World models stopped disclosing — and NVIDIA's open Cosmos is the only window investors still get

On September 18, 2026, TechCrunch reporter Russell Brandom asked a simple question of the frontier world-model companies: what are you building and when will anyone see it? AMI Labs co-founder Michael Rabbat replied, \"We'll talk about it when we're ready to talk about it,\" then clarified by email that AMI is \"still in a research and building phase\" and \"not talking publicly about any product plans or timeline.\" The same pattern held at World Labs, where Marble is the most developed product on offer but commercialization plans stay undisclosed. Physicl, a video-data supplier to the same labs, told TechCrunch the secrecy cuts both ways — the company can build more useful data only if it knows what's being trained on, and CEO Alex de Vigan said he is \"still in the dark.\"

Secrecy hasn't slowed the checkbook. Per Forbes reporting on June 30, 2026, world-model startups absorbed more than $3 billion across the first half of the year, building at least four private companies valued in the billions and pulling a hyperscaler into the bidding. The dollar map of who's hoarding what reads like a list of firms betting that opacity compounds, not expires.

Read one level deeper and the secrecy thesis snaps into focus. A state-of-the-art world model is data-limited far more than chip-limited. NVIDIA's Cosmos 3 ate 20 million hours of video to ship the open version in May 2026; Meta's V-JEPA 2 consumed more than 1 million hours of video plus robot interaction data; GR00T N1.7 added 20,000 hours of egocentric human video. Telling a competitor what you trained on tells them what data is still missing — and which suppliers (like Physicl) to grab next. Closed roadmaps buy closed data advantages.

That reasoning plays out in AMI Labs' deliberate caginess: with $1.03B in seed capital and a $3.5B valuation, LeCun's venture doesn't need to explain its thesis to win funding, but it does need to keep its data mix unobservable long enough for its modular JEPA-style world models to differentiate. World Labs' move is more revealing — Atlas (September 1, 2026) is a multimodal autoregressive diffusion transformer pretrained from scratch on text, images, video, and 3D, capable of up to a minute of 1440p camera-controlled generation. The capability was public on day one, which is unusual for a firm whose peers are circling the wagons.

On May 31, 2026, NVIDIA launched Cosmos 3 — the world's first \"fully open omnimodel\" for physical AI — and almost immediately announced the Cosmos Coalition, a closed-cohort alliance of robotics and AI labs (Agile Robots, Black Forest Labs, Generalist, LTX, Runway, Skild AI) committed to sharing models, research, and evaluation techniques inside the NVIDIA DGX Cloud perimeter. The lever is plain: open weights attract developers, while the coalition traps the highest-value contributors behind a single-stack subscription.

If the secrecy were denting real demand, the foundry and accelerator order books would show it. They don't. NVIDIA's TTM revenue through Q2 FY27 reached $302.97 billion with a roughly 75% gross margin; FY2028 guidance calls for ~70% growth, which implies a much larger data-center quarter ahead. The company has now committed to roughly $160 billion in long-term memory purchase obligations, with TSMC handling advanced-node packaging. The world-model labs may be silent, but the silicon they depend on is sold out for years.

The bulk of the demand these labs are training against is the embedded-AI stack — humanoids, robotaxis, warehouse robotics. Morgan Stanley sizes the humanoid market at up to $5 trillion by 2050; Goldman Sachs puts the global robotaxi market alone at roughly $415 billion by 2035; the combined TAM is a multi-hundred-billion-dollar runway the closed-roadmap labs are targeting without telling customers when. Listed exposure is thin but real: Tesla is the largest humanoid deployment in commercial planning (Optimus Gen 3), and Agility Robotics filed its S-4 in September 2026 showing $1.8 million in 2025 revenue and a $140 million operating loss ahead of a $2.5 billion SPAC merger. The financials are early; the point is who's first to list.

Two near-term catalysts can pry the roadmaps open. First, an Anthropic-, Amazon-, or Meta-scale acquirer closes a deal with a private frontier lab (Anthropic reportedly walked away from a roughly $6 billion Decart offer in September 2026) — and the acquirer must disclose integration plans to its shareholders. Second, a Cosmos Coalition partner graduates from \"development partner\" to \"published enterprise customer\" status — Skild AI is the closest to that threshold, having crossed $100M in ARR and shown a 7x step-success lead over comparable systems in single-video task learning on Sep 10, 2026. Once one of those breaks, the rest follow fast; the closed-roadmap pattern only holds if everyone stays dark.

The secrecy at frontier world-model labs isn't a sign of weakness; it's a sign of optionality that nobody has yet had to monetize. The labs raised $3B+ in H1 2026, ship limited demos on their own schedule, and refuse to reveal commercialization plans — which is the best evidence that the value sits in proprietary data and trained weights, not in any single product SKU. For listed investors, that means two clear things. The substrate (chips, memory, foundry) is already oversold by multiples of the disclosed training effort, and the substrate will keep winning even if the labs stay quiet. The application layer (humanoids, robotaxis, simulation) is the deferred monetization — and the only listed way to ride it before a 2027 IPO window is concentrated exposure to chip leaders and the most credible humanoid vehicle, with selective bets on coalition partners as their ARR prints.