> ## Documentation Index > Fetch the complete documentation index at: https://kensou.mintlify.site/llms.txt > Use this file to discover all available pages before exploring further. # Nvidia cosmos # NVIDIA Cosmos — 世界基础模型作为 GPU 营收战略 (NVIDIA Cosmos — World Foundation Models as a GPU-Revenue Play) > **发布时间**: GTC 2025 / CES 2025（持续发布中） > **公司 / 产品名**: NVIDIA Cosmos · Omniverse · Isaac Lab · GR00T > **覆盖范围 / 公司类型**: 跨域物理 AI 基础模型 + 仿真栈；芯片厂商纵向整合 > **核心定位**: 一句话回答 — Cosmos 表面是"物理世界的 GPT-3"，本质是把 Omniverse → Cosmos → Isaac Lab 串成三级 GPU 消耗漏斗的中段。 **Status:** v1.1 — opinionated draft. Backfilled to AGENTS.md 14-item dissection template 2026-05-21. Roadmap claims marked `UNVERIFIED`. Internal numbers marked `UNVERIFIED — no public source`. **TL;DR:** Cosmos is positioned as "GPT-3 for the physical world" but its function inside NVIDIA is to make Omniverse + Isaac Lab + Cosmos a three-stage funnel that consumes GPUs at every stage. Read the announcement as a *revenue architecture* first and a robotics breakthrough second; the robotics part is real, but it's the smaller story. ### X-Ray（非专家友好开场） （a）机器人公司被"训练数据稀缺"卡住 — 真实数据贵、仿真不够真。（b）Cosmos 把视频世界模型作为"合成数据基座"开源，让客户先用 Omniverse 造场景、再用 Cosmos 跑 rollout、最后 Isaac Lab 训策略，每步跑 NVIDIA GPU。（c）对 spatial / 具身 AI 工程师：Cosmos 是真模型，但商业理由是把"物理 AI 工作流"变成 NVIDIA 销售漏斗 — 技术阅读必须穿过这层。 ### 📍 Cosmos / NVIDIA 物理 AI 产品演进时间线 ``` 2018 Isaac Sim ─► 2020 Omniverse ─► 2023 Isaac Lab ─► 2024 GR00T ─► ★ 2025 Cosmos-1.0 (GTC) ─► Cosmos-1.x? ─► Cosmos-Reasoning? 2026+ │ │ │ └── 视频 WFM + tokenizer + guardrails 开源 └── 人形基础模型（与 Cosmos 双向整合中） ``` NVIDIA 用了 7 年把"造场景→训策略"的全栈搭起来，Cosmos 是 2025 补上的中段世界模型。 ### ⚡ Eureka Moment **Cosmos 的产品定义不是"更好的世界模型"，而是"GPU 漏斗的中段"。** Omniverse、Cosmos、Isaac Lab 各自是真技术，但三者串联把"任何物理 AI 应用"的 GPU 小时数提升 3–10×（合成数据 + 基础模型微调 + 策略训练 + 边缘推理）。技术评价应与商业架构分开读。 ### 📌 Napkin Formula ``` NVIDIA 物理 AI 营收 ≈ ∑(Omniverse GPU-hours) + ∑(Cosmos fine-tune GPU-hours) + ∑(Isaac Lab GPU-hours) + ∑(Jetson 边缘出货) ─ 每一段都是独立 SKU + 独立收入线 ─ ``` *** ## 1 · Where Cosmos sits in NVIDIA's robotics stack NVIDIA's robotics flywheel has three layers, each consuming GPU cycles: ``` ┌─────────────────────────────┐ Synthetic │ Omniverse (3D scenes) │ training-data GPU spend data │ Isaac Sim (physics) │ ◄───────────── generation └─────────────────────────────┘ │ ▼ ┌─────────────────────────────┐ World │ Cosmos (video WFM) │ foundation-training GPU spend models │ pretrained + finetuned │ ◄───────────── └─────────────────────────────┘ │ ▼ ┌─────────────────────────────┐ Policy │ Isaac Lab (RL / IL) │ policy-training GPU spend training │ GROOT (humanoid) │ ◄───────────── └─────────────────────────────┘ │ ▼ Jetson Thor / AGX edge-inference GPU spend ``` Each box is its own SKU and each arrow is its own DGX-hours line item. Cosmos slots into the middle — it's the foundation model the customer fine-tunes for their robot, before policy training. That positioning is what makes it commercially load-bearing for NVIDIA even if the model itself isn't yet better than open alternatives at every task. *** ## 2 · What Cosmos actually is Cosmos WFMs are video-generation models trained on a large filtered corpus of physical-world video, released in two families: **Cosmos-1.0 Diffusion** (high-fidelity rollouts) and **Cosmos-1.0 Autoregressive** (faster + controllable rollouts). Sizes a few B to \~13B params at first release `UNVERIFIED — exact split`. Released with the Cosmos tokenizer and a guardrails wrapper. **Use cases:** synthetic data for robot policy training; text/image/action-conditioned rollouts; pre-training base for embodied AI. **Competes with:** Wayve GAIA, Sora, Genie/DeepMind, Runway. Differentiation is physical-world filtering + NVIDIA stack integration, not raw video quality. *** ## 3 · The GTC 2025 reveal in context Framed as "Project Cosmos / open WFM release" — weights under a permissive-ish license for research + commercial fine-tuning, paired with Omniverse + Isaac Lab integrations. Three messages, three subtexts: | Surface | Subtext | | - | - | | "WFMs are the next foundation model frontier" | Robotics-data scarcity reframed as a model problem NVIDIA sells hardware against | | "Open weights, permissive license" | Adoption flywheel — every fine-tune is downstream GPU spend | | "Integrated with Omniverse + Isaac" | Lock-in — value is the stack, not the weights | The robotics community got real artifacts. The shareholder got "we have the AI factory for physical AI". *** ## 4 · Why this is GPU-revenue-driven Three tells: (1) **Funnel structure** — Cosmos sits between Omniverse and Isaac Lab and *increases* GPU-hour requirements 3–10× for any customer building a physical-AI product. (2) **License shape** — permissive for research, with clauses nudging production users toward NVIDIA hardware / cloud `UNVERIFIED — terms shift across releases`. (3) **Roadmap pacing** — releases align with the Blackwell → Rubin hardware cycle; each WFM generation demands more memory that matches the next chip's selling point `UNVERIFIED — correlational`. None of this means the models are bad. It means they exist publicly to expand GPU TAM, and technical claims should be read through that lens. *** ## 5 · What's actually shipping vs roadmap **Shipping (as of public announcements through 2026-Q1).** `UNVERIFIED — pace of release outpaces this doc` * Cosmos-1.0 Diffusion + Autoregressive weights (multiple sizes) * Cosmos tokenizer (independently useful) * Cosmos guardrails / safety wrapper * Isaac Sim / Omniverse integrations for data generation * Documentation + research papers **Announced / roadmap (read skeptically).** `UNVERIFIED` * Larger Cosmos models (continued scaling) * Action-conditioned variants tuned for specific robot platforms * Cosmos Reasoning — variants that can answer questions about the world they model * Tighter GR00T (humanoid foundation model) integration * Real-time inference variants for on-Jetson deployment The shipping list is real and substantial. The roadmap list is the part to track for *delivery vs slip*. NVIDIA's track record on shipping ML announcements is mixed — the hardware always ships, the software-stack timelines slip. *** ## 5.5 · Worked example — 一家机器人初创如何被漏斗"吸进去" 设仓储抓取初创 X，原本 RGB + 自采 5 万 demo 训策略。接入 Cosmos： 1. **Omniverse 场景** 500 变体 → 几千 GPU-hours。 2. **Cosmos 合成扩量** 50 万条 rollout → 数万 GPU-hours。 3. **Isaac Lab 策略训练** → 数万 GPU-hours。 4. **Jetson Thor 边缘** → 硬件出货。 5. **账单** 从 \~5K → \~50K-100K GPU-hours，策略略好（OOD），但成本 10–20×。 漏斗精明在**每段都给真价值**（场景多样、合成扩量、policy infra）—客户不觉被坑，NVIDIA 的营收账面清晰。 *** ## 6 · How to read Cosmos benchmarks When NVIDIA-published numbers compare Cosmos to other video models: **physical-realism metrics** favor Cosmos's filtered training data and are often the only ones where Cosmos clearly wins; **general video quality (FVD, etc.)** — Sora / closed competitors win `UNVERIFIED`; **downstream policy success rate** — the metric that would matter is rarely published with rigor. Independent reproduction is the safer signal — academic fine-tunes on niche embodied datasets are the cleaner readout. *** ## 6.5 · Hidden Assumptions 战略叙事下的隐含假设，违反任一条都会动摇结论： * **跨域合成数据可替代真实数据** — 学界证据混合，下游策略成功率公开数字稀少 `UNVERIFIED`。 * **GPU 主导未来 10 年物理 AI 算力** — ASIC / 边缘芯片崛起会绕过漏斗。 * **开源 WFM 不反噬** — AMD ROCm + Hugging Face 追赶会削弱护城河。 * **机器人客户接受 vendor lock-in** — 2027 后若有等效开源替代，迁移成本下降。 * **Blackwell → Rubin 节奏不断** — 供应链危机延后芯片即延后 Cosmos。 * **WFM 对具身策略真有用** — Sora 路线 vs 物理仿真路线尚未分出胜负。 * **不出 safety 大事** — 一次合成数据训练策略事故会监管化整个叙事。 *** ## 6.6 · Interview Tip 被问"NVIDIA Cosmos 真正想做什么" — 给两层答案：**表层**是开源视频世界模型，做物理 AI 基础模型；**深层**是把 Omniverse + Cosmos + Isaac Lab 串成 GPU 营收漏斗，让任何机器人工程团队的算力账单上升 3–10×。最后补一句：模型本身是真工程，路线图（Cosmos-Reasoning、GR00T 整合）的兑付节奏才是观察重点，NVIDIA 软件栈历史上交付时间表常 slip。 *** ## 7 · Strategic read NVIDIA has assembled the most credible end-to-end physical-AI stack. Whether it wins the robot foundation model wars is separate — Wayve owns AD-specific WFMs with a tighter feedback loop, Physical Intelligence owns the manipulation policy flywheel, Tesla / Waymo own their own data + sim. Cosmos's bet is that *cross-domain* synthetic data is the moat; whether that pays out depends on downstream users finding WFM rollouts useful enough to fine-tune on instead of collecting more real data. NVIDIA's safer bet: they sell GPUs either way. The team is real engineering; the *product* is a sales motion. *** ## For the reader * **Robotics engineer** — try Cosmos as a synthetic-data engine where real data is scarce; don't expect it to replace domain-specific fine-tuning data. * **AD / driving engineer** — Wayve's GAIA is closer to your use case. * **World-model researcher** — open weights are a real gift; the integration story is sales. * **Investor** — read Cosmos announcements as GPU-demand signals. *** ## References * NVIDIA Cosmos announcement (GTC 2025 / CES 2025). [https://www.nvidia.com/en-us/ai/cosmos/](https://www.nvidia.com/en-us/ai/cosmos/) * Cosmos research paper / technical report. [https://arxiv.org/abs/2501.03575](https://arxiv.org/abs/2501.03575) * Cosmos GitHub. [https://github.com/NVIDIA/Cosmos](https://github.com/NVIDIA/Cosmos) * NVIDIA Isaac Lab. [https://github.com/isaac-sim/IsaacLab](https://github.com/isaac-sim/IsaacLab) * NVIDIA Omniverse. [https://www.nvidia.com/en-us/omniverse/](https://www.nvidia.com/en-us/omniverse/) * Compare: `companies/wayve_world_model.md` (AD-specific WFM strategy) * Cross-ref: `foundations/world-models/` for technical dissection; `embodiments/driving/` for AD deployment context. ## 🤖 Moltbot Updates [//]: # "" *** *Last opinion update: 2026-05-21. v1.1 — backfilled to AGENTS.md 14-item template.* This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.