Oracle FQ2 26 – Reversal of Fortune

The business model does not quite work.

• The real take-home from Oracle’s numbers is the increasing credit risk as the bond market assesses the business model of providing AI compute, which, as it is today, is problematic.
• Oracle FQ2 26 revenue / EPS were $16.1bn / $2.10, which was slightly light on revenue (consensus $16.2bn) but ahead on profit (consensus $1.64).
• FQ2 Cloud revenue grew 34% YoY to $7.98bn while the infrastructure business grew 68% YoY to $4.08bn, but both of these were slightly below what the market was looking for.
• At the same time, the capex bill for FY 2026 is increasing by $15bn to $50bn to support the rollout of AI data centres for which it has cumulative orders of $523bn.
• However, the problem here is not demand but whether or not Oracle will be able to offer this compute at a decent level of profitability.
• This is what the bond market is questioning, and the cost to insure oneself against Oracle going bankrupt in the next 5 years has risen by 0.05% to 1.246% which is a big move.
• This is what is referred to as a credit default swap (CDS) and is an important gauge of both sentiment and a leading indicator of a liquidity crisis (such as the 2008 financial crisis).
• The move in the CDS mirrors some of my own thinking, as when I look at the business model of building an AI data centre and selling the compute, the numbers don’t add up to anything very exciting.
• In Q4 2025, CoreWeave will generate about $1.56bn in revenues with around 590GW of capacity being online.
• This means that the going rate for 1GW of data centre capacity to rent is about $10bn per year.
• 1GW of capacity is going to cost around $40bn and 3 years to build out, and if the depreciation schedule is correct, it will only last for 7 years.
• On top of that, the data centre will cost around $1.6bn in OPEX to run each year.
• Hence, running flat out with no price erosion for the full 7 years, the 10-year return on invested capital is 8.2%.
• The market is currently charging CoreWeave 10.3% to borrow money, as this is the yield to maturity of its 2030 9.25% coupon bond.
• This is bad news for equity holders as the bond holders get paid first, and the bond market is signalling that it doesn’t like the business model as it stands today.
• This is a reasonably generous assessment of CoreWeave, giving me a strong sense that something in the business model needs to change to prevent a liquidity crunch.
• This is precisely why Oracle’s stock is down 10% in after-hours trading, taking the decline since last quarter’s earnings to 40% and it is why the CDS is widening.
• It is a sign that there is a problem with the business model of providing compute, and that something needs to change.
• Hence, one or more of the following needs to happen.
• The cost of a data centre needs to become cheaper or the price for compute needs to rise, or the data centre needs to produce more tokens per $ spent to produce them.
• Nvidia’s answer to this will be: pay 2x the price of Blackwell for Rubin, and we will give you 4x the number of tokens (2x revenue).
• If I plug that into my back-of-the-envelope calculation, I get a return on invested capital of 13.0% which means that CoreWeave would at least be covering its cost of capital, but nothing like the level that will make the equity holders happy, given what they have to pay to buy the shares today.
• Another answer would be to increase the price of compute, which seems unlikely given the amount of capacity currently under construction and the fact that everyone is fighting to grab the nascent market.
• The third would be to reduce the cost of capex, which for a 1GW $40bn data centre built with Nvidia, $19.6bn is Nvidia’s gross margin.
• This is why everyone is looking at alternatives like AMD, Cerebras, in-house solutions (Google TPU etc), Qualcomm, Grok LPUs and so on.
• The net result is that something needs to give, and I hope that the industry can find a way to increase token output without increasing costs, as this is the only way that does not involve a correction of some kind.
• This is why the CDS on Oracle should now be carefully watched, as it was the CDS that heralded the fall of Lehman Brothers, and if it continues to widen, could cause a liquidity crunch at Oracle and more broadly across the sector.
• Oracle’s shares have given up 40% since the last set of results, indicating that the euphoria that greeted the huge spending plans is beginning to be tinged with reality.
• OpenAI is extremely fortunate that it is not public.