Launch: Dana agentic platform For physical AI application development
Phase 1: tools engineers needed Develop and test autonomous systems
Phase 2: OS for the machines Underpinning layer for physical systems
Phase 3: intelligence on machines Models and planning running on hardware
Mission: intelligence on a billion machines Qasar Younis’s stated destination
Bigger than robotaxi tooling Physical AI across many machine types
Collective prediction: few full-stack winners Build, operate, and own the market
Capital went to full-stack robotaxis Each program rebuilt the same stack
Every team hired its own engineers And assembled its own fleet
Same internal tooling rebuilt repeatedly Engineers carried the ritual company to company
Lots of conviction about autonomy Very little shared best practices
Make autonomy best practices a company Sell infrastructure, not robotaxis
Minority / unfashionable positioning Against the vertical-integration fashion
Large L4 programs said no Cruise and peers preferred internal tools
They would not wait on a startup Their engineers planned to build tooling anyway
Many full-stack programs later failed or shrank Cruise sunset after GM acquisition / incident
Applied outlasted most of them Supplier thesis survived the hype cycle
Cars are not winner-take-all software Regulated, capital-intensive, multi-OEM distribution
One maker stays single-digit share Industry structure favors many manufacturers
Other ~90% of cars need intelligence too Via the existing automotive industry
OEMs cannot each rebuild every layer Independent supplier must provide common stack
Wager: diffusion over concentration Incumbents become technically different companies
Founders: Qasar Younis + Peter Ludwig GM, Google, YC experience + safety-critical patience
First customers: smaller autonomy teams Voyage and Bay Area cohort
Planning simulator first Then perception simulation
Then data + mass simulation ops Millions of scenario runs as infrastructure
Deterministic sim correlated to reality Unglamorous but industry-critical
Road event → repeatable test Then implement the fix in the real world
Shared touchstone for “does it work?” Without testing every change in production on roads
GM formal tooling RFP (~2018–19) 28 companies bid, including giants
Applied beat NVIDIA / Ansys on the spec Performance against requirements won the deal
Startup won industrial procurement Despite size and many alternatives
Playbook: small aggressive customers first Then evidence sells to industrial incumbents
Wedge became legitimacy with OEMs Repeatable go-to-market pattern
Defense entry ~18 months after founding Hire people who know the domain from inside
Apply auto pattern to defense systems Simulation, data, integration, validation
Then construction and mining Then commercial trucking
Not “change nouns in the sales deck” Rebuild product around domain constraints
Procurement + safety cases shape the product Each vertical has real physical rules
Lower layer: sim + data Scenarios, logs, reproduce, evaluate
Generate scenarios and collect machine data Reproduce events and score behavior
Middle layer: OS for physical machines Schedule, middleware, memory, comms, safety
OS owns the machine, not a UI app stack Functional safety is first-class
Upper layer: intelligence itself World models and planning systems
Deploy across land, air, and sea Same stack idea, many vehicle types
~$15B valuation (2024–25 round) External measure of the platform thesis
18 of 20 top non-Chinese automakers OEM coverage as proof of diffusion bet
Driverless L4 trucks in Japan Using Applied technology in operation
U.S. Army deployments + Navy OS Defense and maritime production footprint
Mining, construction, agriculture, trucking Physical industries beyond passenger cars
Nearly preserved ~$1B primary capital While nearly doubling revenue for years
Rare SV capital discipline at scale Efficiency as part of the story
Tool solves a bounded task first Then becomes the shared interface
Accumulates integrations and workflows Plus test cases and organizational memory
Stops being an accessory Becomes the environment of production
Browser made the net usable Did not create the internet; enabled the economy
Sim → data → OS followed customer work Software became how they build machines
Agentic layer on Applied’s full system Not a greenfield toy agent
Start from a requirement Connect to relevant code automatically
Run change through sim + defined evals Then test bench or hardware-in-the-loop
Stage for a physical machine Close the loop to deployment
Replace manual handoffs between tools Agent coordinates the path
Keep record of how/why it changed Traceability is part of the product
Old frontier: imitation learning Copy enough recorded human driving
New frontier: end-to-end RL closed loop Fail, sample, train, retest the same case
Applied already has the loop plumbing Sim, synthetic data, eval, deploy
Dana lets agents operate that loop Human+agent orchestration of the cycle
Long hardware cycles + physical risk Mundane until tons of mass near people
Every link from spec to deploy is scrutinized Safety cases must survive review
Belief ≠ permission to ship Regulators need auditable reasoning
Dana built around traceability Evals and records as first-class value
Speed without proof is not enough Physical AI needs both
Model capability arrived faster than deploy Orgs cannot absorb frontier models alone
Intelligence cost falling steeply Autonomy layer trends toward abundance
One OEM only sells what it builds Dana upgrades existing 20-year fleets too
Machines no longer need a human cockpit shape Design space opens without a driver seat
Path to a billion intelligent machines Through manufacturers, not one robotaxi fleet
Bottleneck moved from models to deployment Hardware adaptation + ops far from the lab
Most firms cannot hire 1000-person AV orgs They need a platform instead
Nine years of contact with hardware Now packaged so anyone can move a machine alone

Applied Intuition & Dana

1. Mission & launch Dana and the billion-machine goal

Launch: Dana agentic platform For physical AI application development

Phase 1: tools engineers needed Develop and test autonomous systems

Phase 2: OS for the machines Underpinning layer for physical systems

Phase 3: intelligence on machines Models and planning running on hardware

Mission: intelligence on a billion machines Qasar Younis’s stated destination

Bigger than robotaxi tooling Physical AI across many machine types

2. 2017 industry context Full-stack robotaxi consensus

Collective prediction: few full-stack winners Build, operate, and own the market

Capital went to full-stack robotaxis Each program rebuilt the same stack

Every team hired its own engineers And assembled its own fleet

Same internal tooling rebuilt repeatedly Engineers carried the ritual company to company

Lots of conviction about autonomy Very little shared best practices

3. The unfashionable bet Shared tooling vs vertical integration

Make autonomy best practices a company Sell infrastructure, not robotaxis

Minority / unfashionable positioning Against the vertical-integration fashion

Large L4 programs said no Cruise and peers preferred internal tools

They would not wait on a startup Their engineers planned to build tooling anyway

Many full-stack programs later failed or shrank Cruise sunset after GM acquisition / incident

Applied outlasted most of them Supplier thesis survived the hype cycle

4. Founding thesis Diffusion over concentration

Cars are not winner-take-all software Regulated, capital-intensive, multi-OEM distribution

One maker stays single-digit share Industry structure favors many manufacturers

Other ~90% of cars need intelligence too Via the existing automotive industry

OEMs cannot each rebuild every layer Independent supplier must provide common stack

Wager: diffusion over concentration Incumbents become technically different companies

Founders: Qasar Younis + Peter Ludwig GM, Google, YC experience + safety-critical patience

5. Early wedge Small teams → product demands

First customers: smaller autonomy teams Voyage and Bay Area cohort

Planning simulator first Then perception simulation

Then data + mass simulation ops Millions of scenario runs as infrastructure

Deterministic sim correlated to reality Unglamorous but industry-critical

Road event → repeatable test Then implement the fix in the real world

Shared touchstone for “does it work?” Without testing every change in production on roads

6. GM breakthrough Legitimacy with OEMs

GM formal tooling RFP (~2018–19) 28 companies bid, including giants

Applied beat NVIDIA / Ansys on the spec Performance against requirements won the deal

Startup won industrial procurement Despite size and many alternatives

Playbook: small aggressive customers first Then evidence sells to industrial incumbents

Wedge became legitimacy with OEMs Repeatable go-to-market pattern

7. New verticals Defense, mining, trucking…

Defense entry ~18 months after founding Hire people who know the domain from inside

Apply auto pattern to defense systems Simulation, data, integration, validation

Then construction and mining Then commercial trucking

Not “change nouns in the sales deck” Rebuild product around domain constraints

Procurement + safety cases shape the product Each vertical has real physical rules

8. Product stack ladder Sim → OS → intelligence

Lower layer: sim + data Scenarios, logs, reproduce, evaluate

Generate scenarios and collect machine data Reproduce events and score behavior

Middle layer: OS for physical machines Schedule, middleware, memory, comms, safety

OS owns the machine, not a UI app stack Functional safety is first-class

Upper layer: intelligence itself World models and planning systems

Deploy across land, air, and sea Same stack idea, many vehicle types

9. Results by 2024–25 Valuation, customers, deployments

~$15B valuation (2024–25 round) External measure of the platform thesis

18 of 20 top non-Chinese automakers OEM coverage as proof of diffusion bet

Driverless L4 trucks in Japan Using Applied technology in operation

U.S. Army deployments + Navy OS Defense and maritime production footprint

Mining, construction, agriculture, trucking Physical industries beyond passenger cars

Nearly preserved ~$1B primary capital While nearly doubling revenue for years

Rare SV capital discipline at scale Efficiency as part of the story

10. Tools become platforms Browser-like corollary

Tool solves a bounded task first Then becomes the shared interface

Accumulates integrations and workflows Plus test cases and organizational memory

Stops being an accessory Becomes the environment of production

Browser made the net usable Did not create the internet; enabled the economy

Sim → data → OS followed customer work Software became how they build machines

11. What Dana is Agentic interface over the system

Agentic layer on Applied’s full system Not a greenfield toy agent

Start from a requirement Connect to relevant code automatically

Run change through sim + defined evals Then test bench or hardware-in-the-loop

Stage for a physical machine Close the loop to deployment

Replace manual handoffs between tools Agent coordinates the path

Keep record of how/why it changed Traceability is part of the product

12. Technical shift Imitation → RL closed loop

Old frontier: imitation learning Copy enough recorded human driving

New frontier: end-to-end RL closed loop Fail, sample, train, retest the same case

Applied already has the loop plumbing Sim, synthetic data, eval, deploy

Dana lets agents operate that loop Human+agent orchestration of the cycle

13. Burden of proof Traceability for physical AI

Long hardware cycles + physical risk Mundane until tons of mass near people

Every link from spec to deploy is scrutinized Safety cases must survive review

Belief ≠ permission to ship Regulators need auditable reasoning

Dana built around traceability Evals and records as first-class value

Speed without proof is not enough Physical AI needs both

14. Economics & why now Models cheap; deployment hard

Model capability arrived faster than deploy Orgs cannot absorb frontier models alone

Intelligence cost falling steeply Autonomy layer trends toward abundance

One OEM only sells what it builds Dana upgrades existing 20-year fleets too

Machines no longer need a human cockpit shape Design space opens without a driver seat

Path to a billion intelligent machines Through manufacturers, not one robotaxi fleet

Bottleneck moved from models to deployment Hardware adaptation + ops far from the lab

Most firms cannot hire 1000-person AV orgs They need a platform instead

Nine years of contact with hardware Now packaged so anyone can move a machine alone