Team teardown · Moonlake AI

The barbell approach to building a frontier lab.

Three olympiad champions. Six doctorates. A 25-year NVIDIA Distinguished Engineer. And a 329,000-citation professor — part-time. That is one 30-person startup in San Francisco. Across the 31 profiles we read, only four are anywhere near the middle of their career.

Get the full visual brief as a PDF.
Who we read
31 profiles · 26 current
The bet
World models
The funding
$28M seed
The tell
4 of 31 mid-career
What to steal
Recruit with gravity

0.1 — The read

The most interesting team shape in AI right now: no middle.

Moonlake AI — founded 2025 by two Stanford PhD students, $28M seed from AIX Ventures, Threshold and NVIDIA — is betting that world models close the sim-to-real gap for physical AI. We read all 31 public profiles behind it. This brief reads the company in three acts: the academic bet, the pivot from games to atoms, and the barbell that came out the other side.

3
Olympiad champions
IPhO World Champion · ACM-ICPC first place · IMO-selection gold. Act I: an academic core takes an academic bet.
+76%
Headcount, four months
“A team of 17” in May → 30 on the page by September. Act II: the pivot to physical AI, staffed mid-flight.
4of 31
Mid-career rows
New grads at one end, 10-to-25-year veterans at the other. Act III: the barbell, and what it costs.

Everything here is public — LinkedIn, papers, GitHub, two founder talks. The reading is ours.

§ 01 — Act one of three

The academic bet.

Two PhD students walk out of a Stanford lab to attack sim-to-real, and recruit the only way academics can: through the lab.

1.1 — The founders

Founded by people most startups would call new grads.

No operator co-founder, no adult in the room. The founding credentials are medals and benchmarks, not exits — and one founder was still mid-PhD when the company started.

01
Fan-Yun Sun
Cofounder · CEO
Stanford PhD CS · NVIDIA Research · NTU 4.0. ACM-ICPC first place. Author of Holodeck — 2,509 citations, 60% of them since 2021, the only citation record on the roster. The research brand of the company is, so far, him.
02
Sharon Lee
Co-Founder
Stanford PhD in progress · Knight-Hennessy · Siebel. Left mid-PhD to start the company. Behavior-1K and NOIR — embodied-AI benchmarks with CoRL and CVPR credits. The robotics thread was hers before it was the company's.
03
Yitong Deng
Chief Scientist · promoted Jun 2026
Stanford PhD · Epic Games · Netflix. The first key hire, Chief Scientist within a year. SIGGRAPH Asia Best Paper; technical lead of Reverie, the rendering layer the whole bet rests on.

1.2 — The bet

Sim-to-real, attacked in two layers.

The founder laid the architecture out in a public Stanford talk in May 2026, and it doubles as the org chart. If you want to know who a technical founder will hire next, read the system diagram.

Layer 1 · Reasoning
A codegen agent that builds worlds like an expert developer — picks the engine, models geometry, assigns physics and logic. The Chen · Cao · Hsieh cluster.
Layer 2 · Rendering
Reverie, real-time neural rendering that reskins engine output into real-world pixels: the sim-to-real closer. Deng's layer — and where the veterans land.
The trick
Code is the memory layer, so there is no compounding next-frame drift — “the character changes over time” in autoregressive rivals.
The customers
“Companies with a combined market cap of $8 trillion” — simulations plus naturally-annotated training data. NVIDIA sits on both sides: seed investor and Omniverse partner.
The endgame
A self-improving loop — generate verifiable worlds at the right difficulty, an auto-curriculum for embodied AI. Simulations “as casual as chat.”

Also on record: the demo runs a forked open-source engine made “AI native”; a two-year “ChatGPT moment” prediction; and efficiency named as the bottleneck — autoregressive rivals cost “a few hundred dollars per simulated hour.”

1.3 — Who said yes

Hired out of a university, not out of companies.

No employer appears more than twice. Stanford appears sixteen times — 52% of everyone profiled. Early recruiting wasn't a strategy; it was gravity. The rest of the map is also campuses: Berkeley 5 · CMU 4 · Peking 4 · NTU 3 · Columbia 2.

C1
Lab alumni
10 people
The Stanford bloc: both founders, Manning, the interns, two investors.
C2
Lab collaborators
8 people
Shared a school or a company with someone already inside — Laforte, Buttimer, Chang, Lim.
C3
Domain specialists
8 people
3D/4D reconstruction, video generation, physics simulation, robotics.
C4
GTM / Ops
5 people
Hired late: growth, design, investor ops.

1.4 — The talent proof

Pre-revenue, the pitch is who already said yes.

3olympiad champions
IPhO World Champion (Hsu) · ACM-ICPC first place (Sun) · IMO-selection gold (Lim). Two medals were in the building from day one; the third arrived with the pivot.
329Kcitations, part-time
Christopher Manning — GloVe, CoreNLP, von Neumann Medal — took a Distinguished MTS seat while remaining a Stanford professor and an AIX Ventures GP (the same fund on the cap table). The halo hire: adds no management, signals frontier.
6doctorates
Sun, Chen and Hsu completed; Lee and Deng mid-PhD; the sixth is the advisor's.
19named inventions
Across seven people — Holodeck, Reverie, Behavior-1K, House-GAN, HoliGS, Blender MCP (14.5k★ — which departed with its author). The proof the medals generalize.

§ 02 — Act two of three

From games to atoms.

Over six months the tagline, the peer set, the product and the hires all moved from gaming to physical AI — and the team was rebuilt mid-flight.

2.1 — The pivot, on the record

The company stopped saying “game.”

Founders agonize over pivots in private. From outside, you can watch this one happen across every public surface in six months.

Tagline
“Generative Game Engine” → “Building Frontier World Models.” Spring materials versus the current LinkedIn page.
Headquarters
San Mateo → San Francisco. The on-site-San-Mateo hiring condition is already history.
Employee headlines
“Simulate and deploy robots” · “World Models & Simulation Infrastructure for Embodied AI” · “Building Interactive World Model.”
The peer set
Outsiders now group Moonlake with Mechanize, HUD Evals and Prime Intellect — “RL environments, evals and robotics simulation.”
CEO framing
Gaming is now one vertical among “robotics, mechanical industrial engineering, architecture and design” — a market, no longer the identity.

The tension worth naming: the roster's largest technical cluster is still 3D Vision + Graphics (13), with Game Development third (5). The talent reads as the old thesis, the language as the new one — which is what Act III is about.

2.2 — In their own words

The cleanest pivot test: what do the new hires say they do?

01 · Harry Hsu — MTS, Aug 2026“…accelerate physical AI”Describes the job unprompted: “the simulation and world model infrastructure needed to accelerate physical AI… that help robotics teams train, evaluate, and deploy.”
02 · Ethan Buttimer — Technical Artist“…for robotics simulation”His own words: workflows for “an AI-powered modeling agent to generate 3D scenes for robotics simulation.” The newest title on the page reads simply “Robotics.”
03 · The product moved too3D Agent × NVIDIA OmniverseA June 2026 release into Isaac / Omniverse — assets, digital twins, scenarios at scale. Sun's framing: “use simulation to narrow the solution space… close the sim-to-real gap.”
04 · GTM followedA physical-AI pipelineJennifer Rong's mandate: a “GTM pipeline for physical AI” — while the $30k Creator Cup keeps the consumer motion warm in parallel.

The buyer was on record by April: “folks at NVIDIA are paying a lot of dollars to purchase these types of interactive worlds… for training the robots or policies” — Sun, Latent Space.

Talent brief

Want the full visual breakdown?

Download the PDF version of this teardown, including the seniority histogram, the two ends of the bar named row by row, and the five-move hiring playbook.

§ 03 — Act three of three

The barbell.

Prodigies at one end, veterans at the other, almost nobody in between — the most interesting hiring shape we've mapped this year, and its costs.

3.1 — The ramp

Going to industry? Hire industry.

Selling worlds to robotics companies is an enterprise motion, and the roster moved to match. The five newest members: an NVIDIA Distinguished Engineer, an MIT PhD ex-Google, a Meta/Microsoft/DreamWorks technical artist, a Stanford engineer, and a UPenn roboticist.

17 → 30
“A team of 17” in May, 30 on the page by September — roughly +76% in four months, net of three departures. The ramp began with the pivot, not before it.
26%
Big-tech share, up from 19% — the growth is arriving senior and industry-sourced, not intern-shaped. Three of the five newest members came from big tech.
1st remote hire
Laforte, in Markham, Ontario — the on-site rule bent for a Distinguished Engineer.

Note what they did not hire: mid-level engineers. The new cohort is almost entirely at the veteran end of the bar — which gives Act III its shape.

3.2 — The two ends

Medals at one end. Decades at the other.

— The campus end

Fan-Yun Sun
Cofounder & CEO. Stanford PhD · ACM-ICPC first place · Holodeck, 2,509 citations.
Sharon Lee
Co-Founder. Mid-PhD · Knight-Hennessy & Siebel · Behavior-1K.
Jing Yi Lim
Founding Engineer. Berkeley EECS 4.0 · IMO-selection gold · Voleon quant.
Dalton Omens
Member of Technical Staff. Stanford School of Engineering — the pipeline extends.
The intern bench
Stanford · Berkeley, BAIR-adjacent — the recruiting posture is a CVPR happy hour.

— The veteran end

Christian Laforte
Member of Technical Staff. 25 years — AMD (Navi4/FSR4) · Stability AI · NVIDIA Distinguished Engineer.
Sandi Chakravarty
Member of Technical Staff. 18 years in game engines — Zynga · Playco · GOLF+. The Unity/Unreal hand.
Ethan Buttimer
Technical Artist. Meta Reality Labs · Microsoft · DreamWorks — Xbox, Azure, FX.
Harry Hsu
Member of Technical Staff. MIT PhD · ex-Google · IPhO World Champion — both ends in one hire.
Christopher Manning
Distinguished MTS, part-time. 329K citations — still a Stanford professor, still an AIX GP.

3.3 — The missing middle

Only 4 of 31 are mid-career.

Seniority as classified: Staff 12 · Mid 4 · Executive 3 · Senior 3 · Junior 3 · Lead 1 · Unknown 5. Average experience is 8.0 years — an average of two crowds, not a crowd of averages.

Read this as shape, not census. Public work histories are incomplete for roughly half the team, which is why five rows land in “unknown” — the buckets are noisy, the barbell is not.

3.4 — The barbell's tax

Three exits from seventeen — each end churns its own way.

Nov 2025 · 6 months in
William Liu, Founding Engineer, Stanford CS. The campus end's exit mode: leaves to found. Now “Co-Founder & CEO, Stealth.” High-optionality profiles treat seats as launchpads — price it in.
Feb 2026 · 4 months in
Jitesh Mulchandani, MTS, ex-PlayStation (10 yrs). The middle's exit mode: the thesis moved. Hired for the game-engine chapter, gone in four months — now Staff SWE at Unity, the engine company whose talent Moonlake lacks.
May 2026 · 13 months in
Siddharth Ahuja, Head of Product, Blender MCP author. The costliest one: the product seat emptied exactly as the beta commercialized, and Blender MCP's 14.5k★ left with him. Still unbackfilled on the page.

Counterpoint: the other games veteran of that same hiring wave, Sandi Chakravarty, stayed, and is now the only deep Unity/Unreal hand. The veterans who map to the new thesis hold.

Excluded: Kushal Kodnad's exit is an intern non-conversion (Berkeley ’26 → Google), and he still publicly champions the company. For ~17 people in May, three exits is a real rate.

3.5 — The shared language

What holds a barbell together: one filter both ends respect.

— The stated filter

“People who have the intersection of knowledge within code generation and computer vision and graphics… the majority of the team today do have both backgrounds.”

— Fan-Yun Sun, Latent Space, April 2026. The credentials he names out loud: “if you've written a game engine,” RL-ing coding models on varied objectives, multimodal latent-space alignment. A 22-year-old medalist and a 25-year veteran can both clear that bar — and both respect it.

— Tested against the roster

6 of 26 people (23%) show both backgrounds; 12 show either. Among engineers and researchers the ratio is roughly half — most of the “neither” rows are GTM, finance and investors.

Partial support, not contradiction. Founders round up when describing their own teams; the roster says the filter is real but aspirational. Worth knowing which of your own filters are which.

3.6 — The question

Is this the new barbell of hiring?

What the shape buysFrontier speed + enterprise credibility, day one.The campus end publishes at NeurIPS pace; the veteran end walks into NVIDIA and Meta rooms as peers. No translation layer of managers between them — the two-layer architecture is the org chart.
What it skipsThe people who ship v2.No visible infra owner for the six serving-and-training gaps; the product seat has been empty since May; DevRel absent. The execution middle is exactly what a commercializing beta consumes.
Where it's fragileBoth ends have exit velocity.Prodigies leave to found (Liu, month six). Mid-hires leave when the thesis moves (Mulchandani, month four). The bar only holds while the problem stays frontier, which may be the point.
What we can't tell you yetNo base rate.Zero of five comparable seed-stage physical-AI teams analyzed so far — we can't say whether the barbell is a Moonlake quirk or the new default. We'll re-run this map in two quarters.

Coda — What founders can steal

Five moves to steal. One tax to budget for.

01 · SourcingRecruit with gravity, not job posts.The first ten came from one lab's orbit — 52% Stanford, no employer clearing two. Your founding team is whoever your orbit already contains. Pick the orbit before the company.
02 · SignalMake the halo hire.A part-time legend (Manning, 329K citations) signals frontier harder than a full-time VP — and adds zero management overhead. The catch: you need the orbit first (he advised the co-founder).
03 · ScreeningPublish the architecture; let it recruit.The two-layer talk doubles as five hiring profiles — “if you've written a game engine” is a sharper filter than any JD. The best candidates self-select on the system diagram.
04 · SequencingPivot the roster with the thesis.Going after industry, they hired industry — big-tech share 19% → 26% in one cohort, the first remote exception for a Distinguished Engineer, and a newest title that reads simply “Robotics.”
05 · The tellLet new hires tell you if the pivot landed.Two of five newest members describe the new company unprompted, in their own headlines. If your week-two hires still describe the old company, the pivot is a memo, not a reality.
Budget forThe barbell's tax.Prodigies leave to found; mid-hires churn on pivots; backfill product before the beta commercializes, not after. Three exits from seventeen is the going rate.

The mirror image of our Unconventional AI read, where 43 hires averaged 17.4 years apiece and exactly one was early-career. Same trust-graph recruiting as Elorian AI and Eigen AI; opposite answer on seniority.

Base to Base · Recruiting

A team with no middle is not an accident of hiring. It is a statement about which problems are still frontier — and a bill that comes due when they stop being.

Read the seniority histogram before the org chart, and you can tell which stage a company thinks it is actually in.

— The takeaway

Medals at one end,
decades at the other —
and a bet that the middle can wait.

Brief
Moonlake AI · Talent Brief
Prepared by
Base to Base · Recruiting

Methodology & limitations

Wide coverage — honestly counted.

— Sources

  • Public LinkedIn profiles, posts and work history — 31 profiles read in full.
  • Publications, conference records and citations.
  • Founder media — two public appearances, quoted throughout.
  • The company's People page and investor announcements.

— Limitations

  • Coverage is 26 of 30 · 87%, measured against LinkedIn's own member count of 30. The 31 profiles include people who have since left, and investors — kept in the tables, flagged wherever they skew a number.
  • Four of the 30 are not yet profiled — the GTM and Finance leads among them, and two of the newest members keep light public footprints.
  • Public work histories are incomplete for roughly half the team, so the seniority buckets are noisy. One founder talk (Aug 18) is not yet covered, and there are no comparable-company base rates yet.
  • Data collected September 8, 2026; published September 2026. Teardowns like this are how our searches begin — this one's on us.