel. The ledger entries are SCENARIO estimates.
"""
LINE = 1e25
LEDGER = [("pre-training, option B2", 9.60e24), ("long-context mid-training", 0.30e24),
("RL post-training (environments)", 0.45e24), ("skill fine-tunes (per customer)", 0.02e24)]
def main():
tot = 0
for name, f in LEDGER:
tot += f; print(f"{name:34} {f:9.2e} running total {tot:9.2e} ({tot / LINE:5.1%} of line)")
print(f"\ntotal {tot:.2e} FLOPs: {'ABOVE' if tot > LINE else 'below'} the 1e25 presumption by {abs(tot - LINE) / LINE:.1%}")
need = (tot - LINE) / (6 * 32e9)
print(f"to stay under, cut pre-training by {need / 1e12:.1f}T tokens (from 50T) or drop the RL stage")
if __name__ == "__main__":
main()Cumulative training compute vs the 10^25 line
Real output. Running totals, as a share of the line.
🔎 Go deeper: which stages count?
- The Act speaks of cumulative compute used for training.
- The Commission's guidelines for general-purpose AI providers (2025) explain how to estimate it; whether and how later fine-tuning and synthetic-data generation are included is exactly the detail to confirm against them with counsel ⚠️.
- The team counted conservatively: every stage that changes the model's weights.
- Counting conservatively costs little; being found under-counted later costs the company's credibility.
4. Shave or cross
Tomás wanted to shave. Two trillion fewer tokens and we are under. Nobody will notice the difference in quality.
Elena Brandt, who ran safety, wanted to cross. We already run most of this. Evaluations, red-teaming, incident response. The law is asking us to write down what we do.
Hana gave them both the same answer. If you stay under by 2%, every future stage you add, every retrain for a customer, puts you back over. You will be managing an accounting line instead of a model.
Gopala decided before the meeting ended. They would cross, on purpose, and say so publicly. A company whose product is evidence that changes are safe could not be the company that trimmed its run to avoid evaluation.
| Option | Compute | What it costs | What it risks |
|---|---|---|---|
| Shave 1.9T tokens | ~1.00e25 | A slightly weaker model | Any later stage crosses anyway |
| Drop RL post-training | 9.92e24 | The stage that makes Maestro good at its job | Product quality |
| Cross and comply (chosen) | 1.04e25 | $2.20M, 12 weeks in parallel | A public commitment the team must keep |
5. What crossing costs
"""comply_plan.py — what crossing the line costs.
Obligations for providers of general-purpose models with systemic risk come
from AI Act Articles 52-55 (notify the Commission, documentation, evaluations
including adversarial testing, risk mitigation, serious-incident reporting,
cybersecurity). Owners, weeks and costs are SCENARIO estimates.
"""
OB = [("notify the Commission (Art. 52)", "CEO + counsel", 1, 0.02),
("update technical docs + downstream info (Art. 53)", "research + docs", 6, 0.15),
("update copyright policy + data summary (Art. 53)", "data lead + counsel", 4, 0.10),
("evaluations incl. adversarial testing (Art. 55)", "safety team + external red team", 10, 0.90),
("systemic-risk assessment + mitigation (Art. 55)", "safety lead", 8, 0.35),
("serious-incident reporting process (Art. 55)", "ops + safety", 3, 0.08),
("cybersecurity for weights and infra (Art. 55)", "security", 12, 0.60)]
def main():
for name, who, wk, c in OB: print(f"{name:52} {who:32} {wk:3} wk ${c:4.2f}M")
tot = sum(o[3] for o in OB); crit = max(o[2] for o in OB)
print(f"\ntotal ${tot:.2f}M; longest item {crit} weeks (runs in parallel with the ~12-week training run)")
print(f"share of the Maestro-2 compute bill (~$19.5M on 4,096 GPUs): {tot / 19.5:.0%}")
if __name__ == "__main__":
main()Compliance work, in weeks (parallel with training)
Real output. Total $2.20M, about 11% of the compute bill.
- What surprised the team.
- Most of the cost was work the safety team had planned anyway.
- The real critical path was securing the weights: access controls, encrypted storage and audited transfers for a model that is now a regulated asset.
- They signed the General-Purpose AI Code of Practice (story event), which turns the duties into a concrete checklist.
6. Sandbox: the compute ledger
🎮 Sandbox: the compute ledger
Same ledger as eu_line.py (32B active parameters, 6ND, plus 0.02e24 for customer skill fine-tunes). Not legal advice: how each stage counts should be checked against the Commission's guidelines.
- Try this.
- Drop RL post-training: 99.2%, under the line, but you have removed the stage that makes Maestro good at its job.
- Keep everything and cut pre-training to 48T tokens: 99.9%, a hair under.
- Push tokens to 60T: 123%. Bigger plans cross with room to spare.
7. Five mistakes
- Counting pre-training compute and forgetting the other stages.
- Designing the model around a regulatory line instead of the job.
- Notifying after the run instead of when you know you will cross.
- Treating compliance as paperwork instead of security work on the weights.
- Assuming 10^25 is permanent: the Commission can change it.
8. Delegation box
- The decision to cross, and saying so publicly
- Compute ledger tracking
- Drafting documentation from the run records
- Counsel for the notification
- An external red team for adversarial testing
9. Budget and risk tracker
| Item | Status after Part 12 | Note |
|---|---|---|
| Compute | 1.04e25 planned | 3.7% over the line |
| Status | Systemic-risk GPAI provider | Commission notified (story) |
| Compliance | $2.20M · 12 weeks | in parallel with training |
| Credibility | Strengthened | crossed on purpose |
10. The series
The Maestro, Season 2: 23 parts in five phases (Season 1: 18 parts)
11. Reading the numbers
- Primary / standard
- Regulation (EU) 2024/1689 (the AI Act), Articles 51–55; the General-Purpose AI Code of Practice (July 2025).
- Secondary ⚠️
- Article summaries from artificialintelligenceact.eu ⚠️ (secondary).
- Our assumptions
- The ledger, costs, timings and all story events are the scenario. Not legal advice.