NEW DELHI: The numbers from the supply side of artificial intelligence are by any measure staggering. According to Gartner’s December 2025 forecast, total worldwide AI spending will reach $3.49 trillion by 2027, up from 1.77 trillion in 2025 a 47 percent year-on-year expansion currently underway. The four major American hyperscalers Microsoft, Alphabet, Meta and Amazon collectively committed over 295 billion to AI in infrastructure in 2025 alone, a 32% increase from 226 billion the year before. In venture capital markets, AI commanded 61 percent of all global venture funding deployed in 2025, absorbing $259 billion in a single year.
On the demand side the picture is almost perfectly inverted. McKinsey’s State of AI 2025 report found that while 88 percent of surveyed organisations now regularly use AI in at least one business function, only 6 percent qualify as genuine value generators defined as attributing 5 percent or more of their earnings before interest and taxes to AI integration. Over 80 percent report no meaningful EBIT impact. The PwC CEO Survey 2026 found that 56 percent of global chief executives said their AI investments had generated zero incremental revenue gains.
This is the central economic contradiction of the current technology cycle. Capital is flowing in at a historic rate but returns are not flowing out.
The Arithmetic of Overreach
The structural scale of this mismatch was quantified precisely in June 2024 by David Cahn, a partner at Sequoia Capital in an analysis titled “AI’s $600B Question”. Cahn’s methodology was straightforward. Nvidia’s data centre revenue implied a total cost of ownership accounting for energy, cooling, buildings and networking of approximately $300 billion annually. To justify that infrastructure at a standard 50 percent gross margin for end users, the AI industry collectively needed to generate $600 billion in annual revenue. Cahn’s generous estimate of actual AI-attributable revenues across all major technology companies combined came to roughly $100 billion. The structural revenue deficit: $500 billion.
That gap has not closed. It has widened.
The capital allocation data makes the imbalance visible at a granular level. In Q2 2025, analysis of industry spend found that approximately 98 percent of the $82 billion in quarterly AI expenditure was absorbed by hardware, cloud infrastructure and foundation model layers. Enterprise application deployment the layer where actual business value is created received just 2 percent. The investment pyramid is almost perfectly inverted relative to where economic returns must ultimately originate. Pilot Purgatory Is Not a Phase it Is the Default State.
Multiple independent research streams confirm that the ROI failure is not concentrated in a few poorly managed deployments. It is the norm.
MIT Project NANDA’s July 2025 study of more than 300 public enterprise AI initiatives found that 95 percent of organisations deploying generative AI saw zero measurable return in terms of sustained productivity gains or documented profit-and-loss impact. BCG’s September 2025 survey of 1,250 executives found that 60 percent of organisations generated zero material value and only 5 percent had successfully scaled AI initiatives to capture significant value. Gartner confirmed that at least 50 percent of generative AI projects initiated were abandoned after proof of concept. S&P Global Market Intelligence found that by mid-2025, 42 percent of surveyed American companies had abandoned most of their AI initiatives up from 17 percent the year before.
The IBM Watson Health case remains the most documented cautionary precedent. Launched in 2015 with approximately $4 billion in acquisitions and a heavily publicised $62 million partnership with MD Anderson Cancer Centre, Watson Health took over six years to train on just seven cancer types due to poor interoperability with electronic health records. The MD Anderson partnership was terminated in 2017. In January 2022, IBM divested the entire division to private equity for approximately $1 billion a fraction of invested capital.
Why the Returns Are Not Coming
The failure is not a failure of the underlying technology. It is a failure of integration architecture, data readiness and organisational will.
Gartner estimates that 85 percent of enterprise AI projects fail specifically due to poor data quality, inadequate accessibility or missing data infrastructure. Only 12 percent of Chief Data Officers surveyed in Informatica’s 2025 CDO Insights report said their organisation’s data was of sufficient quality to support AI applications. The problem is not a shortage of data. It is a shortage of AI-ready data clean, integrated, consistently structured and accessible across organisational silos.
The talent dimension compounds this. Thirty-five percent of executives in the same survey cited a shortage of technical skills as a primary blocker. More strikingly, Writer and Workplace Intelligence’s 2025 survey of 1,500 enterprise employees found that 31 percent admitted to actively undermining corporate AI initiatives refusing to use sanctioned tools, slowing project timelines or feeding low-quality data into enterprise models. Change management, treated almost universally as an afterthought, is in practice a primary determinant of whether a deployment survives contact with an actual workforce.
Cost escalation is a further structural drag. Unlike conventional software, which exhibits near-zero marginal cost at scale, generative AI inference costs scale directly with usage volume. Successful pilots regularly become budget deficits in production as API token consumption grows. Goldman Sachs analysts Carly Davenport and Alberto Gandolfi have warned that supply chain constraints in the utilities sector will trigger a severe power crunch for data centres, driving up hosting fees and enterprise total cost of ownership unpredictably.
This Has Happened Before
The current moment is a classic general-purpose technology adoption lag. Nobel Laureate economist Robert Solow observed in 1987 that “you can see the computer age everywhere but in the productivity statistics.” The same paradox held for electrification. Electric motors were commercially available by the late nineteenth century, yet American manufacturing productivity remained flat for nearly thirty years. Real gains materialised only in the 1920s, after engineers completely redesigned factory layouts around the capabilities of the new technology not merely substituting electric motors for steam engines in buildings designed for an earlier era.
Stanford economist Erik Brynjolfsson, Director of the Stanford Digital Economy Lab, frames this as the Productivity J-Curve. Aggregate productivity initially declines after a general-purpose technology is introduced because organisations must divert capital and executive attention toward unmeasured intangible investments process redesign, workforce reskilling, data reorganisation. Productivity growth accelerates only after these complementary assets mature. The J-curve is not a flaw in the technology. It is the cost of genuine transformation.
However, not all economists accept that the upswing is guaranteed or imminent. MIT Institute Professor Daron Acemoglu, in his January 2025 paper The Simple Macroeconomics of AI, published in Economic Policy, estimates that only 23 percent of AI-exposed tasks are cost-effective for corporations to automate within the next decade, implying that AI will materially impact less than 5 percent of all work tasks economy-wide. His central estimate for total factor productivity growth over the next decade from AI: 0.66 percent. Goldman Sachs Senior Economist Joseph Briggs is more optimistic, projecting a 9 percent increase in US labour productivity over ten years. The honest position is that the range of credible expert estimates is very wide and that enterprises making capital commitments today are doing so in genuine uncertainty about the macroeconomic payoff.
Where Returns Are Actually Appearing
The ROI picture is not uniformly bleak. Returns are materialising, but in specific, structurally favourable conditions.
Financial services, with an AI adoption rate of 91 percent, reports the clearest gains. Highly structured data, well-defined use cases in fraud detection and credit underwriting and clear process metrics have produced 20 to 60 percent increases in processing speeds and measurable cost reductions. Customer service operations using task-specific AI agents report cost savings in the 10 to 15 percent range. Software engineering teams using code assistants report coding speed improvements of up to 40 percent, though these remain localised efficiency gains rather than enterprise-level EBIT expansion.
Healthcare, by contrast illustrates the ceiling imposed by structural barriers. Despite a 74 percent pilot adoption rate, 81.3 percent of American hospitals have not deployed clinical AI applications in production, according to research published in Nature Health. Unstructured, legally protected, non-interoperable clinical data makes safe deployment practically impossible at scale under current conditions.
The pattern is consistent across sectors: where data is structured, use cases are narrow and process metrics are clear, AI generates returns. Where data is messy, regulatory exposure is high and workflows are complex, the technology stalls at the proof-of-concept stage.
What Actually Works
McKinsey’s 2025 data identifies one finding that separates high performers from the rest with striking clarity. Organisations that successfully capture financial value from AI are twice as likely to have comprehensively redesigned their end-to-end workflows before selecting a model or vendor. The failure mode of most deployments is the reverse: purchasing a tool and inserting it into a process designed for human workers.
Mature organisations spend 2.1 times more on data integration and orchestration infrastructure than on AI software licences. Those that make these complementary investments achieve an average ROI of 10.3 times, compared to 3.7 times for firms with fragmented data frameworks. Gartner’s 2025 maturity analysis found that 91 percent of high-performing enterprises had appointed a dedicated AI governance leader and 60 percent had fully centralised their AI strategy.
The Bank for International Settlements, in Working Paper No. 1325 published in January 2026, found a modest but statistically significant 4 percent labour productivity increase among AI-adopting firms but found that this gain was driven by capital deepening and intangible investment, not by headcount replacement. The returns exist, they are being unlocked by organisational discipline, not by infrastructure spending.
The Question the Numbers Force
The $3.5 trillion being committed to artificial intelligence between now and 2027 is not obviously irrational. The electrification analogy and the internet analogy both suggest that the productivity payoff from transformative general-purpose technologies arrives on a delay measured in years, not quarters. The infrastructure being built today may prove to be the foundation for genuinely consequential economic gains a decade from now.
What the current data does not support is the implicit premise driving much of the current capital allocation that deployment at scale produces returns at scale, automatically and soon. Ninety-five percent of generative AI deployments showing zero measurable ROI is not a transitional statistic. It is a structural indictment of how most enterprises are approaching the technology.
The corporations that will look prescient in 2030 are unlikely to be those that spent the most on compute. They will be those that spent the most on the unglamorous work of data infrastructure, workflow redesign and workforce capability the intangible assets that no press release announces and no GPU shipment delivers.
The $3.5 trillion question has an answer. It is just not the answer most of the capital is currently pointed at.
About the Author
