By Reginal Campbell · September 16, 2026 · 10-minute read

Who Benefits from the AI Slowdown?

Frontier AI capability keeps climbing and AI capital keeps setting records. What has actually slowed is everyone’s ability to turn either one into dependable, organization-wide value.

The slowdown that isn’t a slowdown

AI has supposedly entered a slowdown. Nobody told the benchmarks, the venture capitalists, the hyperscalers, or the utilities racing to build power capacity for data centers.

The room keeps splitting in two: executives convinced AI has hit its ceiling, and executives racing to keep up with a technology they believe is nowhere close to slowing down. Both are watching the same models, the same investment, the same headlines. The problem is not what they are seeing. It is the question they are asking.

Here is what the evidence actually supports as of September 2026. Frontier AI capability has not plateaued. AI capital has not withdrawn. AI-related power demand and hyperscaler spending are accelerating. What has slowed, and become far more selective, is the conversion of model capability into dependable, organization-wide economic value.

That distinction changes who benefits. This is not a story about AI skeptics being vindicated or AI believers being humbled. It is a story about where bargaining power moves once the binding constraint stops being “can the model do this” and becomes “can this organization deploy it reliably, profitably, and safely.” When capability becomes abundant, the complements become scarce. That doesn’t mean the frontier itself is cheap, the best models are still rate-limited, expensive, and held by a handful of labs. It means that for most of what a business actually needs done, good-enough capability is increasingly available from more than one vendor, so it stops being the thing you can charge a premium for. That is where bargaining power moves instead: compute and power, enterprise data and integration, evaluation and governance capacity, implementation expertise, patient capital.

I am not arguing AI progress is unlimited, or that every enterprise AI investment pays off. The popular framing, a broad AI slowdown, is the wrong lens for what is actually a redistribution: of capability, of capital, and of leverage.

What the evidence actually shows

Let’s kill the plateau myth first, because it’s the easiest one to kill. Stanford’s 2026 AI Index shows a 30-point jump in a single year on Humanity’s Last Exam, and agentic performance is climbing right along with it, OSWorld scores are up to 66.3 percent. METR, the outfit that tracks how long a task a model can run on its own without supervision, still hasn’t found the ceiling. The trend keeps climbing.

None of that means progress is clean. Benchmarks are getting harder to trust, not easier: some widely used evaluations have invalid-question rates as high as 42 percent, which tells you the tests are starting to break before the models do. But “the technology stalled” is simply not a sentence the data lets you write.

Money is telling the same story, with a twist worth sitting with. U.S. venture capital deployed 320 billion dollars in 2025, up 51 percent year over year, and AI captured 65.4 percent of it, according to the National Venture Capital Association. By the middle of 2026, U.S. startups had already raised more than 400 billion dollars, more than any full year on record. Nobody sounds like that while quietly backing away from a technology.

Here’s the twist: almost none of that money is evenly spread. The five biggest 2025 recipients pulled in nearly 60 billion dollars between them. Rounds of 100 million dollars or bigger were just 3.2 percent of deals and swallowed 67 percent of the value, while venture fundraising itself sank to 67 billion dollars, a nine-year low. Strip the mega-rounds out and the rest of the market looks almost boring, close to a normal pre-boom year. So if you are not a frontier lab or one of the five, you might genuinely be living through a tight funding market in the middle of what the headlines are calling a historic boom. Both things are true. That’s not a contradiction. That’s just what concentration looks like from the outside.

The real bottleneck sits downstream of the model

Here’s the bottleneck, and it isn’t the model. Getting from “the model can do this” to “the business makes money from this” means clearing a gauntlet: compute access, data plumbing, reliability checks, redesigned workflows, retrained people, and finally something you can actually point to on a P&L. Every one of those steps is a place value quietly leaks out, and wherever it leaks, somebody downstream is building a business to catch it.

You can watch the gap in real time just by comparing how organizations talk about AI use to how they actually use it. Ask them in a survey and 88 percent say yes, we’re using it, according to Stanford’s aggregated data. Ask the Census Bureau, which samples businesses nationally instead of whoever felt like answering a survey, and the number drops to somewhere between 17 and 20 percent, climbing to 37 percent only once you get to firms with 250-plus employees. Those aren’t contradictory numbers, they’re answering different questions. Stanford is aggregating broad “has your organization used AI at all” surveys. The Census Bureau is asking a nationally representative sample of businesses something narrower: did you use AI in any business function in the past two weeks. Same topic, different bar, very different number. Adoption is a mile wide and an inch deep.

Productivity is the same story wearing a different costume: real, but wildly uneven. Workers surveyed by the St. Louis Fed say AI is saving them time, about 1.6 percent of all work hours, though nobody can prove the tool caused the savings rather than something else going on at the same time. Meanwhile a randomized trial on experienced open-source developers found the opposite: early-2025 AI tools made them 19 percent slower. Both studies are correct. Neither one is the whole truth, and picking whichever one flatters your priors doesn’t settle anything.

None of this is new, by the way. General-purpose technologies have always demanded the unglamorous complementary work first, rebuild the process, fix the data, redefine the job, before the new tool actually pays for itself. What’s different this time is speed. People picked up generative AI faster than they picked up the internet, though that comparison isn’t quite apples to apples, AI launched onto a world that already had broadband and a smartphone in every pocket. So the gap between what a model can technically do and what a company can reliably run keeps widening, because institutions were never built to move this fast.

Who captures value when realization gets selective

So who’s actually cashing in? Four groups, roughly: the people who own the pipes, the people who get paid to unclog them, the companies with their data and governance already in order, and the experienced professionals whose judgment nobody’s figured out how to automate yet.

Start with the pipes. Alphabet spent 91.4 billion dollars on capital expenditure in 2025 and is guiding to 175 to 185 billion in 2026. Microsoft spent 41 billion dollars in a single quarter. Those are company-reported numbers, not proof customers are seeing a return, and spending that much is a bet as much as a moat, it goes badly if utilization or pricing ever softens. But nobody writes checks that size while believing the party’s over.

Then there’s the cleanup crew. Implementation friction isn’t wasted value, it’s somebody’s invoice. Accenture booked 2.7 billion dollars in generative and agentic AI revenue in fiscal 2025 and 5.9 billion in bookings, much of it tied directly to helping clients get unstuck. Every gate I described above, integration, redesign, evaluation, change management, is a line item on somebody’s statement of work. Incumbent software vendors are eating from the same plate for a simpler reason: they already have your data, your security sign-off, and your login. They don’t need to win you as a customer. You’re already theirs.

The most interesting winners might be the people already in the building. Stanford’s Digital Economy Lab tracked payroll data through June 2026 and found no mass layoffs, but 22-to-25-year-olds in the most AI-exposed jobs are sitting about 19 percent below where hiring trends would otherwise put them. The researchers are careful to call this descriptive, not proof of cause. Fair enough. But squint at it and you see something worth saying out loud: companies may be protecting their experienced people, because someone still has to check the machine’s work, while quietly starving the entry-level jobs that used to train the next generation of people capable of doing that checking.

A firm can look fine in this year’s headcount while quietly eroding its own future expertise pipeline.

That’s not a talent strategy. That’s eating your seed corn and calling it efficiency.

The two groups losing ground are easier to spot. Undifferentiated AI startups are stuck between mega-round money going to the giants and incumbents bundling AI features into tools you already pay for. Frontier labs are not casualties here. They’re still hauling in some of the biggest checks in the economy. None of this means Microsoft and Alphabet sat in a room and engineered a slower rollout on purpose. Their capex says the opposite. It just means whoever already owns the plumbing gets paid first, every time.

The infrastructure and measurement constraints that matter

Power is the part of this story with the least room for spin. Data centers ate roughly 415 terawatt-hours of electricity in 2024, about 1.5 percent of the world’s supply, according to the International Energy Agency, and that number is projected to more than double by 2030. A follow-up IEA report found that AI-specific data-center electricity use jumped 50 percent in 2025 alone, and that without grid upgrades, one in five planned data-center projects could get stuck waiting on power that isn’t there yet.

Here’s the part that trips people up: efficiency is improving and total demand is still exploding, at the same time, because usage is growing faster than the efficiency gains can offset. That’s not a contradiction. It’s just what happens when a technology gets cheaper per unit and wildly more popular at once. And this isn’t a permanent global crisis, it’s a local and near-term one, concentrated in specific grids and permitting offices, which is exactly why hyperscalers keep throwing money at expanding supply instead of hedging their bets.

The less visible constraint is evaluation, and it deserves more attention than it usually gets, because it’s a governance problem hiding inside a technical one. I already mentioned the benchmark reliability issue, up to 42 percent of the questions on some widely used tests turned out to be invalid or ambiguous, but it belongs here too, because a shaky benchmark doesn’t just make a leaderboard embarrassing, it means you can’t outsource your evaluation to a public score in the first place. If you can’t trust the leaderboard, you build and govern your own, which is slower and easier to skip when a deadline is bearing down. That’s not evidence AI has stalled. It’s evidence that the tools we use to decide what to deploy haven’t caught up to what there is to evaluate.

What leaders should do differently

Six changes follow from this evidence, and none require betting on where frontier capability goes next.

Strategic ImperativeActionable Execution
Workflow MeasurementPick 5 to 10 actual workflows. Capture baseline cycle time, error and rework rates, cost, and quality before deployment, and hold the initiative to the same standard as any other capital investment.
Bifurcated BudgetingSeparate frontier-optionality spending, exploratory work with no near-term ROI requirement, from production-value spending, initiatives with defined operating KPIs, and give each its own hurdle rate.
Prerequisite InfrastructureFund data readiness, process ownership, and operating-model redesign before buying more model seats.
Capacity ManagementTreat power and compute as long-lead strategic capacity, not an on-demand commodity. Factor energy and interconnection timelines into vendor and location decisions.
Talent Pipeline DefenseProtect entry-level skill formation on purpose. Redesign junior roles around AI-assisted critique and verification rather than eliminating them outright.
Continuous EvalsInstitutionalize independent evaluation across procurement and production, not just the pilot. Treat it as an AI governance function, not a QA afterthought.

For policymakers, the useful move is targeting specific bottlenecks, grid interconnection, workforce transition, competitive procurement, safety evaluation, rather than treating “faster AI” or “slower AI” as the objective. The evidence here does not support either as a coherent target.

The gap is the story

None of this means AI is failing to deliver. Frontier capability keeps rising, capital keeps setting records, and individual use keeps spreading. The “AI slowdown” executives keep describing is not principally a slowdown in AI. It is the widening gap between how fast machine capability improves and how slowly the physical, organizational, and institutional systems around it can absorb that capability into durable value, and that gap is where the real competitive story is being written.

The winners in this phase are not simply the companies with the best models. They are the companies, and the individual workers, who control the scarce complements that turn a capable model into a reliable one: compute and power, integration and data, evaluation and governance, implementation expertise, and the patience to fund all of it before a return shows up in a quarterly number. Stop treating benchmark progress or pilot counts as proof that realization has happened, and start measuring what actually converts capability into an outcome the organization can defend. For the foreseeable stretch, that gap is the entire competitive landscape.

Sources

Official statistics and government data

Technical evaluations and capability research

Capital-markets datasets

Enterprise adoption and productivity studies

Labor-market research

Company disclosures (vendor-reported; not independent audits)

Historical and systems analogues

  • Brynjolfsson, Rock & Syverson, “The Productivity J-Curve: How Intangibles Complement General Purpose Technologies,” 2021.

Note on source quality: capability and capital figures come from official statistics and established industry datasets (Census, IEA, NVCA/PitchBook, Stanford HAI). Company-specific figures are vendor disclosures and are labeled as such rather than treated as independent proof of customer return. Labor findings are descriptive, ADP-based correlations, not causal estimates, per the authors’ own characterization.

About the author

Reginal Campbell writes about enterprise technology, AI governance, leadership, and the systems organizations build to make consequential decisions.

Read more articles · Connect with Reginal on LinkedIn