Author: Raoul Pal, Founder/CEO of Global Macro Investor, Founder of Real Vision
Everyone believes AI trading is over. Massive cloud providers missed earnings expectations, capital expenditures have peaked, round-trip trading is just another Enron, debt is a bubble, and 2000 is repeating itself...
I don't believe a single word. Here's what's actually happening.
The market is pricing in a liquidity pause.
The oil shock triggered by the war in Iran pushed up the dollar and short-term interest rates, both of which constitute financial conditions. Current financial conditions lead liquidity by about a month, and global liquidity growth has been slowing month-over-month since March. The Nasdaq has a 97% correlation with liquidity, so when funding slows, valuation multiples decline regardless of what the underlying businesses are doing—this is “The Everything Code”…
That's it. Everything else you've heard is hindsight storytelling.
And the actual performance of the underlying business is as follows...
Microsoft’s capital expenditures were $5 billion lower than guidance—$45 billion versus over $50 billion—due to suppliers failing to deliver capacity on time. Azure grew 48% for the quarter, exceeding guidance, with next quarter’s guidance returning to over $50 billion. No one lowered their 2027 guidance; Amazon stated it will significantly increase spending. The four major cloud providers hold $2.49 trillion in signed contracts, roughly double the amount from a year ago.
Anthropic reached $46 billion in annualized revenue in five and a half years. Google took 14 years, Amazon 17, NVIDIA 30, and Microsoft 32. This is the fastest revenue growth in business history.
Builders are also profiting from this. The operating profit margins of the three major cloud providers rose from 5% in 2019 to 39%, while expenses surged during the same period; the operating cash flow of the four major cloud providers grew at an annual rate of 34%, compared to a cost of capital of 4.9%. This is the opposite of what happened in 2000.
But hardware is lagging. On our comprehensive capacity metric—built from TSMC and Micron revenues, chip equipment manufacturers’ spending, and capital expenditures by the five major cloud providers—this is the tightest chip cycle since at least 2011.
A company that doesn't even have time to plug in its chips doesn't have a demand problem...
The first product in history with infinite demand
Everything else will saturate. You only need so many cars, so much oil, so many phones. Intelligence won’t saturate, because whenever its price drops, we find more uses for it.
The price of fixed-level intelligence has decreased by approximately 13 times annually, and now a 1% price drop brings 2.2% more usage, with this ratio increasing throughout the year. Tokens on OpenRouter have grown 32 times in a year, reaching 188 trillion per week. Welcome to Jevons’ Paradox…
Humans are no longer the primary customers—machines are. Just agent applications now send 43 trillion tokens per week through OpenRouter, a 26-fold increase in one year.
Last month, coding agents jointly made 20 million commits on public GitHub, up from 3,000 in January 2025, and opened 4.7 million pull requests. Each smarter agent takes on more work, requiring more agents, which demands more compute power, making the next agent even smarter. Repeat infinitely...
Something that is both a product and its own customer. No market in history has ever existed like this, not even close. This is absolutely not broadband or fiber, nor any of the nonsense people tell you...
No one can slow down.
Intelligence per unit of energy doubles every four to five months. Look back just one year, and you’re three doublings behind—your intelligence per watt is only one-eighth of what those who keep running have. You can’t catch up. You’re permanently out.
This applies to Microsoft versus Google, OpenAI versus Anthropic, the U.S. versus China. On the developer marketplace OpenRouter, Chinese models now account for 70% of tokens, up from about a quarter in January.
No U.S. president would allow America to fall behind in this race. Capital expenditure is the entry fee, and it rises every year. The U.S. leads in innovation, while China leads in scale—this is nearly the same pattern seen across all products...
The only real constraint is supply
Installed AI computing power grows 1.8 times annually, while demand increases 7 to 35 times depending on the metric. Gas turbines are sold out through 2030, large transformers take two years to deliver, and chips sit in boxes waiting for power. Demand massively outstrips supply—not the other way around.
Demand doesn't disappear. It's waiting in line.
But the big picture is... AI has already consumed about 10 gigawatts. All human brains on Earth combined operate at roughly 160 gigawatts. On the central path, AI will surpass all of this around 2032. I'm not sure you fully grasp this chart, but by 2032, AI will exceed the total computational power of all humanity.
Also, people like to say that the human brain consumes only 20 watts of energy. Let’s rephrase this, because everyone is wrong…
Anthropic is cash-flow positive, so let’s look at their subscription: $200 per month, and I gain intelligence superior to nearly every existing human across every known discipline. My cost is $2,400 per year. That’s roughly Anthropic’s cost too. But even if it weren’t, suppose Anthropic’s cost is $5,000 per year. Even then, the average human anywhere in the world costs at least twice that much just to sustain — and that’s only the energy input for one human brain.
People confuse training an entire model and serving it to billions of users with your own AI instance—which is equivalent to a single human brain. The overall AI is equivalent to the combined sum of all human brains. And the cost of AI continues to collapse.
Intelligence has infinite demand, and according to Leffler's Law, its cost will continue to decline. Yet, at the same time, leading labs will enjoy profit margins because anyone serious will pay for the most advanced intelligence. That’s why Anthropic has a 65% profit margin—and those margins will persist. Because whoever controls the most intelligence wins. It’s that simple and obvious.
Above.
Liquidity is cyclical; it pauses, as usual.
When the dollar declines, short-term rates fall, and the yield curve steepens, financial conditions ease and liquidity returns. And when liquidity comes back, it meets an unblinked capital expenditure cycle, $2.5 trillion in signed orders, and a product everyone says isn’t enough—stop overinterpreting it.
Twitter: https://twitter.com/BitpushNewsCN
BitPush Telegram community: https://t.me/BitPushCommunity
BitPush TG subscription: https://t.me/bitpush