Why $650 billion in AI investment is not a bubble — and may not be enough. A five-gauge framework for telling the difference.

The question I hear most from boards and investors is some version of: are we watching a bubble? The FT has published over a hundred articles exploring AI-as-bubble. Michael Burry — the man who shorted the housing market — is short Nvidia and Palantir. The sceptics have history on their side: technology booms have a nasty habit of ending badly.

I think they’re asking the wrong question. This isn’t a bubble. It’s a stampede. And the strategic response to each is entirely different.

The five-gauge dashboard

Over the past two years I’ve built an analytical framework — modelled on a pilot’s instrument cluster — to distinguish booms from bubbles. It tracks five gauges: economic strain (AI capex as a share of GDP), industry strain (the ratio of investment to revenue), revenue momentum, valuation heat, and funding quality. The framework draws on 200 years of technology-linked booms and busts, grounded in the work of Carlota Perez and Bill Janeway. You can explore it at boomorbubble.ai.

Today, most gauges read green. Two are amber. None are red. In every historical bubble I’ve studied — railways in 1872, telecoms in 2000 — at least three gauges were flashing red before the crash. We are nowhere close.

Revenue is the gauge that matters most

The single most important indicator is Industry Strain: how much capital is chasing each dollar of actual AI revenue. Five months ago it stood at 6.1×. Today it has dropped to 4.7×, and the trajectory points to crossing below 3× by Q2 this year. For context, the telecoms bubble peaked at just over 4×. AI’s ratio is moving in the opposite direction to every historical bubble — strain is falling, not rising.

The revenue numbers explain why. Monthly AI revenue grew from $772 million in January 2024 to $13.8 billion by December 2025 — an eighteen-fold increase in two years. Claude Code alone has become a $3 billion business, doubling in January this year. Google Cloud grew 48% year-on-year to $17.7 billion. When Pichai, Nadella, and Jassy all attribute cloud growth to AI, the attribution question starts to answer itself.

This is not speculative demand. The share of S&P 500 companies making quantified AI efficiency claims in earnings calls — not vague aspirations, but specific numbers — jumped from 1.9% to 13.2% in two years. Bank of America’s AI coding tools cut development time by 30%, saving the equivalent of 2,000 full-time engineers. Norway’s $2 trillion sovereign wealth fund automated portfolio monitoring, saving $17–32 million annually. As I wrote in my recent analysis: boring adoption is real adoption.

The infrastructure we built was for a chatbot world

Here is what the bubble narrative misses entirely. The infrastructure being built today was designed for a chatbot world — brief exchanges consuming hundreds of tokens. We have already crossed into the agent world, where autonomous AI workflows run for hours and consume millions of tokens per task. Software that would have cost a million pounds to write now costs £500 using AI agents. My own organisation has committed several hundred thousand lines of AI-generated code this year alone.

The demand implications are staggering. A basic chatbot turn involves a few hundred tokens. An agentic workflow that plans, loads tools, and spawns sub-agents can consume tens of thousands — 10 to 40× more. To the user it feels like one question. Under the surface, the token bill is an order of magnitude higher. We have gone from snacking to feasting, and the kitchen was built for snacks.

This is why the constraint has shifted from capital to physics. You can commit $650 billion in capex, but you cannot will a power plant into existence. Data centres take 18–36 months to build. Grid connections in Europe face 7–10 year backlogs. Only 11–14 gigawatts of AI-ready capacity is online against 40–50 gigawatts in the queue. The bottleneck is not money. It is time.

What would make it a bubble

Could I be wrong? Of course. The bubble diagnosis requires intellectual honesty about the conditions that would change it.

Watch for Industry Strain reversing — if the ratio starts climbing back toward 6× rather than falling toward 3×, that signals revenue growth stalling while capex continues. Watch for funding quality deteriorating — CoreWeave’s asset-backed debt and Oracle’s leveraged balance sheet already carry amber flags. And watch for the depreciation bomb: frontier models function as rapidly depreciating infrastructure, their value eroded by competition before costs are recovered. If hyperscalers cannot generate enough revenue from each generation of chips before the next generation arrives, the economics break.

But none of those conditions hold today. Revenue is accelerating. The ratio is compressing. And the balance sheets funding most of this buildout belong to the strongest companies in history.

What boards should actually worry about

The bubble question is a distraction. The real strategic risk is being caught in a famine — a period where demand for AI compute outstrips the physical capacity to supply it. Microsoft has already rationed compute between Azure clients and its own products. AWS lost a $10 million contract because it could not guarantee capacity. Six-year-old A100 GPUs remain in service because every available transistor has paying work to do. In the entire history of computing, that has never happened.

The strategic response to a bubble is caution. The strategic response to a stampede is acceleration — securing capacity, restructuring operations, and building the organisational muscle to absorb AI before your competitors do. The boards that treat this as a bubble will discover, too late, that they were sitting out a stampede.

The real risk is not that we have invested too much in AI. It is that we have not invested nearly enough.

---

Azeem explores the AI investment landscape with leadership teams. Begin a conversation.