Several data points this week have underpinned our long compute thesis across neoclouds, scalers, and powered shells. ORCL reported Q2 earnings, OpenAI, IREN and SPCX presented at Goldman’s conference and MSFT and OpenAI had other data points released into the market.
We believe the tight compute market is here to stay, and CRWV and SPCX are set up for upward earnings revisions as analysts start to flow through $20–50M/MW/year revenue recognition for future compute brought online. A key positive data point for neoclouds was ORCL’s statement that older GPUs re-contracted at an average premium of 20%.
CRWV’s consensus estimates imply just under $12M/MW/year in revenue realization for its entire portfolio by 2030. The missing link is that ~5.5GW of its 8GW target by 2030 is open for pricing, which means the likely average realizations could be north of $15–20M/MW/year. As a reference point, every $1M/MW/year of additional realization at 5x EBITDA would be worth $17/share. An additional benefit from this repricing of the model will be the improved cash flow and debt metrics, which could have an additional positive impact on the equity.
ORCL looks positioned to move higher into its October analyst day after a strong earnings report that showed acceleration of backlog from a wider breadth of customers as well as solid execution commentary. OpenAI’s positive Astra comments and recent CFO comments about accelerating demand also go a long way toward reducing counterparty concerns for ORCL. Capital discipline also appears to be on the horizon, and the FCF flex will be powerful when it happens. This appears to be <2 years away, based on various comments by the CFO.
In addition to neoclouds and scalers, powered shells are the other leg of the bullish compute thesis. Power is a critical resource and near-term access is increasingly hard to find. MSFT’s and OpenAI’s news today (below) would indicate assets long access to power will be in high demand. FRMI will have 4+GW of power available within 24 months, and other select powered shells could provide strong risk/rewards.
ORCL Q1’27 earnings takeaways:
ORCL characterized the quarter by the word “acceleration,” establishing a confident tone around 2027 guidance, with no impact from delays at what it describes as smaller NM and WI projects. ORCL is bringing facilities online and getting acceptance turnover time down to 24 hours, which shows a move up the learning curve of AI factories. OpenAI trained Astra at Abilene, and the first Vera Rubin compute will be installed in Q2.
A positive data point for older-generation GPU value: ORCL indicated that older-generation GPU contracts renewed at an average 20% premium. This indicates continued pricing power for prior-generation GPUs, a topic that has been the subject of significant debate for neocloud earnings, valuation, and financing capabilities.
Capex came in at ~$28B, in line with consensus; the company did not raise capex guidance and notably spent more time discussing the future self-funding nature of the buildout and the emerging FCF profile. The commentary would lead you to believe an FCF moment could be coming in 1–2 years as the major buildout spend peaks.
SPCX, IREN, MSFT and OpenAI data points:
SPCX presented at the Goldman conference and cited increased confidence in reaching a $100B ARR exit rate this year, driven in part by an announced $1B/month compute contract beginning in December. The company also added confidence around its ability to bring on significant additional compute in 2027, noting that it would not have made this statement without line of sight to the power. This was an important statement that can validate the revenue opportunity in a rising compute price environment.
SPCX’s CFO also discussed orbital data centers, noting that the company sees costs converging with terrestrial data centers and plans to bring significant compute to space in 2028. The Starlink business indicated strong demand trends, and there continues to be a growing undertone of competition with AT&T and Verizon, as the company shared that the new Starlink versions will be 5G-comparable.
IREN presented at Goldman and, based on various conference notes posted online, validated rising compute pricing in the market: 3-year deals at over $25M/MW/year and shorter-term contracts (under 1 year) at $40M/MW/year. This continues to validate elevated and rising pricing trends for compute.
A report on MSFT’s compute build strategy came out on Bloomberg after the close. It is from an unidentified company source, but it cites MSFT’s intention to triple its data center capacity to 38GW by 2032 (12GW currently). Apparently only 2GW are AI chips, and that is expected to grow to 12GW by 2032.
In our analysis of MSFT’s U.S. AI compute, we see the ~2GW of compute growing in the U.S. to 7.5GW by 2030. Clearly the article sounds like an acceleration on MSFT’s part, but we don’t know the geographic composition. If it were very U.S.-centric, MSFT could continue to lean in on the Nscale Monarch campus, where upwards of 8GW of capacity has been indicated.
OpenAI: OpenAI is out of compute. According to a tweet by an OpenAI employee (link), the company is rationing its highest-tier subscriptions due to overwhelming demand for Astra. We’ve seen this movie before: it drives prices higher and adds demand for additional compute. This is also good for ORCL sentiment.
