Anthropic – Not Quite Plain Sailing

Some adjustment may be needed at IPO

• Anthropic’s focus on coding and the enterprise is paying dividends, making an IPO viable, but there is great uncertainty regarding H2 2026, which may cause some problems.  
• It has been clear that Anthropic is in much better financial shape than OpenAI for some time, and this has become clear by having a look at some of the other estimates that are available.
• Triangulating RFM research, The WSJ and SemiAnalysis, it is clear that Anthropic is by far the global leader when it comes to generating revenue from AI, that it has broken even and should be generating cash.
• This is almost entirely due to Anthropic’s ability to grow the amount of revenue that it is able to generate relative to what it pays for the compute capacity it uses to create and sell its product.
• By my estimate, revenue was $17bn/GW in Q1 26, $20bn/GW in Q2 26, will be $23bn/GW in Q3 26 and $25bn/GW in Q4 26.
• My rule of thumb for 2026 has been that an acceptable return on investment is achieved at $15bn/GW, indicating that Anthropic is in pretty good financial shape.
• However, the market’s forecast is driven by Anthropic’s cost of inference compute growing much more slowly than its prices, as using SemiAnalysis estimates of compute capacity gives a cost of $6.3bn/GW in Q1 26, $7.2bn/GW in Q2 26, $7.8bn/GW in Q3 26 and $8.5bn/GW in Q4 26.
• If these are accurate, then breaking even is simply a matter of being big enough to cover the cost of training as well as the normal operating expenses of running a company.
• On an adjusted basis, it achieved this in Q1 26 and on a GAAP basis, it looks like this will be achieved in Q3 26.
• However, the cost numbers do not make sense and, in my opinion, are much too favourable to Anthropic.
• Anthropic’s recent deal with xAI was for around 400MW of compute, which I think will start during this quarter and for which it is paying $1.25bn per month or $37.7bn/GW.
• Adding this 400MW of capacity in Q3 26 should lift the average inference cost per GW by at least $4bn/GW, but the public estimate seems to be rising by only $0.7bn/GW.
• This implies to me that the forecasted profitability of Anthropic may be too high, as it does not appear to be taking into account that the scarcity of compute has caused the price to skyrocket.
• This would not preclude Anrthopic from making money, but it would be making less money than perhaps the market has now come to expect.
• Hence, when the figures are fully detailed, there may be some disappointment as Anrthopic’s weakness is that it did not buy enough compute early on when prices for compute were more reasonable, meaning that it has to pay much higher prices now.
• I think that within a few quarters, prices are likely to become more reasonable and with a 90-day cancellation option, Anthropic will have the option to renegotiate.
• Anthropic is clearly in the right place, at the right time and with the best product by a meaningful margin, but competition is rapidly coming.
• The net result is that I think that the market is currently painting an overly optimistic picture of Anthropic’s financial performance, meaning that some adjustment in terms of performance and valuation is going to be required.
• The impact on the size and success of the IPO is therefore uncertain in my view, meaning that yet another delay may occur.