The most honest moment in Jensen Huang's GTC Berlin remarks this week was not the claim that demand for neoclouds is "off the charts" — it was the admission that the companies supplying AI compute are running short of working capital, and that in Europe they are also running short of electricity. For anyone planning AI infrastructure purchases or supplier contracts, the takeaway is blunt: the constraint on AI is no longer chip supply. It is capital structure and grid capacity.
Huang, speaking to CNBC on the sidelines of Nvidia's GTC conference in Berlin on October 22, said demand for the services of neoclouds — the new breed of cloud providers that buy AI servers in volume and lease the compute to AI companies — is "through the roof" and that their prospects are "exceptional." Operators, he said, are "completely sold out." But he added that the missing link is working capital: "Demand for capital far exceeds its availability," while insisting "this problem will be solved."
The market is not convinced, and the gap between the CEO's framing and the tape is the story. Over the past month, CoreWeave — one of the largest neoclouds and a heavy user of debt to finance data-center construction — has lost more than 27 percent of its value, with another 4 percent drop in premarket trading on October 22, per Oninvest. Netherlands-based Nebius fell 26 percent over the month and another 3.6 percent; IREN and Applied Digital each fell more than 4 percent ahead of the open. The sell-off reflects investor anxiety that returns on AI infrastructure spending will take longer than expected, and specific concern about how much of the buildout is being financed with leverage. AInvest's framing is apt: Nvidia's growth now runs through the bond market.
Huang's own actions reveal the shape of the problem. Nvidia has begun guaranteeing utilization floors for neocloud partners such as Sharon AI, agreeing to rent idle capacity if customer demand falls short — effectively backstopping the very business model the market is questioning. The company has also committed nearly USD 10 billion to European companies, universities and research institutes, including equity stakes worth more than USD 1 billion in five British startups announced last week in London, plus investments in French startups.
In Europe the bottleneck is double. Power prices are higher than in the United States, and grid connections are the binding constraint. The day after Huang spoke, Capacity reported that UK grid constraints could push Nscale's flagship Essex data-center project — whose anchor customer Microsoft recently withdrew — into a wait for power that could run into the early-to-mid 2030s; Loughton was first announced in January 2025 as a 50MW site scheduled to go live in Q4 2026. At a Capacity roundtable, 70 percent of infrastructure leaders said they expect grid connections to take more than 18 months. Huang's prescription for Europe — build local neocloud infrastructure so that an AI economy can form a closed loop — collides with a physical reality in which a missed grid date can mean a lost customer. Nvidia's answer on its own side is the DSX pod, a containerized AI factory unit built within the supply chain, with the first unit expected toward the end of the year; Huang noted build timelines have compressed from three years to customers expecting sites running within six months.
Read the whole picture and the B2B implication is clear. The next phase of the AI infrastructure race is not about whether Nvidia can ship chips; it is about whether neoclouds can finance their debt, secure power and convert "completely sold out" demand into delivered capacity. Huang is right that the absence of demand is not the problem — but for the companies holding the debt and waiting on the grid, that is precisely what makes the moment precarious.
There is a second-order risk Nvidia is quietly carrying. If neoclouds run into operating distress, the chipmaker faces a double hit: chip sales slow as capacity stalls, and the financial support Nvidia has extended — utilization-floor guarantees, equity stakes, co-investment — could suffer losses, as Chinese financial media noted in summarizing the Berlin remarks. That linkage turns Huang's reassurance that "this problem will be solved" into a testable statement: it will be solved only if capital markets reopen for neocloud financing and if grid projects stop slipping. The UK Nscale case shows how fragile the second condition is; the CoreWeave sell-off shows how fragile the first has become. For buyers of AI compute, supplier due diligence now has two columns that matter more than GPU counts: the balance sheet (how much of the buildout is debt-financed) and the grid connection (when the power actually arrives). Everything else is marketing. For sellers of compute, the same two columns determine which counterparties will still be standing when the next financing cycle tightens — the best demand in the world is not collectible from a provider that runs out of working capital mid-build.
