The €63.2 million that isn't

The European Commission announced a €63.2 million funding package today to "support AI innovation in health and online safety," timed against the 2 August 2026 deadline for full enforcement of the AI Act.¹ The package is being framed as innovation support. I want to walk through why I think it is mostly an enforcement-softening gesture, and why the arithmetic of it is, by the Commission's own implied numbers, an order of magnitude too small to do the job.

Start with what the Commission is paying for, and then with what compliance actually costs.

The €63.2 million is spread across health AI and online-safety AI, two verticals the Act treats as high-risk or systemically relevant. The Commission's own guidance, read alongside the SME impact assessments that have been circulating since late 2024, puts compliance costs at €160,000 to €330,000 per high-risk system for a European SME, and $8 million to $15 million in initial outlay for a large enterprise deploying across categories.² These are not FLUX numbers; they are the numbers European builders have been reporting to the Commission and the numbers the Commission has, to its credit, broadly accepted.

The €63.2 million is spread across health AI and online-safety AI, two verticals the Act treats as high-risk or systemically relevant.

Take the midpoints. €245,000 per SME high-risk system. €63.2 million divided by €245,000 is 258 systems. That is the entire programme, if it were a direct compliance-cost reimbursement, which it is not, it is innovation funding, which means some fraction goes to salaries, some to compute, some to consortium overhead. Call the effective compliance-offset fraction 30% and you are funding roughly 77 SME high-risk systems across twenty-seven member states across two verticals. This is not a policy response. It is a press release.

Which brings us to what the package is actually doing, which is something different and a little more interesting.

The AI Act's GPAI obligations, the provisions targeting general-purpose models above the systemic-risk compute threshold, came into early enforcement in Q1 2026, and the first-wave targets have been xAI's Grok and Meta's Llama ecosystem.³ Both are American. Neither is going to be solved by a €63.2 million European innovation fund. The Commission knows this. The fund is not trying to compete with Llama; it is trying to produce a European counter-narrative to the "AI Act is killing European AI" line that Mistral, ASML-adjacent lobbying, and a rotating cast of MEPs have been running since roughly Q3 2025.

Read through the AI safety as market position frame, the structure becomes legible. The EU is attempting to hold a regulatory posture, the world's strictest AI framework, while mitigating the predictable consequence, which is that European builders either move compliance-heavy workloads to non-EU subsidiaries or exit the high-risk categories entirely. The €63.2 million is the mitigation. It is small because the Commission does not have large sums of direct industrial policy money; the big money sits in member-state budgets and in the European Investment Bank, both of which move slowly. So DG CNECT does what it can do, which is a targeted innovation fund with a favourable headline number.

This is a case of something I have seen three or four times now across jurisdictions: the regulator announces a compliance regime, the industry estimates the compliance cost, the regulator announces a support programme an order of magnitude smaller than the compliance cost, and the support programme is presented as the answer to the compliance cost. The UK did a version of this with its AI Safety Institute funding round in 2024. The US did a version of this with the NIST AI RMI allocations. The numbers are always off by roughly a factor of ten, and the programmes are always framed as if they are not.

The performativity frame matters here too, in an inverted form. FLUX normally applies performativity to AI capex, the idea that spending at scale makes you a market actor regardless of whether the spend produces returns. The EU is running the inverse play: spending at small scale on a programme designed to make the EU a market actor in AI governance, where the relevant currency is not capex but regulatory credibility. €63.2 million is cheap for a credibility purchase. Whether it works depends on whether European builders actually receive enough of it to stop complaining publicly, which is a political question, not an economic one.

What the primary document does not say, and I looked, is anything about who is eligible, what the compliance-cost-offset ratio is, or how the funding interacts with existing Horizon Europe allocations.⁴ This is normal for Commission announcements at this stage; the operational detail comes in the work programme, which will land in the summer. I would watch for two things in that document. First, whether any portion is earmarked explicitly for AI Act conformity assessment costs, the technical audits, notified body fees, documentation requirements, which are the part of compliance that is pure deadweight from a builder's perspective. Second, whether non-EU-headquartered firms with EU operations are eligible. The latter is the tell for whether this is genuinely industrial policy or is structured to quietly accommodate the American hyperscalers who are the actual enforcement targets.

The structural story, such as it is: the AI Act is doing what it was designed to do, which is impose compliance costs that function as a de facto tax on frontier AI deployment in Europe. The tax incidence is falling on European SMEs at €160K–330K per system and on American large enterprises at $8–15M per deployment. The €63.2 million package is a small offset, primarily political, directed at the former cohort to keep them from becoming a lobbying problem. It does not change the tax incidence on the latter cohort, which is the cohort the Act is actually trying to regulate.

One thing to watch, which nobody seems to be watching yet: the GPAI early enforcement against Grok and Llama is the pattern that will determine how the Act actually operates. If the Commission wins tidy settlements from both, disclosure commitments, training-data documentation, systemic-risk evaluations, the Act becomes a paperwork regime with teeth. If either firm litigates and wins, or exits the European market for frontier deployment, the Act becomes a different thing entirely, and the €63.2 million becomes either inadequate compensation for a successful regulatory wall or irrelevant seed funding for a regulatory regime that has lost its primary targets.

I do not know which of those two paths we are on. I suspect the Commission does not either, which is part of why the support programme is sized the way it is. You do not commit large industrial-policy money to mitigate a regime whose effectiveness you are still discovering.

Footnotes

¶