Company teardown · World Labs
The 77 people worth $8.2B — and why you never pass on an influential AI researcher's startup.
Fei-Fei Li built ImageNet; Justin Johnson was her PhD student; Ben Mildenhall created NeRF. In 32 months their lab raised $1.23B, shipped Marble, bought a robotics team and agreed to sell to AMD for $8.2B in stock. A famous founder explains the first cheque. It doesn't explain the price. We read 77 of 83 profiles to find what does: two research lineages hired on purpose, a platform built before the model, an acquisition made for its people — and a buyer who put in writing that it was buying the team.
0.1 — The case in six steps
A famous founder explains the first cheque. Seventy-seven profiles explain the price.
Everything here is public — LinkedIn, GitHub, OpenAlex, the careers page, five founder appearances with transcripts, and the two deal announcements. The reading is ours.
§ 01 — Part one of six · The idea
Stop predicting words. Start predicting worlds.
Step one of the case: the bet is a scaling bet, made by founders who had already won one — and a lab that publishes at SIGGRAPH as often as at CVPR.
1.1 — The problem they're solving
Why a model that can write still can't see around a corner.
— The founder's framing
“I think three of us have total conviction about the scaling law.”
— Fei-Fei Li, The a16z Show, September 2026.
— World model, translated
A model that represents a scene well enough to render it, simulate what happens next and plan inside it — World Labs' own three functions, which share “largely the same” knowledge. Marble renders, SceniX simulates, Atlas predicts.
1.2 — The idea
Render it, simulate in it, then predict the view nobody photographed.
Johnson's frame: “simple representations and then scale up the model behind the simple representations.” Two tracks ran at once — explicit splats (Marble) and implicit pixels (RTFM, then Atlas) — “and that's exactly why we've done both at World Labs.”
01 · Render
02 · Simulate
03 · Predict
↺ rendering, simulating and planning share one body of knowledge — so one lab, one roster
1.3 — The wager
A research lab that bets compute is the constraint — and agrees to sell itself to a chipmaker.
Johnson, three and a half weeks before the deal: “I think we're basically at the beginning and we're basically limited by compute at this point.” World Labs' answer was not a bigger round. It was to join the company making the GPUs, after a year of training and inference work on AMD hardware. The bottleneck named in September was answered in September.
— The company's own reason
“To accelerate into this future requires scaling our efforts, scaling our reach, and getting closer to the hardware.”
— World Labs blog, September 28, 2026.
— And nobody gave a headcount
No founder stated one in five appearances. Press said about 42 in December 2025 and about 70 at the deal; the roster reads 32 by the end of 2025 and 77 now — 75 without interns, 72 without fractional staff.
§ 02 — Part two of six · The business
Sell the world, $35 a month at a time.
Step two: the revenue doesn't explain the price. A freemium app, a per-world API, a robotics engine for design partners — and investors who look like customers.
2.1 — The product, as described
An app, an API, a robotics engine — and investors who look like customers.
No revenue, retention or API-volume figure has been disclosed. The 100k-user figure comes from an employee profile, not a company release.
2.2 — The money
$230M at launch, $1B seventeen months later — and $8.2B in AMD stock.
2.3 — Who else is trying
Compared with video labs — hired from 3D labs.
Talent brief
Want the full visual breakdown?
Download the PDF version of this teardown — both deals reconstructed from the principals' own words, the six-group org chart, ten names to know, the build order month by month, the SceniX fifteen, and a six-question audit for any star-researcher startup.
§ 03 — Part three of six · The deals
One deal bought a fifth of the company. The next one buys all of it.
Step three: the buyers say what they are paying for. A former postdoc's startup, acquired for its people; a GPU partner that invested, then agreed to buy — and wrote that it was buying researchers.
3.1 — SceniX, reconstructed from the founder's account
Customer, recognition, a talent case — then fifteen people and a New York office.
↺ five weeks later Li calls SceniX “our robotics team”
15 of 77 arrived through SceniX; 9 of the 11 New York rows; 3 Body Labs alumni; 7 robotics venues new to the roster. Reconstructed from the a16z episodes of Jul 28 and Sep 4, 2026, the World Labs blog of Jul 21 and Li's Substack of Sep 28. Founder media is self-report; the price has never been disclosed.
3.2 — AMD, from GPU partner to buyer in about fourteen months
Partner, investor, acquirer — and the deal team was hired along the way.
3.3 — What AMD is buying, in both sides' words
A roadmap and a research org — and not a word about the product.
— The constraint, named
“I think we're basically at the beginning and we're basically limited by compute at this point.”
— Justin Johnson, The a16z Show, September 4, 2026.
— The founder's framing
“My career has led up to this moment — to continue growing and running a frontier research team, within a leading AI company.”
— Fei-Fei Li, Substack, September 28, 2026.
Commentators read the deal as a talent and model buy; the founder's own text supports it. “Continue growing” is the only hiring signal in the announcement, and it is qualitative.
3.4 — The business layer, from 77 profiles
Eight people run the business side — none of them before August 2025.
Nobody is titled sales. Who negotiated with AMD is not on the public record; the join dates are.
3.5 — What the record does not show
Five things we could not verify — and two numbers we rejected.
3.6 — What transfers from the deals
Five moves from two deals worth stealing.
§ 04 — Part four of six · The founders
A professor, her student, and the creator of NeRF.
Step four: the names start the case; they don't finish it. 288K citations between three co-founders, a fourth who stepped back, and a leadership layer hired almost entirely in 2026.
4.1 — The four who founded it
ImageNet's creator, her PhD student, NeRF's first author — and the co-founder who stepped back.
4.2 — The leadership layer
Two leads from day one, a COO in year three — and three Heads hired in 2026.
Below the co-founders, 43 of 77 are titled Member of Technical Staff — a 26-year VMware principal and a 2026 PhD alike. No VP, no director in engineering; the three Heads are all on the business side.
4.3 — How they hire · The a16z Show ×2 · Latent Space · TWIML · Substack
Deep researchers first, then engineers, then business — and a founder who says academia shouldn't train for labs.
§ 05 — Part five of six · The team
Seventy-seven of 83, hired in three waves and one acquisition.
Step five: the roster, hire by hire. Founders and platform in 2024, product in 2025, a doubling in 2026 — and a doctorate rate that didn't move.
5.1 — The org chart, from the bets · inferred from profiles
Six groups — and robotics, acquired whole, is already the fourth largest.
Groups are our reading of 77 profiles, not an org chart World Labs publishes, and they sum to 77. Robotics did not exist in June and is the fourth-largest group by August — fourteen people who arrived in a single transaction.
5.2 — Ten names to know
The hires that set the bar.
5.3 — The shape of the team · 77 profiles
Senior, doctoral, and mostly new.
— Career stage at join · 76 of 77 profiles, 1 unknown
The early-career band is not a junior bench: the pipeline counts post-degree years only, so a 2025 PhD lands there. Among the 45 who joined in 2026, 18 hold doctorates — 40%, the company's own rate; growth has not diluted it.
Schools, by LinkedIn's census over all 104 members: Stanford 11, UIUC 9, CMU 6, MIT 5, Brown 5. Our own read under-counts UIUC because few education pages loaded, so treat the census as the better UIUC number only — it counts all 104 cards, including the 17 investors and 4 advisors this page excludes everywhere else, and an investor bench clusters exactly where a Stanford count would.
5.4 — The build order · first World Labs role, 77 profiles
Founders and platform in 2024. Product in 2025. Everything doubled in 2026.
Departed staff never appear on a current roster, so every count above carries survivorship bias. At least one founding-era MTS left in September 2024 and is now a robotics CTO; he still advises.
5.5 — Where they came from
Two pools — and an acquisition as the third.
Twenty of 77 have Meta on the CV and 16 have Google — the vision pool. Sixteen passed through Pixar, ILM, Epic, Riot, Niantic, Magic Leap or Body Labs — the graphics pool. The third pool arrived in one transaction.
These count prior affiliations — employers, schools and programmes — across the 77 profiles we read. Highlighted chips are the sources that supplied a research bench rather than individuals: the two big-tech vision labs, the acquisition, and Stability's 3D line. Immediate prior employer: SceniX 13, Meta 8, Adobe 4, Google 3. The Runway zero is real — nobody has moved from the video world-model labs Atlas is compared against.
arrived in one deal
The third pool is SceniX — Columbia, robotics and simulation, 15 of 77 in a single July. It is also where the two older pools meet: three Body Labs alumni came through it, and a fourth passed through World Labs as a co-founder.
5.6 — The acquisition as a feeder · 15 of 77
One July, fifteen people — a robotics bench and a Body Labs reunion.
SceniX was twenty months old when World Labs bought it. Its people still list “SceniX (now part of World Labs)” or a World Labs role from July–August 2026.
5.7 — Research firepower · OpenAlex, 29 verified records
One name is 71% of the citations — the rest is a graphics-and-vision lab.
5.8 — The census and the job board
Business arrived late — and the board still asks for researchers.
— The company · our reading of 77 profiles, Sep 2026
— The job board · 11 open roles, worldlabs.ai/careers · Ashby
LinkedIn's census over 104 members reads Engineering 53, Business Development 26 — investors file under BD, researchers under engineering. The roster says the operating layer was built in 2026: COO in January, Head of Strategy in April, Head of People in July, a New York talent lead in August — and the General Counsel is still fractional.
§ 06 — Part six of six · The verdict
Never pass — and here is why the roster says so.
You can't copy ImageNet's author or an $8.2B buyer. What transfers: when a top researcher leaves room in the round, the roster is the reason to take it.
6.1 — The audit · six questions for any star-researcher startup
A name is a reason to look. The roster is the reason to write the cheque.
6.2 — Sourcing patterns
Six patterns — most of them are lineages, not job posts.
6.3 — The lessons
Five moves worth stealing.
— The takeaway
A famous name explains the first cheque.
Seventy-seven profiles explain the price —
and why you never pass when a top researcher leaves room for yours.
Methodology & limitations
A near-complete read — honestly counted.
— Sources
- The LinkedIn company page — 104 associated members, less 17 investors, 4 advisors and 3 cards belonging to other companies named “World Labs”, plus 3 SceniX co-founders who appear in acquisition coverage but not on the page — with profile histories for 77 of 83 staff.
- The GitHub orgs
worldlabsaiandsparkjsdev, and the careers page (11 roles, fetched Oct 4). - Five founder appearances with transcripts — The a16z Show ×2, Latent Space, TWIML, and the founder's Substack.
- OpenAlex citations for 29 members, a floor; a 30-day public-signal scan (Aug 30 – Sep 29); and for Parts two and three the World Labs blog, AMD's release, TechCrunch, CNBC, Yahoo Finance, Quartz, Construction Dive, Radiance Fields, Beri and Seedtable, fetched Oct 4.
— Limitations
- The doctorate rate is a floor. LinkedIn withheld education pages after the first 14 profiles; the other 63 degrees come from profile headers, academic roles, CVs or theses, 35 of them with a cited URL. Read 40% as “at least.”
- 83 is a derived ceiling, not a headcount, and 77 is what we read. The 6 we could not read are anonymised “LinkedIn Member” cards. Press put the team near 70 at the deal, below our own 77, so we quote both and lead with the roster.
- Every join count carries survivorship bias. Departed staff never appear on a current roster. One of four co-founders stepped back in June 2026 and his headline still reads “Cofounder”; at least one founding-era MTS left in 2024.
- The six groups on 5.1 are our reading of profiles, not an org chart World Labs publishes.
- No comparative baseline. World Labs is the first in our world-model cohort; a comparison needs five or more.
- The roster predates the AMD agreement, which had not closed at collection. Collected September 29–30, 2026. Teardowns like this are how our searches begin — this one's on us.