OpenAI – Silver Lining

Early contracts could make OpenAI the cost leader.

• Early financial data from OpenAI makes for fairly grim reading, but despite clear problems with scale, its early purchase of capacity at far lower prices than the going rate today may give it an edge.
• Financial data on OpenAI has leaked either from an investor or as a result of the confidential IPO filing with the SEC and found its way into the hands of an independent journalist (see here).
• Perusing through these numbers, a few things immediately become clear.
  • First, growth: which was 253% YoY in 2025, far lower than Anthropic’s, explaining how Anthropic is now a larger company than OpenAI.
  • I suspect that this is a factor of Anthropic being right at the current AI sweet spot, which is AI-generated coding, but it is still not a good look for OpenAI.
  • OpenAI needs to grow much faster if it is not going to be left behind.
  • Second, operating leverage: which is simply not improving quickly enough.
  • Revenues grew by 253% YoY in 2025, but operating expenditure grew by 170%, meaning that operating margins improved but not by nearly enough.
  • Operating margin in 2025 was -160%, better than 2024, which was -237%, but given that revenues were up 253%, this is not nearly enough.
  • This is driven by the short life span of models and the cost to build them, as well as a gigantic increase in sales and marketing (see below).
  • Ideally, growth at this scale and in a software business should deliver very rapid improvement in operating margins as operating expenditure should not need to keep growing at this rate.
  • Third, sales and marketing: which increased by 416% YoY to $5.73bn (44% sales) up from 33% sales in 2024.
  • This makes for particularly concerning reading as it implies that OpenAI has to spend to keep its users, implying that its product is inferior to those of its competitors.
• The net result is that this does not read like the income statement of a going concern, as the financials look unlikely to improve at a rate that would allow it to break even anytime soon.
• These kinds of figures increase the risk of the market saying no when asked to pay more than $1tn for OpenAI at IPO.
• This kind of refusal could easily trigger a valuation reset across the entire industry and is something I worry about when thinking about the valuation of the industry more generally.
• However, there is a silver lining in these figures which may just allow OpenAI to regain its lost momentum and improve its financial performance.
• This silver lining involves the cost of compute, where I estimate that in 2025 it paid around $14bn per GW for the compute it purchased from its customers.
• OpenAI has been very early into the compute scale game and has secured a number of contracts with Oracle and others that appear to be priced at levels way below the $37bn / GW that Anthropic has signed for with xAI and the $55bn / GW that Google has just signed for.
• Oracle appears to be earning around $11bn to $12bn per GW and looking at the figures that OpenAI has paid to Microsoft, its contracts with Azure look to be not far from this level.
• If OpenAI has secured pricing for the next few years, then it could suddenly become by far the cost leader in the AI race, meaning that it can price much more aggressively than its competitors, putting it in a very good position.
• This will become much clearer once the filing documents are finalised and have been made public but it is possible that OpenAI’s early move to secure compute capacity has put it in a commanding position.
• Either way, this is yet another sign that market power has switched to the providers of compute and remains strong with the picks and shovels of the AI boom.
• Hence, I remain very comfortable with both Samsung and Qualcomm and am inclined to take a look at anyone who has new capacity to sell over the next year or so.