NextFin News - Pierre Ferragu, the managing partner at New Street Research who has been among the most prominent bulls on Nvidia, is flagging a threat that runs the other way: rising local and political opposition to AI data center construction could slow the very infrastructure buildout that has powered the chipmaker's record revenue. The warning, delivered in a television interview on August 27, lands as the four largest hyperscalers are on pace to pour roughly $725 billion into AI infrastructure this year and as Nvidia trades in the low $210s, below the all-time closing high of $235.47 set in May.
Ferragu's point is simple in structure and unsettling in implication. Nvidia's data center revenue rose 117% in the quarter reported August 26, to $89.0 billion, on total revenue of $96.2 billion. That growth is not an abstraction: it is the mechanical result of cloud providers, AI labs and sovereign buyers taking delivery of Blackwell Ultra racks and installing them in new facilities. If the facilities cannot be permitted, powered or connected — because voters, city councils or state executives say no — then the chips have nowhere to go.
The opposition is measurable, and it is spreading faster than most investors appreciate. Gallup's first poll on the subject, conducted March 2-18 among 1,000 U.S. adults with a margin of error of plus or minus 4 percentage points, found 71% of Americans oppose building an AI data center in their local area, with 48% strongly opposed; only about one in four favor it. Opposition to a nearby data center ran higher than opposition to a nearby nuclear power plant, 71% to 53% — a striking reversal of decades of public attitudes toward nuclear energy.
The political response has moved from protest to policy. On April 7, voters in Port Washington, Wisconsin — a lakeside town of about 12,000 people — approved the nation's first anti-data center referendum with 66% support. The measure requires voter approval of tax incremental financing for projects with a base value or cost above $10 million. The Vantage Data Centers campus under construction there is set to receive $458 million in tax benefits. In June, Holyoke, Massachusetts became the first community in that state to ban AI data center development over 12 megawatts, in a 9-4 city council vote; Seattle's council passed an emergency one-year moratorium on facilities above 20 megavolt-amperes; Socorro County, New Mexico adopted a yearlong moratorium; and Monterey Park, California passed three ordinances blocking data center development.
The resistance has now scaled to the state level. A state-level policy tracker updated August 24 counted 18 active state bans or moratoriums, with 25 of the 50 states either restricting data centers or considering restrictions. New York became the first state to impose a statewide pause when Governor Kathy Hochul issued Executive Order No. 62 on July 14, halting permitting reviews for data centers consuming at least 50 megawatts while regulators develop a comprehensive framework. On August 18, Pennsylvania Governor Josh Shapiro signed Executive Order 2026-05, barring the state environmental agency from reviewing a data center permit unless the developer has secured local approval and committed to responsible-infrastructure standards. Roughly a dozen other state legislatures are considering similar measures.
The Transmission Mechanism: No Power, No GPU
The chain from a town-hall protest to Nvidia's revenue line is longer than a single link, and each link matters. First, a project is delayed or blocked — a project tracker maintained by AI company 10a Labs counted 20 data center projects worth about $98 billion blocked or delayed between March and June 2025 alone. Second, delays raise the cost of capital and force developers to re-bid power contracts. Third, hyperscalers that have been committing to ever-larger capital programs begin to sequence deployments more cautiously when the physical sites are not ready.
The capital numbers involved are large enough that even a modest slowdown would matter. The four largest hyperscalers are on pace to spend approximately $725 billion on AI infrastructure in 2026, up 77% from about $410 billion in 2025 and roughly three times the $226 billion spent in 2024, with Amazon, Alphabet, Meta and Microsoft accounting for the bulk. Consensus 2027 forecasts put the four at a combined $934.5 billion — Alphabet $284.8 billion, Amazon $256.5 billion, Microsoft $207.6 billion, Meta $185.6 billion — within reach of a $1 trillion annual milestone.
But the financing is getting harder. Alphabet raised its 2026 capital expenditure forecast to $205 billion, and the consensus figure for Amazon climbed to $207.4 billion. Amazon's long-term debt rose 81% to $119 billion between December 31 and March 31, and Alphabet's debt rose 111% to $98 billion in the first six months of 2026, while Alphabet turned cash-flow negative in the second quarter for the first time. When the companies that buy Nvidia's chips are funding their expansion with more debt and thinner free cash flow, the tolerance for delayed sites shrinks.
Cyclical Wave or Structural Constraint?
This is the central call, and it cuts against the easy read. The surface interpretation is cyclical: a political wave will crest, utilities will build more capacity, and the AI buildout will resume its trajectory. That view has history on its side. Infrastructure booms — railroads, highways, telecom fiber — have always run into local resistance, and most projects eventually get built somewhere else.
The structural counter-argument is stronger. Data centers are not footloose in the way factories are. They need three things simultaneously: cheap, abundant power; large volumes of water for cooling; and proximity to fiber trunks and low-latency networks. A site that satisfies all three is scarce. When a community vetoes one, you cannot simply move the project to the next county without re-doing the interconnection queue, the water rights and the power purchase agreement — processes that run in years, not months.
The polling shift supports the structural read. Echelon Insights found in June that voters opposed an AI data center in their community by 62% to 27%, and still opposed it 58% to 31% after being shown messaging about economic and technological benefits. A representative of the firm said opposition to new data centers rose to 61% in its June-July survey from 49% in February-March — a 12-point move described as the largest shift on any AI question the firm has measured, and not confined to one party: 69% of Democrats, 54% of Republicans and 53% of independents oppose new local data centers.
When opposition survives pro-growth messaging and crosses party lines, it is not a transient sentiment. It is a new constraint on the industry's geography.
The Second-Order Effect: Permitting Becomes the Capex Valve
The first-order effect of opposition is delayed projects. The second-order effect, which the market has not fully priced, is a change in how hyperscalers commit capital. For the past two years, the dominant assumption has been that if Nvidia can ship chips, the buyers will find a place to install them. Ferragu's warning inverts that: if the buyers cannot secure sites, their ordering discipline tightens first, and Nvidia's order book — which Chief Executive Jensen Huang said in a GTC address reflects "at least $1 trillion" of demand through 2027 — becomes the variable at risk, not just the installation schedule.
"I see, through 2027, at least $1 trillion," Huang said at the company's technology conference, citing cumulative demand for Nvidia compute.
There is already evidence of the tension. A May industry report showed $46.5 billion in U.S. data center construction starts through March 2026 and a $63.2 billion late-stage pipeline expected to start between April and December — activity that remains elevated. But in Seattle, two developers withdrew from AI data center plans amid grid limits and community pushback, and the city council approved a one-year moratorium. The market is simultaneously pricing a trillion-dollar buildout and watching individual projects die at the zoning board.
New Street's own research has tracked the scale of the commitment: the firm noted that Nvidia had added $500 billion of orders since October 2025, putting the company on a run rate of more than $1 trillion per year in orders. With that much contracted demand already in hand, a handful of local referendums cannot change the revenue trajectory for the next several quarters. The chips that will ship in fiscal 2027 are already ordered; the sites for them are already being prepared.
There is also a substitution argument. If hyperscale campuses face resistance, the industry can move workloads to smaller, distributed facilities — in early-stage experiments, even to fractional nodes attached to residential construction. And sovereign and enterprise demand, the fastest-growing customer segment, is less exposed to local U.S. zoning politics than commercial hyperscale campuses.
The Counter-Thesis: A Speed Bump, Not a Wall
The strongest case against Ferragu's warning is straightforward. Nvidia's order book is deep, Blackwell Ultra is ramping, and the revenue for fiscal 2027 is largely locked in by commitments already made. A wave of town-hall opposition does not cancel a purchase order signed six months ago. On the near-term horizon, the counter-thesis likely wins.
But it depends on one assumption: that the opposition stays local and project-specific. If it scales to the state and federal level — as New York's statewide executive order, Pennsylvania's recent action and roughly a dozen pending state bills suggest — the substitution argument weakens, because there is no jurisdiction left to substitute into. A federal moratorium proposal announced in March by Senator Bernie Sanders and Representative Alexandria Ocasio-Cortez and introduced in the House in June would pause construction of AI data centers with peak loads above 20 megawatts or liquid cooling systems until Congress passes comprehensive AI legislation.
The Falsifying Signal
The specific signal that would prove the structural-constraint thesis wrong is concrete: if, over the next two quarters, the count of blocked or delayed data center projects falls back toward its pre-2025 baseline while hyperscaler capital expenditure guidance for 2027 is revised upward by more than 10% from current consensus, then opposition has proven to be a negotiating cost rather than a binding constraint. Until that happens, the burden of proof sits with the bulls.
Outlook: Three Horizons, One Divergence
The near-term, medium-term and long-term readings point in different directions, and that divergence is the story.
In the short term — the next one to two quarters — the counter-thesis likely wins. Nvidia's order book is deep, Blackwell Ultra is ramping, and fiscal 2027 revenue is largely locked in by commitments already made. A wave of local opposition does not cancel a purchase order signed six months ago.
In the medium term — fiscal 2028 and beyond — the permitting constraint begins to bite. Hyperscalers facing debt-funded capital programs approaching $1 trillion annually cannot absorb indefinite delays. If sites are not ready, deployments slip, and new orders get pushed out. This is where Ferragu's warning carries the most weight: not as a call on next quarter's earnings, but as a call on the durability of the order cadence.
In the long term, the structural question is whether the industry finds a political accommodation — state preemption laws that override local vetoes, or a federal permitting framework — or whether the buildout is permanently rerated to a slower, more contested path. New York's statewide pause and the dozen states considering similar action suggest the accommodation is not yet in place.
The beneficiaries and the exposed are not symmetrical. Companies that own power assets, transmission infrastructure and cooling technology stand to gain from a world in which power, not chips, is the scarce input. Nvidia's ecosystem — the company most directly exposed to the pace of data center deployment — carries the other side of that trade. A slowdown in facility construction does not hurt the chip designer immediately; it hurts the order book before it hurts the income statement.
Three signals are worth watching: the quarterly count of blocked or delayed projects; the 2027 capital expenditure revisions from the four largest hyperscalers; and any state or federal permitting legislation that preempts local vetoes. If the first two rise while the third fails to materialize, the structural constraint is confirmed. If permitting reform passes, the thesis breaks.
The market has spent two years treating data center construction as an engineering problem. Ferragu's warning is that it has become a political one — and politics does not respect a deployment schedule.
Explore more exclusive insights at nextfin.ai.
