Complex work instructions
Technicians run multi-step procedures on high-mix parts. A missed step shows up weeks later at the fab.
Hasta Labs
Foundation model for GPU manufacturing
The operating intelligence for the factories that build AI.
The problem
The factories that build AI still run on tribal knowledge
Almost nothing is recorded
Which step, which defect, how long, who decided: very little of it becomes data a model can learn from.
General models are blind here
Frontier models have read the internet. None has watched a chamber part get cleaned or a GPU rack get built.
The data gap
The process data AI's supply chain runs on has never been collected
What a frontier model knows
What a GPU factory needs
Textbook semiconductor processHow this site cleans this part number, step by step
Public spec sheetsDefect types per part, per customer, per site
Generic robot demosThe hand motions of expert inspectors and technicians
Average answersDecisions that hold under cycle-time and zero-escape limits
Proof today
The first layer is running on a real floor
- Working with a Tier-1 supplier that cleans chamber parts for leading fabs and tool makers.
- AI visual inspection, version 2, tested on their real parts at their development site.
- Paid pilot being scoped at a production site.
- Every inspector's accept or reject call becomes a training label.
- In talks with a GPU lifecycle company's teardown and test floor.
>25%
faster inspection cycle, with detection the same or better
Pilot success target set by the customer
The model today
We start on frontier models and earn our own
Frontier VLMs in our harness
It reads the part's procedure, defect library and material cards, then checks the part on video.
Zero escapes to ship
No change ships unless it catches every defect the last version caught: 20 held-out parts, 3 runs each.
Labelled floor record
Each inspector call is tied to the video, the part number and the procedure step.
Our own model
Trained on labelled floor hours, shipped when it beats the frontier harness on the same gate.
The data engine
Every deployment teaches the model
Verify
AI checks each part and step against the procedure. Inspectors approve or reject.
Train
The same recordings train new hires, so they work alone sooner.
Automate
Expert recordings become robot demonstrations for the same task.
The verifier we sell today becomes the robot's acceptance test tomorrow.
The vision
A foundation model that understands how GPUs get made, and makes better operating decisions than any general model can
Sees
Video from the floor: head cameras, CCTV, inspection stations.
Knows
Procedures, defect records, part histories, supply events.
Acts
Directs people, software agents and robots on one plan.
Where we start
One model across the whole GPU supply chain
- 01Later
Wafer fab
TSMC, Samsung, Intel
- 02Piloting
Tools and chamber parts
Applied Materials, Lam, their parts cleaners
- 03Next
Packaging and test
Advanced packaging, OSATs
- 04Next
Board and rack build
Foxconn, Wistron, rack integrators
- 05Later
Data center build
Hyperscalers and their builders
- 06In talks
Second life
GPU teardown, test and resale
Companies named show who works at each stage. They are not Hasta Labs customers.
Use cases
Run today's factories, stand up the next ones
Run a GPU factory
- Zero defects on high-mix parts
- Checks each part and step against its procedure before it ships to the fab.
- Every step followed, on record
- Sees skipped steps and queues building, and moves people before the line stalls.
- New technicians up to speed sooner
- Expert recordings become training, then robot work as the mix changes.
Stand one up anywhere, fast
- Give it the goal, the budget and the deadline.
- It lays out steps, stations, skills, robots and checks.
- It runs the ramp with what it learned at every site before.
Earth today. Orbit when it's cheaper there.
Why now
Four things changed at once
$151B
AI capex is rebuilding fabs
Forecast 300mm fab equipment spend in 2027, the first year above $150B.
67,000
People are the bottleneck
US chip technician and engineering jobs at risk of going unfilled by 2030.
1 model
VLMs remove per-line setup
A new part is set up from its procedure and defect library. Custom models per line sank the last wave.
$13.8B
Robot models need floor data
Raised by robotics startups in 2025, up from $7.8B. They need demonstrations and an acceptance test. We collect both.
Sources: SEMI 300mm Fab Outlook, Apr 2026 · SIA and Oxford Economics, Jul 2023 · Crunchbase, 2025
How we differ
Others inspect a line. We learn the process
| Who | What they do | Where they stop |
|---|---|---|
| Machine vision Cognex, Keyence | Fixed stations on high-volume lines | Each new part needs an engineer |
| AI inspection Landing AI, Elementary, Instrumental | Defect models trained per product | One model per line; no procedure steps |
| Connected worker Tulip, Augmentir | Digital work instructions | Records clicks, not what hands did |
| Robot model labs Physical Intelligence, Skild AI | General robot policies | No factory floor data or acceptance test |
| Hasta Labs | Checks parts and steps against the procedure | Every call trains people, then robots |
Team
Built by people who shipped vision AI at scale
Bhashwar Ghosh
Co-founder
Ex-NVIDIA, 2021–25. Senior engineer on Metropolis: vision-language models, 3D VLMs, agentic video systems. 4 patents filed. IIT Madras.
Jaivir Parmar
Co-founder
Prior founder with a hardware exit. Runs work on the customer floor.
Yashica Patodia
Co-founder
Working on Gemini Robotics at Google.
Roadmap
From one inspection station to robots on our own model
Verify and record
AI inspection on the floor. Every call logged, every expert motion recorded.
First robot cell
A robot picks the part and runs the inspection our software already judges.
Our robot foundation model
Trained on expert demonstrations from every site, with the verifier as its acceptance test.
Autonomous factories
One model plans, stands up and runs GPU factories, on Earth and beyond.
Contact
Running a floor that builds AI? Let's talk.
We work with parts cleaners, tool makers, packaging and test houses, rack integrators and GPU lifecycle teams.
hello@hastalabs.com