AI Is “Overhyped in the Short Term and Underestimated Over the Long Term”: Demis Hassabis Said It at Cannes — and Two Years Ago

At Cannes Lions 2026, Demis Hassabis — Nobel laureate and co-founder and CEO of Google DeepMind — told a room of the world’s marketers that AI is “overhyped in the short term” and “underappreciated in the medium to long term.” If the line sounds familiar, it should. He said almost exactly the same thing to Bloomberg in May 2024. Two years on, both halves of his prediction are coming true at the same time — and the gap between them is the most important thing any business leader can understand about AI right now.

By Toni Dos Santos, Co-Founder, Spicy Advisory

What Hassabis actually said — twice

On stage at Cannes Lions 2026, in a fireside chat billed as “The Future of Creativity,” Hassabis framed the moment plainly. AI, he said, is “overhyped in the short term” but “underappreciated in the medium to long term.” Over “the next 10, 15 years,” he argued, the technology will open a “new golden era of discovery” across medicine, energy and science — “almost a new human era.” In the same breath he warned of an “over-correction” in AI startup valuations, with vast amounts of capital flowing into pre-revenue companies. You can watch the clip on Instagram here.

Rewind to 8 May 2024. Speaking to Bloomberg’s Tom Mackenzie — the same day Google DeepMind and Isomorphic Labs unveiled AlphaFold 3 — Hassabis said it first:

“I think AI’s overhyped in the short term and probably underestimated over the long term … what it’s going to bring. And I think that’s probably true of a lot of breakthrough technologies.”

He went further. Because of AI’s sudden popularity, he said, “lots of people” were rushing into the space “who maybe haven’t thought about this as long as people like us who’ve been in it for decades,” and so “we’re going to see a sort of rationalisation process happening.” But what AI ends up delivering, he predicted, would be “even beyond what the most optimistic end of things… in the near-term” suggested. The full 2024 Bloomberg interview is here.

Two interviews, two years apart, one thesis. That consistency is the point — and it’s testable now in a way it wasn’t in 2024.

He made the same call two years ago. Here is how it aged.

In 2024 Hassabis made three concrete predictions. All three have now largely landed — which is precisely why the “underestimated long term” half of his statement deserves more attention than the hype-bubble half that gets the headlines.

What Hassabis said in May 2024	Where it stands in mid-2026
Agentic systems that “plan and act in the world and solve goals” are “one to two years away” from real utility	Agentic AI is the defining enterprise theme of 2026. Gartner puts it at the “peak of inflated expectations”; ~17% of organisations have deployed agents and 60%+ intend to within two years
Excited about AlphaFold 3 and what Isomorphic Labs “can do with drug discovery”	AlphaFold won the 2024 Nobel Prize in Chemistry; Isomorphic Labs raised $2.1B and is advancing AI-designed drug programmes
A “rationalisation process” would shake out the newcomers chasing the hype	~95% of enterprise GenAI pilots show no measurable return (MIT); Hassabis himself now warns of a valuation “over-correction”

Read the table top to bottom and the structure of his argument is unmistakable. The short-term froth and the long-term substance are not contradictory forecasts — they are the same forecast, describing two different clocks running at the same time.

The short term really was overhyped — and now there is data

In 2024, “overhyped in the short term” was a feeling. In 2026, it’s a number. The most cited figure comes from an MIT study (Project NANDA) which found that roughly 95% of enterprise generative-AI pilots delivered no measurable impact on profit and loss. More than 80% of organisations had piloted tools like ChatGPT or Copilot and nearly 40% had deployed something — yet the value mostly stopped at individual productivity and never became enterprise outcome. MIT named the phenomenon the “GenAI Divide.”

Gartner tells the same story in different language: generative AI has slid into the trough of disillusionment, the predictable phase where inflated expectations meet integration costs, governance gaps and the hard reality of change management. Meanwhile agentic AI has rocketed to the peak of inflated expectations — the next wave of hype is already cresting before the last one has paid off.

This is the “rationalisation process” Hassabis named in 2024, arriving on schedule. His Cannes warning about an “over-correction” in valuations is the capital-markets version of the same thing. None of this means AI doesn’t work. It means the short-term expectations were wrong — that buying a licence is not the same as capturing value. We have written before about exactly why enterprise AI adoption fails, and the MIT data reads like a footnote to it.

The long term is still being underestimated

Here is the half almost everyone skips. While the market argues about whether AI is a bubble, the long-term curve Hassabis pointed to keeps compounding — quietly, and faster than the consensus expected.

AlphaFold predicted the structure of essentially every known protein and put that knowledge in the hands of more than two million researchers across 190 countries; it earned Hassabis and John Jumper the 2024 Nobel Prize in Chemistry. Isomorphic Labs, his drug-discovery company, raised $2.1B to turn that science into medicines. The agents he sketched in 2024 as “one to two years away” are now real enough that the question for most teams is no longer “can we?” but “on which workflows, and with what guardrails?” — a question we unpack in our guide to enterprise AI agents and autonomous workflows.

That is what “underestimated over the long term” looks like in practice: not a single dramatic moment, but a steady accumulation of capability that the quarterly-results conversation systematically under-weights.

Our take: the gap between the two timelines is the adoption gap

At Spicy Advisory we sit inside this gap every week, and from where we stand Hassabis is describing something very specific: the distance between when a technology becomes available and when an organisation actually absorbs it. That distance is the AI adoption gap, and it is the reason the same enterprises both over-buy and under-deliver.

The numbers are stark. Around 90% of companies are investing in AI, but only about 20% of employees actively use the tools. That is not a model-quality problem — the models are extraordinary and getting better monthly. It is an organisational problem: workflows that were never redesigned, managers who were never enabled, governance that was never written, and pilots that were never connected to a number anyone in finance cares about. MIT’s conclusion is identical: enterprise AI failure is “primarily organisational and strategic, not technical.”

“The short term is overhyped because companies buy tools. The long term is underestimated because value comes from behaviour — and behaviour is the part nobody budgets for.”

So the leaders panicking about a bubble and the leaders quietly compounding value are often looking at the same technology. The difference is whether they treated AI as a procurement event or as a change-management programme. This is the entire thesis of our founder’s book, Teach Them to Drive: you don’t hand someone a faster car and call it training.

What to do while everyone else is stuck in the trough

Gartner makes a counter-intuitive point that maps perfectly onto Hassabis’s: the trough of disillusionment is often the best time to invest, because the technology is more stable than at the peak, vendors are more flexible, and the hype premium has evaporated. Translated: while your competitors wait for the noise to die down, the long-term curve is open for the taking. Here is the sequence we run with enterprises across the UK, France and Portugal.

• Start with a readiness baseline, not a tool. Find where value and risk actually sit in your workflows before you scale anything. Our AI strategy and readiness work exists for precisely this, and you can self-diagnose in 20 minutes with the free AI Adoption Scorecard.
• Align the executive layer first. The adoption gap is set at the top. Build an AI charter, define your value and risk zones, and get leaders speaking a common language — the focus of our C-level AI programmes and our analysis of the C-level readiness gap.
• Enable behaviour, workflow by workflow. Generic “AI 101” is why pilots stall. We rebuild specific workflows with the teams that own them — the approach behind our 4-phase enterprise adoption framework and our work moving projects from pilot to production.
• Tie everything to a number. Pilots die when no one in finance can see the return. Decide the metric before the rollout, as we set out in the CFO’s guide to measuring AI ROI.
• Add agents last, once the foundations hold. Agentic workflows are powerful and, today, over-hyped — the discipline that makes them pay off is covered in our guide to production-ready agentic workflows.

The bottom line

Demis Hassabis has been right about the same thing for two years running, and the proof has only just arrived. The short term was overhyped — the failed pilots, the trough, the looming valuation correction all confirm it. The long term is being underestimated — the Nobel, the agents, the golden era of discovery all confirm that too. The companies that win the next decade won’t be the ones that called the bubble. They’ll be the ones who used the quiet part of the cycle to close the gap between owning AI and actually using it.