On the close of US stocks on November 4, the market sent a set of signals that pointed in opposite directions. According to trading data from Gate Stock, Microsoft rose 6.68% to $555.80, hitting an all-time high and lifting its market cap to $4.13 trillion. At the same time, Amazon and Alphabet also moved higher, and the five tech giants combined added nearly $1 trillion in market value in a single day. In contrast on the same day, the Roundhill Neocloud ETF fell 9.10%, CoreWeave dropped about 9.3%, IREN’s intraday decline exceeded 10%, and Nebius was down 6.4% in pre-market trading. The broader market was not weak. The Dow recorded its best single-day performance since April 2025, but capital clearly rotated within the compute-asset complex: data centers that had already been built were repriced, while expansion plans that had not yet been completed were reconsidered.
Source: Gate Stock market data
A referendum turns "Can we build it?" into "Will they let us build it?"
On November 3, local referendums related to data centers were held simultaneously across multiple US states. Ohio alone had 21 such referendums, and the US total reached 41. The results were almost unanimously against building new data centers. In Texas’s governor election, the Democratic candidate who supported pausing new data center permits won. The market quickly realized that the core assumptions that had supported the AI infrastructure narrative over the past two years—"If you have money, you can build"—were now being overlaid with a new constraint: approvals, interconnection, and permits were becoming a scarcer hurdle than capital.
This constraint did not appear overnight. As early as September 21, Texas Governor Abbott instructed the Texas Commission on Environmental Quality to pause the permitting of statewide data center projects until ERCOT completes an audit of the interconnection queue. The audit was expected to finish in mid-December. At that time, ERCOT interconnection requests had reached about 474 gigawatts—more than five times its historical peak load, around 90% of which came from data centers. In other words, Texas was not "refusing data centers." Instead, it was using the audit to separate "credible projects" from "speculative applications." But for companies that rely on continuous new construction to support valuations, the act of separating them alone was already a shock.
Existing assets turn into "annuities," while incremental expansion becomes "options"
To understand the opposite moves between Microsoft and Neocloud on the same day, you need to split data center assets into two layers. One layer is capacity that has already been built and is generating revenue. The other is planned capacity that has not started yet and depends on future financing and approvals to become real. The first layer is close to an annuity: as long as demand does not disappear, cash flows are highly predictable. The second layer is close to an option: it only has value once it is built and leased out. Otherwise, early-stage investments and financing costs will be sunk.
In the valuation of large operators, the annuity-like portion dominates. Microsoft’s Azure, Amazon’s AWS, and Google’s cloud and advertising businesses are all built on data centers that have been operating for years. When approval thresholds rise and the growth rate of new supply gets constrained, the scarcity of existing capacity increases instead. Baird analyst Michael Antonelli’s description was widely quoted in the market: these companies "already have compute power, and they have data centers. Pull up the drawbridge and you’ve basically handed over a moat for free." The core of that line is not emotion—it is a verifiable accounting reality: when supply is constrained, the reset cost and the difficulty of substitution for already capitalized assets both rise.
Neocloud’s valuation structure is almost the opposite. In pricing from companies like CoreWeave, IREN, and Nebius, a large part of the value embeds assumptions that they will keep opening new campuses next year and the year after. When both Texas’s audit and referendum outcomes point to "new permits may slow down," those assumptions’ discount rates are rapidly increased. IREN has three campuses in Texas totaling about 2.75 gigawatts, roughly half of its already locked-in power capacity; CoreWeave’s expansion model also relies heavily on debt financing. Their decline that day reflects the market repricing the probability of "executing incremental options."
The unexpected "discipline" of capital expenditures: large operators save cash, while Neocloud faces financing barriers
There is a second layer behind the rally by large operators. It can also be tested with data rather than staying at the narrative level. In 2026, the four companies—Microsoft, Amazon, Alphabet, and Meta—are expected to collectively spend between $740 billion and $800 billion in capital expenditures. That is nearly a fivefold increase from the roughly $140 billion figure in 2023. Among them, the free cash flow for Amazon, Alphabet, and Meta has already turned negative, meaning operating cash flow cannot cover capital expenditures. The gap must be filled through the debt and equity markets.
Within these numbers, Microsoft is a structural exception. In the most recent fiscal quarter, Microsoft’s free cash flow was $18.6 billion. Even though it fell 27% year over year, it was the only one among the four that kept positive free cash flow. When regulatory signals cooled expectations for "endless expansion," the market’s first reaction to large operators was not "the growth story is damaged," but "cash flow discipline is back." Dakota Wealth’s Robert Pavlik summed it up like this, and multiple media outlets quoted him: if ultra-large cloud providers no longer need to expand compute at astronomical scale, free cash flow should recover; and the party slamming the brakes is not the companies themselves, but the government.
For Neocloud, the same "brake" means something entirely different. Large operators reduce the marginal outflow of their own cash. Neocloud, meanwhile, faces a narrowing in marginal financing channels. CoreWeave’s expansion depends heavily on the credit market, IREN’s capex guidance requires continuous support from external funding, and Applied Digital still has hundreds of megawatts of capacity waiting for financing to translate into builds. As regulatory uncertainty rises, the risk premium that lenders and equity investors demand for "planned capacity that has not generated revenue yet" rises in tandem. What large operators save is cash; what Neocloud loses is leverage.
Demand hasn’t disappeared—what changes is "who pays" and "who bears the risk"
A counterexample that’s easy to overlook is that on November 4, the data center leasing market did not freeze. Riot Platforms leased its entire 910-megawatt Corsicana campus in Texas for a 15-year term to a cutting-edge AI lab. The initial contract value was about $28 billion. Riot handles the land and power infrastructure, while the tenant invests in and operates the compute modules. On that basis, Riot’s investment comes to about $6 million per megawatt. On the same day, Google agreed to provide credit support for Anthropic’s 401-megawatt, 20-year leasing obligations tied to TeraWulf’s Kentucky project, but the construction funding and construction risk still fall on TeraWulf.
The structure of these two deals points to the same trend: demand has not vanished, but the parties "paying to build compute power" and "bearing approval and financing risks" are separating. The model that provides only land and power has a lower per-unit investment and a smaller risk exposure—but it also caps upside. The party that takes on construction financing can lock in long-term leases, but it must absorb capital costs amid regulatory uncertainty. This is one side of the "license business": raising the bar affects not only "whether you can build," but also "who has the balance sheet to sit through the time required to obtain approvals."
Conclusion
Microsoft’s historical high and Neocloud’s broad-based selloff should not be reduced to a simple style rotation where large-cap stocks outperform small-cap stocks. A more accurate description is this: when data center construction shifts from being capital-driven to being permit-driven, the value gap between existing compute capacity and incremental plans widens systematically. Assets that are already built, already interconnected, and already generating revenue receive a scarcity premium. Expansion stories that depend on continuously securing new permits are forced to confront higher discount-rate hurdles. The next checkpoint is clear: ERCOT’s audit conclusion in mid-December will determine whether Texas’s pause is just short-term friction or a longer-cycle constraint. The financing cost and contract-signing pace in Neocloud’s next earnings report will test whether "incremental options" remain executable. Signals from the demand side have not reversed yet, but the center of pricing power is shifting—from "who can borrow money" to "who already has permits."
FAQ
Why did Microsoft hit a new high when expectations of tighter regulation increased?
It’s not about "regulation being positive for Microsoft." It’s about Microsoft’s valuation being driven primarily by cash flows from already built assets. When the growth rate of new supply gets suppressed by higher approval thresholds, the scarcity of existing compute increases. At the same time, expectations for capital expenditure expansion cool, improving the visibility of free cash flow. Microsoft is the only one among the four ultra-large cloud operators that still has positive free cash flow; in the most recent quarter it was $18.6 billion, putting it in the best position for the "spending discipline" narrative.
Is Neocloud’s drop because its data centers are no longer valuable?
No. IREN does have about 2.75 gigawatts of power capacity already locked in in Texas, and CoreWeave also has operating contracts. The decline reflects the market repricing the incremental assumption of "keeping opening new campuses." When the pace of permits becomes uncertain, future revenues tied to planned capacity that has not started must be evaluated with a higher discount rate, and that incremental portion makes up a much larger share of Neocloud’s valuation than it does for the large operators.
What are the key timing points for the Texas audit?
ERCOT’s audit of the interconnection queue is expected to be completed in mid-December 2026. The goal of the audit is to distinguish credible projects from speculative applications. The findings will determine whether the pause in permits is only short-term friction or a longer-cycle constraint. Before that, existing interconnection contracts and projects already under construction are generally not affected by a pause in new permitting, but the queue time for new applications will be extended.
Does Riot’s $2.8 billion lease indicate data center demand is still strong?
It shows demand exists, but the deal structure has changed. Riot provides only the land and power infrastructure. The tenant invests and operates the compute modules. Riot’s investment is about $6 million per megawatt, far lower than the capital intensity of building compute in-house. In this expansion, more risk is shifting to the tenant and the party providing construction financing, rather than being borne by the land and power provider. Demand hasn’t disappeared, but the distribution of "who pays" and "who bears approval risk" is changing.
