• Applied Compute's reported valuation went from $500 million in September 2025 to $1.3 billion in an $80 million round led by Kleiner Perkins in April 2026, and it is now reportedly in talks for a further round near $3 billion led by investor Elad Gil, per SiliconANGLE and AI Weekly.
• The product mechanism is reinforcement-learning post-training scored against checkable outcomes — does the code compile, does a parsed menu match a schema — rather than static human-labeled data.
• Named production customers include DoorDash (menu-image-to-schema conversion for over 100,000 merchant onboardings a year), Cognition's Windsurf (sub-two-second bug flagging), Ramp (spreadsheet search) and Mercor, per Podcast Alpha and the Generalist interview.
• At the $1.3 billion valuation, Applied Compute's annualized revenue was reported at $12.8 million — just over 100x ARR, a multiple that outpaces demonstrated enterprise spend, per AI Weekly.
• All three founders — Yash Patil, Rhythm Garg and Linden Li — previously worked at OpenAI on post-training infrastructure and the o1 reasoning model before starting the company.

Applied Compute, the AI infrastructure startup three OpenAI alumni founded in 2025, has seen its reported valuation climb from $500 million in September 2025 to $1.3 billion in an $80 million round led by Kleiner Perkins in April 2026, and it is now reportedly in talks for a further round near $3 billion led by investor Elad Gil, per SiliconANGLE and AI Weekly. The pitch behind that climb: general-purpose frontier models are a commodity, and companies that don't post-train their own specialized model on their own workflows will lose ground to competitors that do. The revenue disclosed alongside the April round — $12.8 million in annualized revenue — is the number that tests whether the pitch is ahead of itself.

The Mechanism: Post-Training Against Checkable Outcomes

Applied Compute's core technique, as Patil describes it, is reinforcement learning that scores a model's output against a verifiable, checkable condition rather than static human-labeled examples: whether generated code compiles, whether a parsed document matches a target schema, whether a flagged bug is a real one. That structure lets the company train narrow models for a single enterprise workflow instead of shipping a general assistant, per Patil's interviews with the Generalist and Podcast Alpha.

Three Founders, One OpenAI Pedigree

Patil spent two years at OpenAI working on post-training infrastructure and the Codex coding assistant before co-founding Applied Compute, per the Generalist. His co-founders, Rhythm Garg and Linden Li, previously helped build OpenAI's o1 reasoning model and its training infrastructure, respectively, per SiliconANGLE. All three left within roughly the same period to build a company betting that the specialization work they did inside a frontier lab is now a standalone enterprise product.

The Money Moved Fast

Applied Compute closed a $20 million seed round in July 2025, without a disclosed lead investor or valuation, per SiliconANGLE. An $80 million round followed in October 2025 with Benchmark, Sequoia and Lux participating, reported at a $500 million valuation. A second $80 million round in April 2026, led by Kleiner Perkins with Elad Gil, Lux, Greenoaks, NEO and Hana Bicapital joining, valued the company at $1.3 billion, per AI Weekly. Talks for a further round near $3 billion, led by Gil, followed — not yet closed.

Four Customers, One Narrow-Task Pattern

Applied Compute's named customers are deployed narrowly by design. DoorDash uses a post-trained vision-language model to convert merchant menu images into its structured data schema across more than 100,000 merchant onboardings a year, per Podcast Alpha. Cognition, maker of the Windsurf coding tool, uses a model that flags bugs in under two seconds. Ramp is reported to use a post-trained model for spreadsheet search. Mercor is also named as a customer, though the workflow it uses hasn't been detailed in available reporting.

What Patil's DeepSeek Example Actually Shows

On Podcast Alpha, Patil illustrated the cost structure of post-training by pointing to DeepSeek's own published numbers: DeepSeek V3's pretraining ran roughly 2.5 million H800 GPU-hours, while the reinforcement-learning stage that produced DeepSeek R1 used about 150,000 H800-hours — roughly 5% of the pretraining figure. That's a comparison between two stages of related model releases about a month apart, not a measured year-over-year decline in industry-wide training costs, and it should be read as Patil's illustrative example rather than an independently verified trend.

The Revenue Gap

The clearest test of the pitch is revenue, and here the numbers are less flattering. Documentation tied to the April 2026 round put Applied Compute's annualized revenue at $12.8 million, per AI Weekly — against a $1.3 billion valuation, a multiple just over 100x ARR. That gap isn't unusual for an AI infrastructure startup riding investor appetite for anything post-training-adjacent, but it means the growth story is still running well ahead of demonstrated enterprise spend.

The Verdict

Applied Compute's mechanism is credible: named, production customers with specific, checkable wins — DoorDash's onboarding pipeline, Cognition's bug-flagging latency, Ramp's spreadsheet search — are evidence most AI infrastructure pitches don't have. But the valuation trajectory is pricing in an outcome, that most companies will eventually need to own a specialized model rather than rent a general one, well before $12.8 million in ARR demonstrates that outcome is arriving at scale. The mechanism is proven in narrow deployments; the multiple is still a bet on how fast that narrows out into everyone else.