Thursday's session erased roughly $2.3 trillion in market value — the worst single day on record — and not one data-center permit had yet been denied.
That is the whole paradox in one line. Voters approved dozens of ballot measures calling for limits on data-center expansion; Democrats swept both chambers of Congress on a platform that included slowing construction. But a ballot measure is a signal, not a stop-work order. The market repriced the political risk anyway, in hours, before a single regulator acted.
Last week this column argued that the binding constraint on the AI buildout had migrated from chip supply to credit structure. This week it moved again — to the permitting regime. The $800 billion in 2026 capex that Big Tech guided toward did not disappear. It acquired a price list.
The $800 billion in 2026 capex did not disappear. It acquired a price list.
The credit market delivered a split verdict before the equity market finished vomiting. It cost roughly 1,100 basis points a year to insure CoreWeave's debt against default, against 139 basis points for Nvidia and 298 for Oracle, according to pricing CNBC reported. Those spreads are not pricing a demand divide. They are pricing a political one.
Nvidia sells into whoever builds. CoreWeave has to finance the building. If a Texas county commission can add a construction condition and push a 2028 energized date into 2029, the cash-rich hyperscaler absorbs the delay and the debt-funded neocloud eats it. The credit market figured that out before the analysts did.
Morgan Stanley's own numbers make the mechanism legible. The bank told clients to buy the drop and kept its hyperscaler capex forecasts unchanged at $1.6 trillion for 2027 and $1.8 trillion for 2028. But it also expects 20% of US data-center projects planned for 2028 to slip into 2029 — about 5 gigawatts of capacity.
US additions now decline 4% in 2028 before jumping 38% in 2029. Globally, additions grow 9% in 2028 and 20% in 2029. And a shortage of more than 30% in powered data-center capacity persists through 2028.
Read those four numbers together and the thesis writes itself. Political friction is being absorbed as delay, not cancellation. That puts a floor under lease rates for existing compute and a ceiling on new volume. It extends the investment cycle rather than ending it.
The friction has an address. Morgan Stanley pins the concern largely on Texas and Ohio, where newly elected governors could introduce additional requirements for data-center construction. Projects already under construction are not expected to be disrupted, which is why the 2027 forecast holds.
The 2028 cohort is the one that pays. A project that has not broken ground by the time a new governor signs an executive order is a project that can be renegotiated, conditioned, or simply left in the queue while the developer waits for a friendlier jurisdiction.
Which is how Tennessee and Louisiana become the new permitting arbitrage. Morgan Stanley flags both as destinations for projects displaced by less accommodating states. The bank also sees as much as 20 gigawatts of additional development potential across Asia and EMEA — the offshore escape valve for the same political logic.
This is not a story about AI demand. Cloud infrastructure spending hit $168 billion worldwide in the third quarter, up 55% year over year. There were 7,481 megawatts of North American capacity under construction, more than 80% of it already preleased.
AI capex added 0.9 percentage points to annualized US real GDP growth in the third quarter — a little over a quarter of the entire 3.4% pace, by Morgan Stanley's estimate. That contribution was 0.2 points in 2024 and 0.5 in 2025.
The demand is not the variable. The permit is.
The timing of the shift matters. FactSet puts calendar 2026 capex guidance across the largest cloud and AI spenders close to $800 billion including finance leases and customer prepayments. Alphabet raised its guidance to $180–190 billion. Meta raised to $125–145 billion. Cash capex for fiscal 2026 is expected to exceed $690 billion — growth of more than 80% year over year, after roughly 70% the year before.
That buildout is now outrunning internal cash. Incremental debt rose to 32% of capex in the twelve months to June 2026, from 9% in fiscal 2024, taking aggregate debt across the group to about $700 billion. Free cash flow is expected to approach zero or turn negative this fiscal year for all of them except Alphabet and Microsoft.
So the hyperscalers are borrowing to build into a permitting regime that can add a year to any given project, while the marginal cost of that debt is set by credit markets that have already started to discriminate. That is the squeeze. Not a demand problem — a sequencing problem with a financing accelerator attached.
The concession economy is the answer, and its currency is not dollars alone.
Every project now carries a community bill: water rights, grid upgrades, local tax abatements, and something newer and harder to price. The OpenAI-style cost-allocation and financial-backstop promises — the arrangements in which the company standing behind the compute commits to cover the obligations if the offtaker cannot — turn a hyperscaler's balance sheet into municipal insurance.
That is the trade. The state grants the permit; the developer absorbs the grid upgrade, guarantees the water, accepts the tax terms, and signs a backstop that converts political risk into contingent corporate liability. Tennessee and Louisiana are not winning because they love AI. They are winning because they will write a cheaper version of that contract.
The tier that cannot write it gets shaken out. Morgan Stanley expects AI-related debt issuance to rise in 2027 but warns that new data-center developments and neocloud operators will face greater scrutiny from lenders. Lenders like ING have already been distancing themselves from data-center loans.
A hyperscaler with $700 billion of aggregate sector debt and near-zero free cash flow can still roll a delayed project into the next fiscal year. A neocloud paying 1,100 basis points for default insurance cannot roll a delayed energization into anything except a restructuring.
This is where the political ceiling and the credit floor meet. The delays Morgan Stanley forecasts are "ultimately healthy for the AI investment cycle," in the bank's framing — they smooth the supply-demand imbalance, support lease rates, and extend the investment period. That is true for the survivors. It is a different sentence for the leveraged.
One cross-reference is worth noting: the 5 gigawatts of 2028 projects slipping to 2029 is the demand-side mirror of the power-project permitting gap covered separately in this issue. The difference is that here the cause is electoral. The interconnection queue is a mechanical delay. A governor's construction requirement is a political one, and political delays are negotiated, not engineered.
The market took one session to reprice that distinction. It will take the industry until 2029 to pay for it.