Open-weight models are taking a growing share of token demand and closing the gap on the frontier. So what happens to trillion-dollar labs if millions of customers move to close-to-frontier open models? And what does it mean to build a modern AI economy on models that, so far, have mostly come from China?
Patrick Doran, Investor at Cherry in London, has set out Cherry's house view in a new long read here. Here are five key thoughts from it.
1. Open-sourcing is tactical
Few big tech players believe purely in open or closed. Each backs the model of AI that suits its own position. NVIDIA supports a wide field of new labs, many of them open-weight, to keep the model layer fragmented to protect its chips business as OpenAI moves down the stack. Chinese labs use open weights as a distribution play. Backed by significant state subsidies, China is good at nurturing technologies that stay negative-margin for years, as it did with EV batteries.
2. The AI economy is bifurcating
Most buyers care about the price and quality of a token, not where it comes from. That puts European open-weight labs up against Chinese labs that have structural cost advantages: distillation from frontier models, state-subsidised compute and cheap, abundant energy. Europe can't copy that playbook, despite equivalent AI researcher talent. One of the biggest levers it has is lowering the cost of electricity over time, which is why companies like Proxima Fusion, BEO and Fuse Energy matter to AI as much as to the traditional energy markets.
3. The market for sovereign AI is emerging
Enterprises want ROI from AI, and as open models approach the frontier, more tasks can run on them at about a third of the inference cost. But security and control matter too. Sensitive industries and sovereign users, such as defence, critical infrastructure and government, shouldn't hand their intelligence or decades of proprietary data to companies outside their alliances, or to labs that want to eat their lunch. Open weights from sovereignty-aligned labs, deployed on infrastructure the customer controls, are the answer. In those verticals, a sovereign model company is best placed to own the full customer relationship rather than sit beneath a separate application layer.
4. Infra wins as models fragment
The whole stack benefits when there are more model labs. We expect a steady state where open-source tokens make up most of total consumption, with frontier models kept for a narrow set of high-stakes tasks. As the performance gap narrows, that split looks something like this: frontier tokens still take most of the economic value (on one estimate, 60–90%), but only 15–25% of the tokens. That creates room for companies that make open models perform: fine-tuning and reinforcement learning as a service (such as Remanence), harness-level optimisation (such as Weco), and the European cloud and chip capacity that sovereign deployment requires (Lyceum, Mistral, Fractile, OLIX).
5. Europe's edge is the hard stuff
Expect some Chinese labs to start closing new releases over the next six months, as security concerns grow and the performance gap narrows. That would open a multi-billion-dollar opportunity for Western open-weight labs, and Europe can be an early mover. Some labs will raise big from day one. Others will start narrow, for example a grid fault diagnosis model fine-tuned on utility data, and become deeply sticky. Cherry is looking for founders on that second path, with world-class technical depth. Just as important, Europe should lean into what it has always done well: the scientific and engineering breakthroughs in energy – both generation and surrounding infrastructure – and chips that AI depends on.
Mistral and Aleph Alpha have already shown that Europe can build world-class open-weight models, and we shall see who captures the value as open weights spread, and what Europe builds underneath them.
Read Patrick's full piece here. If you're building in open-weight models, AI infrastructure or the deep tech underneath them, we'd love to hear from you.
