The Stampede Ahead
Everyone's distracted by bubbles and missing the stampede. That opening line from Azeem Azhar cuts through months of market anxiety about whether artificial intelligence represents another dot-com-style bubble. Azhar's piece matters because it shifts the question entirely — from whether we've invested too much, to whether we've invested enough to handle what's coming.
Evidence Over Vibes
Azhar writes with a framework built on five indicators: economic strain, industry strain, revenue momentum, valuation heat, and funding quality. Five months ago, his analysis concluded generative AI is a boom, not a bubble. But he insists on evidence-driven conclusions.
"If evidence changes, we change our minds," Azhar writes. That intellectual honesty separates serious analysis from the alarmist chorus. The Financial Times has published over a hundred articles invoking the "AI bubble." Hedge fund investor Steve Sornoff disclosed shorts on Nvidia and Palantir, claiming "almost all AI companies will go bankrupt, and much of the AI spending will be written off." Fund managers surveyed by Bank of America cite AI overexposure as their top tail risk.
The bear case sounds plausible: capital expenditure growing faster than revenue, model costs falling dramatically, and most enterprise AI still停留在 chatbot-level outputs. If enterprises aren't getting results and efficiency gains mean you need less infrastructure, the capex overhang could collapse.
The real risk isn't that we've invested too much in AI. It's that we haven't invested nearly enough.
But while the bubble narrative gained momentum, reality moved the other way. Azhar shows industry strain — the ratio of investment to revenue — dropped from 6.1x to 4.7x in five months. If it holds, the ratio drops below the 3x threshold by Q2 this year, signaling revenues beginning to "carry" the installed base.
Boring Adoption Is Real Adoption
Monthly AI revenue grew from $72 million in January 2024 to $3.8 billion by December 2025 — roughly an eighteen-fold increase in two years. The hyperscalers drive this: Google Cloud grew 48 percent year-over-year to $7.7 billion, AWS expanded 24 percent to $5.6 billion, Azure grew 39 percent with its contracted backlog expanding 110 percent to $25 billion.
As Azhar puts it, "When the CEOs of the three largest cloud companies all tell you the same story — that AI is what's driving their growth — the attribution question starts to answer itself."
But the real signal isn't in startup hype. It's in boring, quantified claims from established institutions. Bank of America reports AI coding tools cut development time by 30 percent, saving the equivalent of 2,000 full-time engineers. Norway's $1.4 trillion sovereign wealth fund automated portfolio monitoring with Claude, saving roughly $7-32 million per year in labor costs. Meta reported a 30 percent increase in engineering output since January 2025. Western Digital reports AI tools improving yield and detecting defect patterns with productivity gains up to 10 percent.
"Boring adoption is real adoption," Azhar writes. That's the point. Percentage efficiency gains and operational savings aren't sexy, but they're real.
Critics might note that earnings-call claims remain self-reported and unverified. A company saying it saved money doesn't prove the savings materialized on the balance sheet. And the survey data shows only 25 percent of organizations currently have 40 percent or more of AI projects in production — though 54 percent expect to reach that level within six months.
The Agentic Threshold
Something changed at the end of 2025. Models passed a threshold of coherence — they can work reliably on tasks lasting an hour or two, and somewhat less reliably on longer tasks. Claude Code, Anthropic's tool for running software agents to write software, became the first beneficiary.
Azhar explains: "Claude Code is a really good software engineer. The developers building it use Claude Code to code itself." An Anthropic engineer used it to build a C compiler for $10,000 in API costs. A comparable project built by humans would typically require five to ten engineers over 18-24 months, around $2-3 million in fully loaded labor costs.
Claude Code became a $1 billion revenue business for Anthropic six months after launch. Today, it's likely generating more than $2 billion in annual recurring revenue. The implications extend beyond software — these agentic systems can tackle research, writing, data analysis, and operational workflows.
Physics, Not Capital
Here's where the argument pivots from boom to stampede. Capital expenditure is accelerating because the Big Tech companies have massive demand and are turning business away. AWS lost a $50 million contract to host Fortnite because it couldn't guarantee compute capacity. Microsoft had to choose between allocating compute capacity between first-party products and third-party business, arresting Azure's growth rate at 39 percent rather than 40 percent and wiping around 13 percent off Microsoft's share price.
Collectively, the big four hyperscalers — Microsoft, Alphabet, Amazon and Meta — have announced commitments of $500 billion this year, four times where they were before ChatGPT. But you cannot will infrastructure into existence.
"You can commit $500 billion in capex, but you cannot will a power plant into existence," Azhar writes. Building a data center takes 18-36 months from planning to power-on. Chip fabrication capacity, concentrated in TSMC's most advanced nodes, is allocated years in advance. Power infrastructure is the tightest bottleneck: some US grid operators report new data center interconnection requests face wait times of three to five years.
The chatbot era required modest compute — a few hundred tokens per interaction. The agentic era requires orders of magnitude more: sustained sessions running hundreds of thousands or millions of tokens per task. AWS CEO Matt Garman said current demand means they have never retired a six-year-old Nvidia A100 GPU. That's a supply constraint, not even accounting for agentic use.
Azhar estimates there is 40-50 gigawatts of capacity in the queue, and only 11-14 gigawatts online. OpenAI claimed that 1 gigawatt translates to roughly $10 billion in annual recurring revenue. Using that ratio, there is $110-140 billion in potential revenue online now.
"The bottleneck isn't capital, it's physics: power generation, chip supply, construction timelines, and the trained workforce to build."
Critics might note that Azhar's framework assumes continued demand growth without addressing what happens if enterprise AI projects fail to deliver promised returns. If the 67 percent of CEOs expecting AI returns within one to three years don't see them, funding could evaporate quickly. And the model providers' economics remain precarious — OpenAI's GPT-5 bundle achieved roughly 48 percent gross margins, well below the 70-80 percent typical of mature software. Frontier models function as rapidly depreciating infrastructure, their value eroded by competition before costs are recovered.
Bottom Line
Azhar's argument reframes the AI debate from financial speculation to physical constraint. The question isn't whether valuations are inflated — it's whether power grids, chip factories, and construction crews can build fast enough to meet demand. That shifts risk from bubble dynamics to stampede dynamics: too many buyers chasing too few servers. The companies that control compute capacity, not just models, will capture the value.