Investors who have poured billions into generative models are beginning to demand a real return on investment (ROI), but the corporate sector is running up against harsh physical constraints: power shortages, a lack of data centers, and inflated financial reports that, upon closer inspection, turn out to be far from reality.
As Logos Press reported, it recently came to light that the Stargate UK project—which was announced with great fanfare last year and was supposed to become the UK’s largest artificial intelligence infrastructure initiative with a stated investment of up to £30 billion— has clearly fallen behind schedule and is unlikely to receive the promised investments.
But there are many such projects. And this is increasingly frustrating investors and government authorities, who need swift and effective projects to build the infrastructure of a “sovereign AI economy.”
The Illusion of Billions: The Guardian’s Investigation and “Phantom Investments”
One of the serious blows to the reputation of the AI hype frenzy was The Guardian’s March investigation, which exposed the problem of so-called “phantom investments.”
Using the example of the United Kingdom, where the government claimed to have attracted more than £100 billion in private capital to the AI sector, journalists discovered that high-profile press releases often mask a lack of actual infrastructure and jobs.
At the center of the scandal were projects linked to the American tech giant Nvidia, specifically the companies CoreWeave and Nscale. According to land registry and construction reports, the new data centers in London and Crawley promised by the government were never actually built—the companies simply leased existing facilities or retrofitted old sites.
Moreover, one of the key supercomputing centers, which was supposed to become the “heart of British AI,” is still nothing more than a regular construction site with scaffolding.
“These are phantom investments. Major tech companies are artificially inflating job creation figures and the economic impact of data centers to please governments that are desperately trying to claim economic growth,” says Cecilia Rikap, a professor of economics at University College London (UCL).
The UK Department for Science, Innovation, and Technology (DSIT) avoided providing detailed answers to questions from The Guardian, acknowledging that the department “does not play an active role in auditing these commitments.” And in a number of cases, press releases announcing the signing of contracts worth billions of dollars were issued without any legally binding agreements at all.
A Hard Landing: Statistics on Failures and Wasted Budgets
While venture capitalists continue to proclaim an inevitable revolution, the statistics on the actual implementation of AI in the corporate sector look disheartening.
According to a February 2026 study by Pertama Partners, based on data from the RAND Corporation and the MIT NANDA project, the industry is facing a crisis of inefficiency:
– More than 80% of AI projects are shut down without delivering the promised value to the business (this is twice the failure rate of typical IT initiatives).
– 95% of pilot projects in the field of generative AI do not show a measurable return on investment (ROI) in P&L statements (profit and loss statements).
– U.S. companies lose an average of 2.4% of their annual revenue on failed AI initiatives, according to a July report by CIO Dive citing analysts at Emergn.
Analysts note that companies are rapidly burning through their annual AI budgets on expensive “tokens,” yet at the end-product level, productivity gains are minimal.
Customers are beginning to realize the law of diminishing returns: each new and larger language model requires exponentially more money, yet solves problems only marginally better than the previous one.
The Energy Impasse: AI Is Limited by the Power Outlet
The main obstacle facing the AI industry has shifted from the realm of algorithms to that of fundamental physics. Scaling up AI requires enormous amounts of electricity, for which the world’s power grids are ill-prepared.
Analysts at Goldman Sachs Commodities Research predict that the electricity demand of U.S. data centers will more than double—from 31 GW in 2025 to 66 GW in 2027. The share of data centers in peak summer energy consumption in the U.S. will jump to 8.5%.
Due to a severe capacity shortage and delays in obtaining permits, Goldman Sachs expects that only 50–60% of planned data centers will be able to open on schedule over the next two years.
Morgan Stanley Research estimates a potential shortage of available capacity for data centers in the U.S. at 49 GW by 2028.
The problem of abandoned or idle data centers—those not connected to the grid—has become apparent even to the industry’s leading visionaries.
“I think the fundamental limiting factor for AI deployment is electrical capacity. It’s clear that very soon—perhaps even by the end of this year—we’ll be producing more chips than we can plug into an outlet,” said xAI founder Elon Musk during a panel discussion at the World Economic Forum in Davos.
Market Imbalance: Hardware Without Applications
Another paradox of the AI market: the production of silicon chips and purchases of GPU accelerators (mainly Nvidia) continue to break records, but they do not correlate with the market’s actual needs or its ability to utilize this resource.
The largest cloud providers (hyperscalers) are investing more than $1 trillion, but these costs are immediately reflected as revenue for hardware suppliers, while the data center operators themselves are forced to amortize these massive capital expenditures over years, which will inevitably hurt the sector’s margins, warns investment firm DWS.
The “Big Tech” business model—based on the principle of “release a product cheaply, hook the customer, and then raise prices”—does not work in AI.
Price wars have already begun in the market, and the cost of AI models is falling due to fierce competition, including from cheap and effective open-source models from China (such as DeepSeek and MiniMax).
It turns out that for most everyday office tasks (writing an email or processing text), models with 10 billion parameters are perfectly adequate, and using behemoths with 2 trillion parameters (such as the GPT-5 generation) — makes no economic sense.
What does this mean?
Of course, this in no way suggests that the “AI boom is cooling off,” much less that the market for artificial intelligence and all related industries will cease to attract investors anytime soon.
It simply means that the AI market has entered a phase of a sharp correction in expectations. Investors’ rose-colored glasses have been shattered by a lack of power grids, phantom financial reports, and a simple lack of business cases capable of justifying trillion-dollar infrastructure costs.
But investors will become much more cautious and discerning. Governments around the world realize that a high-tech AI economy cannot be built on high-profile declarations and “paper investments.” AI companies will adjust their plans and ambitions to reflect reality.
Those who can offer energy-efficient, localized, and low-cost solutions will survive. Projects built solely on venture capital billions and inflated press releases are headed for an inevitable collapse.
