The AI Slowdown: What the Evidence Shows as of October 2026

"AI slowdown" gets used for three different things, and arguments go in circles when people mix them up. One is money: the companies building AI may soon spend less. Another is progress: the models may stop getting much better. The third is a deliberate choice: some people want AI development slowed or stopped on purpose. The evidence on each is different, and so are the people arguing.

Meaning one: the spending may slow

The largest cloud and AI companies have raised their capital spending every quarter for about three years. The April 2026 guidance table in Who Builds Them put the four biggest U.S. hyperscalers above $700 billion for 2026. As of October 2026, the question on Wall Street is less whether spending growth slows than when, and how hard.

Source	What they say	Leans toward
Stijn Van Nieuwerburgh, Columbia Business School, Brookings paper	On current estimates, 2026 capex at Oracle, Microsoft, Amazon, Meta and Alphabet passes their combined operating cash flow for the first time; internal cash is no longer enough without outside money or off-balance-sheet financing. He also notes cash generation "remains substantial."	Strain, not collapse
Ohsung Kwon, Wells Fargo	Expects the market to start thinking about 2027 as a possible peak year; says most investors agree capex will slow meaningfully in 2028	Slowdown in 2027 to 2028
Samuel Tombs and Oliver Allen, Pantheon Macroeconomics	Hyperscaler capex growth set to moderate in 2027; constraints in chips, electrical equipment, grid capacity and construction labor; but AI-linked investment probably keeps growing strongly	Slower growth, still growth
Dan Ives, Yorkville Ives	Infrastructure stays supply-constrained through at least 2027; supply, not demand, is the limit	No slowdown in demand
Bain and Company, September 2025	AI needs about $2 trillion a year in revenue by 2030 to fund the compute it expects; even with AI savings, the world is about $800 billion short each year	A revenue gap that has to close

All of these come from Fortune's October 7, 2026 roundup except Bain. Note what nobody in that list predicts: spending falling to zero. The disagreement is between "growth slows from very fast to fast" and "a peak followed by a real decline." A decline would hurt companies that borrowed heavily against future AI demand, which is why the cash-flow point matters.

Meaning two: the models may stop improving as fast

This is the harder question, because "better" is hard to measure.

• The case that progress is slowing. Gary Marcus, a cognitive scientist and longtime critic of large language models, called GPT-5 "overdue, overhyped and underwhelming" in August 2025 and argued that scaling alone will not produce general intelligence. The MIT NANDA initiative's 2025 report, based on 150 interviews, a 350-person survey and 300 public deployments, found about 95 percent of companies in its dataset saw little or no measurable profit impact from generative AI pilots. The headline number got more attention than its caveats: small sample, an undefined measure of failure, and purchased tools doing much better than in-house builds.
• The case that progress continues. METR, a nonprofit that tests AI models, measured the length of tasks frontier AI agents can complete with 50 percent reliability, judged by how long the tasks take human professionals. That length doubled about every 7 months over roughly six years, and the 2024 to 2025 data alone suggested faster doubling. METR is careful: results depend on which tasks are chosen, and extrapolating the line forward assumes it continues.

Both can be true at once. Benchmarks of what models can do in controlled tests keep rising, while many businesses struggle to turn that into profit. Slower adoption and continued capability gains are separate trends. The lesson Jobs and the Economy covers the measured effects on work.

The physical limits

Even with unlimited money, the buildout runs into hard limits: advanced chips, electrical equipment such as transformers and turbines, grid connections and construction crews. Pantheon names all four. DOE's July 2025 report says load growth from data centers is outpacing planned firm generation. Chip export rules shape who can buy the best chips, as covered in Chips and Compute.

Federal policy pushes both ways here. The permitting and federal land orders try to loosen the limits. EO 14421's restrictions on foreign grid equipment could tighten one of them, depending on how DOE writes the rules due in December 2026.

Meaning three: slowing it down on purpose

Some researchers and public figures want frontier AI slowed or stopped until it is shown to be safe. The Future of Life Institute hosts a Statement on Superintelligence calling for a prohibition on developing superintelligence, "not lifted before there is" broad scientific consensus it can be done safely and controllably, and strong public buy-in. It listed more than 77,000 signatures when checked in October 2026, including 5,000 gathered through a petition by the group Ekō.

The administration's position runs the other way. The AI Action Plan calls for dismantling regulatory barriers and frames AI as a race to win. EO 14409 adds government testing of the most capable models but makes it voluntary and states that it creates no licensing. The positions in that debate, from present-day harms to catastrophic risk, are laid out in Safety and Risk.

Where the evidence points

• Spending growth: mainstream forecasters expect it to moderate in 2027 or slow in 2028. That is a forecast, not a measurement, and forecasters have raised their numbers repeatedly since 2023.
• Financing: the biggest spenders are reaching the point where cash flow alone does not cover capex. That makes the buildout more sensitive to interest rates and investor patience.
• Capability: measured task performance keeps rising on METR's metrics. Business returns are uneven.
• A deliberate pause: there is public support for one in petitions, and no U.S. policy moving toward one.

What a slowdown means locally

A county does not approve AI. It approves a building, often the first of several phases. If the developer's parent slows spending, later phases can stall for years. Property tax projections based on full build-out then run high, and power lines or plants sized for the full campus may sit partly idle. That is why minimum bills, exit fees and collateral in utility contracts matter, as worked out in Ratepayer Protections at the Utility Commission, and why tax deals should be judged phase by phase, as in Taxes and Incentives.