Viewpoints

Is the AI bubble bursting? The buildout is wobbling. The demand isn’t.

Richard White · · 5 min read

This week’s tech selloff, a chip CEO calling data centers a bubble and the election backlash against new sites all point to strain in the AI buildout. But the demand data says enterprise use is still growing fast. What is likely to correct is the financing of capacity, not the use of AI, and firms should plan for that difference.

What happened

Tech shares had their sharpest one-day fall in more than a year on 5 November, a day after the midterm results, Fortune reported. Chipmakers and neocloud providers, the companies that borrowed heavily to build AI data centers, fell furthest. Meta’s president, Dina Powell McCormick, said building in the US “is going to be harder,” and that hyperscalers are weighing building outside the country. One analyst now expects tech capital spending growth to peak next year, and at least 10% of the 800 to 1,000 planned US AI data centers to be dropped, while still expecting Big Tech capital spending to top $1 trillion in 2027.

The same evening, Microchip Technology’s chief executive, Steve Sanghi, told analysts on the company’s earnings call: “I have been saying it, that data center is in the bubble.”

The demand side reads differently. In results reported the week before, Microsoft said Azure revenue grew 48% and that it still has “demand that exceeds supply,” with calendar 2026 capital spending of about $170 billion. Google Cloud revenue more than doubled. Amazon expects about $220 billion of capital spending this year, and its free cash flow has been negative for two straight quarters. Menlo Ventures, a venture firm, estimated on 3 November that enterprise spending on AI deployments reached about $177 billion by September, more than five times the 2025 level, with coding and chat tools the largest categories.

My view

Both things are true, and the useful question is which part is the bubble.

The use is real

Enterprise use of AI is growing faster than any technology I’ve worked with, and it is moving from pilots into daily work, especially in software development. Providers say they can’t build capacity fast enough. Those aren’t the signs of demand that exists only in a pitch deck. When people ask me whether AI is a bubble, they usually mean “will this go away.” It won’t.

The financing is where the risk sits

What looks stretched is the way capacity is being paid for. Spending is running ahead of the cash it generates at some of the largest builders. Smaller providers have taken on debt against long contracts with a handful of customers. Many sites assume cheap power, fast permits and local consent, and the election just made all three less certain. If demand grows a little slower than the most optimistic plans, or a site takes two years longer to build, the debt doesn’t wait. That is how buildouts usually correct: not because the technology fails, but because the capital arrived faster than the revenue.

Value is the part that’s lagging

There is a second gap, and it’s inside customers. Spending on AI is rising faster than most firms can show a return on it. I wrote this week about Apple putting staff on daily AI allowances, and earlier about firms deferring AI budgets because they can’t see the payback. That gap is not a bubble in the market sense. It is a management problem: the tools have arrived, and the work hasn’t been redesigned around them. If spending outruns measurable value for long enough, budgets get cut, and that is what would turn a financing correction into a demand slowdown.

What a correction would look like for a buyer

For a firm putting AI into production, a correction wouldn’t mean the models stop working. It would mean three practical things. Some smaller providers could be acquired, merged or fail, and their customers would need to move workloads at short notice. Prices could fall for general capacity and rise for scarce capacity in constrained regions. And the buyers who know their cost and value per task would be able to negotiate, while the rest would be told what they’re paying.

Why it matters

For the industry. Don’t plan around a crash, and don’t plan around uninterrupted growth either. Plan for a bumpy middle: capacity that is cheaper in some places and harder to get in others, providers whose finances matter as much as their models, and boards asking harder questions about return. None of that argues for slowing down good AI work. It argues for doing it with an exit plan and a scorecard.

How it connects to our work. AI Token Economics measures cost per completed task and separates committed capacity from cash, which is what lets a firm negotiate when prices move. Choosing an Enterprise Platform for AI Agents sets hard gates before commitment, including portability and the provider’s staying power. Why AI stalls covers the value gap: ownership, redesigned work and honest measures. Four checks I’d run now:

  • Know your exposure. List which AI providers each production system depends on, and how hard each would be to replace.
  • Keep workloads portable. Avoid building on features only one provider offers unless the value justifies the lock-in.
  • Read the contract for failure. Check notice periods, data return and exit terms if a provider is sold or closes.
  • Measure value per task. Be ready to show which uses pay back before the budget conversation starts.

This is commentary on technology and business strategy, not investment advice.

Sources

AI economicsData centersAI adoption

Linked sources are the record. Check them for the full reporting.

Read next

Viewpoint · · 4 min read

Apple puts its staff on an AI allowance. Budgets need a unit of value.

Apple has cut some engineers’ daily AI spending limits from $300 to $200 and may cut further, Bloomberg reports. Daily caps control spending, but they say nothing about value. The firms that get AI economics right will measure cost per completed task and fund the work that pays back.

AI economicsAI adoption

Viewpoint · · 4 min read

The AI return is missing because the work hasn’t changed

Firms are delaying AI spending because they can’t see the return. The models aren’t the problem. Most companies added AI to the old job instead of changing the job, and the return lives in the change.

AI adoptionChange management

Viewpoint · · 5 min read

Where your AI runs is now a business decision, not just your provider’s

Local opposition, power limits and the election have made new data center capacity slower and less certain in some of the largest US markets, and providers are weighing building abroad. For firms putting AI into production, the location, capacity and continuity of the compute behind their AI now belong in the same plan as their data residency and resilience rules.

Data centersAI infrastructure

All viewpoints Discuss this with Richard