World Labs
Freemium SaaS subscriptions for Marble (Free / $20 Standard / $35 Pro / $95 Max per month; commercial rights at Pro and above), with an emerging enterprise/partnership layer (Autodesk research-and-model-level collaboration, starting with entertainment use cases) and likely future API/licensing for robotics-simulation and 3D-content pipelines
Both marks are press-reported, not company-confirmed - World Labs declined to confirm the Feb 2026 valuation. Trail shows a ~5x step-up in ~17 months on the world-model thesis rather than disclosed revenue.
Earnings, margins, COGS & capex
No audited financials exist publicly. Total capital raised is $1.23B across two stages: $230M announced Sep 13, 2024 at a reported ~$1B valuation (co-led by a16z, NEA, and Radical Ventures, with Nvidia's NVentures, AMD Ventures, Adobe Ventures, Databricks Ventures, Marc Benioff, and Ashton Kutcher participating), and $1B announced Feb 18, 2026 - anchored by a $200M strategic commitment from Autodesk with AMD, Emerson Collective, Fidelity, Nvidia, and Sea participating - at a reported ~$5B-$5.4B valuation the company declined to confirm. Revenue generation started only on Nov 12, 2025 with Marble's paid tiers; the company is in the invest-ahead-of-revenue phase typical of frontier AI labs, with valuation resting on the world-model thesis rather than current financial performance.
Income statement — where each revenue dollar goes
% of revenueOf every $1 of revenue, ~97¢ is cost of goods and ~2¢ operating expense, leaving ~1¢ of operating profit.
Revenue trend
Margins
3D world generation is inference-intensive; unit economics depend on GPU cost curves and generation caching
presumed deeply negative; research headcount + training compute dominate spend
COGS structure
Undisclosed. Structurally: GPU inference for world generation (each Marble world is a heavy 3D/Gaussian-splat-style generation job), cloud hosting/storage for persistent downloadable environments, and model-training compute amortization. Google Cloud is the primary compute provider (Oct 2024 deal); investor relationships with Nvidia and AMD suggest preferential access to accelerators.
Capex
Undisclosed. As a venture-funded lab, most compute is opex (Google Cloud GPU rental) rather than owned data centers; the $1B Feb 2026 raise was for scaling model training and the Marble platform.
Latest earnings
n/a
none issued
- Total capital raised
- $1.23B ($230M Sep 2024 + $1B Feb 2026)
- Last round
- $1B, announced Feb 18, 2026
- Reported valuation
- ~$5.4B post-money (Forbes); Bloomberg reported ~$5B talks Jan 2026; company declined to confirm
- Valuation step-up
- ~5x in ~17 months (~$1B Sep 2024 -> ~$5-5.4B Feb 2026, as reported)
- Marble pricing
- Free (4 generations) / $20 Standard / $35 Pro / $95 Max per month; commercial rights from Pro ($35)
- Anchor strategic investor
- Autodesk ($200M + research-and-model-level collaboration)
Growth drivers
- Marble commercial rollout — conversion of free users to $20-$95/mo paid tiers with commercial rights at Pro and above
- Autodesk partnership — research-and-model-level collaboration starting with entertainment use cases, with a path into Autodesk's large installed professional design/engineering base
- Robotics simulation demand — persistent, geometrically consistent worlds as synthetic training environments for embodied AI
- Gaming/VFX pipeline adoption — downloadable 3D assets slot into existing engines (Unreal, Unity) unlike video-only world models
- Category tailwind — VCs put ~$3B into world-model startups in the year to mid-2026 (Forbes, Jun 2026); spatial intelligence is the consensus 'next frontier' narrative after LLMs
Bull & bear
World Labs is the credible independent lab in the next platform shift: if AI's frontier moves from language to spatial intelligence, Marble's persistent, pipeline-compatible 3D output plus Autodesk distribution makes it the picks-and-shovels winner across gaming, design, and robotics simulation - and the ~$5B entry looks cheap against frontier-lab comps.
- Structural differentiation: downloadable, editable, geometrically persistent worlds integrate into Unreal/Unity/Autodesk workflows today - most rivals' streamed video worlds do not; this is the difference between a demo and a product
- Autodesk's $200M plus research-and-model-level collaboration is a distribution path into millions of architecture, engineering, and entertainment seats
- Robotics simulation is a real, budgeted enterprise need (synthetic training environments); Marble's consistent geometry is precisely what robot training requires, and investor Nvidia validates the market
- Team quality compounds: Fei-Fei Li's lab lineage keeps attracting researchers a startup otherwise couldn't hire, sustaining model-quality parity with bigger rivals
- Comps support upside: Decart marked at ~$4B (May 2026) with a narrower product focus, and world-model startups absorbed ~$3B of VC in the year to mid-2026; a category-leading position could re-rate World Labs toward the $10B+ frontier-lab bracket in the next round
A reported ~$5.4B valuation on undisclosed (likely negligible) revenue, in a category where Google and Nvidia are the pace-setters and Chinese labs are open-sourcing the substrate - the risk is that world models commoditize before World Labs finds an enterprise business model, leaving late investors underwater.
- Revenue reality: prosumer subscriptions at $20-$95/mo cannot carry a $5B-class valuation; no disclosed enterprise contracts, no disclosed ARR - the entire case is narrative
- Compute asymmetry: DeepMind's Genie line and Nvidia's Cosmos iterate on hyperscaler compute budgets; a $1.23B war chest buys limited frontier training generations, then World Labs must raise again into whatever market exists
- Commoditization from below: Tencent's HY-World 2.0 (Apache 2.0, Apr 2026) already open-sources engine-importable 3D world generation - Marble's core output format - and Alibaba is shipping open world models too, compressing pricing before the market matures
- Absorption risk: if world simulation becomes a mode of general multimodal models (Gemini, Sora line), standalone world-model companies become features, not platforms
- Exit math: at a ~$5.4B post, an acquirer short of Big Tech is hard to find, and Big Tech builds in-house; an IPO requires a revenue story that does not yet exist publicly
What it is worth
Last priced round + frontier-AI-lab comparables (no revenue multiple possible - revenue undisclosed)
$1.5-3B effective value
if world models commoditize (Tencent-style open releases, Genie-class consumer bundling) before enterprise revenue materializes, forcing a flat/down round or strategic sale
$5-8B
modest step-up on continued product momentum and category tailwind, with monetization still early and undisclosed
$12-20B next-round mark
if Marble converts the Autodesk channel into disclosed enterprise revenue and robotics-simulation contracts land, re-rating it into the category-platform bracket
Last mark: ~$5.4B post-money as reported by Forbes (Bloomberg reported ~$5B talks in Jan 2026; the company declined to confirm) from the $1B round announced Feb 18, 2026 - up ~5x from the reported ~$1B mark of Sep 2024. With no disclosed revenue, the mark is underwritten on team, category leadership, and strategic partnerships. Comps: Decart ~$4B (May 2026), Odyssey $1.45B (Jun 2026), versus general frontier labs at far higher marks - World Labs sits at the top of the world-model cohort but well below general-AI labs.
SWOT
Strengths
- Founder credibility — Fei-Fei Li (ImageNet creator, Stanford HAI co-director) plus co-founders Justin Johnson, Christoph Lassner, and Ben Mildenhall (NeRF co-author) - among the strongest research pedigrees in AI, aiding recruiting and fundraising
- Differentiated product architecture — Marble outputs persistent, editable, downloadable 3D environments (usable in existing engines/pipelines) versus most competitors' ephemeral streamed video worlds
- Deep-pocketed strategic cap table — Autodesk, Nvidia, AMD, Fidelity, a16z, NEA, Radical Ventures - aligning distribution (Autodesk) and compute supply (Nvidia/AMD) with equity incentives
- $1.23B raised gives multi-year runway to out-spend most world-model startups on training compute
Weaknesses
- Effectively pre-revenue at a reported ~$5.4B valuation - essentially all value is thesis, not financials
- Competing directly with Google DeepMind (Genie 3 / Project Genie) and Nvidia (Cosmos), which have orders-of-magnitude more compute and distribution
- Prosumer subscription pricing ($20-$95/mo) is a slow path to justifying a $5B-class valuation; enterprise/robotics monetization is still unproven
- World-model quality is compute-bound; a startup's cost of frontier training scales worse than hyperscalers'
Opportunities
- Robotics/embodied-AI simulation — physically consistent synthetic worlds as training data - a market Nvidia (Isaac/Omniverse) validates and Autodesk's industrial base can channel
- Autodesk integration could make Marble the generative layer inside architecture, construction, and manufacturing design tools (partnership starts with entertainment)
- Gaming and VFX cost curves — studios seeking far cheaper environment creation; Marble's engine-compatible output is the natural fit
- Category formation — if 'spatial intelligence' becomes the post-LLM platform shift, the category-defining independent lab commands platform-scale value (the Anthropic/OpenAI playbook)
Threats
- Genie 3-class models from DeepMind improving fast — Project Genie shipped to consumers via Google AI Ultra in Jan 2026, bundled into Google's ecosystem
- Well-funded direct rivals — Decart ($300M at ~$4B, May 2026), Odyssey ($310M at $1.45B, Jun 2026), Runway - plus Tencent open-sourcing HY-World 2.0 (Apr 2026), which also outputs engine-importable 3D and commoditizes the layer
- World models may get absorbed as a feature of frontier multimodal models (OpenAI's Sora line, Gemini) rather than remaining a standalone category
- GPU cost and availability risk; a downturn in AI venture funding would hit a cash-burning lab hardest
Moats, dependencies & bottlenecks
Moats
Research talent and founder brand (Fei-Fei Li / Justin Johnson / Ben Mildenhall) Real but people-dependent; frontier-lab poaching is intense
persistent, exportable 3D worlds vs ephemeral video Current technical differentiation, but already narrowing - Tencent's open HY-World 2.0 also exports engine-ready 3D assets
Equity-aligned; integration into design tools would create workflow lock-in
3D data is scarcer than text/video - a genuine edge if they compound it
None evident yet; worlds are downloadable, which aids adoption but weakens lock-in
Dependencies
primary compute provider Oct 2024 deal makes Google Cloud the main training-compute supplier - while parent Alphabet competes directly via DeepMind's Genie line
GPU supplier + investor GPU access for training/inference; Nvidia also competes via Cosmos/Omniverse - partner and rival simultaneously
accelerator supplier + investor Secondary accelerator relationship diversifies GPU dependence
strategic partner + anchor investor $200M anchor and integration path; concentration of the enterprise story in one partner
Pre-profit frontier lab; future raises depend on the AI funding climate staying open
ecosystem/distribution Marble's value is partly realized inside third-party engines it does not control
Advantages
- Only independent world-model lab with an equity-aligned strategic distribution partner (Autodesk) plus both major GPU vendors on the cap table
- Engine-compatible, downloadable output shipped first commercially (Nov 2025) - though Tencent's open HY-World 2.0 (Apr 2026) now offers a free engine-ready alternative
- Category-defining brand — 'spatial intelligence' is substantially World Labs' framing, giving it the OpenAI-of-the-category narrative position
- ~5x reported valuation step-up in ~17 months signals continued access to capital at improving terms
Weaknesses
- Effectively pre-revenue at a reported ~$5.4B valuation - essentially all value is thesis, not financials
- Competing directly with Google DeepMind (Genie 3 / Project Genie) and Nvidia (Cosmos), which have orders-of-magnitude more compute and distribution
- Prosumer subscription pricing ($20-$95/mo) is a slow path to justifying a $5B-class valuation; enterprise/robotics monetization is still unproven
- World-model quality is compute-bound; a startup's cost of frontier training scales worse than hyperscalers'
Bottlenecks
- Training compute cost and GPU supply for scaling world-model quality
- Scarcity of high-quality 3D/spatial training data versus abundant text and video
- Enterprise sales motion is nascent - no disclosed named enterprise customers yet
- Inference cost per generated world constrains free-tier economics and pricing headroom
- Research talent supply amid frontier-lab compensation escalation
Top signals & trends
Top signals
Strategic validation and a concrete enterprise distribution path, starting with entertainment
Top-tier and strategic investors underwriting the thesis at a reported $5-5.4B
Fastest world-model lab to a shipping, paid, engine-exportable product
Absence of traction disclosure at this valuation warrants caution
Google compressing World Labs' technical lead and bundling world generation into its subscription ecosystem
Validates the category while funding a crowded competitive field
Trends
The core thesis; robotics and physical AI demand 3D-consistent world understanding
Multiple well-funded entrants; open releases (Tencent HY-World 2.0, Apache 2.0) risk commoditizing the generation layer
Marble's persistent geometry is purpose-fit for embodied-AI training environments; Decart's Oasis 3 pivot to driving simulation validates the demand
Incumbents integrating rather than building gives World Labs a partner-led path
Directly improves world-generation unit economics and free-tier viability
A funding-climate turn would hit pre-revenue $5B-class labs hardest
Ecosystem & competitor graph
Suppliers feed the company; customers pull from it. Line thickness shows the strength of each tie (supply-chain dependency, customer earnings contribution). Hover to isolate a tie.
Primary compute provider under an Oct 2024 deal reported to be worth hundreds of millions; parent also competes via DeepMind
GPUs for training/inference; also an investor via NVentures
Alternative accelerators; also an investor via AMD Ventures
Environment creation via Marble's engine-compatible 3D output; individual customers not publicly named
Concept-to-3D workflows; the Autodesk channel targets this base
Synthetic simulation environments for robot training; not publicly named
Freemium Marble tiers, $20-$95/mo
Real-time interactive photorealistic world generation (720p, 20-24fps); Project Genie shipped to Google AI Ultra subscribers Jan 29, 2026 - the compute-rich pace-setter
World foundation models for physical AI plus the Omniverse simulation stack; also a World Labs investor and supplier
Real-time interactive generative worlds, now targeting AV/driving simulation with Oasis 3; raised $300M at ~$4B (May 2026), led by Radical Ventures with Nvidia participating
Interactive world models; $310M Series B at $1.45B (Jun 2026) led by Natural Capital with Amazon and AMD Ventures; founders Oliver Cameron (Voyage/Cruise) and Jeff Hawke (Wayve)
General world models aimed at creative/media production, adjacent to Marble's creator segment
Open-sourced (Apache 2.0, Apr 2026) a world model that outputs engine-importable meshes/Gaussian splats - direct commoditization pressure on Marble's format (context only, not a buy/own call)
Open world-model drops in 2026 add to commoditization risk (context only, not a buy/own call)
LeCun-lineage world-model research (V-JEPA line) plus metaverse 3D-content ambitions
Generative gameplay and video-world capabilities that could absorb world modeling into general models
Incumbent 3D engine adding generative tooling; also a potential Marble output destination
Generative 3D creation inside its own closed platform