I was reading about the recent Anthropic announcements on Claude pricing and usage when a striking detail stopped me. One company reportedly spent $500 million in a single month after failing to set limits on employee usage. Microsoft subsequently cancelled Claude Code licences for its workforce and redirected them to homegrown tools. Sam Altman publicly called runaway AI spend one of the most "fair criticisms" of the industry right now.

The AI cost explosion is no longer theoretical. It is showing up in earnings calls, investor disclosures, and executive statements across the globe. And yet the investment commitments keep climbing.

$700 Billion. And Counting.

Amazon, Alphabet, Meta, Microsoft, and Oracle have committed a combined $700 billion to AI infrastructure in 2026 alone. By 2027, the figure likely crosses $1 trillion. Jensen Huang calls it the largest infrastructure buildout in human history. He is right. What he has not said, and what too few investors are asking, is whether it makes economic sense.

Here is the uncomfortable answer: for most participants, it does not. Not at current revenue levels. Not with honest depreciation. Not when you run the actual numbers.

The Five-Layer Stack and Why the Bottom Three Are Traps

AI is not one business. It is a five-layer stack: energy, chips, infrastructure, models, and applications. Jensen Huang himself has said the economic payoff concentrates at Layer 5, the application layer closest to the customer. Every layer below exists to serve it.

The problem is that capital has been deployed across all five layers as if each earns software-like margins. The lower three layers carry heavy capex, rapid obsolescence, and limited pricing power. They are not software businesses. The P&L mathematics that apply to them are fundamentally different, and far harsher.

The Depreciation Question Nobody Asks Loudly Enough

Consider a straightforward revenue scenario: 200 million subscribers at $200 per year plus substantial enterprise revenue puts you in the range of $100 billion annually. That sounds impressive, until you account for what it costs to build the infrastructure those customers are using.

At 4-year depreciation, which reflects actual technology cycles, $700 billion of capex generates $175 billion in annual depreciation. Every year. Before you earn a dollar of profit.

Add financing costs on that debt and the numbers are stark: on a reasonable revenue base, the industry is structurally loss-making at present scale. The more you build, the deeper the hole.

Hyperscalers depreciate AI servers over 5 to 6 years. These schedules are not dishonest. They are corporate standard. But they are not economically realistic for hardware that is two generations obsolete in three years. Choose your depreciation schedule and you choose your reality.

CoreWeave's depreciation and amortisation costs ballooned fivefold in a single year. IBM's CEO put it plainly: at the industry's current capex trajectory, the interest burden alone makes a return on investment structurally difficult to achieve.

More Revenue, More Loss, Until Something Breaks

This is the paradox most AI narratives skip past. Revenue is real and growing fast. The problem is that every dollar of new revenue requires more compute, which requires more capex, which compounds the depreciation base. The cost curve does not flatten until the buildout slows. The buildout does not slow because demand keeps accelerating.

OpenAI's own financial projections, reported by the Wall Street Journal, show billions in losses extending through 2028 even as revenue triples, with cumulative cash burn running deep into the nine figures before any path to profitability emerges. This is not a struggling startup. This is the most successful AI revenue business on the planet.

The Wrong Business Case Is Costing Companies Real Money

The dominant AI investment thesis inside most organisations today is: reduce headcount. That logic will not close the gap. Reducing a thousand jobs saves perhaps $100 million. The infrastructure required to deliver that outcome costs orders of magnitude more.

The real AI value case must be built on revenue growth, margin expansion, improved ROIC, faster innovation, and new business creation. Cost reduction is a byproduct, not a justification. Companies building their AI investment thesis on headcount savings alone are solving the wrong equation.

Where Value Will Actually Land

Industry research points to roughly $2 trillion in annual revenue being needed by 2030 to sustainably fund the compute infrastructure being built today. The question is not whether that revenue arrives, but how many current participants survive the cumulative losses in the interim.

Survivors will almost certainly concentrate at the application layer, where AI is embedded directly into healthcare, financial services, manufacturing, and commerce. These businesses own the customer, price against outcomes, and carry none of the infrastructure burden. Every other layer serves them, and is priced accordingly.

The most vulnerable are the infrastructure and model layers, which carry the capital cost without the pricing power to match. Some will consolidate into the hyperscalers. Others will not make it.

The Liquidity Squeeze Nobody Is Talking About

When capital gravitates to one place at this scale, it starves everywhere else. The $700 billion being channelled into AI infrastructure in 2026 alone is not sitting idle in sovereign wealth funds or cash reserves. It is being drawn from the same global pools of credit and equity that fund hospitals, ports, housing, grid upgrades, and the ordinary capex of industries that keep economies running.

The consequence is a quiet but deepening liquidity squeeze. Smaller technology firms, mid-market infrastructure developers, and emerging market sovereign borrowers are already finding credit more expensive and harder to access. When the world's largest balance sheets compete for debt financing simultaneously, spreads widen and available capital shrinks for everyone else.

Capital is not infinite. Every dollar committed to a GPU cluster is a dollar not building a power grid, a factory, or the next generation of companies that have nothing to do with AI.

For participants deeper in the AI stack, the risk compounds. Companies that raised at elevated valuations during the 2023 to 2025 build-out cycle now face refinancing in a tighter environment, with revenue projections that remain aspirational and depreciation schedules beginning to bite. Some will find strategic buyers. Others will simply run out of runway. The bankruptcies BlackRock's Larry Fink described as inevitable are not a distant possibility. For several infrastructure and mid-tier model providers, the clock is already running.

My Take

AI is not a bubble in the traditional sense. The technology is real. The demand is real. The value creation at the application layer will be real and sustained.

But the economics of building the infrastructure beneath it are broken for the majority of current participants. Cumulative losses measured in the trillions over five years, even as revenue grows aggressively, is not a rounding error. It is a structural problem that will force consolidation, write-downs, and in some cases, the bankruptcies that market observers have flagged as inevitable.

A compelling technology does not guarantee compelling returns. Confusion between excitement and economics is not a strategy.

The companies that understand this distinction, and build accordingly, will be the ones left standing when the correction completes.