SpaceX just signed a binding $40 billion GPU finance lease with Apollo. Here's what you need to know. SpaceX disclosed the deal in an 8-K filing on Monday, Oct 26. An entity managed by Apollo Global Management will buy GPU pods and related data center equipment, then lease them to SpaceX under a master finance lease. Elon Musk said SpaceX closed $40B of debt for GPUs. The equipment is NVIDIA's next-generation Vera Rubin chips. They will power SpaceX's terrestrial compute buildout, meaning ground-based AI data centers. Reports say BlackRock, Goldman Sachs and KKR are participating. The equipment will be delivered in tranches, and each tranche gets its own lease. Each runs roughly five years. Lease schedules start in the first half of 2027, once the GPUs are delivered. The commitment is binding, but the deal is still subject to closing conditions. SpaceX expects to close before the end of 2026. The price per GPU has not been disclosed. The market reaction was mixed. Shares were up about 1.2% premarket and traded near $147, up about 1.9%. Bernstein raised its price target to $267 from $248 and kept an Outperform rating. It now models 13 GW of ground-based AI compute by the end of 2027, 5 GW above its prior estimate, and 20 GW in 2028 at about $35B of capex per GW. Some commentators call the deal a big vote of confidence in the AI buildout, since Apollo had just published a note questioning AI spending payoffs. One post claimed SpaceX failed to find a buyer for a separate debt offering. Key numbers: - About $40B in total scheduled principal payments - Lease term of about 5 years per tranche - Lease schedules start in 1H 2027 - Closing expected before end of 2026 - Bernstein target $267, up from $248 - Bernstein models 13 GW by end of 2027, 20 GW in 2028 SpaceX gets NVIDIA's Vera Rubin hardware without paying for it upfront, and Apollo carries the purchase.

Cua just released Cua-S1-4B-0.2 today. Here's what you need to know. The Cua team, makers of the computer-use driver behind Hermes Agent, has open-sourced this multimodal decision model under Apache-2.0. It's described as the first multimodal decision model trained with RLOO (Reward-Weighted Leave-One-Out) on live computer-use tasks, using task-completion rewards. The model runs on a frozen Qwen3.5-4B base with LoRA adapters. At each step it takes in the screen state, the task goal, and a fixed set of candidate actions, then returns one action, no accessibility tree required. Training happened in two stages: supervised fine-tuning to teach the decision format, then agentic RL in live cua-bench-basic environments where only a fully completed task earns reward. On the GUI-360 benchmark, it scored 92.9% versus 60.1% for comparison. Text and multimodal adapters, training code, and benchmark results are all merged into the Cua GitHub repo and live on Hugging Face. This is version 0.2, up from 0.1 released the day before. Key numbers: - 4B parameters (Qwen3.5-4B base) - 92.9% vs 60.1% on GUI-360 benchmark - Apache-2.0 license - v0.1 to v0.2 in one day The model is small enough to run on-prem, and developers are already testing it on local hardware for business workflows.