CUDA 13.5 makes Vera Rubin GA; the midterms hit data-center power; memory and GPU supply stay tight

Sources: NVIDIA: CUDA 13.5 brings Vera Rubin GA (lab watcher) · The Information: Blue wave could engulf data center industry, Chip stocks and neoclouds sell off, Apollo outspending neoclouds, Microsoft lowered capex after Iren delays, Lumentum after Marvell offer · The Decoder: OpenAI's community compact · X feed: @rohanpaul_ai on RTX Pro 5500 orders, Apple margin, SambaNova + Supermicro, Qualcomm accelerators, AMD CPU demand · Algorithmic Bridge

TL;DR. CUDA 13.5 moves Vera Rubin from early access to general availability. The release is mostly about sharing a GPU well: locality-domain SM partitions can now include all SMs, MPS v3 adds weighted timeslicing, SM borrowing and lending, and cgroup routing, CUDA Tile kernels gain NVLink/NVSwitch multi-GPU pointers and can call SIMT libraries (NVSHMEM, CCCL) mid-kernel, and CCCL 3.5 adds batched Top-K integrated into FlashInfer. On the same day the US midterms turned power into the scarce input: in 12 states the winning or leading governor candidate called for a data-center pause, 22 of the other 24 back at least one limit, and voters rejected new data centers in 40 of 41 local ballot measures. CoreWeave fell 9.9% and Nebius 6.7% while hyperscalers rose, because the incumbents already hold built capacity. Supply stayed tight: Chinese buyers reportedly ordered ~~6M RTX Pro 5500 cards (~~$70B) that Nvidia can only fill ~1% of within four months, and Apple guided its gross margin down to 45% on memory prices.

Software sharing gets better as new power gets harder

CUDA 13.5 squeezes more tenants per GPU; voters made each new megawatt more expensive.

flowchart LR
  V["Midterm vote<br/><small>12 states favor pauses</small>"] --> P["New power<br/><small>harder, slower</small>"]
  P --> N["Neoclouds<br/><small>CoreWeave -9.9%</small>"]
  P --> H["Built capacity<br/><small>hyperscalers up</small>"]
  C["CUDA 13.5<br/><small>MPS v3, SM lending</small>"] --> U["More per GPU<br/><small>tenants share safely</small>"]
  U --> H
  classDef input fill:#d0ebff,stroke:#1971c2,color:#1b1b1b,stroke-width:2px
  classDef core fill:#e5dbff,stroke:#6741d9,color:#1b1b1b,stroke-width:2px
  classDef exit fill:#d3f9d8,stroke:#2f9e44,color:#1b1b1b,stroke-width:2px
  classDef err fill:#ffe3e3,stroke:#e03131,color:#1b1b1b,stroke-width:2px
  class V,C input
  class P err
  class N err
  class U core
  class H exit

Blue is the driver, red is the constraint and who it hurts, purple is the software response, green is who benefits.

Other supply-side items

Apollo is financing chips at a scale beyond the neoclouds: it buys the Nvidia Vera Rubin chips SpaceX leases ($40B deal) through a subsidiary.
Microsoft spent ~$5B less capex in Q3 than planned because rented capacity (one supplier: Iren) was late.
SambaNova + Supermicro: first batch of 10,000 SN50 chips in disaggregated systems (GPUs prefill, SN50 decodes).
Qualcomm custom AI accelerators began shipping in October; AWS connectivity ramps this month; targets $15B+ data-center revenue by FY29.
AMD expects server CPU revenue to double next year, because agents drive CPU demand for orchestration, databases and tool calls.
OpenAI's National Community Compact and Grid Protectors program (up to 1,000 MW of demand response at its Georgia site) only cover sites OpenAI operates itself, not partner-built ones like Oracle's.

How it relates to prior wiki pages

Extends the 11-04 SpaceX analysis (SemiAnalysis: ~13GW planned for 2027, 91-day retrofits): the vote raises the value of speed and behind-the-meter power exactly where SpaceX is strongest, and Texas holds ~27% of planned US capacity.
Confirms the 11-03 memory squeeze (45-week GPU lead times, SK hynix capacity contracted through 2028) now reaching consumer margins.
MPS v3 and SM lending are the hardware-side answer to Together's "polymorphic inference" (today): many workload shapes on one GPU.