Oracle FQ3 26 – Outlook Overstated.

Revenue is not the problem. 

• Oracle reported excellent results and guided higher for revenue, but with demand for free AI services virtually infinite, it is profitability, not revenue, that is what everyone needs to focus on.
• FQ3 26 revenues / Adj-EPS were $17.2bn / $1.79, nicely ahead of forecasts of $16.9bn / $1.70 and the company guided higher for both the next quarter and for the medium term.
• Here, FQ4 revenues / EPS are expected to grow by 19% – 21% YoY / $1.92 – $1.96 ($1.94), which is pretty much in line with consensus.
• Revenues in FY2027 are now expected to be $90bn, which is slightly ahead of consensus of $86.6bn.
• This is mostly being driven by Oracle’s cloud business and its $300bn deal with OpenAI in particular, which is where the remaining performance obligations (RPO) of $553bn are mostly coming from.
• Oracle is also confident that it can earn 40%+ operating margins on the cloud infrastructure that it supplies to OpenAI, Meta, etc., and this is where the bond market and I have concerns.
• FQ3 26 operating margin on the cloud business was 46%, which is very healthy but I am not convinced that this will be maintained should Oracle build and sell the full $300bn (6GW) to OpenAI.
• This is because the going rate for compute at the moment is around $10bn per GW and with an Nvidia-powered data centre costing around $40.8bn to build and around $1.3bn to run, operating margins are around 13% for the first 5 years.
• This all depends on the depreciation schedule that is applied to the IT infrastructure and here there is a lot of scope for manipulation.
• Google would argue that the IT infrastructure should be depreciated over a longer period of time, as it has TPUs way older than 5 years that are still running flat out.
• This is why the only way to look at the business model of compute is to look at on a cash flow basis and ignore the accounting treatment of depreciation and amortisation.
• Here, the numbers are also not great as RFM has calculated that the internal rate of return on a 1GW Nvidia-powered data centre is just 1%.
• This means that any rational investor would be better off buying “risk-free” US treasuries at 4% than putting his money into a data centre.
• This is what leads be to continue to argue that the business model of compute currently does not work and that something needs to change.
• A 20% increase in revenues per GW or a 20% decrease in the cost to build the data centre would bring the IRR to 8%, but this would still be very far adrift of what Oracle seems to think it can achieve.
• This is why the shares are very far away from their all time high and why the credit default swap on Oracle’s bonds remains at an all time high of around 160bp.
• We are now also seeing the beginning of similar signals coming from the lenders to SoftBank, which is now heavily exposed to the performance of OpenAI, which only serves to increase my worries.
• The bond market has been signalling for some time that there is a problem with the business model of compute and the signals being sent now are similar to the signals that it sent prior to the collapse of Lehman Brothers.
• This is why I remain concerned with the medium term outlook for Oracle, and I derive no comfort from the commentary that the data centre builds are now fully funded thanks to the recent fundraising carried out by OpenAI.
• Oracle earning revenue for the data centres that it builds is not the problem; it is the profitability of those revenues that dictates both the value of Oracle’s bonds and its equity.
• RFM’s calculations using the benchmarks of OpenAI and CoreWeave suggest that Oracle’s guidance for what it can earn from providing computer capacity is overstated, which is why I don’t like the shares even down here.
• Instead, Adobe and ServiceNow (both of which I own) look like better ways to play the software-apocalypse-is-overstated theme but even Salesforce, IBM and many others are likely to fare better than Oracle.
• If I owned Oracle, I would use this bounce in the share price to exit.