AI Is Expensive And It's Not Working (Yet)
The numbers being committed to AI are genuinely staggering. Microsoft pledged $80 billion for AI data centres in fiscal year 2026. Amazon has earmarked tens of billions for AWS AI infrastructure. Google, Meta, and OpenAI are collectively spending at a rate that exceeds every prior technology investment cycle in history. The Stargate project a joint venture between OpenAI, SoftBank, and Oracle announced $500 billion in AI infrastructure investment over four years. Against this backdrop, MIT reviewed over 300 publicly disclosed AI implementations in 2025 and found that most had yet to deliver measurable profit-and-loss impact. Just 5% of the studied pilots generated millions of dollars in value. A Harvard Business Review survey of over 1,000 global executives found companies are cutting based on AI's potential, not its performance meaning the job losses are real, the investment is real, but the productivity gains that justify both are still largely aspirational. The gap between what AI costs and what it currently returns is the most important and least discussed business story of this moment.
Microsoft spent $80 billion on AI data centres in a single fiscal year. An MIT review of 300+ enterprise AI implementations found only 5% generated measurable P&L impact. The gap between the investment and the return is the defining business story of 2025 and 2026.
The Investment Numbers Are Real
| Company | AI Investment Commitment | Period | Stated Purpose |
|---|---|---|---|
| Microsoft | $80 billion (data centres) | FY2026 | Reimagining every product for the AI era |
| Amazon | $75+ billion (AWS AI infrastructure) | 2025–2026 | Most transformative technology since the internet |
| Google/Alphabet | $75 billion (capex 2025) | 2025 | AI-first transformation of Search, Cloud, Workspace |
| Meta | $65 billion (capex FY2025) | 2025 | Building infrastructure for AGI |
| OpenAI + SoftBank + Oracle | $500 billion (Stargate) | 2025–2029 | Physical infrastructure of the intelligence age |
The Return Numbers Are Not
MIT's GenAI Divide: State of AI in Business 2025 surveyed companies across industries on AI implementation outcomes. The finding that received the least coverage was the most important: the primary reason AI initiatives fail is not model quality, not legal risk, and not data limitations. It is execution. Most AI tools fail to learn over time, remain poorly integrated into day-to-day workflows, and stall in the proof-of-concept stage with no clear owner, no economic model for scaling, and unresolved data quality problems. Companies that built AI tools entirely in-house were twice as likely to fail as those that used external platforms. The 5% of implementations that generated meaningful P&L impact were not distinguished by better models. They were distinguished by better problem definition, cleaner data, and named accountability for outcomes.Deloitte's 2025 Emerging Technology Trends study found that while 30% of organisations are exploring agentic AI and 38% are piloting it, only 14% have solutions ready to deploy and a mere 11% are actively using them in production. Forty-two percent are still developing their strategy roadmap. Thirty-five percent have no formal strategy at all. The enterprise that has committed budget to AI and has no formal strategy for deploying it is not an edge case. It is the majority.The National Bureau of Economic Research surveyed nearly 6,000 C-suite executives across the US, UK, Germany, and Australia and found that approximately 90% reported AI had no measurable impact on employment over the prior three years, and 89% reported no change in productivity. These are not sceptics who haven't tried. They are executives who have been trying often with significant budgets and are not seeing the results that justified those budgets.
Where the Money Is Going Wrong
Investing in infrastructure before use cases
The most common AI spending pattern in large enterprises is to commit to cloud infrastructure, GPU capacity, and platform licences before defining the specific business problems the investment is intended to solve. The logic is that the infrastructure needs to be in place before use cases can be built. The consequence is that organisations spend 18 to 24 months building AI infrastructure and then discover that the use cases they assumed would emerge did not emerge, or emerged in forms that the infrastructure is not well suited to support. Infrastructure-first AI investment is the enterprise equivalent of building a stadium before choosing a sport.
Piloting everything, scaling nothing
Gartner estimates that over 40% of agentic AI projects will fail by 2027 specifically because legacy systems cannot support modern AI execution demands. But a significant fraction of the projects that never reach legacy system integration fail earlier, for a simpler reason: they are pilots that were never intended to scale. Organisations run AI pilots to learn, to satisfy board curiosity about AI, to generate press releases, and to check the box of AI investment. The pilot produces interesting results that nobody acts on. The pilot team moves to the next pilot. Nothing ships.
Solving the wrong problem faster
Henry Ford's observation that many people are busy finding better ways to do things that should not have to be done at all applies with particular force to enterprise AI in 2025. The most common AI use case in enterprise deployments is automating a reporting or summarisation process that exists because the underlying business process generates too much noise for humans to synthesise manually. AI makes the summarisation faster. It does not ask whether the underlying process that generates the noise is necessary. Faster noise is not intelligence.
What Is Actually Delivering ROI
The 5% of AI implementations generating measurable P&L impact share a consistent pattern. Walmart saved $75 million through AI-driven supply chain and inventory optimisation a highly structured, data-rich environment with clear measurable outcomes and a specific pre-defined problem. BMW reduced manufacturing defects by 60% using AI visual inspection again, structured, industrial, measurable. JPMorgan automated 360,000 staff hours using AI for document review in compliance processing routine, repetitive, high-volume, rule-based tasks. Every one of these cases starts with a specific problem that was already costing the business a measurable amount of money.Microsoft's own Q1 2025 market study found AI investments returning an average of 3.5x, with 5% of companies reporting returns as high as 8x. The 5% figure appearing in both MIT's implementation failure data and Microsoft's high-return data is not a coincidence. It is the same 5% of organisations the ones that defined a specific problem, assigned clear ownership, invested in data quality, and measured outcomes against a baseline. The outliers are extraordinary. The median is not. And the median is where most enterprise AI investment currently sits.
The Honest Projection
AI will deliver on its productivity promise. The question is not whether but when, and whether the organisations spending now will be the ones who benefit. The current gap between investment and return is characteristic of every major technology transition at the infrastructure buildout phase the internet in 1999, cloud computing in 2008, mobile in 2011 all showed the same pattern of investment preceding return by two to four years. The organisations that invested in those transitions early and correctly captured disproportionate long-run advantage.The difference between those transitions and the current AI investment cycle is the scale of the bet and the speed of the expectation. No technology in history has attracted $500 billion in committed investment over four years based on productivity gains that have not yet materialised at scale. If the productivity gains arrive on the timeline the investment assumes, the returns will be extraordinary. If they are delayed by three years which the current implementation failure data suggests is plausible the organisations that spent first will have carried significant costs without commensurate benefit. That is the risk that almost nobody in the AI investment conversation is stating plainly.