Major neo-cloud vendor securing large-scale inference contracts

Expert Takeaways: Meta > Deployment timeline (30% weight) is the #1 procurement criterion, outranking unit economics (25%), meaning the buyer happily pays a premium for capacity delivered 6–12 months ahead of internal build schedules. > While 72% (4.3GW) of its 6GW footprint is self-operated for frontier training, a single neo-cloud vendor (such as CoreWeave, which holds massive multi-billion-dollar inference contracts with Meta) accounts for 9.2% of total compute—nearly a third of all outsourced capacity (1.7GW). > On-prem clusters are reserved for sensitive, foundational model training, while external clouds absorb volatile workloads like inference, R&D experiments, and demand spikes. > By 2028, inference will grow to 53% of total compute compared to 27% for training, even as absolute training capacity continues to expand. > Flagship chips like the H100 have only a 2–3 year shelf life for cutting-edge training before cascading down to inference and fine-tuning tasks. > Next-gen architectures (Vera Rubin) command 36-month commitments with 90–95% renewal rates, whereas legacy H100s see short 6-month renewals (40–50% rate) at steep 30–40% discounts. > The rollout of Rubin GPUs—delivering vastly higher inference performance per megawatt alongside higher rack density—will lower per-GW leasing costs, compounded by ~15% long-term concessions and legacy Hopper discounts. > Rolling out liquid-cooled Blackwell systems carries significant delivery friction, with 30% missing initial go-live dates and 14% delayed by more than a month. > Although neo-clouds get 92.5%–95% take-or-pay floors to finance buildouts, payments only begin after formal acceptance, and suppliers face 0.75% weekly delay penalties (up to 12.5% of ACV) plus cancellation rights at 90–120 days. > At massive 100,000+ GB200 GPU scales, Oracle Cloud Infrastructure (OCI) matches neo-cloud pricing, while traditional hyperscalers can still win deals at a premium if they deliver faster, broader geographic reach, or superior uptime. Reminder: not all expert provided information may be accurate. Please be advised. $NVDA $NBIS $CRWV $ORCL $META