On October 21, Ken Griffin said AI is "almost certainly" a bubble. Chamath Palihapitiya replied with arithmetic.
Palihapitiya
"1% of companies represent 90% of AI spending. Of that 90%, 80% of this spend are from unprofitable VC backed companies. So, roughly, $0.70-75 of every dollar of AI spending right now comes from AI companies spending VC investment dollars."
If that holds, roughly three quarters of the AI industry's revenue is recycled investment rather than end demand. Which would mean the revenue growth everyone is pricing off is substantially a measure of how much venture capital was raised, not how much value was delivered.
It is a serious claim and it deserves a serious look at where the numbers came from. They came from nowhere.
72%
Of every AI dollar he says is venture money. His own arithmetic, on his own inputs.
Zero
Sources cited for the 1%, the 90% or the 80% that produce it.
600x
The concentration gap that is measured, by a16z and separately by Ramp.
Three figures, no source, one of them doing all the work
Chamath cites nothing. Not a dataset, not a report, not a methodology. Secondary coverage checking the post found the same thing: the 1%, the 90% and the 80% are stated and not attributed.
Only the final figure is his own work, and it is division. 90% times 80% is 72%, which is where "$0.70-75 of every dollar" comes from. The arithmetic is correct and it is the least interesting part, because an unsourced number multiplied by an unsourced number produces an unsourced number with more decimal places.
So the useful question is not whether 72% is right. It is whether the shape survives contact with data somebody actually collected.
Real data exists nearby, and it measures something else
Ramp runs a corporate card. It can see what companies actually pay AI vendors, which makes it the closest thing to a census of real AI spending. Its numbers rhyme with Chamath's and they are not the same numbers.
Look at what the second row does and does not say. 80.4% of top-1% AI spenders are VC-backed is a headcount, not a dollar share. It tells you the heavy spenders are overwhelmingly venture-funded companies. It does not tell you what fraction of total AI dollars those companies represent, which is the entire load-bearing claim.
The third row is narrower still. Kharazian's 80% is 80% of spending on OpenAI and Anthropic, not of all AI spending. Different denominator again.
And the concentration itself is not in question from any direction. a16z, measuring independently, found the top 1% of AI spenders outspending the median company by more than 600x, and outspending even the top 10% by roughly 8x. Our own write-up of $12 per employee against $7,400 is the same gap from Ramp's side.
The circularity question is the real one
Set the sourcing aside for a second, because the mechanism Chamath is pointing at does not depend on his exact percentages being right.
If a large share of AI revenue is venture-funded companies buying compute from other AI companies, then some of that revenue is the same capital counted more than once as it moves down the stack. A fund invests in a startup, the startup spends it on inference, the inference provider books it as revenue, and that revenue supports the provider's valuation, which supports the next round.
That is not fraud and it is not unusual in an infrastructure buildout. It is how railways and fibre were financed. It matters because revenue quality differs: a dollar from a profitable enterprise renewing a contract and a dollar from a startup burning its Series A are the same dollar this quarter and very different dollars in eighteen months.
Chamath's own phrase for where this goes is the one to keep. He has warned separately about the fallout from tokenmaxxing, which is the same argument at company scale: spending that was never tested against value does not survive the first quarter anyone tests it.
What a CEO should take from this
Do not quote the 72%. It is unsourced, and anyone who checks will find that in about four minutes. The number will be used against you in the room where you use it.
Do quote the concentration. It is measured, from more than one instrument, and it is the genuinely useful fact: AI spending is not broadly distributed across the economy. It is a very small number of companies spending enormously, and the median company is barely participating.
Which reframes the usual anxiety. Most boards worry they are behind the curve on AI spending. The data says the curve is almost entirely one percent of companies, and most of those are venture-funded firms whose spending is a function of their last round rather than their customers.
Being outspent by that cohort is not evidence you are losing. The comparison that actually matters is against companies with your revenue model, and the most advanced adopters have started spending less, not more.
Keep reading.
Teardowns, benchmarks and playbooks on building AI-native products.
Browse the Playbook →