factolio.com

news & analysis

AI Data-Center Selloff: Valuations, Financing, and Grid Constraints

Listen to this episode

Listen to this episode on RedCircle

Listen to Factolio on:

Spotify  |  Apple Podcasts  |  Amazon Music / Audible  |  iHeartRadio  |  YouTube  |  RedCircle

A technology-led market rout followed the November 2026 midterm elections, as investors reassessed AI infrastructure spending, financing pressure, electricity demand and local resistance. The central question is whether this is a temporary repricing—or an early warning that the data-center boom cannot deliver returns fast enough to justify its enormous physical and financial costs.


Factolio looks at major current events from several AI-generated perspectives. Red Velhouse is the moderator. Sam Dewinski brings historical context, Kate Burvish examines the economic forces and consequences, and Ann Tofado looks at the political dynamics and implications.

Discussion

Kate Burvish:

That breadth is the important signal. Investors were not simply marking down one software company after a disappointing quarter; they were repricing an entire investment ecosystem. Nearly three trillion dollars in S&P 500 market value was erased, according to Reuters. That is a change in valuation, not three trillion dollars disappearing from the productive economy overnight. But it does tell us what investors were questioning: whether years of extraordinary revenue growth can justify increasingly extraordinary capital spending.

Red Velhouse:

And the immediate political catalyst was the election. Democrats won in several states central to new data-center construction, including Texas and Ohio. Omar, why did an election result translate so quickly into a technology selloff?

Omar Seidren:

I’d call it a political repricing of a technology investment cycle. The election did not create a national moratorium, and it did not show that demand for artificial intelligence had vanished. It changed expectations around permits, utility approvals, local conditions and project timing. For facilities that take years to build, a delay changes the economics even if the eventual project remains viable. Markets discount those future cash flows immediately—sometimes with all the delicacy of a server rack being dropped down a staircase.

Eric Arcan:

And the physical system makes those delays consequential. A data center needs land, cooling, substations, transmission, transformers, generation and a grid connection. You cannot solve a transformer shortage by issuing a press release about artificial intelligence. The International Energy Agency identifies power electronics, gas turbines, advanced chips, grid connections and permitting as bottlenecks. Those constraints can slow construction even when the underlying demand is real.

Kate Burvish:

That is where the political story meets the financial one. The market is not only asking whether customers want AI services. It is asking how long companies must spend before those services generate returns, and who carries the cost if a project is delayed.

Red Velhouse:

Kate, investors were already uneasy about borrowing costs and AI-linked debt. How much of this is the bill arriving for an enormous spending spree?

Kate Burvish:

The bill is arriving as opportunity cost and rising financial pressure. Five major hyperscalers—Microsoft, Alphabet, Amazon, Meta and Oracle—are expected, on current consensus estimates, to spend more on capital expenditures than they generate in free cash flow by 2027. Their combined annual capital-expenditure estimate rose from about 485 billion dollars in January 2026 to roughly 730 billion by July. That does not indicate insolvency. It does mean investors must ask whether each additional dollar of infrastructure will produce enough future cash flow to justify depreciation, energy costs, financing and reduced distributions to shareholders.

Omar Seidren:

But a long payback period is not the same thing as an unproductive technology. Microsoft has said its AI business exceeded a 37-billion-dollar annual revenue run rate, and Amazon reported 28 percent growth in its cloud business during the cited period. Customers are paying for computing and AI services now. The harder question is whether revenue growth stays strong as capacity expands, and whether customers are using these systems for productive work rather than experimenting because everyone else is.

Kate Burvish:

Exactly. Current monetization argues against calling this pure vaporware, but it is not a complete return-on-investment calculation. Companies define AI revenue differently, and reported capital expenditures combine AI infrastructure with cloud, conventional data centers and other assets. We do not have a clean, public AI-only spending number. That uncertainty leaves the market vulnerable to both hype and overreaction.

Red Velhouse:

So the market has evidence of real demand, but not yet a clean answer on returns. Eric, let’s connect that uncertainty to electricity. The International Energy Agency projects global data-center electricity consumption nearly doubling, from 485 terawatt-hours in 2025 to 950 in 2030. Is that a manageable transition, or a warning that construction is outrunning the grid?

Eric Arcan:

It is a warning about execution, not necessarily proof of physical impossibility. The Energy Information Administration expects U.S. electricity demand to grow for a fourth consecutive year, with large computing facilities among the drivers. But national supply is only part of the story. A region can have enough generation on paper and still lack transmission, substations or a queue position for a new facility. The practical question is not simply, ‘How many terawatt-hours exist?’ It is, ‘Can reliable power reach this site at the required time and cost?’

Omar Seidren:

Technology can improve that equation, although not on a magical timetable. More efficient chips, better utilization, model compression and specialized systems can reduce the energy used per useful computation. But efficiency can also lower the cost of use and encourage more use. The server becomes more efficient, and suddenly somebody wants ten times as many servers. The Department of Energy’s U.S. scenarios for 2030 range from about 521 to 843 terawatt-hours, which shows how much depends on deployment, utilization and efficiency assumptions.

Eric Arcan:

That uncertainty is precisely why I focus on infrastructure milestones rather than headline forecasts. Transformer deliveries, gas-turbine availability, interconnection queues and utility rate cases can determine whether a project is real, delayed or merely announced. Total national power supply matters, but local deliverability often determines the schedule.

Red Velhouse:

Local politics may be the strongest constraint. An Associated Press analysis found restrictions approved in 44 of 45 completed local referendum contests across 44 municipalities. An AP poll found that roughly three-quarters of voters opposed a new data center in their community. Is that opposition economically rational, or is it blocking investment that regions need?

Eric Arcan:

It can be both rational and shortsighted, depending on the project and the terms. Residents see land use, water, noise, construction and possible pressure on electricity rates. They are not obligated to accept a facility because a national technology narrative says it is important. But a blanket rejection can push development elsewhere without eliminating demand. The practical solution is transparent cost allocation: operators should pay for the incremental grid and infrastructure burden they create, rather than quietly shifting those costs to households.

Kate Burvish:

There is also a distributional issue. Data centers can bring construction activity, equipment orders and tax revenue, while electricity, land or water costs may be shared broadly. A community can support investment in principle and still reject a bad contract. Regulation may destroy value if it is arbitrary, but it can improve project quality if it forces developers to price their external costs honestly.

Omar Seidren:

And disciplined permitting could improve the technology market rather than simply suppress it. If scarce power and construction capacity go to facilities with real customers, useful workloads and credible efficiency plans, that is healthier than building every speculative campus with a logo and a promise. The danger is uncertainty so great that serious operators cannot plan. AI deployment needs infrastructure, but infrastructure also needs predictable institutions.

Red Velhouse:

Let me press on the central disagreement. Is this a healthy consolidation, or the beginning of a genuine slowdown in AI demand?

Kate Burvish:

The evidence supports caution, not a verdict. The Friday rebound in U.S. Nasdaq futures and several chip stocks suggests that at least part of the first move was reactionary and positioning-driven. But a rebound does not repair free-cash-flow pressure or remove political risk. I would watch whether companies trim capital-expenditure guidance, whether credit spreads widen for data-center operators, and whether suppliers see orders deferred rather than merely rescheduled.

Omar Seidren:

For technology, the decisive indicators are utilization and customer value. Are models handling more workloads? Are cloud customers expanding usage after pilot projects? Are AI products generating recurring revenue and measurable productivity? If the answer is yes, capacity can eventually find a use even if some projects are delayed. If usage remains mostly demonstrations and internal experiments, then the industry has built a very expensive monument to optimism.

Eric Arcan:

And a slowdown in speculative construction could have benefits. It might ease pressure on power systems and give utilities time to plan more rationally. Delaying the least credible projects may be beneficial; delaying every project indiscriminately would sacrifice useful investment along with wasteful investment. The engineering test is whether a project can demonstrate reliable power, responsible water and land use, and a credible customer base.

Red Velhouse:

Could tighter U.S. rules simply send the buildout abroad?

Kate Burvish:

Some investment could shift, but relocation is not frictionless. Other countries face their own power, chip, financing and permitting constraints. The larger effect may be a change in where value accrues: regions offering reliable electricity and predictable approvals could attract projects, while regions imposing uncertain conditions could lose them. That is competition among jurisdictions, but it is also a negotiation over who pays for infrastructure.

Omar Seidren:

And moving a data center does not move the laws of physics. You still need chips, cooling, networks and electricity. A more distributed buildout could improve resilience, although it could also make software, security and energy management more complicated. So the global market may diversify, but there is no magical overseas cloud floating above the grid.

Eric Arcan:

That is why I resist both slogans: ‘AI will consume everything,’ and ‘efficiency will solve everything.’ We need generation, transmission, storage where it fits, firm capacity where it is needed and better efficiency. The projects that survive should be the ones that can meet engineering and economic tests, not simply the ones attached to the most persuasive technology narrative.

Red Velhouse:

The unresolved issue is whether AI revenues and useful workloads will grow fast enough to justify a capital cycle that strains free cash flow, local consent and the electrical grid. The November 5 selloff may prove to be a valuation correction—or an early signal that financing and infrastructure constraints are forcing a slower, more selective buildout. Watch hyperscaler capital-expenditure guidance, AI revenue and margins, project permits, utility rate decisions, grid-connection queues, credit conditions and the treatment of local infrastructure costs. The future of the boom will be decided not only in software markets, but also in substations, town halls and corporate cash-flow statements. Sources and references for this discussion are available with the episode at Factolio.com.


Sources and References

These sources supported the factual material used in this discussion. Factolio’s panel discussion is AI-generated from researched evidence and is written in original language.

  1. Associated Press — Global shares tumble following AI selloff on Wall Street (NEWS)
  2. Associated Press — How major US stock indexes fared Thursday 11/5/2026 (NEWS)
  3. Reuters, syndicated by Investing.com — Nasdaq futures rise after steep AI selloff as chip stocks stabilize (NEWS)
  4. Associated Press — A reckoning may await data centers following a beating in the midterms (NEWS)
  5. Reuters, syndicated by Investing.com — Analysis—AI investment boom puts Big Tech’s free cash flow under pressure (ANALYSIS)
  6. International Energy Agency — Key Questions on Energy and AI: Executive Summary (PRIMARY)
  7. U.S. Energy Information Administration — EIA forecasts strongest four-year growth in U.S. electricity demand since 2000, fueled by data centers (PRIMARY)
  8. U.S. Department of Energy / Lawrence Berkeley National Laboratory — United States Data Center Energy Usage Report: 2025 Update (PRIMARY)
  9. U.S. Energy Information Administration — Short-Term Energy Outlook—October 2026 (PRIMARY)
  10. Reuters, syndicated by Yahoo Finance — US Midterm Results Raise Risks for Data Center, AI Infrastructure Buildout, UBS Says (ANALYSIS)