Vision-Language-Action Models

Nvidia's stack underpins VLA from silicon to models

Nvidia appears in 50 deals on this theme, far ahead of Samsung (11), Blackstone (9), Google (7) and Tencent (6). Real-time VLA and whole-body controllers are being benchmarked on its hardware, such as a 50Hz diffusion controller on an RTX 5090 with 11.8ms median inference [11]. Nvidia is also moving up-stack with open multimodal models (Nemotron 3.5 Super VL) [8] and partner-built models like Salesforce's [27]. It is planning stakes in four French startups including Genesis AI [34] and has pledged $2B to Europe [12]. GR00T-N1 sits in the tracked company set, so Nvidia is chips, open robot foundation model and investor at once.

Physical Intelligence is repeatedly cited as the reference research-first robotics lab, alongside Generalist [19][26], and Dyna frames its approach as 'product research' that works backward from an end-state product [45]. Dyna-1, a napkin-folding robot, is its commercial proof point [44]. Google DeepMind (RT-2) is the other comparison for robot-brain research [46].

Why it matters · Differentiation is shifting from model architecture to data loops and deployed products, which favors teams with real-world task deployment.

A recent paper shows four different frozen VLMs (PaliGemma, pi0.5, Qwen2.5-VL-7B, DM0.5) all yield usable task memory with the same actor architecture. LIBERO success clusters at 94.9-95.95% while robustness varies from 73% to 79% [33]. This fits the ecosystem of open models (OpenVLA, SmolVLA, Octo) and add-on layers such as PhysVLA and MemoryVLA.

Why it matters · As raw success rates converge, edge value moves to robustness, memory and correction tooling rather than the base model.

Papers argue that humanoid players like Figure, Boston Dynamics and Apptronik over-focus on end-effector dexterity, when using the whole body as a contact surface is the bigger unlock [16]. New frameworks such as Dexterity from Supervised Play [17] and a diffusion whole-body controller [11] point the same way.

Why it matters · Action-head designs will need to extend beyond grippers, opening room for new entrants against humanoid incumbents.

Strategic corporates are leading rounds in AI infrastructure and deep tech [38], and Nvidia's own equity plus Blackstone and Apollo credit are financing adjacent compute buildouts [5][7]. Samsung, Tencent and Google round out the top investors. Weekly capital fell from $11.5B (late July) to $50M in the week of Oct 12, even as stage mix is dominated by unknown-stage ($25.6B) and seed ($5.5B) rounds.

Why it matters · Capital is lumpy and strategic-driven, so VLA startups must court corporate partners or face a thinner venture market.