Three Ex-OpenAI Researchers Are Selling Cheap AI to the Enterprise

When DeepSeek released R1 in January 2025, Silicon Valley read it as a warning shot. Marc Andreessen compared it to Sputnik; Alexandr Wang of Scale AI called it a wake-up call. Three young researchers inside OpenAI read it as a business.

Five months later Yash Patil, Rhythm Garg and Linden Li left to found Applied Compute in San Francisco. The company takes open models — Alibaba's Qwen, Moonshot's Kimi — fine-tunes them on a customer's own data, and runs the result on its own cloud.

The economics are the argument

The gap is not subtle. Top proprietary models such as Anthropic's Fable or OpenAI's Sol can cost close to forty times more per million tokens than some open alternatives. Patil says his platform can cut a customer's cost by as much as ten times against using the strongest frontier models — and enterprises have started to notice.

The round is being led by angel investor Elad Gil, with Lux Capital and Kleiner Perkins returning. Most of the money, Patil says, goes to chips: the cloud behind the service is the product.

Customers arrived before the valuation did

DoorDash uses the platform to generate accurate menus for restaurants joining its delivery network. Nvidia came to continue training Nemotron, its open-weight model. Harvey, the legal AI company, built an agent that reviews litigation files and contracts in under two months — work its research lead, Niko Grupen, says used to take a team and years.

Patil, 23, grew up in Austin in an Indian immigrant family; his father designed circuits at Qualcomm and ARM. He met Sam Altman at a Stanford dinner, emailed him months later, and joined OpenAI at twenty — where he worked on training infrastructure and Codex.

The question the pitch does not answer

Every one of those deployments moves a company's most sensitive material — menus, legal files, internal documents — into a training pipeline run by a twenty-person startup. That is the trade: you stop sending data to a frontier lab per token and start handing it to a smaller vendor in bulk.

Cheaper inference is real, and it is the right call for many workloads. But cost is the easy half of the decision. The other half is who can read the data, who can copy the model, how the tenant boundary is enforced, and what happens the day the vendor is acquired — because a $3.3bn valuation is exactly the kind of number that ends in an acquisition.

ABN Asia builds software for organisations that cannot afford to be the case study: threat models before design, least privilege in every integration, isolation proven by tests, secrets in a managed store, and a written response plan. You not only get software — you get secure software.

If you are moving internal data into an AI platform, talk to us before it ships: [email protected].

Source: Forbes Vietnam, November 2026 issue, reporting on Applied Compute.