Datacentre Update – Still Adrift

Q3 2026 needs to be amazing.

• A quick glance at CoreWeave and SpaceX confirms that while both talk very confidently about the future, the current financials leave a lot to be desired, meaning that Q3 26 must see a large price improvement or really hard questions will begin to be asked.
• When looking at these businesses, it is important to remember that adjusted EBITDA is nothing more than a fantasy.
• This is because the vast majority of the costs that enable datacentres to operate are depreciation, which this measure explicitly excludes.
• Therefore, adjusted EBITDA is not a true and fair reflection of how the business is operating, and so this analysis will focus on income from operations (EBIT) and net cash flow.
  • First, SpaceX: which reported earnings on 4^th August 2026.
  • Here, xAI generated $2.56bn in revenues and lost $1.26bn in income from operations (-49% margin).
  • SpaceX ended the quarter with 1.4GW of compute capacity (10Q page 36) up from around 1.0GW at the end of Q1 26 giving an average capacity of 1.2GW for Q2 2026.
  • On this capacity, it earned $2.56bn, giving an average price of $8.5bn/year/GW for its compute capacity.
  • This is why xAI has negative margins of 49%, as RFM Research has long calculated that using Blackwell, datacentre AI compute is barely profitable at $10bn/GW and needs to be around $12bn/GW to generate a return of 8%, which is just about acceptable.
  • However, on the call, SpaceX mentioned that datacentre pricing for Blackwell and Rubin is between $30bn-$50bn per GW, meaning that there should be much better times ahead.
  • If xAI had had that pricing locked in in Q2 2026, it would have generated revenues of $12bn in Q2 with operating margins of 68% and substantial cash generation.
  • The contracts that SpaceX is referring to with this pricing are supposed to start this quarter, meaning that there should be a substantial lift in revenue/GW the next time it reports.
  • Failure will not be well received.
  • Second, CoreWeave: which reported Q2 26 on 11^th August 2026.
  • Here, CoreWeave generated $2.56bn of revenues and lost $49m (-1.9% margin) in income from operations.
  • CoreWeave ended the quarter with 1.5GW, having added 500MW during the quarter, meaning that it had an average of 1.25GW for the quarter.
  • This means that it had an average price of $8.2bn/GW, broadly in line with SpaceX but still below what would be considered a level that would generate an economic return.
  • However, CoreWeave is doing better than SpaceX, but I suspect that this is because xAI is using some of its capacity for internal uses, meaning that my estimate for how much capacity it had to sell is overstated to some degree.
  • Either way, it is clear that CoreWeave also needs to see a substantial increase in pricing in Q3 2026 to begin to see returns that will justify the current valuation of its equity.
• Despite the fact that these businesses are currently not operating economically, I am confident that this is going to rapidly change in H2 2026.
• This is because compute is so scarce that pricing for capacity available now has risen to a level where supernormal profitability is on offer.
• This is the $30bn – $50bn / GW that Mr Musk was referring to on the call and matches the terms of the deals that he has signed with Anthropic and Google.
• If this level of pricing is sustainable for a few quarters, then xAI and CoreWeave will quickly generate substantial amounts of cash, but I suspect that this bonanza will be short-lived.
• This is because any rational person with capacity at the moment will be parking their own activities and making their capacity available for sale (e.g. Meta).
• This, in turn, will cause a normalisation of pricing as more capacity becomes available, and I expect to see it settle somewhere between $20bn-$25bn / GW.
• This still offers good returns but far short of the spectacular returns currently available.
• This is why I am expecting that Meta will make its capacity available for a short period and then internalise it again once pricing has normalised.
• The other good news is that older capacity is running for much longer than anticipated, with A100 (3 generations old) still in use today, as CoreWeave has been able to contract it out until 2029.
• This capacity should have already been fully depreciated on a 5-year time horizon, meaning that it will be able to contribute to improving profitability going forward.
• The net result is that there is still no sign of datacentre overbuild, but with the plans that are in place for 2027, this might become a problem.
• This is especially the case as, to make a return on the capacity that it is contracting for, Anthropic will need to substantially increase its prices yet again.
• This could impact demand for its services, and there are already signs that the pricing for Fable 5 is beginning to weaken demand given how expensive it is to use.
• I think that for the next 6 months, the datacentre bonanza is going to continue as capacity takes around 2 years to build and so shortages are likely to persist for a while.
• This is why I have a position in Meta (which is fairly cheap anyway) but could offer further upside if it rents out its capacity into this market on a short-term basis.
• I am looking for around $770 per share on this position.