Last week I discussed the unprecedented investment Europe is making in sovereign cloud infrastructure. This week, I want to focus on the AI side of that equation.
At its core, the current AI race is an infrastructure problem. Compute is the main constraint on AI development, which explains why the investment figures matter so much. But here's the thing: Europe can't outspend the US, and it doesn't need to. The real opportunity seems to shift to cost-efficiency, regulatory compliance, and building on domain expertise.
Before diving into where Europe actually competes, let's look at the current market landscape across three dimensions: spending, performance, and cost.
AI Investment (B USD)
Top Model Arena Elo
Cost per M Tokens (USD)
Sources: Stanford HAI AI Index 2026 (investment figures from pp. 182-183, Arena Elo from p. 78), Anthropic Pricing, Mistral Pricing, DeepSeek Pricing. Cost comparison uses output tokens for flagship models: Claude Sonnet 4.6 (US), Mistral Large 3 (EU), DeepSeek V3 (China). Arena Elo represents top model per region: Anthropic (US), Alibaba (China), Mistral AI (EU).
The Spending Gap
US venture capital dominates AI investment by a staggering margin. At $285.9B, US CapEx is nearly 14 times Europe's $20.9B. Anthropic's latest funding round alone raised more than three times what the entire EU collectively invested in AI startups.
This money funds not just frontier model development, but an entire ecosystem: startups, research communities, datacenter buildouts, and talent pipelines. The consequence for Europe, and basically the whole world, is a persistent brain drain as top AI researchers and practitioners migrate to where the funding and compute resources are.
But what we see now is that raw spending doesn't translate linearly to capability. DeepSeek and other Chinese models proved this when they achieved near-frontier performance while operating under US chip export sanctions. Constraints, it turns out, can breed efficiency.
The Performance Surprise
This is, honestly, the most surprising aspect of the current landscape. The performance gap between US flagship models and their Chinese or European counterparts doesn't remotely match the 14:1 investment differential.
US models hold a clear lead — Anthropic's top model scores 1,503 on Arena Elo — but the margins are tighter than the funding gap would suggest. Chinese models from Alibaba sit at 1,449 (a 3.6% gap), while Mistral's best reaches 1,416 (5.8% behind).
Companies like Mistral and Aleph Alpha have made remarkable progress in a short time. This lends weight to what some AI researchers have been arguing: we may be hitting diminishing returns on the current paradigm. Yann LeCun has been vocal that LLMs alone won't get us to AGI. Ilya Sutskever, after leaving OpenAI, headed back to R&D and hinted that the next breakthrough requires fundamentally different approaches.
For enterprise use cases, the question isn't whether US models are better, they are. The question is whether that marginal improvement justifies a 10x price premium.
The Cost Equation
US models run 10-15x more expensive than Chinese or European alternatives. Claude Sonnet 4.6 costs $15 per million output tokens. Mistral Large 3 costs $1.50. DeepSeek V3 costs $1.10.
How did this happen? The Chinese labs, forced to work around chip sanctions, had to invent new training algorithms and optimization techniques. Necessity drove efficiency. Meanwhile, US labs operated on an "all you can eat" compute buffet, optimizing for capability over cost.
The result: for the vast majority of enterprise applications, a model at $1.50/M tokens is functionally indistinguishable from one at $15/M tokens. The "good enough" threshold could well be the largest obstacle for widespread adoption of frontier models.
Where Europe Actually Competes
I don't see Europe winning frontier model race anytime soon. But that's not the only game in town. Three areas offer genuine competitive advantage:
1. The Regulatory Moat
The EU AI Act creates compliance complexity that European companies are uniquely positioned to navigate. For critical and highly regulated industries — finance, healthcare, public infrastructure — there's increasingly no choice other than adopting European-operated and deployed AI. What looks like bureaucratic overhead is seemingly the feature here!
2. Enterprise Integration Is the Actual Prize
Increasingly, the ground is shifting from the hype around foundational models to what I call "Functional AI" — the stuff that actually works, integrates within complex enterprise environments, and gets real things done.
For Functional AI, the model is the easy part. The hard part is putting it to use: enabling companies to leverage their existing systems of record, optimize workflows, redesign broken processes, and capture valuable insights they can act on. Take SAP ERP as an example — the data is already there, the processes are already defined, but unlocking AI value requires deep integration expertise.
This is where Europe's industrial giants — SAP, Siemens, Bosch, Daimler — are positioned to take advantage. They've spent decades learning how to deploy complex technology in enterprise environments. That muscle memory matters more than raw model capability.
3. Domain-Specific Applications
| Domain | European Leaders | AI Application Focus |
|---|---|---|
| Automotive | BMW, Mercedes, VW | Autonomous systems, manufacturing |
| Industrial | Siemens, ABB | Predictive maintenance, robotics |
| Healthcare | Philips, Roche | Diagnostics, drug discovery |
| Finance | Deutsche Bank, ING | Risk modeling, compliance |
| Enterprise | SAP, Dassault | Business process automation |
It's well established now that not all models are equally good at all tasks. Some excel at coding, others at reasoning, others at specific vertical applications. And even for coding, the model alone isn't enough — you need tooling around it to maximize benefit. Take Claude Code as an example: same model, but the tooling transforms what's possible.
The same principle applies here. European companies lead in verticals where domain expertise matters more than raw capability. An AI system for predictive maintenance in manufacturing requires deep understanding of industrial processes — something Siemens has that OpenAI doesn't. An AI system trained on decades of ERP transaction history and process data will be unmatched for business process optimization, regardless of which foundation model powers it.
The Token Economics
Here's the question worth asking: if performance between models is this tight, and cost is this far apart, what percentage of use cases actually justify paying for the latest and greatest?
For most enterprise workloads — summarization, classification, extraction, basic reasoning — the performance difference between a $15/M token model and a $1.50/M token model would likely be imperceptible. Mistral Large would handle these as well as Claude for the majority of cases.
Vercel's recent AI Gateway Production Index puts concrete numbers to this. Their analysis of production AI workloads reveals a telling pattern: spend follows the cost of being wrong.
| Workload Type | % of Tokens | % of Cost |
|---|---|---|
| Personal assistants | 40% | 20% |
| Coding agents | 20% | 22% |
| Back office agents | 15% | 6% |
| App generation | 11% | 7% |
The pattern is clear: high-volume, low-stakes workloads run on cheaper models because individual errors are quickly corrected. Coding agents, where mistakes carry higher consequences, justify premium spending. B2B applications spend roughly double per token compared to B2C — not because they need better models, but because the cost of being wrong is higher.
This data validates the European positioning: when the vast majority of enterprise API calls (like back-office automation) require only "good enough" reasoning, hyper-efficient models from providers like Mistral or Aleph Alpha become the most financially viable choice for scaled deployment.
What This Means
Europe's path forward isn't to out-invest the US or out-optimize the Chinese. It's to leverage what it actually has:
- Enterprise deployment expertise from decades of complex system integration
- Regulatory compliance that becomes a competitive advantage as AI governance tightens globally
- Domain depth in industries that run the physical economy
- Sovereignty demand from organizations that need local, compliant AI infrastructure
The scene is divided — physically and increasingly politically. Europe is losing the frontier model race decisively, but it's not clear that this matters as much as the headlines suggest. The race to have useful, compliant, cost-effective AI in enterprise environments is a different game entirely — and one where Europe has home-field advantage.
What's your organization's approach to AI vendor selection? Are you optimizing for capability, cost, or compliance? I'd be interested to hear how others are navigating these tradeoffs.