Has AI Passed Peak Acceleration? The AI boom is still enormous. But some of the signals underneath it are beginning to change.
The AI boom is still enormous. But some of the signals underneath it are beginning to change.
For the better part of three years, the artificial intelligence industry has operated according to a remarkably simple equation: more compute, more data centers, more chips, more power, more capital.
Every generation of frontier AI required larger training clusters. Hyperscalers responded with enormous capital expenditure programs. Semiconductor companies expanded production. Utilities received staggering requests for electricity. Data-center developers raced to secure land and power. Private-credit firms and bond investors helped finance increasingly ambitious projects.
The result has been one of the largest infrastructure investment cycles in modern history.
And by most conventional measurements, that boom is still very much alive.
But there is another possibility worth examining.
The AI boom may not be ending.
It may simply have passed, or be approaching, peak acceleration.
That distinction matters.
An object can continue traveling extraordinarily fast even after it stops accelerating. The same thing can happen to an industry. Construction can remain at record levels, semiconductor demand can remain enormous and hyperscalers can continue spending hundreds of billions of dollars even while the forces driving the next round of acceleration begin weakening.
To determine whether that is happening, we need to look beyond Nvidia earnings or the latest frontier model.
We need to examine the entire machine.
The Boom Is Still Booming
Start with the strongest evidence against an AI slowdown: actual data-center demand.
CBRE reported that construction across the eight primary North American data-center markets increased 24.8% year over year during the first half of 2026, reaching a record 7,481 megawatts under construction.
Even more striking, 80.4% of that capacity is already preleased, compared with 74.3% one year earlier.
Vacancy sits at only 1.4%.
Less than 1,500 MW of future capacity remains available for preleasing across those primary markets, equivalent to approximately six months of demand at the current absorption rate.
Those are not the numbers of an industry experiencing collapsing demand.
Follow the supply chain another layer down and the picture looks equally strong.
AI data centers require much more than GPUs. They require transformers, switchgear, cooling systems, generators, cables, pipes, concrete and enormous quantities of electrical equipment.
Those companies are still seeing extraordinary demand.
South Korea's HD Hyundai Electric reported an order backlog of approximately $8.5 billion, up 23% in six months. The backlog represents more than three years of production, with discussions already underway for equipment scheduled for delivery as far out as 2030.
China's Hainan Jinpan Smart Technology reported that new data-center orders more than quadrupled during the first half of 2026.
Cooling companies are experiencing similar demand as AI racks consume increasingly large amounts of electricity and generate corresponding amounts of heat.
Even companies far removed from Silicon Valley are feeling the effects.
Wisconsin-based Generac, historically known for residential backup generators, is spending approximately $250 million expanding production of industrial generators for data centers. Its data-center generator backlog reached roughly $1.6 billion, while the company plans to add approximately 1,000 employees. Generac subsequently announced a long-term agreement with Amazon that could be worth as much as $8 billion.
The semiconductor supply chain isn't flashing recession either.
Samsung said September 29 that high-bandwidth memory, a critical component of AI accelerators, could consume nearly 30% of global DRAM wafer capacity next year, compared with approximately 20% today. Expanding HBM production is now threatening to constrain supplies of conventional memory.
These are important counterweights to any AI-bubble narrative.
The infrastructure machine is still running extremely hot.
But Something Changes When We Look Forward
The interesting signals appear when we stop examining equipment already ordered and facilities already under construction and instead examine projects that are supposed to come next.
Utilities have received extraordinary requests for electricity from prospective data centers.
Perhaps too extraordinary.
A Reuters examination of utilities and grid operators found more than 700 gigawatts of requested electricity from very large users, predominantly data centers, across portions of the Midwest, Mid-Atlantic and South.
That is more than ten times estimates of current U.S. data-center electricity consumption.
Nobody seriously expects all of those projects to be built.
Many applications appear to represent duplicate requests, speculative projects or developers attempting to reserve scarce electrical capacity before securing customers and financing.
And utilities are beginning to force developers to prove that their projects are real.
The results have been revealing.
When Exelon imposed stricter collateral requirements, its tally of high-probability data-center demand dropped by approximately 40%.
After Ohio adopted stronger requirements including grid-connection study fees, AEP Ohio's data-center power pipeline fell by more than half.
Pennsylvania had more than 100 proposed data centers, yet only 20 had applied for permits necessary to move their projects forward.
Texas has gone even further, freezing new grid connections while authorities investigate whether proposed projects have legitimate ownership, financing and development plans.
Call it the phantom boom problem.
The demand for AI compute may be completely genuine while some portion of the projected infrastructure surrounding that demand is not.
For several years, merely possessing land near electrical infrastructure could create enormous theoretical value. Developers rushed to reserve power, sometimes long before financing, customers or permits existed.
Now the industry is beginning to distinguish between announced megawatts and financeable megawatts.
That distinction could dramatically reshape projections for the second half of this decade.
Money Is Starting to Ask Harder Questions
The next signal comes from credit markets.
For much of the AI boom, financing was almost an afterthought. The largest technology companies generated enormous cash flows, while investors eagerly financed infrastructure connected to companies such as Microsoft, Amazon, Alphabet, Meta, Oracle and OpenAI.
That environment is changing.
Reuters reported in September that spreads on debt from AI-related issuers were running around 115 basis points, compared with approximately 78 basis points for the broader investment-grade market.
The difference does not mean investors believe the hyperscalers are suddenly in danger of default.
The problem is supply.
The amount of money required to finance the AI buildout is becoming so enormous that even global credit markets must decide how much exposure they are willing to absorb.
Goldman Sachs estimates hyperscaler debt issuance could reach $420 billion in 2027, approximately 60% above estimated 2026 issuance.
Some investors are already approaching internal concentration limits for exposure to individual technology companies once conventional bonds and project-financing structures are combined.
AI companies can still borrow.
They are simply beginning to pay more for the privilege.
That is an important change.
Capital has not disappeared.
Capital has started asking questions.
The Blue Owl Warning
One particularly interesting case involves Oracle, Blue Owl Capital and Project Jupiter, a massive New Mexico data-center campus being developed by Blue Owl's STACK Infrastructure to support OpenAI.
Oracle recently invoked a force-majeure provision related to delays securing power for the site.
The project remains active, and the companies say their financial commitments have not changed.
But the delay matters.
Blue Owl reportedly has approximately $3 billion of equity invested in Jupiter, with returns structured to increase once construction is completed. Delaying completion therefore delays those higher returns.
More importantly, lenders are studying Jupiter as a real-world test of who absorbs financial losses when enormous AI infrastructure projects encounter problems outside anyone's immediate control.
The consequences are already spreading beyond one project.
Reuters reported that Jupiter's difficulties are influencing discussions around other data-center financings, while SB Energy delayed an IPO as financing conditions changed.
During the second quarter alone, at least 45 data-center projects representing approximately $68 billion faced community opposition. That followed 75 projects worth roughly $130 billion disrupted during the first quarter.
Meanwhile, Blackstone's QTS abandoned its planned Digital Gateway project in Virginia after years of planning and regulatory review.
One delayed project doesn't make a trend.
But when financing structures across an industry begin changing because lenders are reconsidering risk, the signal becomes more interesting.
The Neocloud Question
Another potential fault line lies among the rapidly expanding AI-focused cloud companies sometimes called neoclouds.
These businesses require extraordinary amounts of capital because GPUs must be purchased and data centers constructed long before much of the associated revenue arrives.
CoreWeave recently announced another $3 billion convertible-debt offering, illustrating just how capital-intensive this business has become.
Nscale provides an even more dramatic example.
The company has contracted approximately $103 billion of future business and is seeking a public valuation reportedly around $30 billion.
Yet as of late September, only about 5% of the 461,000 chips it ultimately plans to operate were actually running.
Less than $3 billion of its $103 billion contracted revenue base was associated with infrastructure already online and generating revenue.
The remaining approximately $100 billion depends on infrastructure that still must be constructed.
Nscale's own IPO filing says it has not yet secured binding financing commitments for all of the infrastructure required to fulfill those agreements.
That doesn't mean the business model fails.
It demonstrates how much of the AI industry's future exists in the gap between contracted demand and constructed capacity.
Closing that gap requires enormous amounts of capital.
And the price of that capital is rising.
The Most Important Number May Be Acceleration
Perhaps the clearest illustration of the changing environment comes from projected hyperscaler capital expenditures.
UBS estimated in July that hyperscaler capex would rise approximately 76% in 2026, reaching about $673 billion.
Then something interesting happens.
UBS estimates growth falling to approximately 25% in 2027.
And approximately 6% in 2028.
That does not mean spending collapses.
It means the acceleration collapses.
Imagine an industry whose spending trajectory resembles:
$100 ? $176 ? $220 ? $233.
Every number is larger than the previous number.
The industry never contracts.
But something fundamental has still changed.
Companies throughout the supply chain that were valued under assumptions of continuously accelerating infrastructure spending may eventually encounter a very different environment when spending merely remains enormous.
That is why peak acceleration may prove more useful than "AI bubble" when describing the next phase of the industry.
Even Frontier AI May Be Changing
There is another signal emerging farther upstream.
Frontier AI developers themselves have begun publicly discussing slowing the pace of capability development for safety reasons.
In September, Anthropic CEO Dario Amodei proposed greater coordination among frontier developers and mechanisms intended to moderate the pace of capability advancement. The proposal subsequently received public support from other major AI leaders.
Whether those efforts ultimately produce a meaningful slowdown remains uncertain.
But the economics are interesting.
A slower frontier-model cycle would not necessarily mean lower AI usage.
Quite possibly the opposite.
Inference, the process of actually running AI models for users and businesses, is expected to become an increasingly large share of data-center demand. Slower frontier development could give companies additional time to commercialize existing models, improve efficiency and extract returns from infrastructure already constructed.
Safety and economics therefore do not necessarily point in opposite directions.
They may increasingly point in the same direction.
AI laboratories may want more time to evaluate increasingly capable systems.
Hyperscalers may want more time to monetize existing infrastructure.
Utilities need more time to construct electrical capacity.
Investors want clearer evidence of returns.
Communities want greater control over development.
And lenders increasingly want to know exactly who absorbs the risk when something goes wrong.
None of those forces requires the AI boom to end.
Together, however, they could slow its acceleration.
What We Should Watch Next
If the AI infrastructure cycle really is approaching an inflection point, the earliest evidence probably won't appear in Nvidia's quarterly GPU revenue.
By then, much of the change would already have worked its way through the system.
Instead, watch the edges of the machine:
• New project financing: Are data-center borrowers paying materially higher rates or providing stronger guarantees?
• Utility deposits: How many proposed projects disappear when developers must put substantial capital behind their power requests?
• Project cancellations: Delays caused by transformers or grid connections are different from projects being abandoned because their economics no longer work.
• Hyperscaler capex guidance: Slower growth is expected. Actual reductions in spending plans would represent a much stronger signal.
• Transformer and generator orders: These companies currently have enormous backlogs. Cancellations would provide an unusually clear downstream warning.
• GPU lead times: Easier access to frontier accelerators could indicate that supply is finally catching demand.
• HBM pricing and allocations: Memory remains constrained today. A sudden easing would deserve attention.
• Neocloud financing: Companies such as CoreWeave and Nscale sit directly between enormous AI demand and enormous capital requirements.
• Credit spreads: The cost of financing may ultimately matter as much as the cost of compute.
• Model-release cadence: If frontier developers begin extending model generations rather than racing immediately toward larger successors, the economics of infrastructure could change significantly.
These are the industry's seismographs.
Maybe 2026 Is the Inflection Point
There is currently little evidence that AI demand is collapsing.
Quite the opposite.
Data-center vacancy is extraordinarily low.
Construction remains at record levels.
Equipment suppliers have years of backlog.
High-bandwidth memory remains constrained.
Hyperscalers continue spending sums that would have seemed almost unimaginable only a few years ago.
But underneath those numbers, something has changed.
Speculative projects are disappearing when utilities require financial commitments.
Lenders are demanding greater compensation.
AI-related borrowers are encountering increasingly selective bond markets.
Community opposition is delaying billions of dollars of construction.
Power availability has become a fundamental constraint.
Project contracts are being rewritten to determine who absorbs delays.
And forecasts suggest the extraordinary acceleration in hyperscaler capital spending could moderate dramatically over the next two years.
Perhaps this is simply the maturation of an infrastructure boom.
Perhaps another breakthrough will ignite another round of acceleration.
Or perhaps historians will eventually look back at 2026 and recognize something that wasn't obvious while we were living through it.
The AI boom didn't end.
It simply stopped accelerating.
And sometimes you don't recognize the top of a hill because you're still traveling very, very fast.