The question has been asked at every major financial conference, technology summit, and earnings call for the past two years. Analysts are split. Founders are defensive. Investors are cautious but still writing the largest checks in the history of private markets. And somewhere in the middle of all of it, the actual answer has gotten lost.
The honest position is this: there are bubble-like conditions in AI markets right now. But calling it an AI bubble the way people mean it, the 2000 dot-com kind that collapses and takes a decade to recover, misreads what is actually happening.
01The numbers that look like a bubble
The scale of spending is genuinely difficult to comprehend. The four major hyperscalers, Alphabet, Amazon, Meta, and Microsoft, are expected to collectively spend between $660 billion and $690 billion on AI infrastructure in 2026 alone. That is the largest corporate investment programme in history outside of wartime mobilisation.
Venture capital has followed suit. In 2025, AI startups attracted $258.7 billion globally, representing 61% of all global venture capital. That figure doubled from 30% of global VC just three years earlier. OpenAI raised $122 billion in new funding in early 2026 at a $730 billion pre-money valuation, anchored by Amazon, NVIDIA, and SoftBank. Anthropic raised $30 billion at a $380 billion valuation around the same period.
And then there is the gap. For every dollar of AI software revenue being generated today, roughly $20 is being spent on the infrastructure to support it. OpenAI's annualized revenue is approximately $25 billion. Anthropic's run rate has crossed $30 billion. The entire pure-play AI software sector combined generates around $35 billion. The infrastructure spending to support it is nearly twenty times that figure.
MIT research from 2025 found that despite $30 to $40 billion in enterprise GenAI investment globally, 95% of organizations are reporting zero measurable return. The spending-to-revenue gap is real, and anyone who dismisses it entirely is not paying attention.
02Why this is not 2000
The dot-com bubble is the most common comparison and the least useful one.
In 2000, the top four technology leaders traded at roughly 70 times two-year forward earnings. The companies driving the rally were, in aggregate, destroying capital. The entire thesis rested on future revenue from an internet economy that, while real in hindsight, was years from generating the cash flows the market was discounting. In the three years following the peak, net income for the sector collapsed by 65%.
The situation today is structurally different. The companies driving the AI rally are among the most profitable in corporate history. Apple, Microsoft, Alphabet, Amazon, and Meta generated a combined $350 billion in free cash flow in their most recent fiscal years. The AI hyperscalers' average two-year forward price-to-earnings ratio sits at around 26 times elevated, but not comparable to Cisco trading at 200. This is balance sheet spending by companies that generate more cash than most sovereign wealth funds.
The demand is also real in a way it was not in 1999. By 2025, 71% of organizations were regularly using generative AI in at least one business function. OpenAI alone reports 900 million weekly active users. Enterprise now makes up more than 40% of OpenAI’s revenue and is on track to reach parity with consumer by the end of 2026. Codex has 3 million weekly active users. The platform processes more than 15 billion tokens per minute. These are not projected future users or aspirational metrics. They are current operating numbers from a company that raised $122 billion in new funding as recently as March 2026.
03The supply shortage nobody is talking about
In May 2026, BlackRock CEO Larry Fink made one of the more clarifying statements on this topic at the Milken Institute Global Conference: "There is not an AI bubble. There is the opposite. We have supply shortages. Demand is growing much faster than anyone has ever anticipated."
His specific framing: "We're short power, we're short compute, we're short chips." And then: "A new asset class will be buying futures of compute."
That last sentence matters. Fink is describing a world where access to computing resources becomes a tradable commodity, similar to oil futures or electricity contracts. The scarcity is not manufactured. The IEA projects data center electricity consumption will double to roughly 945 TWh by 2030. Chip demand has been outpacing supply for multiple consecutive quarters. TSMC's advanced packaging capacity, even at full utilization, may not meet projected global demand this year.
When supply is genuinely constrained against real and growing demand, the economics look different from a speculative bubble. Bubbles are characterized by demand that is fictional. What Fink is describing is demand that is real and supply that cannot keep up with it.
04This is more likely a war than a burst
The framing that gets closest to what is actually happening is not bubble. It is arms race.
The competition between OpenAI, Anthropic, and Google is not three companies fighting over the same finite market. It is three entities, backed by unprecedented capital, racing to build infrastructure and capability that will define how economies, businesses, and governments operate for the next several decades. Palantir has embedded AI directly into defence and military applications. Governments are watching. It is a matter of time before national AI strategy becomes as central to foreign policy as nuclear doctrine was in the mid-20th century.
Arms races do not follow bubble logic. They do not burst when sentiment changes or when a few high-profile failures shake confidence. They accelerate when one side appears to be pulling ahead. The DeepSeek release in January 2025, which briefly caused NVIDIA to lose $588 billion in a single day, was not evidence of a bubble popping. It was evidence of competition intensifying. The reaction from U.S. companies was not to reduce investment. It was to increase it.
BlackRock, through its Global Infrastructure Partners consortium, agreed to acquire Aligned Data Centers for approximately $40 billion. This is institutional capital building physical infrastructure with 20-year lifespans. That is not the behaviour of money that expects a bubble to burst in the near term.
05My Take
There are two things that can be true at the same time.
AI valuations in parts of the market are stretched, and some of the capital being deployed will be wasted. Larry Fink acknowledged this directly: some big winners, some big losers. That is capitalism accurately described. The capex cycle has barely started producing returns, and the 18 to 36 months between infrastructure investment and proportional revenue means the full picture is not yet visible.
But the technology is not speculative. The demand is not imaginary. The companies writing the largest checks have the cash flow to back them and the strategic imperative to keep going regardless of short-term market conditions.
Bubbles burst when the thing underneath them turns out to be fiction. When the internet bubble burst in 2000, the internet itself did not disappear. It took another decade to produce the revenue the market had priced in prematurely. AI may follow a similar trajectory: overhyped in the short term and underestimated over the longer one. But mistaking a competition of historic scale for a speculative mania misses the more important dynamic.
The infrastructure being built right now is not going away. The countries, companies, and institutions that control it will have structural advantages that compound over time. That is a different kind of risk than the one people are trying to avoid by asking whether there is a bubble.