China supplied the proof of concept. Meituan released LongCat-2.0, a 1.6-trillion-parameter model it says was trained entirely on domestically developed chips, with performance it claims matches Google’s Gemini 3.1 Pro. Whether or not the hardware claim fully holds up, the signal is clear: export controls are pushing rivals to build their own stack rather than wait for access.

Here is the finance angle. All of this lands during a global chip shortage that already forced Apple and Microsoft to raise device prices, while US tech giants alone plan to spend around $650bn on AI infrastructure this year. Capital is being committed at national scale, and it is increasingly duplicative by design – every major bloc now wants its own models, its own chips and its own data centres. For businesses, that means a future with more suppliers and more political risk baked into the AI supply chain, not less. The question for any company building on top of AI is no longer only “which model is best,” but “whose infrastructure am I standing on, and what happens if that access changes.”

Money is starting to follow that logic. Stripe, Anthropic and the OpenAI Foundation backed Intercept, a $500M nonprofit funding tools to block respiratory viruses and clean indoor air, with the stated goal of making routine colds and flu far rarer. The framing is deliberately economic: the group estimates people lose 15-25 days a year to routine infections, adding up to about $600B in global productivity losses. The $500M is meant to carry early research far enough that pharma firms and investors take over the expensive final stretch – a structure worth watching, because it treats public health as an infrastructure investment with a return.

The third story is the most technical, but maybe the most important. Today, there are more than 120 different AI models for biology, and until now they could not work together – each spoke its own format and refused to connect. Stanford’s Brian Hie released Proto, an open tool that finally lets researchers plug these models into one pipeline instead of running each on its own. The result is faster and cheaper: in testing, Proto found working designs 32% of the time after trying just 65 options, where the old approach succeeded only 7% of the time after trying around 1,000. And researchers do not have to write it all by hand – AI agents can write Proto programs themselves, with the team using Claude to do part of the work.

For finance and business leaders, the takeaway is not that AI cures disease this quarter. It is that the ROI clock for AI is splitting in two. One track is the fast, visible productivity story inside everyday software. The other is this slower one: research instruments that compress years of trial and error into months, in healthcare, drug discovery and biology. The second track will not show up in next quarter’s numbers, but it is where some of the largest long-term value – and some of the biggest capital commitments – are quietly being placed.