The $690B Question Wall Street Can't Answer
Wall Street thinks AI infrastructure is overbuilt and AI is powerful enough to destroy $800B in SaaS market cap. Both cannot be true.
Wall Street thinks AI infrastructure is overbuilt and AI is powerful enough to destroy $800 billion in software market cap. Both cannot be true. The arithmetic picks a winner.
Bank of America called it last week. The SaaS selloff is “internally inconsistent.”
Here is the contradiction: Wall Street simultaneously believes AI is powerful enough to destroy $800 billion in software market cap AND that the infrastructure to run that AI is overbuilt. Both things cannot be true. If AI is a parlor trick, SaaS companies are fine. If AI is the real thing, you need every GPU they are buying.
Pick one.
The Arithmetic
The Big Four will spend roughly $690 billion on AI infrastructure in 2026. Google: $185 billion. Amazon: $200 billion. Microsoft: $145 billion. Meta: $115 to $135 billion.
Those are large numbers. They are also, measured against history, entirely sane.
AI capital expenditure in 2026 will equal approximately 2.5% of US GDP. The Railway Mania of the 1840s hit 8%. The fiber optic boom reached 1.2% before it collapsed. We are three times below the threshold where historical infrastructure booms became manias.
Google is spending more on AI in a single year than the GDP of Ukraine. But Google also generates enough revenue to do that without borrowing. So does Microsoft. So does Amazon. This is not speculation funded by junk bonds. It is retained earnings pointed at a bet these companies can see working inside their own walls.
The Demand Proof
Now look at the other side of the ledger.
Thomson Reuters fell 16%. Intuit dropped 34%. LegalZoom lost 20%. Eight hundred billion dollars evaporated from companies whose business models assume humans do knowledge work manually.
What triggered the LegalZoom collapse? Roughly 200 lines of structured markdown. That is the size of the Anthropic Claude plugin that automates legal document review. Two hundred lines.
The demand signal is not theoretical. KPMG forced Grant Thornton’s audit fees down from $416,000 to $357,000, a 14% cut. KPMG did not deploy AI to do it. They used the existence of AI as leverage. The threat of automation is already repricing professional services, and the automation itself has barely started.
Anthropic went from $1 billion to $14 billion in annual recurring revenue in 14 months. Three hundred thousand business customers. Deloitte is deploying Claude to 470,000 employees. That is not a pilot. That is an operating decision.
The Structural Break
Here is where the bears miss the story entirely.
Previous infrastructure booms built dumb pipes. Fiber was fiber. Rail was rail. AI infrastructure is different because the demand curve compounds.
In 2023, inference was 33% of total AI compute. Today it is 67%. The ratio is projected to reach 118:1. That shift alone would consume the entire planned buildout and then some.
Why? Agents. A chatbot consumes a handful of tokens per interaction. An AI agent working through a complex task consumes 100 to 1,000 times more. The per-token cost has fallen 280 times since GPT-3.5. But total bills are going up because consumption is exploding faster than prices are falling.
The OpenRouter study across 100 trillion tokens tells the same story: programming tasks went from 11% to 50% of all usage. Reasoning model calls went from zero to over 50%. The workloads are getting harder, longer, and more compute-intensive every quarter.
This is not a pipe. It is a flywheel.
The Historical Pattern
Every infrastructure boom in modern history looked insane in real time and obvious in retrospect.
Railroad shares fell 66% during the Railway Mania. The companies that survived built 6,000 miles of track that turbocharged the Industrial Revolution. Ninety-five percent of fiber optic capacity went dark after the telecom crash. YouTube and Netflix launched on top of it.
AWS generated $21 million in revenue in 2006. Today it runs at over $100 billion annually. The infrastructure preceded the demand by years. That is not a bug. That is how platforms work.
The pattern is always the same: overbuild, crash, consolidation, then a wave of value creation that dwarfs the original investment. The question is not whether this cycle repeats. It is whether you position for the crash or the creation.
What This Means Monday Morning
If you run engineering teams: The inference cost curve is your friend, but only if you are building for agent-scale workloads now. The gap between “chatbot bolted onto existing workflow” and “agent replacing the workflow” is a 100x difference in compute, value, and competitive distance.
If you run a budget: AI infrastructure spend is not a line item to cut. It is the line item that determines whether every other line item still makes sense in 18 months. KPMG did not even use AI and still extracted 14% from a vendor. Your competitors will not be so gentle.
If you set strategy: The $690 billion is not a bubble. It is a down payment. The companies spending it can see the demand curves inside their own data centers. The question for everyone else is simpler: what happens to your business when 470,000 Deloitte consultants have an AI co-worker, and your team does not?
The bears think this is a supply story. It is a demand story. And demand, unlike speculation, does not pop.
