SpaceX has a binding commitment for an approximately $40B master finance lease with an entity managed by Apollo, which Apollo calls the largest GPU financing ever. The lease will supply Nvidia Vera Rubin chips for its terrestrial compute buildout, with Goldman Sachs as sole structuring agent and anchor commitments from funds managed by BlackRock and Goldman Sachs Asset Management and accounts advised by KKR. The lease is expected to close before year-end 2026, and lease schedules begin when the equipment is delivered in the first half of 2027. SpaceX's CFO said the company sees payback periods under one year on its AI infrastructure investments. (Apollo press release)
Private Companies
Unconventional AIother
Unconventional AI agreed to acquire NeuReality, an Israeli AI inference systems and silicon company, with both boards approving the deal. Terms were not disclosed, and closing is expected in the fourth quarter of 2026. NeuReality CEO Moshe Tanach and its founding team will join Unconventional, which plans to keep NeuReality's engineering presence in Israel. Unconventional CEO Naveen Rao said the company is chasing "a 1,000x improvement in AI compute efficiency." (Unconventional AI)
Classiqother
Intel signed a definitive agreement to acquire Classiq, a quantum software company whose platform builds and runs quantum programs across different hardware. Intel did not disclose a price, and SiliconANGLE reported about $2.2B. Closing is expected in the first half of 2027, subject to regulatory approvals. Intel CEO Lip-Bu Tan said the deal turns Intel's Quantum Computing Group into "a full-stack quantum computing business." (Intel)
Public Markets
Meta launched a $35B equity offering, $22.5B of Class A stock through forward sales and $12.5B of mandatory convertible preferred, and plans a separate $60B at-the-market program of Class C shares. Nvidia, AMD and Broadcom indicated non-binding interest in up to $5.0B of the common stock. (SEC filing)
Emerging
KV cache, decompressed in place: A paper from Yuetao Chen, Peng Wang and Hong Xu, ZipFusion: Fusing Lossless KV Decompression with Attention for Efficient LLM Decoding, has the attention kernel read a compressed KV cache, the stored attention state that grows with every token, and rebuild the values in registers instead of decompressing it back into GPU memory first. On an Nvidia H100 it reports about 1.25x faster decode attention than FlashInfer, and up to 26.3% higher decode throughput than uncompressed SGLang at 8K context under the same memory budget. The figures are the authors' own. (arXiv)