The artificial intelligence boom has been one of the most extraordinary spectacles in recent business history: billions poured in, valuations soared, and the breathless promise of a technological revolution captured boardrooms and headlines alike. But quietly, underneath the noise, a different story is emerging. The gap between what AI was supposed to deliver and what it is actually delivering is widening, and 2026 is shaping up to be the year that gap becomes impossible to ignore. The question is no longer whether AI will change the world; it probably will, but whether the businesses built on top of it will survive long enough to see it happen.
The ROI Problem Nobody Wants to Talk About
Start with the numbers, because they are sobering. Less than 1% of executives report returns of 20% or greater from their AI investments, while 53% report returns of just 1 to 5%. Roughly 60% of organisations see minimal or no meaningful value from their AI deployments, and approximately 30% of generative AI projects are being quietly abandoned after the proof-of-concept stage — that expensive, optimistic phase where things look promising before the real costs of integration and scaling kick in. Companies are spending heavily and showing very little for it. The fundamental problem is that AI has proven far easier to demonstrate than to deploy at scale in ways that meaningfully move the needle on revenue or efficiency. A convincing demo is not a business model, and the boardroom is starting to notice.
The Economics Are Not Adding Up
The infrastructure underpinning the AI industry is staggering in scale, with data centres, chips, and power contracts, but the revenue required to justify that infrastructure is not keeping pace. Usage-based pricing models are now emerging partly because flat-fee subscriptions were quietly bleeding money: a single developer using more tokens than anticipated could flip a profitable account into a losing one. The economics of inference, actually running these models at scale for millions of users, are trickier than the economics of training, and the industry is still working through that reality. Meanwhile, the physical world is pushing
back in ways that cannot be engineered away quickly. The most serious bottlenecks in scaling AI infrastructure are thermodynamic, power generation, grid capacity, and the hard physics of getting electricity to where it needs to go. These are not software problems with software solutions. They are infrastructure problems measured in years and billions of dollars.
Markets Are Adjusting, and So Is Public Opinion
Financial markets are doing what they always do: repricing assets as narratives shift. With rate cuts ruled out for 2026 and 10-year yields climbing, the long-duration assets that make up the AI investment complex are being mechanically devalued. As one analyst framed it, this is gravity, not a verdict on the underlying technology. But gravity is still gravity, and AI companies with stretched valuations and unproven unit economics are feeling it acutely. The investor mood has shifted from funding possibilities to demanding profitability, and many startups that raised at peak euphoria are now discovering that their runways are shorter than their roadmaps. Compounding this, public sentiment is cooling; a recent poll found that over half of Americans now believe AI will do more harm than good, a shift that will inevitably shape regulation and slow enterprise adoption in ways that optimistic forecasts did not account for.
What Survives and What Doesn’t
None of this means AI is a fraud or that the technology is failing. The models keep improving, the capabilities keep expanding, and the genuine use cases, in drug discovery, materials science, software development, and logistics, are real and growing. What is failing, or at least struggling, is the assumption that transformative technology automatically translates into transformative business value on a venture-capital timeline. The companies that will survive the next two years are likely those solving narrow, specific, high-value problems rather than chasing the broadest possible market. Verticalized AI tools built deep into a specific workflow for a specific industry are where the early ROI stories are actually emerging. General-purpose AI assistants competing on the same ground as the foundational model providers themselves face a much harder road.
The AI reckoning is not the death of artificial intelligence; it is the death of magical thinking about it. The technology is real. The value, when it arrives, will be real. But the industry is passing through an uncomfortable but necessary phase:
The gap between narrative and numbers is closing, not because the numbers are catching up to the narrative, but because the narrative is finally coming down to meet the numbers. The startups and enterprises that approach AI with discipline, clear use cases, measurable outcomes, and sustainable unit economics will be positioned to build something durable. Those still chasing the hype, burning capital on proofs-of-concept that never convert, are running out of time. The bill for the boom is coming due.
Sources
- Forbes Research – 2025 AI Survey (October 2025)
- BCG – “Are You Generating Value from AI?” (October 2025) • Gartner – Data & Analytics Summit, Sydney (July 2024)
- Deloitte – AI ROI: The Paradox of Rising Investment and Elusive Returns (October 2025)
- Pew Research Centre – How Americans View AI (June 2025, 5,023 US adults)
- Ipsos Global AI Survey (2025)
- Bloomberg Intelligence / Wall Street Journal analyst commentary (2025– 2026)
- McKinsey – State of AI Global Survey (November 2025)