AI News

US Data Center Construction Slows: Power, Ballots, and Cash Flow—Three Constraints Rewrite the AI Race

TMTPost · AGIUpdated Nov 16 21:06

AI summary

A TMTPost analysis points out that US data center expansion is simultaneously hitting three constraints—power, public opinion, and cash flow. In the 2026 midterm elections, all 27 restrictive proposals passed, and the Philadelphia Semiconductor Index fell 12.25% in a single day on November 5. The article states that AI training racks consume 30 to 100 kilowatts or more, and that by 2030 US data center demand will be about 75 gigawatts while committed supply is only 58 gigawatts—a 17-gigawatt gap; Microsoft has already been unable to deliver part of an $80 billion Azure order due to power shortages. The author argues that the construction slowdown turns a possible stall into a controlled descent, and whether capital expenditure growth in 2027 is 54% or 12% will become a key indicator to watch.

Why read this

AI score 70

The article links three constraints—power, community votes, and cash flow—into a single thread, explaining why US data center expansion collectively decelerated in 2026.

Topics

Tags

Body · AI translation

Over the past two years, the scarcest resource in the AI industry was chips; starting in the fall of 2026, the answer became electricity and votes.

On the night of the U.S. midterm elections in early November, something rare occurred: for the first time, data centers became a nationwide campaign issue on a large scale. Of the 41 local ballot measures directly related to data centers, all 27 restrictive measures passed, and all 13 pro-development measures were rejected, with the industry winning only one race, in Harwood, North Dakota. The capital market's reaction came fast and hard: on November 5, the Philadelphia Semiconductor Index plunged 12.25%, its largest single-day drop since March 2020, while the Nasdaq fell 5.5% the same day, and the so-called "Magnificent Seven" shed roughly $1 trillion in market value in a single day.

The market's violent reaction easily brings to mind the familiar "AI bubble bursting" narrative. But looking a bit further reveals another thread: this is not a simple valuation correction, but the first time the expansion of AI infrastructure has simultaneously run into the triple constraints of electricity, public opinion, and cash flow.

A shift in constraints: from scrambling for chips to scrambling for power

The power consumption of AI racks is a full order of magnitude higher than traditional loads: a single AI training rack draws 30 to 100 kilowatts or more, whereas traditional server racks typically draw only 5 to 15 kilowatts. The International Energy Agency (IEA) estimates that global data center electricity consumption will more than double by 2030, reaching about 945 terawatt-hours, slightly more than Japan's current annual electricity consumption; in the United States, data centers will contribute nearly half of the growth in electricity demand over the same period.

The problem is that the chip industry iterates at the pace of Moore's Law, while the power system expands at the pace of infrastructure construction. Lead times for key grid components such as transformers and cables have doubled over the past three years; in 2026, delivery cycles for substation transformers are generally more than three years, and building a new transmission line often takes years.

The speed gap shows up directly on project books: part of Microsoft's $80 billion Azure order backlog cannot be delivered precisely because of power shortages, as Nadella has confirmed on earnings calls. The scale of the supply-demand gap is not small: according to industry estimates, U.S. data center demand is expected to reach 75 gigawatts by 2030, while committed supply is only 58 gigawatts, a gap of 17 gigawatts. This means the power problem is not a failure at some single link, but a systemic bottleneck running through the entire chain of generation, transmission, and distribution.

An easily overlooked detail is the efficiency paradox. Google disclosed that the median energy consumption of its Gemini text prompts fell 33-fold from a year earlier to 0.24 watt-hours; meanwhile, its disclosed monthly token processing volume has risen from 480 trillion at the 2025 I/O conference to more than 3,200 trillion by the 2026 I/O conference. A single inference has become cheaper, yet total demand is larger—this is precisely the Jevons paradox replaying in the computing power domain.

Hoping that technological progress alone will close the power gap is unrealistic at present. In fact, efficiency gains change unit costs, not the total demand curve; when costs fall, previously uneconomic application scenarios are activated, and demand is released even faster. To judge whether this round of expansion has bottomed out, the key is not energy efficiency figures, but the speed at which the supply-demand gap narrows.

A veto at the ballot box

If the power gap is a technical problem, community opposition is turning the technical problem into a political one.

The ranking of objections is clear: rising electricity bills, noise pollution, and strain on public infrastructure, with electricity bills first.

Under current U.S. electricity rules, the cost of grid expansion is shared by all users in the region, and the cost of computing power expansion ultimately shows up on ordinary households' electricity bills. Capacity auction prices in the PJM grid rose from $28.92 per megawatt-day to $269.92, and estimates show that about 63% of the increase was driven by computing power demand. When costs are deducted directly every month, voters naturally want to reclaim pricing power through their ballots. According to the research firm DataCenterWatch, active anti-data-center groups in the United States increased from 396 at the end of 2025 to 833 in the first quarter of 2026, spread across 49 states; in the first quarter alone, at least 75 projects involving about $130 billion in investment were blocked or delayed, matching all of 2025.

Voting results on data center-related measures in U.S. local referendums on November 3, 2026 (data sources: Ballotpedia, CNBC; under another tally, 44 of 45 measures leaned toward restrictions)

The reason this referendum result matters is that it established the precedent that a community veto can override capital budgets. Ohio was the focus of this round, with 18 municipalities voting on 21 measures, and the results leaned almost entirely toward restrictions; local media vote counts showed most restrictive measures won more than two-thirds of voter support. A Gallup survey in March 2026 showed that 71% of Americans opposed building AI data centers near their homes, with 48% strongly opposed, and this cut across party lines; in NBC's exit poll on Election Day, November 3, the figure was 76%. What people voted against was data centers, not AI itself; residents do not reject the technology, they simply do not want to bear its costs alone.

Institutional contraction unfolded in parallel. After November, community resistance escalated from scattered complaints into an organized, institutionalized force. New York State has already implemented a statewide moratorium on facilities of 50 megawatts or more; as of early November, related restrictive measures had appeared in 46 states. Goldman Sachs estimates that states potentially facing policy constraints account for about 45% of planned data center capacity nationwide, with Texas the largest concentration, accounting for about one-fifth of planned capacity nationwide. The focus of the contest is shifting from whether to build to who pays and how compensation is made.

Nuclear power and gas: two paths through the transition period

Faced with grid queues, hyperscale cloud providers have turned their attention to self-supplied power, and nuclear is the most favored option among them. A report released by the International Energy Agency at the end of October noted that "queue inflation" in grid connection applications has become widespread across major global computing markets, with the cumulative capacity of projects awaiting connection in several regions exceeding the local existing system peak load; the IEA estimates that if data centers were allowed to curtail electricity use during periods of tight power supply, an additional 750 to 900 gigawatts of grid hosting capacity could be freed up globally (IEA "Energy and AI" report).

Microsoft signed a 20-year power purchase agreement with Constellation to restart Three Mile Island Unit 1, with 835 megawatts of capacity expected to reconnect to the grid in 2027; Google committed to purchasing up to 500 megawatts of small modular reactors (SMRs) from Kairos Power; Meta reached cooperation agreements totaling up to 6.6 gigawatts with Vistra, TerraPower and Oklo, among which the 20-year power purchase agreement with Vistra covers two operating nuclear power plants. The nuclear cooperation announced by multiple companies totals about 10 gigawatts. The Palisades plant in Michigan has already reconnected to the grid and resumed generating power, becoming the first operating unit in the United States to be restarted; together with Three Mile Island, which is scheduled to connect to the grid in 2027, it forms the main force of this round of nuclear restarts.

But nuclear power faces a timing mismatch: large-scale power supply from SMRs will not arrive until the 2030s, and before the end of this decade, the only real path to connecting data centers is basically restarting existing units. The task of filling the gap has fallen to natural gas. According to a report in July this year by the environmental group Environmental Integrity Project, the United States has planned at least 74 behind-the-meter gas-fired power plants dedicated to data centers, totaling about 143 gigawatts; once built, they will emit 662 million tons of greenhouse gases per year, roughly equivalent to adding 140 million cars.

Nuclear power is the long-term answer, gas is the transitional cost, and the connecting window between the two is precisely 2027 to 2030. The risks in this window are equally obvious: if nuclear power delivery falls short of expectations and gas-fired units are slowed by environmental pressure, the power gap will widen further, and the construction cycle of data centers will be forced to synchronize with the grid investment cycle—a scenario no cloud provider wants to see.

Capex shifts gears: from a sprint to an orderly slowdown

The constraints of power and public opinion will ultimately be transmitted to the financial statements.

The third-quarter reports have already given the signal: Alphabet's quarterly capital expenditure rose 114% year on year to $51.4 billion, and free cash flow turned from positive to negative $9.1 billion; Meta's free cash flow turned negative, and it raised up to $95 billion in additional stock; Amazon's trailing-12-month free cash flow gap widened to $42.4 billion. Among the five major companies, only Microsoft maintained positive cash flow.

The market is also repricing: Microsoft announced before the market open that it would cut its 2026 capital expenditure forecast by $5 billion to about $170 billion, and was reviewing and exiting multiple third-party computing projects; the primary market tightened in tandem, with OpenAI lowering the target for its latest funding round from $30 billion to $20 billion and its target valuation from $1.4 trillion to $1.25 trillion; rating agencies also shifted, with S&P downgrading Meta's rating outlook to negative. The stage of funding infrastructure with operating cash flow is over, and the era of financing support has begun. According to Morgan Stanley estimates, AI-related companies issued about $466 billion in bonds this year, far above about $101 billion in the same period of 2025.

As financing structures such as bond issuance, additional share offerings, SPVs and lease commitments are activated one by one, the market has begun to ask the same question: where exactly is the payback period for these investments?

Free cash flow of hyperscale cloud providers disclosed in the latest financial reports (USD 100 million). Note: Amazon is on a trailing-12-month basis (data source: company financial reports, Nomura Securities)

At such a rate of cash burn, a slowdown in construction may not be a bad thing.

Hyperscale cloud providers' capital expenditure plans for 2026 have been raised to about $800 billion (Goldman Sachs basis), nearly double the about $410 billion in 2025, while the global figure tracked by the International Energy Agency exceeded $700 billion for the first time; continuing to expand at this slope would only make the financing cliff and inflationary pressure greater. The constraints have precisely acted as a speed-reduction valve: the triple limits of politics, power and financing spread investment over more years, turning a possible stall-induced turn into a controlled descent and buying time for the monetization of AI applications.

Data centers were always heavy assets requiring long-cycle payback, and building them more slowly does not change their economic model. More importantly, the slowdown alleviates two risks previously masked by high growth: first, it avoids excessive supply of computing power before demand matures; second, it leaves time for supporting infrastructure such as grids and nuclear power to catch up, allowing the supply curve and the demand curve to meet more smoothly.

Scale of U.S. hyperscale cloud providers' capital expenditure and institutional forecasts for 2027 (data source: company disclosures, Goldman Sachs, UBS; 2026 is an approximate Goldman Sachs figure)

Two paths, the verdict in 2027

Pulling back the lens, China and the United States are answering the same question in two ways.

The United States relies on ballots and market bargaining, with costs priced after repeated tug-of-war between capital and public opinion; China takes the administrative planning route: the National Development and Reform Commission and three other departments require that new data centers at national hub nodes draw more than 80% of their electricity from green sources, the China Electricity Council estimates that nationwide computing electricity consumption will be about 800 billion kilowatt-hours in 2030, accounting for about 6% of total social electricity consumption, and CGN has reached green power cooperation agreements totaling more than 3 billion kilowatt-hours with Huawei, Tencent and others. The efficiency and costs of the two paths will determine the direction of the global computing landscape in the coming years.

Combining the two threads of constraints and funding, the trajectory in 2027 roughly has three possibilities.

The baseline scenario is an orderly slowdown. Goldman Sachs forecasts that capital expenditure growth will fall from nearly 100% in 2026 to 54% in 2027 and 12% in 2028, with total investment still climbing (about $1.2 trillion in 2027) but the slope clearly flattening. The upside scenario is that monetization catches up, with AI revenue growing faster than capital expenditure contracts, forming a positive cycle; the downside scenario is that the pillar gives way—AI infrastructure is already the largest single increment in U.S. corporate fixed investment, and if capital expenditure decelerates sharply rather than slowing, the economy will lose its most critical source of new demand.

Policy is also sending signals: the U.S. Senate has held a procedural vote on a bill requiring new data centers to bear the cost of added grid capacity; once enacted, cost internalization will further raise the threshold for expansion. As for which scenario holds, it is enough to watch two numbers: if the annualized growth rate of U.S. data center electricity demand falls below 5%, the constraint thesis is disproved; if AI capital expenditure growth slides below 25%, the probability of the downside scenario rises significantly.

In 2027, whether the U.S. data center power gap narrows or widens, and whether AI capital expenditure growth is 54% or 12%, are hard indicators that can be verified quarter by quarter. The AI race has no pause button; only the scoring method has changed: the deciding factor in the second half is in power plants, in communities, on balance sheets—not in GPUs. (This article was first published on TMTPost APP; author | Silicon Valley Tech-news, editor | Jiao Yan)

Source

TMTPost · AGI