Everyone is looking for an AI bubble. The data says stampede. The real danger is timidity, not exuberance.

Every board meeting I attend starts with the same question: are we in an AI bubble?

It’s the wrong question. The data points to something different — a stampede. And if you’re a business leader, the distinction matters enormously, because the strategic response to a bubble is caution, while the strategic response to a stampede is acceleration.

Let me show you the numbers.

AI revenue grew eighteen-fold in two years. Claude Code — Anthropic’s coding agent — went from zero to $3 billion in annualised revenue in under a year. Thirteen per cent of S&P 500 companies are now making quantified AI efficiency claims in their earnings calls, up from 1.9% just months ago. This isn’t speculative froth. This is customers spending real money and reporting real results.

I developed what I call the “Industry Strain” metric — the ratio of investment to revenue in the AI sector. Think of it as a stress test: how much capital is chasing each dollar of actual income? In early 2025, that ratio stood at 6.1×. Five months later, it had fallen to 4.7×. It’s heading toward 3×, which is the level where investment is firmly revenue-supported. During the dotcom crash, telecom peaked at 4× and collapsed. AI is moving in the opposite direction.

The broader “Economic Strain” measure — AI sector capex as a share of GDP — sits at 1.6%. That’s amber, worth watching, but comfortably below the 2% historical redline where technology buildouts have caused genuine macroeconomic distortion.

Now look at the supply side. Microsoft is rationing compute. AWS is turning away business. Six-year-old Nvidia A100 GPUs — hardware from 2020 — have never been retired. In the entire history of computing, this has never happened. You don’t hoard six-year-old chips when demand is speculative. You hoard them when demand is so real that every available transistor has paying work to do.

The $650 billion in committed hyperscaler capital expenditure is driven by genuine supply constraints, not speculative excess. Hyperscaler requests have escalated from 100 megawatts to 1 gigawatt. These aren’t moonshot bets. They’re infrastructure responses to customer demand that already exceeds capacity.

I use a five-gauge dashboard to track this — economy-wide strain, revenue coverage, customer demand, market rationality, and funding quality. If three gauges flash red, history says you’re in trouble. Two amber gauges warrant caution. Right now? Most gauges are green. Two are amber. None are red.

Could I be wrong? Of course. I’m open to the possibility that this is a speculative bubble that works out over twenty years, like the dotcom — which was both a genuine bubble and, over a longer horizon, the right bet. The two are not mutually exclusive.

But here’s what keeps me up at night. I’ve run the numbers on my own organisation. The software we’ve built with AI agents would have cost roughly a million pounds traditionally. We spent about £500. A C++ compiler was built by sixteen AI agents for $20,000 — work that would traditionally require millions of pounds and years of engineering time. Once this realisation moves from early adopters to the mainstream, the $650 billion won’t look like reckless spending. It will look like they didn’t spend enough.

The real risk for your organisation isn’t that you’ll over-invest in AI. It’s that you’ll under-invest — and watch your competitors pull away while you’re still debating the business case.

Azeem explores this topic with leadership teams. Begin a conversation.