In recent years, anyone wanting to know how far AI development had progressed looked to benchmarks. Anyone wanting to know today must read financial statements, five-year plans and local politics. The first week of September provided three pieces of evidence for this, which at first glance appear to come from different sectors but, taken together, answer the same question: what is driving the expansion of computing capacity, and where are the bottlenecks?

Europe is investing capital

On 8 September, Mistral announced a funding round worth three billion euros, with a post-round valuation of over 21 billion euros. According to the company, this is the largest equity funding round ever completed by a European technology company. It is led by Samsung Electronics, together with the Scaleup Europe Fund managed by EQT and existing investor PSG Equity. Other participants include Advent, funds from BlackRock and Luxembourg, as well as previous investors a16z, ASML, Nvidia and Salesforce Ventures.

More interesting than the sum involved is the composition of the investors. A Korean electronics conglomerate is leading a funding round for a European modelling provider, whilst a Dutch engineering firm is on both the investor and client sides. Mistral states that it has more than 125 corporate clients, including Airbus, ASML and HSBC, and intends to use the funds to finance a full-stack strategy with open model weights across models, infrastructure and computing capacity.

Europe is therefore organising its technological sovereignty via the capital market. This is a swift process and does not tie up tax revenue. However, because it depends on valuations that can fluctuate within a single quarter, it is also the most fragile of the three models, which can be summarised as ‘capital for Europe’, ‘kilowatts for China’ and ‘consensus for the US’.

China pays with planning

One day before Mistral announced the funding round, China’s Ministry of Industry and Information Technology presented a plan for the years 2026 to 2030. The national AI computing capacity is set to more than quadruple to 9,800 exaflops by 2030, up from 2,185 exaflops in June – a figure that already represented a 177 per cent increase on the previous year. Plans include a cumulative 3.8 trillion yuan in infrastructure investment – equivalent to around 532 billion US dollars – as well as the systematic development of clusters, each comprising more than 10,000 accelerator cards. This will be built upon the existing network of data centre hubs, in which, according to the Ministry, 52 such facilities are already operational.

Computing power is treated here like an electricity grid or a railway line: as infrastructure whose expansion is proactively driven rather than passively awaited. The price of this approach is well known and is called misallocation. Its advantage is that no single local authority can stop it.

The US is paying the price with its consent, and that consent is running out

This is precisely where the problem now lies in those areas where expansion has proceeded most rapidly to date. According to research by the Wall Street Journal, more than ten US states have begun to withdraw tax breaks for data centres, which amount to more than one billion dollars a year across individual jurisdictions. Maine has scrapped its incentives entirely; the governors of Illinois, Massachusetts and Ohio have suspended their programmes; whilst Nebraska, Washington and North Carolina have restricted or allowed exemptions to expire. The reasons cited include lost revenue, rising electricity prices for private consumers and opposition in the affected communities.

The American data centre boom was never purely a private-sector affair. It ran on a layer of subsidies that is now thinning – not as a result of a decision in Washington, but through numerous decisions taken at the local level. That is the real takeaway from this week: it is not chips or models that are currently the critical factor, but local acceptance and the electricity price that communities are willing to pay.

What this means for the industry

Three consequences are concrete enough to be incorporated into planning. Firstly, the price assumptions: anyone calculating quotes for AI services today is working with terms that presuppose a subsidised infrastructure. If this layer is removed, the effect will, with a delay, be passed on to token and instance prices. Contracts with long terms and price escalation clauses are currently worth more than they were a year ago.

Secondly, vendor lock-in: the three funding models imply three different probabilities of failure. A state-funded provider goes under for different reasons than a venture-capital-funded one. For procurement, this does not mean choosing one camp over another, but rather keeping the switching costs between camps low. Open model weights are precisely for this reason a procurement consideration and not an ideology.

Thirdly, a company’s own sites: companies that wish to process plant and design data in-house for data protection reasons will, in future, face the same decision as the hyperscalers, only on a smaller scale: connected load, approval times and grid charges will become location factors for their own AI capacity. Anyone wishing to carry out computations in 2028 will need to speak to the grid operator in 2026.

These three reports do not provide a single answer. However, they pose the same question, and they pose it to anyone who is firmly incorporating AI into their value chain.