OpenAI – The Numbers Game

Everything depends on revenue per GW.

• OpenAI’s new revenue forecast should be enough for the company to start generating cash, but this depends on revenue per GW increasing substantially from $10bn / GW where it has been stuck for the last 3 years.
• OpenAI’s latest projection has it earning $100bn from advertising by 2030, which presumably means that the other $180bn will be coming from subscriptions and enterprise sales.
• For once, I actually think that the advertising forecast is a perfectly reasonable one for it to have made (in size, perhaps not in time).
• I suspect that there are over 1bn users who access ChatGPT at least once per month (MaU) but do not pay it anything.
• OpenAI does not have an efficient search algorithm upon which to fall back on, and so the compute that these users consume is both large in size and expensive.
• Hence, some form of monetisation is essential, and Google, Apple and Tencent have built highly profitable businesses on the back of monetising the digital consumer ecosystem.
• Hence, given the size of the user base, there is no reason why OpenAI cannot monetise these users using either advertising or subscription and achieve the advertising forecast.
• The net result is that I think that $100bn in advertising revenues is possible, and so I am willing to entertain the $280bn 2030 revenue forecast.
• OpenAI has committed to spending $600bn by 2030, which will get it roughly 10GW of compute capacity (RFM Research estimates that Rubin costs $60bn per GW to roll out).
• Assuming that the forecasts are met, this means that in the year 2030, OpenAI thinks it will be generating around $28bn per GW.
• This is where things get awkward, as the industry standard at the moment is that 1GW will generate $10bn of revenue, which would leave OpenAI only able to generate $100bn of revenue, a miss of 64%.
• Furthermore, the company would still be burning cash and so would have an equity value of $0.
• This is why revenue per GW is so important, meaning that Jensen’s comments during his keynote at GTC 2026 were the most important of the whole conference.  
• Here, he claimed that the combination of an increase in token average selling price (ASP) and higher token production per GW made possible by Vera Rubin would allow revenue per GW to rise by 5x to $50bn.
• OpenAI is effectively estimating a 2.8x rise, which, with Vera Rubin, would offer a 26% annualised return on a 1GW data centre that costs $60bn to build.
• If we take OpenAI’s 10GW forecast, that would mean that in 2030, it would have revenues of $280bn and a profit of $149bn or 53.9%.
• However, this does not take into account sales and marketing or research and development, which in most technology companies is usually somewhere around 25% of revenue.
• This would leave $79bn in operating profit or something like $65bn after tax.
• Put a 20x PER multiple on that, and one can see how one gets to a $1.3tn valuation.
• However, the industry has stubbornly been stuck at $10bn per GW for 3 years now, despite promises that the H100 and Blackwell would increase revenue per GW materially.
• Hence, the financial future of OpenAI and everyone else (except equipment suppliers) rests on the historical trend being broken and revenue per GW beginning to grow.
• To break even on its compute, OpenAI needs revenue per GW of $12.7bn, and to generate cash, I estimate it needs $17bn / GW or more.
• That is why revenue per GW is the metric to watch, and so far, there are very few signs that it is moving in a positive direction.
• Hence, I remain very cautious about OpenAI and continue to think that it and Anthropic will offer very little to investors at IPO.
• Stick with the vendors who are relatively cheap and are making money today rather than jam tomorrow.