No, it ain't over...
文章目录
Everyone has decided the AI trade is finished. The hyperscalers missed, capex has peaked, the circular deals are Enron, the debt is a bubble and it's 2000 all over again...
I don't buy a word of it. Here is what is actually going on.
The market is pricing a pause in liquidity
The oil shock from the Iran war pushed the dollar and short-term rates higher, and those two are financial conditions. Conditions lead liquidity by about a month right now, and global liquidity growth has slowed every month since March. The Nasdaq runs at a 97% correlation with liquidity, so when the money slows the multiples come down with it, whatever the businesses are doing underneath. This is The Everything Code....
That's it. Everything else you are hearing is narrative fitted to the price after the fact.
And underneath, the businesses are doing this...
Microsoft's capex came in $5bn under its own guide, $45bn against more than $50bn, because its suppliers couldn't deliver the capacity in time. Azure grew 48% in the same quarter, ahead of guidance, and the next quarter is guided back above $50bn. Nobody guided 2027 lower, and Amazon says it will spend significantly more. The four big clouds hold $2.49trn of signed contracts, roughly double a year ago.
Anthropic hit $46bn of annualised revenue in five and a half years. Google took 14 years to get there, Amazon 17, Nvidia 30 and Microsoft 32. It is the fastest revenue ramp in the history of business.
The builders are earning on it too. Big-3 cloud operating margins have gone from 5% in 2019 to 39%, while the spending soared, and operating cash flow at the big four is growing 34% a year against a 4.9% cost of money. That is the opposite of 2000.
And the hardware can't keep up. On our capacity composite, built from TSMC and Micron revenue, the chip-equipment makers and Big-5 capex, this is the tightest chip cycle since at least 2011.
A company that can't get its chips plugged in fast enough doesn't have a demand problem...
The first product in history with infinite demand
Everything else saturates. You only need so many cars, so many barrels of oil, so many phones. Intelligence doesn't, because every time its price falls we find more things to do with it.
The price of a fixed level of intelligence is falling about 13x a year, and every 1% off the price now brings 2.2% more use, a ratio that has risen all year. Tokens through OpenRouter are up 32x in a year, to 188 trillion a week. Welcome to Jevon's Paradox...
And the humans are no longer the main customer... the machines are. Agent apps alone now send 43 trillion tokens a week through OpenRouter, up 26x in a year.
Coding agents co-wrote 20 million commits on public GitHub last month, up from 3,000 in January 2025, and opened 4.7 million pull requests. Every agent that gets smarter takes on more work, which needs more agents, which need more compute, which makes the next agent smarter again. Repeat until infinity....
A product that is also its own customer. There has never been a market like it. Not even fucking close. This sure ain't broad band or fibre or whatever BS people tell you...
Nobody can slow down
Intelligence per unit of energy is doubling every four to five months. Step back for a year and you're three doublings behind, with an eighth of the intelligence per watt of whoever kept going. You don't catch that up. You're out of the race for good.
That goes for Microsoft against Google, OpenAI against Anthropic, and the US against China. On OpenRouter, the developers' market, Chinese models now carry 70% of the tokens, up from about a quarter in January. No American president is going to let the US step back from that race. Capex is the entry fee, and it goes up every year. The US holds the frontier and China holds the volume, pretty much as it's always been in all products ...
The only real constraint is supply
Installed AI compute is growing 1.8x a year, while demand is growing somewhere between 7x and 35x depending on the meter. Gas turbines are sold out to 2030, the big transformers take two years and the chips are sitting in crates waiting for electricity. ) Demand is massively overwhelming supply and not the other way around.
The demand isn't going anywhere. It's queuing.
And the big picture... AI already draws about 10 gigawatts. Every human brain on Earth together runs on about 160. On the central path, AI passes all of them around 2032. I'm not sure you quite understand this chart, but all of human compute will be surpassed by AI by 2032.
Also, people like to say, well, humans only take 20 watts of energy to run a brain. Let's explain it in a different way, because everybody is wrong...
Anthropic is cash flow positive, so let's take their subscriptions. For $200 a month, I get intelligence better than pretty much any human in existence in every subject known to humanity. That's $2,400 a year is my cost. That is pretty much Anthropic's cost too. But even if it's not, let's say Anthropic's cost is $5,000 a year. Well, the average human anywhere in the world costs double that at least to feed, and that is the energy input for one human brain.
People confuse training the entire model and serving it to a billion people versus your instance of AI, which is the equivalent to a human brain. AI in total is the equivalent of all of humanity's brains. And the costs of AI keep collapsing.
Intelligence has infinite demand, and it's going to follow Wright's law in terms of cost. But at the same time, the frontier labs will command a margin because anybody who's serious needs to pay for the maximum intelligence. This is why Anthropic has 65% margins. And margins will hold. Because he/she who commands the most intelligence, wins. Simple . And fucking obvious.
Put it together
Liquidity is cyclical and it has paused, the way it always does.
When the dollar rolls over, short rates come down and the curve steepens, conditions ease and the liquidity comes back. When it does, it meets a capex cycle that never blinked, $2.5trn of signed orders and a product nobody can get enough of.
Try not to over think it.
Every chart here comes off the GMI dashboards, live and updated as the data lands, and our Super Cycle Monitor, eight measures from token demand to credit, still reads ACCELERATING. The full work, and what we're doing about it, is in the November Global Macro Investor.
