There's a tension in how we talk about AI that I've been chewing over for a while. On one side: the $700 billion question — the combined capital expenditure the hyperscalers are heading for this year, and the widening gap between what they're spending and what the ecosystem is actually generating in revenue. On the other: every serious software team I work with now uses code generation daily, and in the past eighteen months I've seen more genuinely useful AI deployments in UK public-sector work than in the previous three years combined.
Demis Hassabis, CEO of Google DeepMind, said it about as clearly as anyone has. In a DeepMind podcast interview in December, he described AI as "overhyped in the short term and still underappreciated in the medium to long term." He also said parts of the AI ecosystem are "probably in bubbles" — referring to the startup valuations where billion-dollar companies emerge from nothing but a pitch deck. But he drew a distinction between that froth and the "lot of real business" supporting Big Tech's AI investment.
I think he's right on both counts. And I think the question "is it a bubble or isn't it?" largely misses the point.
The Short-Term Case Is Real
Let's do the bubble case properly.
Sequoia's David Cahn first flagged the revenue gap in September 2023 as "AI's $200 billion question." By June 2024, it had ballooned to $600 billion. The arithmetic is simple: you take what the industry is spending on Nvidia GPUs, you factor in the total cost of AI data centres and the margins the cloud providers need to make, and you get an implied revenue requirement that dwarfs what the AI ecosystem is actually generating. OpenAI was doing $3.4 billion annualised at that point. Nvidia's data centre revenue alone was running at many times that.
The numbers have only gotten louder. The four hyperscalers — Microsoft, Alphabet, Amazon, Meta — spent around $226 billion combined on capex in 2024, per CNBC and Axios reporting earlier this year. That roughly doubled to $410 billion in 2025. For 2026, the guidance range is $610 billion to $725 billion. Amazon is looking at negative free cash flow as a result. Alphabet's free cash flow is projected to drop almost 90%.
MIT's Daron Acemoglu, who won the 2024 Nobel in economics, ran the productivity numbers through his framework and came back with a modest conclusion: AI will boost US GDP by roughly 1% over ten years — a far cry from the 7% Goldman Sachs forecast or the double-digit trillions McKinsey projected. His reasoning is that only about 5% of tasks can be profitably automated in that timeframe. Most enterprises are not Google or OpenAI. The cost of implementation eats into the benefit.
Goldman's own June 2024 report was titled "Gen AI: too much spend, too little benefit?" — and that was before the current spending levels. Jim Covello, their head of global equity research, argued that AI isn't designed to solve the complex problems that would justify its costs.
Sam Altman told The Verge in August last year that yes, we are in a bubble. "When bubbles happen, smart people get overexcited about a kernel of truth," he said. "Someone is going to lose a phenomenal amount of money."
Bill Gates said the same thing two months later on CNBC, comparing it to the dot-com era: "Absolutely, there are a ton of these investments that will be dead ends."
Even Yann LeCun is warning that frontier AI labs face a "big bubble explosion" if costs don't come down fast enough.
Taken together, that's not a straw-man case. That's a serious one.
The Long-Term Case Is Also Real
Here's the part that gets less attention.
The same people who say we're in a bubble also say something else. Altman, in the same interview: "AI is the most important thing to happen in a very long time." Gates, in the same CNBC appearance, called it "the biggest technical thing ever in my lifetime."
McKinsey's 2025 State of AI survey found that 88% of organisations now regularly use AI. That number is up dramatically from previous years. Twenty-three percent are already scaling agentic systems. And the share of organisations McKinsey calls "high performers" — the ones actually capturing meaningful EBIT contribution from AI — has roughly doubled year-on-year. The gap between "everyone's trying it" and "it's creating real value" is narrowing faster than the critics acknowledge, and the trajectory matters more than the absolute number.
What's actually working? Code generation is the most mature — GitHub Copilot, Cursor, the tools that developers have adopted as naturally as they adopted Stack Overflow a decade ago. Document workflow automation — summarising, extracting, generating — is mainstream in any organisation with decent data hygiene. Customer service automation is generating real ROI, unglamorous as it sounds. Agentic operations are early but scaling.
And then there are the domain-specific breakthroughs. AlphaFold, which Hassabis won a Nobel Prize for, is not hype. It solved a fifty-year problem in biology. The same underlying capability is being applied to drug discovery, materials science, and energy research. A lot of what gets dismissed as "AI hype" is actually working, quietly, in labs and factories and server rooms that nobody tweets about.
Why Both Can Be True
The dot-com comparison is the right one, but only if you do it properly.
In 2000, Pets.com went under. Webvan went under. A lot of money was incinerated. But Amazon survived. Google survived. And the dark fibre that was laid during the boom — the infrastructure that everyone said was over-built and wasted — became the substrate for everything that followed. Streaming, social media, cloud computing, the entire internet economy of the 2010s ran on cables that were laid during the bubble.
The hype cycle and the structural shift ran on different clocks. The first was measured in quarters. The second was measured in decades.
I think we're in a similar moment with AI. The froth in startup valuations, the trillion-dollar capex numbers, the pricing distortions, the "every PowerPoint deck has AI on it now" phenomenon — all of that is real. Some of it will end badly. Companies will fail. Capital will be destroyed. The B100 will make the H100 look expensive, and the H100 will depreciate faster than anyone planned.
But the underlying technological shift is also real. The fact that a bubble and a transformation can coexist is not a contradiction. It's a pattern we've seen before.
What This Means on Monday Morning
Here's where the abstract debates meet the actual work.
I spend most of my time with public-sector and mid-market organisations in the UK. None of them are building foundation models. None of them are raising billion-dollar rounds. They are trying to make their services better, their operations cheaper, their data more useful.
For those organisations, the bubble question is a distraction. The right question is not "is it a bubble?" but "what are you doing about it?" — and the good ones are answering that question in three ways.
First, they're finding the narrow things that work now — not waiting for AGI, not chasing the keynote demo, but deploying AI for specific, bounded, well-understood problems. Code generation. Document processing. Customer triage. Data extraction. The boring stuff. The stuff that pays.
Second, they're investing in foundations first. Data quality. Governance. Security. The organisations that will capture value from AI are the ones that have clean data, clear policies, and people who understand both the technology and the domain. If your data is a mess, AI won't fix it. It will amplify the mess.
Third, they're building the capability, not just buying the tool. The 6% of organisations that McKinsey calls "high performers" share a pattern: they treat AI adoption as an organisational change, not a technology procurement. They build feedback loops. They measure outcomes, not activity. They have someone whose actual job is to own this.
Whether we're in an AI bubble or not is genuinely interesting — for analysts, for investors, for anyone whose job it is to worry about asset prices. But for everyone else — the people who actually have to make technology work on Tuesday morning — it's the wrong question.
The right one is: what are you actually doing with it?
And the answer, for anyone serious, has nothing to do with whether the bubble pops.
Start with what works
If you'd rather talk about the practical next step than the asset-price debate, that's exactly the conversation we like having — bounded, honest, and focused on what pays.
Book a discovery call Explore AI Strategy & Roadmap
Gary Willock is an independent IT consultant and enterprise architect based in Liverpool, working across cyber security, AI, cloud, and data with UK public-sector and mid-market clients. He runs Magma Cloud, NeuraSec, and Digital Scaffold. These are his own views.