The title of this post is taken from an apocryphal story from the South Sea Bubble in 1720, where promoters lured investors with vague promises of high profits without revealing the actual business. A cautionary tale of speculative frenzy.
The Anthropic IPO is tomorrow, and I understand the final price has been set as I’ve been writing this at $83 per share, giving a total valuation at $1.5 trillion. I have read that the IPO is oversubscribed 10 to 20 times so it seems there was no trouble placing the shares, but the real question here is not the value tomorrow, but what the shares are likely to be worth in a year, or two, or three. That’s what I’ll be talking about today, so let’s start with some numbers.
The IPO tomorrow is projected to raise about $90bn in new capital for Anthropic but in the context of capex plans for $1 trillion in spending over the next few years that’s not a huge number. It will give a decent capital cushion in the short term at least.
Anthropic made $16.96bn in revenue in Q3 2026, up from $11.56bn in Q2. Overall these were the key numbers for income and cash for the year to the end of September:
In terms of just revenue, these were the revenue numbers for the first three quarters of 2026.
What about liabilities? Here the obvious big numbers are what Anthropic will be paying for compute costs. My thanks to 9fin.com for crunching those numbers to produce the chart below showing Anthropic’s 2026 revenue vs compute costs & calculating the likely annual compute cost for 2027. This is likely a lowball number, or at least is likely to rise considerably after 2027 but is a decent working minimum:
So, just crunching those numbers, the lowball annual contracted compute cost is likely at least $152bn in 2027. Overall revenue for 2026, extrapolating from the three quarterly numbers so far would come in at perhaps $55bn. At the absolute minimum I’d expect that to make an operating profit in 2027 Anthropic would need to triple revenue from 2026 levels. How likely is that?
Well how are token prices looking? Hmmm, not great, so that revenue estimate for Q4 2026 may be a serious overestimate:
How far are US AI models ahead of China & (less so) Europe? It looks like they are staying about four to six months ahead of China. A consistent but not a commanding lead for a level of AI sophistication that is demonstrably useful for hackers and criminals, but less obviously useful to businesses so far.
Which models are developers using? Hmmm, that can’t be good.
There are bigger picture issues here worth mentioning.
AI is likely to be big and an important tool for improving productivity for decades to come, even if productivity gains from AI so far have proved hard to identify. With that said:
The research I’ve seen suggests strongly that most AI customers aren’t much interested in AI frontier models, which is where competition seems to be fiercest. What they are looking for is AI that is cheap and good enough. That gives the Chinese open source models a big advantage as now, and longer term, they are likely to be both cheaper and good enough.
There is a common view that AI will be the usual Tech winner takes all scenario leaving one or two big players dominating the market worldwide. That seems unlikely for the following reasons:
Open Source - Chinese open source models will deliver the ability for most regions and countries and larger and perhaps medium-sized companies to run their own AI locally as a commodity that is good enough. No need to buy tokens elsewhere. Any uses requiring advanced frontier models may go to the frontier leaders but seem unlikely to be a very large or particularly loyal segment of the market. Most AI capacity will likely grow locally. China and to a lesser extent so far Europe are moving in this direction already, others will too.
Security - AI is likely to be strategically important. Other big Tech monopolies with any strategic importance will likely see regional competitors created and be replaced by them in non-US markets. Those non-US competitors may well displace their US competitors in markets that don’t trust the US not to weaponize their dependence on it, which is increasingly everybody outside the US and Israel, though that may improve after 2028. That obviously starts with Starlink, with a European alternative coming as soon as possible, but won’t end with them.
The years of the big US worldwide tech titans may well be drawing to a close, not so much for consumer products like Apple or Netflix or Amazon or computer chips and memory like NVidia, but in areas like operating systems, business processes, social media and data processing, where US Tech titans like Microsoft, META & Google in their search engine business will likely come under increasing pressure. Europe has already started to gradually move over to cheap and reliable Linux and generic business process applications and is considering banning some social media networks, notably X and Facebook, for unacceptable content, poor controls, negligible child protection, and political interference on behalf of the US administration. Expect this trend to continue in Europe and elsewhere. The world is changing.
Liability - When AI goes wrong, who is liable? The liabilities are potentially in the trillions, as we should all be aware in a week where many big networks in the US have been failing with rogue AI as a possible suspect. Who bears the liability when AI agents go wrong? The good news is that if rogue AI eliminates humanity then there will be nobody left to file a lawsuit, but there are a lot of other liability issues here, particularly as frontier AI has repeatedly escaped testing sandboxes to penetrate outside systems. This is a big imponderable that could wipe out big AI companies if big things go wrong, and even if a lot of small things go wrong. These potential liabilities are huge and uninsurable.
Overall buying a share of Anthropic at the IPO price tomorrow seems like a strange kind of lottery ticket. If anything goes wrong, as normal you lose the price of the ticket, if everything goes right then you might get back the price of the ticket. Is it ever going to be valued at more than $1.5 trillion? Perhaps. Is it ever going to be worth $1.5 billion to anyone other than a greater fool? Possibly. If the AI bubble bursts is it then likely to go out of business within a year or two? Probably.
I’m doing a three part series on the (likely) AI bubble and published the first instalment late last week on my The Bigger Picture substack. If you missed it you can see that here. The second instalment in the AI bubble three part series, should be posted on Friday, and will be looking at why the AI bubble is likely to be within a year or two at most, and which Tech companies are most likely to survive the bursting of the bubble.
If you like my analysis and would like to see more, please take a free subscription at my thebiggerpicture substack, where I publish these posts first and for members (from next week) also bi-weekly videos looking at equity indices, bonds, currencies and commodities. Those videos are posted on my Youtube channel after a seven day delay. Links to all my posts from my charting substacks are also always posted on my twitter.










