If your team builds robot-training scenes or 3D content on World Labs' Marble or its World API, your supplier just agreed to join a chipmaker — and nobody has said what happens to the product. AMD announced a definitive agreement to acquire World Labs for approximately $8.2 billion in stock, closing by the end of 2026 subject to regulatory approval. Neither AMD's release nor World Labs' own announcement says a word about Marble, the World API, pricing or existing customers. The terms you accepted already let World Labs change or end the service at any time. You have roughly three months to turn silence into something written down.
This is not a story about a chip company buying a startup. It is a story about a hardware vendor buying the workload it wants its silicon to be designed around — and about the customers who happen to be standing inside that workload.
What AMD Actually Bought
AMD bought a research team and a model roadmap, and said so plainly. Lisa Su's quote in the AMD release frames the rationale as needing "a deep understanding of how models are evolving" to build compute platforms, and praises the team's "research leadership and model expertise." Fei-Fei Li becomes AMD's executive vice president and chief scientist, reporting to Su. World Labs' post adds that Justin Johnson and Ben Mildenhall will keep leading the team inside AMD as "a world leading frontier research organization."
A world model is an AI model that generates a persistent, navigable 3D environment — layout, depth, lighting and geometry — from text, images or video, rather than a flat picture or a clip. That is the capability World Labs sells today, through two front doors:
- Marble, the app, with plans from free to $35 a month for Pro, plus Max and custom Enterprise tiers.
- The World API, launched January 21, 2026, which named Preview, Fenestra, Escape.ai, Interior AI and xFigura as early builders and pitched robotics training environments that work with NVIDIA Isaac Sim, MuJoCo and RoboSuite.
The price is steep for a company whose $1 billion round closed on February 18, 2026 — seven months ago — with AMD, Autodesk, NVIDIA, Fidelity and others on the cap table. TechCrunch reports the two companies already had a training-and-inference optimisation partnership from 2025, and that AMD sees the work as part of competing with NVIDIA's ecosystem, which already ships open world models under the Cosmos name.
It is also AMD's second model-side purchase in two months, after it bought Taalas and its model-specific silicon in August. The pattern is a hardware company buying its way up the stack.
What the Announcements Leave Out
Both announcements are silent on the product, and in an acquisition, silence is the default setting of a roadmap in review. The AMD release describes World Labs as a developer of spatial-intelligence models for "robotic learning and simulation" and commits only that the team will "continue advancing AI model research." Research is not a service-level agreement.
Steel-man the reassuring reading first. World Labs' post says the combined company wants "an end-to-end open AI ecosystem spanning hardware, software, platforms, and widely accessible open models." AMD's strategic interest is in more developers building on world models, not fewer, and a paid API is the cheapest way to keep a developer base warm. A plausible outcome is that Marble keeps running, gets cheaper, and gains open weights.
Here is why that reading should not be your plan. A hardware vendor's incentive is to make the workload run best on its hardware. It has no particular reason to keep running a consumer-priced creation app, or to prioritise a metered API over a research agenda that now reports to the CEO. We have watched this exact silence before: when Adobe closed Topaz Labs, its standalone pledge named apps and models but never the API, and when NVIDIA bought Hugging Face, the open question was whether the accelerator-neutral tooling stayed neutral.
What Your Contract Already Allows
Your existing contract gives you almost no protection, because the World Labs terms were written to make exactly this transaction frictionless. From the World Labs Terms of Service, last updated January 21, 2026:
- Assignment without consent. Section 12.7 says "World Labs may assign or transfer this TOS without your consent," while you cannot assign yours without World Labs' written consent. Your agreement moves to AMD automatically.
- Change or stop at any time. Section 12.2 reserves the right "to modify our Services or to suspend or terminate providing all or part of our Services at any time."
- Termination for convenience, one way. Under Section 11.2(a), World Labs may terminate "for convenience at any time." Paid accounts keep access to the end of the billing period and get a pro rata refund for the unused part of it. That is the full remedy.
- Liability capped at twelve months of fees, or $100 if that is greater, under Section 9.1.
- Training rights on your content. Section 3.6 grants World Labs a licence to use User Content for "training of World Labs AI Models." For free accounts it is irrevocable; paid users may opt out, but only prospectively.
Output ownership is where teams should read slowly. The terms say paid-account users own their outputs, and paid users who put the API into customer-facing apps may license output to their end users "subject to specific terms including restrictions on downstream sublicensing, attribution requirements, and usage volume limitations." Yet the Marble billing documentation lists commercial rights only from the $35 Pro plan up, not on the $20 Standard plan. If your production assets came out of a Standard seat or a free account, find out which document governs before the owner changes, not after.
None of this is unusual for a startup. It is exactly what a change of control exposes. We made the same point when Databricks bought Row Zero under 30-day cancellation terms: the clause that did not matter last quarter is the only one that matters now.
What It Costs to Walk Away — and What It Costs to Stay
The good news is that getting your assets out is cheap today, so do it while pricing is known. The World API pricing page sells credits at $1.00 per 1,250. A standard marble-1.1 world costs 1,500 credits, or $1.20; a marble-1.0-draft world costs 150 credits, or $0.12. A mesh export costs 3,500 credits per world — $2.80. API credits do not expire, but on the Marble app side, top-up credits expire after a year and cancellations get no pro-rated refund. Under Section 11.2(b) of the terms, if you terminate, "no refunds will be provided for prepaid fees, unused Service Credits, or Top-Up Credits."
That makes the arithmetic simple. A library of 2,000 simulation scenes costs about $5,600 in mesh-export credits to convert to GLB. That is a rounding error against the engineering time it would take to regenerate them from scratch on a different model.
Export formats are open, which is the other piece of good news. According to the export specifications, Marble exports Gaussian splats as SPZ or PLY at roughly 500,000 or 2 million splats, a 100,000-200,000-triangle collider mesh in GLB for physics, a high-quality GLB mesh of 600,000 textured or 1 million vertex-coloured triangles, and the source 360° panorama as PNG. A Gaussian splat is a 3D scene stored as millions of small, coloured, semi-transparent ellipsoids rather than as triangles — photorealistic to render, but still unfamiliar to many physics engines. That is why the collider mesh matters for robotics. High-quality mesh generation is rate-limited to four requests per hour per user, so a large backlog takes days, not minutes. Start now.
The renderer is the soft dependency people forget. World Labs' Spark library renders splats in THREE.js and reads PLY, SPZ, SPLAT, KSPLAT and SOGS files. If your web viewer is built on it, add it to the list of things you need a maintenance answer for.
Is the Hardware-Neutrality Risk Real?
Yes, but it is a medium-term risk, not a closing-day one. Today's World API output is plain files that run anywhere, and the launch post specifically positioned it for NVIDIA Isaac Sim. The risk is in the next model: if future Marble versions ship as open weights tuned for AMD Instinct and ROCm, or the hosted API's price favours workloads that run on AMD capacity, your simulation stack inherits a hardware preference you never chose.
The hedge already exists. NVIDIA made its Cosmos world foundation models openly available under its open model licence permitting commercial use in January 2025, with 1X, Agility Robotics, XPENG, Uber and Waabi named as early users. Cosmos is not a drop-in replacement for Marble's exportable scenes, and you should test that claim on your own workflow, but it means a robotics team is not choosing between World Labs and nothing. If silicon choice is already a live debate on your team, our comparison of NVIDIA alternatives for inference is the other half of this decision.
What to Do Before the Deal Closes
This Week:
- Inventory every Marble and World API account — including the free and Standard seats designers opened on a personal card. List the plan tier for each, because output rights and training rights differ by tier.
- Turn on the training opt-out on every paid account (Section 3.6). It applies only from the day you set it.
- Start the export job. Pull every production world as SPZ plus collider GLB, and high-quality GLB where you need it, into storage you control. Budget $2.80 per mesh and the four-per-hour rate limit.
This Month:
- Ask World Labs, in writing, four questions: Will the World API and Marble continue after close, and for how long? Will current pricing hold through a stated date? Will you get advance notice — at least 90 days — before discontinuation or a material price change? Will outputs generated on the current plans keep their current licence terms?
- If you spend meaningfully, ask for an Order Form that overrides Sections 11.2(a) and 12.2 with a notice period and a migration window. The terms allow for Order Forms; a custom Enterprise plan exists. Now is when your leverage is highest — nobody wants a customer complaint on file during regulatory review.
- Run a two-week Cosmos pilot on one synthetic-data workflow, so a replacement exists as a tested option, not a slide.
Before Close (end of 2026):
- Decouple your viewer and pipeline from World Labs-specific formats. Treat PLY and GLB as the canonical storage format; SPZ as a cache.
- Add a hardware-neutrality question to your robotics-stack review: if the next world model you depend on is optimised for one accelerator, what does that cost you in capacity you already own?
The Bottom Line
The last decade had a version of this story. When design and simulation tools consolidated into a few big suites, customers who had kept their geometry in open formats changed vendors on their own schedule; customers whose work lived in a proprietary scene graph changed vendors on the acquirer's. World models are at the start of the same arc.
AMD may well keep Marble alive and make it better. It has not said so. Your contract does not require it to. And the export that protects you costs less than three dollars a scene.
Export first. Ask questions second. Silence is not a roadmap.
Continue Reading
- AMD Bought Taalas. Now Name the Model You'd Freeze.
- Nvidia Buys the Bridge to Trainium. Go Grep Your Imports.
- Adobe's Topaz Pledge Names Apps and Models but Never the API
- Databricks Buys Row Zero, Whose Terms Allow 30-Day Cancellation
- NVIDIA Alternatives for Inference: Only Trainium Pays Off
- ABB and NVIDIA Close the Sim-to-Real Gap: 80% Faster Robot Deployment