Team teardown · Unconventional AI

How to build a team that redesigns silicon from the circuit up.

Naveen Rao's third company re-hires the benches of his first two, then bolts on the analog crew that built Wi-Fi. Forty-three people, twelve months, and almost no one junior — all on one bet: let the chip's own physics do the math, instead of logic.

Get the full visual brief as a PDF.
Who we read
43 of 54 members
The mission
1000× less energy
The funding
$475M seed
The tell
17.4 yrs average
What to steal
The two-pool method

0.1 — The read

Forty-three people, twelve months, seventeen years of experience apiece. Not a startup staffing up — an industry re-forming.

This brief reads Unconventional as a talent-flow story, in five questions — who they assembled, what they're after, who's writing the checks, where it points, and what the rest of us can steal.

QThe questionThe short answer
Q1Who they assembledTwo alumni networks with zero overlap, held together by the founder himself.
Q2What they're afterSwap the substrate: physics does the math, at a thousandth of the energy.
Q3Who's writing the checksa16z, Lightspeed, Lux, DCVC, Bezos — $475M at seed, priced at $4.5B.
Q4Where this pointsThe hiring order points at a first chip. The gaps say software comes next.
Q5What others can learnThe two-pool method, and the check size it assumes.

§ 01 — Question one of five

Who did they assemble?

Forty-three people in twelve continuous months — most hired as veterans, a third holding PhDs, led by five VPs of engineering covering five different problems.

1.1 — The standouts

Ten hires that set the bar.

01
Naveen Rao
CEO & Co-Founder
Third act: founded Nervana (→ Intel) and MosaicML (→ Databricks, $1.3B), then ran AI at Databricks. Put $10M of his own into the round.
02
Michael Carbin
Co-Founder
MIT professor; author of the Lottery Ticket Hypothesis — 5,807 citations on sparsity. MosaicML founding advisor, Databricks principal scientist.
03
Sara Achour
Co-Founder · Research Fellow
Stanford professor of analog & unconventional compilation — how to program computers that aren't digital.
04
MeeLan Lee
Co-Founder & VP Engineering
ex-Google engineer; aboard at incorporation and the first VP of Engineering. The least public of the four founders.
05
Srenik Mehta
VP of Engineering
A long career in Atheros–Qualcomm analog silicon — the anchor of the second talent pool, and the reason the rest of that bench picked up the phone.
06
Sriram Krishnamoorthy
VP, Engineering
Two decades of supercomputing at PNNL, a US national lab; now hiring the compiler leads who sit between models and hardware.
07
Carlos Morales
VP Eng · System Modeling
Nervana alum from Rao's first company. Owns the simulation layer where the physics meets the model.
08
Chris Kim
Compute Hardware Research
Minnesota chip-design professor: 7,776 citations, h-index 45 — memory devices and circuits beyond the standard transistor.
09
Giacomo Pedretti
MTS · AI Hardware
ex-HPE; 3,129 citations in memristive, in-memory and analog computing.
10
Saavan Patel
AI Hardware Architect
Hired in June from rival Normal Computing; Berkeley PhD on neuromorphic and Ising-machine silicon.

1.2 — The shape of the team

They skipped the junior tier entirely.

17.4
Years average experience at join
35%
Hold PhDs — 15 of 43
63%
Big-tech background
11of 43
VP or executive level
1
Early-career hire, total

No intern cohort, no visible training pipeline, one Junior title on the roster. They are buying experience, not developing it, and the five VP-of-engineering-level owners each take a distinct problem: silicon, compilers, system modeling, software, and the build itself.

Schools underneath: Stanford ×7, Berkeley ×7, Harvard ×3 — the faculty ties are structural, not decorative. Rao's own bar: “after the 20th hire I probably would never have been hired.”

1.3 — The build order

Executives first. Research second. Silicon third. Software last.

Sep 2025
The nucleus. CEO and two co-founders — a VP Engineering and a Research Fellow — incorporate.
Oct 2025
The VP layer. Third co-founder, two more VPs of Engineering, Head of IT, the first circuit designer, ops.
Nov – Dec 2025
VP Engineering (System Modeling), Operations Manager, Head of Talent — the machine that hires the rest.
Jan – Feb 2026
CFO and the research bench opens: memory devices, AI theory, equilibrium propagation, compute hardware, analog design.
Mar – Jun 2026
Silicon implementation: SoC design, ASIC design, physical design, hardware validation, plus more AI research.
Jul – Aug 2026
Month ten+: VP Software Engineering, device-physics researcher, first marketing hire.

Three to four hires a month, without pause (40 of 43 join dates recovered). The entire VP layer predates the first research hire — an ordering only a pre-raised, multi-year plan can afford.

1.4 — Hiring topology

Two alumni networks that had never met.

One pool is the founder's own career, three companies deep — 11 of 43 people (26%) worked inside it. The other is the analog radio-chip world of the 2000s. They share no employer and no era: the only thing joining them is Rao himself, who has built in both.

Databricks · 8Google · 8Intel · 7MosaicML · 6Qualcomm · 5Atheros · 4Nervana · 3
The founder's ledger
Naveen RaoMichael CarbinCarlos MoralesRichard MahShawn Flood
Nervana → Intel → MosaicML → Databricks
no shared employer
no shared decade
The analog bench
Srenik MehtaMichael MackJason HouV. Gutnik
Atheros → Qualcomm · radio & analog chips

Plus one hire from neither pool: Saavan Patel, raided from direct rival Normal Computing. Nobody builds an AI company out of Wi-Fi-chip alumni. That's the tell: this bench exists to get an analog chip built.

1.5 — The sourcing moves

Three patterns a sourcer should take notes on.

01 · The raidsGroq · SambaNova · Normal · RainSix hires — 14% of the roster — came from AI-chip rivals: Patel (Normal Computing), Kwon & Wright (SambaNova), Flood & Thompson (Groq), Mah (Rain). The digital-accelerator generation is staffing the physics bet — and Patel came from a direct physics rival.
02 · The intact pairPNNL, recruited wholeVP Sriram Krishnamoorthy and computer scientist Ang Li co-authored QASMBench together at Pacific Northwest National Lab. Hiring a working collaboration skips the forming stage — and marks PNNL's quantum/HPC group as a live channel.
03 · The professorsMIT · Stanford · UMNCarbin, Achour and Chris Kim joined while still holding faculty posts (per LinkedIn) — three universities' worth of compiler and device-physics depth — without anyone having to quit their day job.

Network density overall: 7 lab alumni, 15 collaborators, 14 domain specialists, 7 GTM/ops — a real network structure, not a job-board team.

§ 02 — Question two of five

What are they after?

Pulling the digital computer out from under AI — a three-part physics 101, then the bench hired to build it.

2.1 — The thesis, in one line

The GPU isn't the enemy. The digital substrate is.

Every rival answer to AI's energy wall is a better digital chip. Unconventional's bet is that simulation itself is the waste: let a physical system — analog circuits that swing like pendulums and pull each other into step be the neural network, and the energy bill falls ~1000×.

“It's time to stop simulating neural networks on digital logic and start building hardware that actually behaves like them.”

Company launch, December 2025

Rao's public frame: AI scaling goes energy-limited at the global level within 3–4 years. The method is “neural co-evolution” — hardware and models designed together from day zero, rather than models written for chips that already exist. His stated odds: 50× “with a very high chance” on the first swing — with a path on to 1,000× and beyond.

2.2 — Physics-chip 101 · The wall

Three numbers explain the whole company.

80%+

Of all compute becomes one workload — deep neural nets — within 3–5 years. Heading for 99%.

109–10×

How far today's best GPU sits above the Landauer limit — physics' floor on energy per computation.

~10×

The brain's distance from that same floor — on 20 watts. The existence proof walks around.

Computers stayed general for eighty years because workloads were diverse — ENIAC priced artillery tables in 1945, then came accounting, databases, video. That diversity is ending, and the grid can't follow: AI demand wants ~400 GW of new US capacity this decade; the grid adds ~4 GW a year. The binding constraint is power generation, not transistors.

“We're running against fundamental limits of power generation. We need to be ready with a solution.”

Naveen Rao · “The Investment Memo,” Lightspeed · 2026

2.3 — Physics-chip 101 · The mechanism

Stop simulating the network. Be the network.

GPUs · every AI chip today

Digital — simulate it

A neural net, written in software
Frameworks · kernels · instruction sets
Arithmetic — trillions of multiply-adds
Transistors used as on/off switches

Four rungs of translation so one machine can run anything, and most of the watts aren't even math: they're spent shuttling state to memory and back.

Unconventional's wager

Physics — embody it

A neural net, as circuit behavior
Coupled oscillators settle into agreement — settling is the inference
Transistors used as dynamical systems

“Here's the initial state — kick it and let it run.” No memory round-trips. “The brain is not doing linear algebra. The brain is just doing brain things.”

Bounded on purpose: standard silicon, existing fabs, a five-year clock — near-frontier pieces that already work, engineered together. Hence 1,000×, not a billion-× — and no exotic-materials wing on the roster.

2.4 — Physics-chip 101 · Why these hires

“People who know circuits and silicon don't know anything about AI. People who know AI don't know anything about circuits and silicon.

Naveen Rao — on why this team can't be hired off a job board · “The Investment Memo,” 2026

No one spans the stack, so they hired one deep specialist per layer and tied them together with code and simulation — “herding cats and squirrels and frogs,” per the founder. Q1's two pools aren't a quirk; they're this stack, staffed.

L1
Circuits that behave
The Atheros–Qualcomm analog bench
Careers spent forcing Wi-Fi silicon to act linear, now aimed the other way: at the “ugly,” nonlinear regions where the compute lives.
L2
Devices & memory
Kim · Pedretti · device physicists
Which structures oscillate, couple and remember — inside a standard fab process.
L3
The iteration machine
Morales · system-modeling org
No breadboards at a million oscillators: propose → simulate → tape out → learn, repeat. The loop the $475M buys.
L4
Compilers, sans instruction set
Achour · Krishnamoorthy · Li
Someone must map PyTorch-shaped intent onto settling physics — Achour's entire academic field.
L5
Models physics can run
Carbin · Laydevant · the Databricks pool
Networks redesigned for noise and sparsity. Un-0 is this row's first receipt.

2.5 — Staffed by its authors

They hired the people who wrote the literature.

01
Chris Kim
7,776 cites · h-45
VLSI and memory devices at Minnesota — the beyond-CMOS circuits the substrate needs.
02
Ang Li
5,884 cites · h-37
Quantum and parallel computing at PNNL — how unconventional hardware meets real workloads.
03
Michael Carbin
5,807 cites · h-33
The Lottery Ticket Hypothesis: networks stay trainable when most connections go. Un-0's sparsity result replays his thesis on oscillators.
04
Giacomo Pedretti
3,129 cites · h-26
Memristive and in-memory computing at HPE — analog matrix math in real silicon.
05
Jérémie Laydevant
Nature Communications
“Training an Ising machine with equilibrium propagation” — now literally the day job.
06
Matthew Bull
Nature Reviews Physics
The multiscale physics of cilia and flagella — AI theory hired from soft-matter biophysics, not from ML.

26,196 combined citations across 12 resolved researchers (a floor) — the publication profile of a semiconductor lab that happens to target AI: Nature Materials, Nature Physics, PNAS, IEEE JSSC, SIGPLAN.

2.6 — The scoreboard

Nine months from founding to a shipped model.

No revenue, no customers — by design. The scoreboard is existence proofs: does physics actually do the computing?

Sep 2025
Incorporated. Rao, Lee and Achour aboard; Carbin joins weeks later.
Dec 2025
Out of stealth: $475M seed at $4.5B — three months old, priced entirely on the team.
Jun 2026
Un-0 ships: an image generator where a lattice of coupled oscillators does the work — a 6.74 FID score on ImageNet 64×64, where lower is better, and “the most capable physics-based image generator to date.” ~88% of parameters live in the oscillators, not the decoder.
Aug 2026
The sparsity result: cutting 93.25% of the connections improved the score by ~1.9 against a like-for-like dense baseline — fully-connected grids collapse into lockstep and stop being useful. How sparsely you wire it is a design knob.
Summer 2026
First silicon in flight: “no team in January” to a working prototype design in six months, tape-out — sending the design to the factory — set for this summer, with TSMC named as the intended manufacturer.

Un-0 is fully open: weights, training scripts, test results. ~320 GitHub stars — and effectively one committer. And it runs in simulation, on GPUs: the chip comes later. A proof, not yet a platform. Their yardstick, per Rao: energy per word generated, at equal output quality — “very hard to game.”

Talent brief

Want the full visual breakdown?

Download the PDF version of this teardown, including the build order, the two-pool topology map, and the five-move hiring playbook.

2.7 — Un-0, explained

Two metronomes on a table, scaled to 16,384.

Set two metronomes on a shared table and they drift into lockstep — that's coupling. Un-0's engine is 16,384 oscillators (the Kuramoto model, if you want the term) where every connection between a pair is a learned weight: 322M parameters, ~88% of them in the oscillator fabric, plus a small ordinary decoder that turns the result into pixels.

01 · The seed
Every oscillator starts at a random position — the equivalent of the random noise an image model starts from. Different seed, different image.
02 · The prompt
A small extra group of oscillators is wired in, nudging the swarm toward “daisy” or “volcano.”
03 · The compute
Released, the oscillators pull on one another and settle. No layers, no matrix multiplication — the settling is the computation.
04 · The snapshot
At a set moment, read every position once — a grid of angles that encodes the image.
05 · The render
A conventional decoder — under 13% of the parameters — turns phases into pixels.

— The honest nuance

Un-0 is a simulation of the physical system, run on ordinary GPUs — 640 B200-hours for the largest run. It proves the math computes; making it cheap is the chip's job, and the chip does not exist yet.

— The receipts

Controlled tests show the decoder alone fails, and untrained oscillators do clearly worse than trained ones. The oscillators are doing the computing. The physics supplies variety; the decoder supplies polish.

2.8 — The lane

Three companies are trying to re-found computing.

01
Normal Computing
Nearest rival · thermodynamic SPU
Same wager, different physics — and already a talent donor: one architect crossed to Unconventional in June.
02
Extropic
Same cohort · thermodynamic TSU
Announced its Z1 chip and Torx software in August — the chip has taped out, with first deliveries slated for 2027. Its CEO publicly claims the whole category. The loudest brand, the smaller published bench.
03
Intel Loihi · IBM TrueNorth
The ghost · neuromorphic lineage
The research programs that never left the lab. Rao ran Intel's AI products group next door to Loihi — this team is built to outrun that history.
04
NVIDIA + digital accelerators
The default · incumbent substrate
The actual enemy is the energy bill of digital simulation. Rao's frame: deep networks become a dominant global workload — and power is the limiter.

For a sourcer, Normal, Extropic and Unconventional are one market — physical-substrate AI compute. The qualified pool is dozens deep worldwide, and all three companies know each of them by name.

§ 03 — Question three of five

Who's writing the checks?

A $475M seed at a $4.5B valuation, three months after incorporation — priced entirely on the team.

3.1 — The cap table

Seed money at Series-C size.

$475M in December 2025 — and more coming: Rao's math is “probably a billion and a half” to first product. Sized like three companies because the plan is three disciplines: models, compilers, tape-outs.

01
a16z · Lightspeed
Co-leads · institutional
a16z put Rao on its NeurIPS stage the week of the announce; Lightspeed walked him through its own investment memo on camera. Both funds' AI-infra theses converge here.
02
Lux · DCVC
Deep tech · hard science
The signature underwriters of physics risk — the names that make a substrate bet legible to later investors.
03
Jeff Bezos · Naveen Rao
The angels · personal checks
Bezos personally, and the CEO's own $10M. A founder pricing his own conviction.
04
Sequoia · Databricks · Felicis · E14
The quiet bench · formation & strategic
Sequoia's Buhler calls himself “step zero” — in at formation. Databricks is on the cap table — the acquirer bankrolling its own alumni raid. Nine investor-advisors on the People page; a16z led yet appears nowhere on it.

The money is the hiring strategy — Rao's first chip company ran on $24M lifetime; this seed is 20× that.

§ 04 — Question four of five

What future does this point to?

The build order says a chip is coming. The absences say the software company hasn't been hired yet.

4.1 — Read the absences

What's missing is the roadmap.

01 · No ML infraNobody owns training at scaleNo one visible owns large-scale training or model operations — despite a shipped open-source model. The repo agrees: 22 of 25 release commits came from one person.
02 · No product, no GTMNothing is for saleThe first marketing hire landed in month eleven; there is no product or sales engineering at all. Until the chip works there is nothing to sell — and they aren't pretending otherwise.
03 · One designerThe audience is physicistsA single (anonymized) design hire. The constituencies that matter for two years: researchers, chip partners, and the talent they're still recruiting.
04 · Software arrived lastA tools org, pre-productEight-plus chip-implementation roles against a software group that only began in July, when the VP of Software Engineering arrived — month ten of twelve.
05 · No junior tierEvery gap is a senior reqNobody grows into these roles from inside — which keeps the hiring bar, the compensation, and the sourcing difficulty exactly where they are.
The readAbsences here are choices.At a $475M seed, nothing is missing by accident. Each gap dates the roadmap: infra and product hiring will announce the chip better than a press release will.

4.2 — The recruiting signals

Where the reqs point next.

01
Open now
AI Theory · System Modeling
The two posted roles (Mountain View): simulating physical systems, and the theory behind them — the core skills the whole bet rests on.
02
Hiring in public
ML compilers
Krishnamoorthy on LinkedIn: compiler-optimization leads — “technical leaders, not necessarily managers” — mapping models to novel hardware.
03
Building from zero
Software engineering
The VP of Software Engineering arrived in July — month ten. The whole software layer is being stood up now, after the chip work, not before it.
04
Confirmed: silicon
Tape-out · now public
The Mar–Jun implementation wave pointed at tape-out — Rao has since said it aloud: first silicon “this summer,” “maybe the largest analog chip people have ever built.” Hardware-test and lab bring-up roles should follow.
05
After silicon
Training infra · DX
Once hardware exists, the missing ML-infra and developer layers get hired, and the Databricks pool is where they'll shop first.

Sourcing against them? The analog pool is small, mapped and loyal — reach the Atheros–Qualcomm alumni before the next wave does.

4.3 — The 30-day pulse · Aug 1–31

A loud category, a quiet team.

01 · The noiseAugust was silicon-announcement monthExtropic announced its Z1 chip and Torx software — deliveries slated for 2027; Mythic's M-1 hit 80M analog weights at 3W; BrainChip boarded the AKD1500. Unconventional's August artifact: a model.
02 · Two generationsThey hired from the quiet oneZero hires from Extropic, Mythic or BrainChip. And the four chip companies that do feed the roster — Groq, SambaNova, Normal, Rain — made no category noise in 30 days. Disjoint sets.
03 · ReachabilityOne public voice: the CEOOf five people checked, only Rao posts. Carbin is silent at 5,807 citations. This roster is reached through LinkedIn and referrals, not through anyone's feed.
04 · The IP wallTooling policy, stated twiceRao publicly bans any AI tool that retains company data — no outside model gets trained on their work. An engineer echoes the same concern five days later.
05 · ConvictionThe thesis runs after hoursPaul Kwon keeps a personal repo probing synchronization in Sakana's Continuous Thought Machines — oscillator dynamics on his own time, at an oscillator company.
The readThe bidding may be widening.An account posting as ChronoPhase began recruiting the same profile on Aug 28 — one post, company unconfirmed. A watch item, not a bidder: today that's still Normal, Extropic and Unconventional.

§ 05 — Question five of five

What can the rest of us learn?

You can't copy the check or the ledger. The method underneath transfers.

5.1 — The lessons

Five moves worth stealing.

01 · SourcingWork the acquisition chainEvery acquisition mints an alumni pool that has trusted each other twice. Nervana→Intel, MosaicML→Databricks: Rao re-hired through both. Map who your founder has already paid.
02 · ScreeningBuy experience when physics is the risk27 of 43 joined as veterans; one early-career hire. When the open question is “does the physics work,” you don't apprentice — 17.4 years average is the point. The screen on top: “What do you do outside of work?” — listening for risk tolerance.
03 · SequencingHire the org chart top-downFive VP-level owners — silicon, compilers, system modeling, software, the build — before the first researcher. Only pre-raised capital makes it possible; if you have it, use it.
04 · The poolFind the bench nobody else wantsThe Atheros–Qualcomm analog crowd is on no AI recruiter's list. The rarest skills often sit in an industry everyone wrote off, and they stay loyal to whoever finds them.
05 · Units of hireRecruit relationships, not résumésA pair of co-authors from PNNL, three professors with their fields attached, a rival's architect. The unit of hiring is the working relationship — one warm introduction deep, every time.
The catchThis playbook costs $475M.Rao raised the entire multi-year plan up front precisely so he could hire like this. Without that check, sequence tighter and buy less — the patterns still hold at a tenth the scale.

Same shape as our Elorian AI and Eigen AI reads: the trust graph does the recruiting. Here it just happens to run through two industries instead of one lab.

Base to Base · Recruiting

The hardest teams to hire aren't assembled from one pool. They're welded from two industries that never met — by the one person who has worked in both.

Read the build order and the absences, and you can date a company's roadmap before it announces anything.

— The takeaway

They didn't assemble a team.
They re-convened an industry —
and aimed it at physics.

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

Methodology & limitations

How this read was assembled.

— Sources

  • The LinkedIn People page (54 associated members), with full profile histories for all 43 employees.
  • GitHub and OpenAlex — repo activity, citations, published research.
  • A 30-day public-signal pulse across X, GitHub and Hacker News (Aug 1–31, 2026).
  • Funding disclosures, founder podcasts and company blog posts.

— Limitations

  • The 54 associated members break down as 43 employees + 9 investor-advisors + 1 unrelated listing + 1 anonymized profile; the investor-advisors are excluded from team statistics, and 40 of 43 join dates were recovered.
  • Citation totals are a floor — 12 of 43 researchers resolved to public profiles.
  • This is not a complete org chart. The goal is a pattern-level read on how the team was formed. Data collected August 2026; published September 2026. Teardowns like this are how our searches begin — this one's on us.