bubble boi
Semiconductor Market Commentary, Thru Project Mascot
Nebius deliberately avoided locking in long-term contracts for 2026 and 2027, likely because short-term surge demand is strong enough to command higher prices, and that bet should prove right. Per @pequityresearch, Nebius's CEO says demand visibility has improved from 18 months to 24 months, pointing toward early 2029. x.comÂ
Nebius's second major price increase in four months shows insane pricing power, and current estimates could prove conservative if demand and utilization hold at these prices. It raised on-demand pricing 30% in June and another 17-21% from October 1, lifting H100, H200, B200 and B300 rates 53-56% cumulatively. x.com x.comÂ
Nebius's price increases show the GPU compute deflation thesis is empirically dead. It just raised H100, H200, B200 and B300 pricing 17-21%, its second hike in four months, leaving those GPUs up 53-56% since June. Owners of uncontracted capacity don't raise prices twice in a cycle when supply is loose; compute deflation was a model assumption, not a market observation. x.comÂ
Nebius's recently signed higher-margin deals are predominantly with pre-revenue neolabs, making it aggressive to project a few small, short-term contracts' economics across all available capacity for years. Short-term economics are very attractive now, and Nebius management has better insight into demand, but assuming all these customers renew at similar or better terms indefinitely seems unrealistic. x.com x.com x.comÂ
Rising GB300 rental prices make the 30% drop in GPU-access providers IREN, Nebius and CoreWeave hard to justify. @ViralMuzik1989 reports a 32% increase in GB300 rental prices over 30 days, from $9.42 to $12.42. x.comÂ
Lower AI token prices drive more usage, making GPUs scarcer and raising rental rates for neocloud capacity. Gavin Baker reports rising GPU rental prices even as token prices fall; the GPU constraint never left, and higher rental prices make the ARR math work again. x.com x.comÂ
Hardware suppliers and VC keep financing cash-hungry labs and neoclouds, making compute rental prices at CoreWeave, Nebius and IREN look artificially supported. Lab revenue growth is finally plateauing at 800% year over year after roughly tenfold equity gains across the stack over the last decade. The buildout is also crowding out the economy's appetite for debt. x.com x.com x.com x.com x.com x.comÂ
Nebius's already-raised cash and expected customer prepayments should buffer its 2027 plans against a short-term debt-market freeze. Rising GPU rental prices could put new deals at $35m+/MW annually, up from last quarter's $20m-$25m, reducing estimated payback from 22 months to around a year. If Nebius deploys 1 GW next year at a one-year payback, equity value creation could be tremendous. x.comÂ
Neocloud valuations will have to be rerated as GPU rental prices rise, something analysts find extremely hard to predict; B200 rental rates rose 80% last quarter. Nebius's earnings could jump when contracts start renewing at prices 5x or more above existing ones. x.comÂ
CoreWeave and Nebius would reprice first whenever AI labs stop paying up for certainty, because the compute premium sits on energized contiguous capacity, not GPUs. Per a Google CTO on an AlphaSense expert call, SpaceX charges $30-50M per MW, 2-3x other neoclouds, for energized gigawatt blocks. CoreWeave and Nebius are writing shorter tenors, and no lab wants to be locked into 2026 clearing prices past 2028. Hyperscalers with their own silicon have the most cushion. x.com x.comÂ
AI agents' productivity over human labor should lift GPU demand and rates for neoclouds. The bottom looks in for CoreWeave, Nebius, Boost Run, Sharon AI and IREN. Gavin Baker says token prices can fall as revenue per GPU rises; paraphrasing Elon, even $38 per GPU-hour can pay off. CoreWeave's five-year fixed contracts may lag current rates; Boost Run has less capacity locked in. If SpaceX takes more growth, slower cash burn could rerate neoclouds, though smaller customers may pay less. x.com … Show more