For most of the past two years the AI story has had a single, simple shape: spend whatever it takes on compute, because demand is effectively unlimited. The first week of July complicated that, as a report that Meta was preparing to sell its spare computing capacity to outside customers, the first serious hint that a hyperscaler might have built ahead of its own needs, was enough to knock Micron and SanDisk down by around 11% in a session, and Samsung by 8%. The same week offered the opposite signal, as memory was still sold out into 2027, Washington had just lifted a national-security ban on Anthropic’s two most capable models, and the US administration was clearing gigawatts of new data-centre power by emergency order.
AI and Semiconductor Shares One-day Move (1 Jul 2026, %)
Source: Company data, InterCapital Research
The four largest US cloud builders, Amazon, Alphabet, Meta and Microsoft, together with Oracle, are on course to commit between USD 660bn and 690bn of capital this year, more than 60% above 2025, with something close to three-quarters of it going into AI compute, data centres and networking. A growing share of that is now funded with debt rather than out of operating cash flow, and much of it circulates within a closed loop: the chipmakers take stakes in the model developers, the developers sign multi-year cloud contracts, and the cloud providers spend the proceeds on more chips. Estimates put the interlinked commitments across that chain at well over USD 800bn. As long as demand keeps outrunning capacity the arrangement is self-reinforcing. Its weakness is that it concentrates the risk in a handful of names, so a single credible sign of over-capacity, which is what the Meta report amounted to, can re-rate the whole chain in an afternoon. That is why the claim that the industry is supply-constrained rather than demand-constrained has become the load-bearing assumption of the entire trade, and also its most exposed one.
Total Yearly Capital Expenditures (2017-2026, USD bn)
Source: Company data, InterCapital Research
The second constraint is physical, and memory is where it shows most plainly. High-bandwidth memory, the DRAM that sits next to the processor in an AI accelerator, uses far more silicon than conventional memory, so every wafer turned over to it takes two to three wafers of ordinary supply out of the market. With Samsung, SK Hynix and Micron between them controlling more than 90% of production and showing no appetite to repeat the capacity gluts of previous cycles, prices have behaved accordingly: DRAM contract prices rose 80 to 90% in the first quarter and a further 50% or so in the second, and the three producers’ combined operating profit is set to rise several times over this year. Capacity is already booked into 2027, the hyperscalers have moved from one-year to five-year supply agreements, and Intel’s chief executive has told investors there is no relief until 2028. The effect is reaching the shelf, with PC makers lifting prices 15 to 20%. This is not a shortage produced by financial positioning; it is a physical one, and it is the clearest evidence that the bottleneck has shifted from capital to capacity.
The third constraint is the state, which has started to shape the build-out from both directions. On one side it acts as an accelerant, treating AI infrastructure much as it would treat energy or transport. In the United States the Stargate programme is working towards roughly 10 GW of capacity, and the administration has used emergency powers to short-circuit permitting and connect gigawatt-scale loads to the grid, with one Michigan site cleared for 1.4 GW; the same instinct runs through the drive to bring chip production onshore, of which Apple’s USD 30bn commitment to US-made Broadcom silicon is the clearest example, and through Europe’s answer in its Chips Act and sovereign-cloud plans. On the other side the state has become a gatekeeper. When Anthropic released Fable 5 and Mythos 5, its two most capable models, in early June, the US Commerce Department imposed an export-control ban within days, restricting access even for the company’s own foreign staff, on national-security grounds and after Amazon had warned that the models’ safeguards could be circumvented. The order was lifted on 1 July, but only once Anthropic agreed to detect security risks proactively, to help write the standards for future models, and to report malicious use.
That points to a fourth constraint, the models themselves. When Fable was switched back on, users were quick to note that it had been quietly pared back from the version released three weeks earlier, a reminder that a frontier model now sits inside three tightening limits at the same time, the compute it can be allocated, the safety work regulators expect of it, and the plain cost of running it at scale. The economics underneath still look impressive, with the leading platforms handling trillions of tokens a day at unit prices that keep falling. But capability, safety and cost increasingly pull against one another, and the distance between what a model can do in a demonstration and what it can do cheaply and safely for millions of users is where the next disappointment, or the next positive surprise, is most likely to come from.
For our region the change worth noting is that Central and Eastern Europe is no longer only watching. Croatia is the obvious example: the proposed Pantheon AI campus at Topusko, a EUR 50bn, 1 GW hyperscale project backed by the transatlantic Pantheon Atlas group and unveiled at the Three Seas summit in Dubrovnik, would make the country a serious European AI node for the first time, and would expose it to all four constraints at once, from the memory bill inside the racks to the power drawn off the grid and the rules governing what is allowed to run on it.
Comparison of Selected Data Centers by IT Power (MW)
Source: Company data, InterCapital Research