Is AI in a Bubble? What the Numbers Say

In the first quarter of 2026, four funding rounds — OpenAI (122,000M), Anthropic (30,000M), xAI (20,000M), and Waymo (16,000M) — absorbed 65% of all global venture capital. Not 65% of tech capital. 65% of all venture capital, across all sectors, across the entire planet.

AI captured 80% of global VC that quarter. Three years ago it was 30%.

The question is not whether there is exuberance. The question is whether it is rational.

The short verdict: a selective bubble within a real transformation

The AI of April 2026 is not a simple bubble. It is a real overinvestment in a real technology, but with prices and expectations that, in a significant portion of the market, are already at bubble levels.

The technology works. The leaders’ revenues are real. The mass adoption is unprecedented. But there are three curves advancing at very different speeds: technical improvement, enterprise adoption, and sustainable monetization. And capital is running ahead of all three.

AI is experiencing a selective bubble within a real technological transformation. The risk is not that the technology doesn’t work — it’s that the returns arrive, but too late to justify what has already been invested.

The numbers that matter

Capex: 660 billion in a single year

The five major hyperscalers have committed roughly 660 billion dollars in capex for 2026, 36% more than in 2025. Approximately 75% is earmarked exclusively for AI infrastructure:

• Amazon: ~200,000M (AWS infrastructure for AI)
• Alphabet: ~175,000-185,000M (TPUs, Gemini, doubling budget for the second year in a row)
• Meta: ~115,000-135,000M (+87% year-over-year, for Llama and Superintelligence Labs)
• Microsoft: ~100,000-120,000M (Azure AI, contractual compute commitments)
• Oracle: ~20,000-50,000M (massive data center hosting)

Capital intensity has reached critical levels: Amazon allocates 57% of its revenue to capex, Meta 52%, Microsoft 48%, Google 45%. In 2025 alone, the sector issued 108 billion in new debt — x3.4 the historical annual average of 32,000M.

Revenues: real but insufficient for the scale of the bet

The leaders’ revenues are no longer a promise:

• OpenAI: 13,100M in revenue in 2025, run-rate of ~25,000M as of April 2026, 900M+ weekly active users, ~50M paying subscribers
• Anthropic: ARR jumped from 14,000M (February) to ~19,000M (March 2026) following its 30,000M Series G round at a 380,000M valuation
• NVIDIA: 215,900M in revenue in FY2026 with gross margins of 71.1%

But the gap between revenue and valuation is enormous. OpenAI has a run-rate of ~25,000M against a valuation of 852,000M — a multiple of ~x34 on annualized revenue. And it projects losses of 14,000M in 2026, with an expense budget of 38,000M.

The terms of its latest round reveal the fragility: of Amazon’s 50,000M, only 15,000M are unconditional — the remaining 35,000M are contingent on achieving AGI before the end of 2028 or completing an IPO.

The Sora case: when the demo doesn’t become a business

The shutdown of Sora 2 on March 24, 2026 is the most important case study of this era.

• Inference cost: 15M dollars per day (~5,400M annualized)
• Total lifetime revenue: 2.1M dollars
• App downloads: dropped 66% three months after the peak
• Disney deal for 1,000M: canceled (executives were notified less than one hour before the public announcement)

Bill Peebles, head of Sora, internally admitted in October 2025 that the economics were “completely unsustainable.” The Wall Street Journal described it as “a money pit nobody used.”

Sora proves that impressing is not the same as monetizing. And that the inference cost of the most intensive modalities — video, deep reasoning, autonomous agents — can grow faster than the ability to charge for them.

The circularity nobody wants to see

A significant portion of the AI ecosystem finances itself in a loop:

• Microsoft invests in OpenAI, which buys infrastructure from Oracle/Stargate, which buys GPUs from NVIDIA, which reinvests in models that drive demand for more GPUs
• CoreWeave — a key intermediary — has 21,000M in debt at 11% interest, allocates 25% of its gross revenue just to pay interest, and 70% of its revenue depends on Microsoft. It has accumulated 20,000M in backlog from OpenAI
• According to a Seaport analyst, between 12% and 25% of NVIDIA’s revenues have a circular component — x2-x3 what Cisco had before the dot-com burst

A CoreWeave default would flood the secondary market with tens of thousands of used GPUs, collapse collateral values, freeze NVIDIA sales, and drag down its stock price.

Capital circularity is not a theory. It is a documented mechanism that simultaneously distorts revenues, demand, and valuations.

The ROI that doesn’t show up

This is where the industry narrative collides with the evidence:

• NBER (February 2026, 6,000 executives in the U.S., UK, Germany, and Australia): more than 90% reported that AI had no impact on hiring or employment; 89% said the impact on real labor productivity was zero. Executives use AI an average of 1.5 hours per week.

• MIT Media Lab (mid-2025): after auditing large corporations, 95% of organizations get zero ROI from generative AI. Only 5% move beyond the “experimental pilot” stage.

• S&P Global (451 Research): in March 2025, 42% of companies had to abandon the majority of their AI initiatives — compared to 17% abandonment a year earlier.

And yet, 94% of companies surveyed by BCG and McKinsey say they will continue investing “even if the technology doesn’t generate immediate returns.” The driver is not ROI. It’s FOMO.

Is this the dot-com bubble all over again?

The comparison is tempting but imprecise:

What does repeat: excess capital deployed ahead of time, infrastructure built before proving returns, infinite TAM narratives, circularity and mutual validation.

What does not repeat: today’s megacaps make a lot of money, the technology already has mass adoption, the product exists and is used daily. As Jerome Powell said in October 2025: “this is different… these companies actually have profits.”

What could happen in the next 12 months

The canary in the coal mine is Oracle: BBB rating with a strongly negative outlook, negative free cash flow, leverage >x4 EBITDA projected for three years, and credit default swaps at their highest since 2009. Its stock price has already fallen ~58% from its highs.

What can I do?

• If you’re an investor: Don’t confuse “the technology works” with “the investment is good at any price.” The megacaps’ fundamentals are solid, but the private valuations of OpenAI/Anthropic price in years of perfect growth that may not materialize.

• If you run a company: Before allocating budget to AI, measure the real ROI — not the theoretical potential. 95% of organizations don’t get past the pilot stage. If your use case has no impact metrics after 6 months, reconsider the scale of the investment.

• If you’re a developer: Open source and API price compression are your best allies. DeepSeek-V3.2 operates at $0.028 per million tokens. Flash/lite models cover most use cases. You don’t need the most expensive model for most tasks.

• If you work in regulation: Capital circularity in the AI ecosystem is a systemic risk that financial regulators should monitor. A default in the middle layer (CoreWeave, Oracle) can have cascading effects across the entire tech sector.