From the latest financial reports of CRWV and NBIS, we can see five clear development trends in the CSP (AI Cloud Computing) sector: 1. The demand still significantly exceeds the recent deliverable supply. At the end of the second quarter, CoreWeave had approximately 1.5GW of operational power and 3.7GW of contracted power, with a backlog of orders of approximately $104 billion, and did not include the additional commitment of over $25 billion at the beginning of the third quarter; Nebius plans to deploy over 1GW annually starting from 2027, with a mid-term goal of 5GW contracted capacity. The industry bottleneck has shifted from "whether there are customers" to "when can power be turned on, GPUs delivered, and billing started". 2. The pricing power has indeed been strengthened, but not all GPU rental prices have increased indiscriminately. The management of CoreWeave stated that prices for various SKUs were generally increased by about 25% in July; Nebius disclosed that the price of old generation GPUs has increased by more than 30% compared to the first quarter. In the second quarter, the average annualized revenue from newly signed contracts exceeded $20 million per MW, with some reaching $20-25 million. In the early third quarter, the short-term emergency capacity even reached $40-50 million per MW. The most obvious price increases are in short-term capacity, new generation GPUs, large-scale clusters, and production level inference, rather than traditional low priority, long-term locked in raw computing power. 3. The project level ROI can be accounted for, but the company level ROIC has not yet been proven. Nebius has disclosed for the first time a relatively clear project payback period; CoreWeave covers GPU investments with five-year contracts and asset level financing. But both companies are currently in a stage where capital investment far exceeds profit release, and depreciation and interest still consume most of their operating profits. But the fact that the ROI of a single project can already be accounted for is a signal that the business is further improving. Until more and more single project ROIs can be accounted for and company level ROICs can prove it in the future, 4. Both companies are upgrading to "AI infrastructure operating systems". Future competition is no longer just about renting out a GPU every hour, but an integrated platform that provides training, inference, storage, networking, model deployment, monitoring, governance, agent runtime environment, and cross cloud operations. That is to say, in mid July, here is https://(x.com)/qinbank/status/2076839550119444702? The AI operation operating system and solid intermediate layer discussed by s=46&t=k6rimWs Ebo2D2TXolYcM-A 5. The two companies have started using third-party capital, but the methods are not entirely the same. 1) Nebius' light asset model is closer to "funding the construction of AI factories by capital partners, and Nebius outputting operating systems and operational capabilities"; 2) CoreWeave Omni is closer to fully deploying CoreWeave cloud to customers' own data centers and GPUs. The direction is the same, but currently Nebius emphasizes reducing capital investment, while CoreWeave emphasizes hybrid cloud and sovereign AI delivery.