“You don’t need to beat every rival”: NVIDIA’s Ming-Yu Liu on Cosmos and making markets

Liu rejects the idea that model competition is a Squid Game. If everyone tried to solve exactly the same problem, it would be; but there are countless problems and tastes. Should every company train its own model? Treat it like electricity: early factories built power plants to guarantee supply, but once power is abundant you don’t. And people don’t stay aligned forever—he notes that Dario Amodei left OpenAI and asks whether everyone at Anthropic believes the same thing. Anyone with conviction and resources can build.

His controversial forecast: model capabilities will converge, and differentiation will come from what companies build around them. Leaders have more time to find that edge, and strong companies can widen their ecosystems. Software was once rare too; the survivors found other ways to stand out. He likens deep learning to traditional Chinese medicine—a black box refined by many trials into a body of knowledge that spreads.

Is world modeling a war NVIDIA can’t lose? Not if others also open-source models, he says. Physical AI is a huge market: if every home had two or three robots, each needing compute, demand for compute would soar, which is enormously good for NVIDIA. The goal is to make the market, as CUDA made today’s AI market. He quotes Jensen Huang’s idea of a “zero-billion-dollar business”—no business today, a billion-dollar one if it succeeds; if someone already does it, maybe don’t. A $100-million-a-year business would be a distraction for NVIDIA. That is risky: CUDA meant selling programmable GPUs at a premium, but without it GPUs become a commodity.

Liu believes he is the first NVIDIA researcher promoted to vice president. He is VP of research and of the Cosmos lab; 80–90 people report to him directly, and 200–300 work on Cosmos in total through dotted lines, with him as the mission owner. He still reads papers, reviews code and argues details—he fears losing touch with the problems as a manager.

Researchers who later built Sora once interned on his team. He calls that a source of pride, not regret, while asking himself what he failed to see at the time. Why did OpenAI largely shut Sora down? He frames it as focus: Sam Altman runs a portfolio like an investor, and over the past year Anthropic’s coding agents showed enormous economic value—chat produces a minute of tokens, coding agents work for 24 or 48 hours. OpenAI concentrated on coding. Sunzi’s warning applies: prepare everywhere and you are weak everywhere.

Why does an infrastructure company do research? To build infrastructure applications actually need, you must walk their path; chips and system architectures take years to plan. He also wants ambitious young people everywhere to have tools, which is why NVIDIA open-sources models—mostly free, with a long road left to real deployment, especially in embodied AI. Shipping generation after generation signals commitment so partners can invest in their own hard problems. On whether he would rather build the number-one model or avoid competing with customers: “number one” is fleeting. He wants the capability to build a frontier model, used for the greatest social contribution—not a model for a few people that concentrates power.

His group has a single project: Cosmos. After version 1 he asked Huang whether to continue; Huang said to go all the way to Cosmos 97. The number was offhand, but the commitment was real.

Why world models for physical AI, not LLMs or Sora-style creative video? His background is computer vision and media; NVIDIA has an LLM team (Nemotron). Sora was a world simulator aimed at creators, but many social problems need physical tools: self-driving cars that keep older people mobile, robots for dangerous or labor-intensive work, help with chores. So Cosmos went toward physical AI.

His homepage vision—“a Matrix for robots”—is about generalization. Cosmos helps in three ways: better data (synthetic data covering what the real world didn’t capture), a better starting point (a world model’s representation bootstraps robot policies), and better environments (many realistic worlds to practice in at once, the hardest part). The project was greenlit around March 2024, after Sora made the case obvious to Huang, who is disciplined about pacing. First-mover advantage depends on your position; the most profitable video-generation business today, he notes, is probably ByteDance’s Seedance.

Huang insisted Liu propose a name himself because naming defines positioning: Cosmos, a universe, aimed at solving physical problems rather than content creation, and at developers rather than consumers. Physical AI is a necessity: everyone ages, labor is short everywhere. His definition of world models is practical: models built for prediction, for understanding why things happen (like diagnosing a stalled production line), or for 3D reconstruction. He no longer works on reconstruction. The term is too broad to define precisely, which he doesn’t mind. Cosmos was named a “world foundation model” so others can build their own world models on it. NVIDIA also invests in LeCun’s AMI and Fei-Fei Li’s World Labs; he sees NVIDIA as an enabler, and its real competitors as companies building an entire AI ecosystem.

Cosmos moved fast, inspired by DeepSeek’s three iterations in a year. Working with partners, the team built models for prediction, for making simulator footage more realistic (Transfer), for explaining events (Reason), and then a robot policy model, after realizing that how pixels change is close to how actions change. Because all of these model the same world, a foundation model should absorb all their data—so they moved toward a single model.

He once told Huang he would need about 22 models. Huang praised the effort, then told him Louis Vuitton’s sales rose after it cut the number of products on display: too many choices confuse customers. Liu took it to heart; each model must be maintained, and spreading effort slows iteration. Cosmos 3, started around October last year after Cosmos 2 and 2.5, merges language, video, audio and action in one model. A Chinese founder said the paper had few big new ideas but verified many engineering details; Liu largely agrees and says details are what make models better—Mixture-of-Transformers and related ideas existed, but scaling them to combine audio, video and action is the hard part.

NVIDIA released the model, training framework and some data, because many frontier labs no longer explain key steps. Action is a first-class citizen because physical agents change the world. Modalities help each other: predicting video helps action learning, audio marks contact moments, detailed language helps video. Two towers handle discrete signals (better for understanding) and continuous ones (better for generation), mainly so customers can post-train one side without degrading the other; a single tower is the next goal, but usability comes first. Would it beat Claude at coding? Again: spread yourself thin and you are weak everywhere—strategy is deciding what to give up.

Evaluation needs benchmarks, arena-style comparisons and, above all in physical AI, experts with real problems. Data is split into navigation (easier to share across robots) and manipulation (harder); the trend toward human-like hands led them to add first-person video of human hands to Cosmos 3. NVIDIA’s edges: cheaper compute, years in physical AI (Thor chips, Isaac simulation), and a genuine interest in customers’ success. Compute for Cosmos is in the tens of thousands of GPUs. He is proud Cosmos 3 beat all its predecessors; the regret is shipping before it fully converged, which 3.1 and 3.2 will address.

The Matrix metaphor is about learning faster in simulated worlds, not about machines enslaving people. Society is built and operated by people, and with consensus, catastrophe can be avoided. He doubts chips feel the pain and suffering that drive human progress, while supporting guardrails.

To rival world-model teams: world models are tools; work with NVIDIA and go further and faster. Yes, NVIDIA wants to sell you GPUs—you need compute anyway, and cloud companies bought lots and made more. Could Cosmos be the next CUDA? CUDA turned application logic into GPU computation; future software is AI built on AI, so Nemotron and Cosmos could be part of it—but only as part of a larger stack. CUDA’s path, he thinks, can’t simply be copied.

On China’s models: very impressive. U.S. frontier labs rarely take interns now, keeping know-how inside; Chinese companies take many, spreading knowledge. On Xie’s “hypnotized” claim: he disagrees—today’s leading companies all started from language models, and he can’t imagine working without coding agents. He expects 2026 to be a year of upheaval: SpaceX’s IPO, OpenAI and Anthropic heading toward listings, soaring San Francisco housing. His prediction: a ChatGPT moment for physical AI is coming. On robots: in the U.S., only Tesla and Figure build full robot bodies; China’s manufacturing allows tighter hardware–software integration, which is worth watching.

In his ten years, NVIDIA went from unknown to his parents to synonymous with AI. There were hard times—when Google’s TPU arrived and the stock fell, some colleagues left; NVIDIA kept beating TPU on MLPerf until it was the only one still entering. Huang’s hair has turned white, but he still reads employees’ top-five priority emails for signals. The company has more than doubled; GPUs moved from desktops to data centers.

Huang says, every time, that the company has only 30 days of cash left—not literally, but to force the right decisions. During the pandemic downturn, NVIDIA cut travel and tools (even considering dropping Slack or a cloud provider) rather than people. Liu hasn’t seen a layoff in his decade; there is no forced ranking, only performance reviews. The downside is slower transitions when technology shifts; the upside is trust and cooperation. NVIDIA runs marathons—CUDA took ten years—and doesn’t pit teams against each other; it trusts people with a mission and sets a very high bar, in Huang’s words torturing people to greatness.

NVIDIA’s distinctive practices, as he lists them: no layoffs, top-five emails that surface signals and duplicate work, radical transparency (Apple is the counterexample), and “the mission is the boss,” with teams aligning like an organism. Huang shaped him: he sent papers for Liu to summarize, learned computer graphics and then deep learning from scratch, decided from first principles, and prioritized ruthlessly. Liu now switches topics every half hour. His computer-science lesson: be lazy—don’t compute what you don’t need. Once, when Liu blamed an unfair experimental setup for losing to a rival, Huang asked, “Are you a quiet baby?” There are no excuses; get the conditions you need.

Is his low ego real or a mask? Confidence, he says—the belief that what they build helps others. As a young researcher, trained to prove state-of-the-art results, he was highly competitive; now he knows every state of the art is soon surpassed, and he cares about developing people and hearing that Cosmos helped a partner. You don’t need to beat your rivals; he prefers to call them partners.

He wakes around 4–5 a.m. and sleeps around 10–11. He has practiced Chinese martial arts since 18 under Xu Ji, in a lineage traced to Li Shuwen, known for Bajiquan and Piguazhang, and recently took up ice plunges. Martial arts, inspired by Jin Yong’s novels, taught him control: a punch’s power runs from the feet to the hand in an instant, which demands constant self-correction—“not to show destructive power.” Favorites: beef noodle soup and the Bay Area’s weather. His advice to young researchers: stay calm and confident, don’t fear that AI will end next year, and look at physical AI, materials and drug discovery. How should history remember him? Don’t talk about him—he hopes Cosmos helps a few important physical-AI builders reach their ChatGPT moment.