Applied Compute is an American artificial intelligence company, based in San Francisco, that trains custom AI models for businesses on their own data and then runs those models on its own cloud.[1][2] It builds special-purpose models and agents by post-training open-weight models such as Alibaba's Qwen and Moonshot AI's Kimi, using reinforcement learning to fit them to a customer's tasks, an approach the company calls "Specific Intelligence".[2][3] It was founded in 2025 by Yash Patil (chief executive), Rhythm Garg (chief technology officer) and Linden Li, three former OpenAI researchers who became friends as students at Stanford University.[2][3][8]

Applied Compute logo, from the official website appliedcompute.com. Trademark of its owner.

Applied Compute started by sending its own engineers and researchers to work inside customer teams. In August 2026 it opened AC2 (Applied Compute Agent Cloud), the training and inference platform its researchers use internally, to outside teams in a private beta.[4] Customers it has named include DoorDash, Cognition AI, Mercor, Harvey, Microsoft and Nvidia.[4] In April 2026 the company announced $80 million in financing at a $1.3 billion post-money valuation, bringing its total funding to $160 million.[5] In September 2026 it told Forbes that it was in talks to raise $350 million at a $3.25 billion valuation in a round led by investor Elad Gil, and cautioned that the figures could change before the round was finalized.[2]

Attribute	Details
Type	Private company
Founded	2025 (May 2025, according to The Information)[2][9]
Headquarters	San Francisco, California[2]
Founders	Yash Patil (CEO), Rhythm Garg (CTO), Linden Li[2][6][8][13]
Main activity	Reinforcement-learning post-training and inference of custom models built on open-weight base models
Products	AC2 (Applied Compute Agent Cloud), Ari research agent, Shift-1 model router
Announced funding	$160 million in total as of April 2026[5]
Last announced valuation	$1.3 billion post-money (April 2026)[5]
Reported talks	$350 million at a $3.25 billion valuation, per the company (September 2026)[2]
Website	appliedcompute.com

History

Founders at OpenAI

The three founders became friends at Stanford, according to Forbes.[2] Patil, who grew up in Austin, Texas, met OpenAI chief executive Sam Altman at a dinner Altman hosted for Stanford engineering students. After ChatGPT launched, Patil emailed Altman asking to work on it, was invited to apply to OpenAI's residency program, and joined the company at age 20. He then referred Garg to OpenAI, and Garg referred Li.[2] Patil left Stanford without finishing his degree and spent two years at OpenAI working on post-training infrastructure and on Codex.[2][6] Garg worked on reasoning models and Li on machine learning systems.[2]

The company's launch post describes Patil as a key member of OpenAI's agentic software engineering effort (Codex), Garg as a core contributor to o1, which it calls the first RL-trained reasoning model, and Li as a core contributor to ML systems and infrastructure for RL training.[3] OpenAI's published list of o1 contributors names Garg among the core contributors and Li among the contributors to the model's reasoning research.[7] Li's profile on X also lists earlier work at Databricks Mosaic AI and Nvidia.[14] In 2025 Garg and Li gave a talk titled "Efficient Reinforcement Learning" at the AI Engineer Code conference.[8]

Forbes traced the company's origin to the January 2025 release of DeepSeek-R1. The three researchers were on a ski trip at Lake Tahoe when the open-weight model came out, and the episode persuaded them that open models, paired with the right infrastructure for optimizing them, could support a large business.[2] Patil told Forbes that their view differed from that of many OpenAI colleagues: instead of a few labs training one model on all of the world's data, he expected large companies to train their own models on data they would never hand to an outside lab.[2] About five months after the R1 release, the three left OpenAI to start Applied Compute.[2]

Founding and seed round (2025)

Applied Compute was founded in May 2025, according to The Information.[9] On June 27, 2025, Upstarts Media reported that the company, then not yet launched, had raised $20 million in a round led by Victor Lazarte of Benchmark that valued it at $100 million. Sequoia Capital, Conviction, Hanabi Capital, Definition and investor Zach Frankel also took part, and Lazarte joined the board.[10] Forbes later described this as a $20 million seed round.[2] In late September 2025 The Information reported that Applied Compute was in talks to raise more money at a $500 million valuation, about five times the valuation of three months earlier.[11]

Public launch (October 2025)

Applied Compute announced itself publicly on October 29, 2025, in a post titled "It's time to get specific". It said it was backed by $80 million in funding from Benchmark, Sequoia, Lux, Hanabi, Neo, Definition, Elad Gil, Victor Lazarte, Omri Casspi and others, and it named Cognition, DoorDash and Mercor as early customers.[3] The post said the company's engineers embed inside customer teams, that its training stack, agent platform and tools were all built in-house, and that two-thirds of its staff were former founders.[3] The Information later reported that the round announced in October had valued the company at about $500 million.[12]

Kleiner Perkins round (April 2026)

In January 2026 The Information reported that Applied Compute was in early talks to raise as much as $70 million at a $1.3 billion valuation including the new money, with Kleiner Perkins in talks to lead.[12] On April 8, 2026, the company announced $80 million in new financing at a $1.3 billion post-money valuation. Kleiner Perkins led the round, with participation from Elad Gil, Lux Capital, Greenoaks, Neo, Hanabi and others, and the company's total funding reached $160 million.[5] Latham & Watkins represented Applied Compute in the financing.[15] According to Forbes, Kleiner partner Leigh Marie Braswell led the firm's investment.[2] In the announcement the founders said they had worked with enterprises across the Fortune 500, and that their agents are trained on a company's own workflows and run inside the company's secure environment, with each decision fed back into training.[5]

AC2 and new funding talks (2026)

On August 10, 2026, The Information reported that Applied Compute was raising a new round at a valuation of about $3 billion, that Elad Gil was in talks to lead it, and that the company had recently been generating about $50 million in annualized revenue.[16] Dealroom's summary of the report described the financing as a Series C likely to total hundreds of millions of dollars and said it had not closed.[17] On August 25, 2026, the company launched AC2 in private beta.[4]

On September 2, 2026, Forbes reported that Applied Compute had told it that it was in talks to raise $350 million at a $3.25 billion valuation. Gil was the lead investor, with existing backers including Lux Capital and Kleiner Perkins also participating, and the company said some of the numbers could change as it finalized the round.[2] Patil told Forbes that most of the new money would go to buying chips to build out the company's cloud with a "massive fleet of compute".[2] Forbes noted that the valuation would be nearly double the one set four months earlier.[2]

Funding

Date	Round or event	Amount	Valuation	Investors	Source
June 2025 (reported)	Seed round, before launch	$20 million	$100 million (reported)	Led by Benchmark (Victor Lazarte); Sequoia, Conviction, Hanabi Capital, Definition, Zach Frankel	Upstarts Media[10]
September 2025 (reported)	Talks on new funding	Not reported	About $500 million (talks)	Not specified	The Information[11]
October 29, 2025	Public launch	$80 million raised in total to that date	About $500 million, according to The Information[12]	Benchmark, Sequoia, Lux, Hanabi, Neo, Definition, Elad Gil, Victor Lazarte, Omri Casspi and others	Company[3]
April 8, 2026	New financing	$80 million ($160 million in total)	$1.3 billion post-money	Led by Kleiner Perkins; Elad Gil, Lux, Greenoaks, Neo, Hanabi and others	Company[5]; Latham & Watkins[15]
August-September 2026	Talks, not reported as closed	$350 million, per the company	$3.25 billion, per the company; about $3 billion, per The Information	Elad Gil (lead); Lux Capital and Kleiner Perkins	The Information[16]; Forbes[2]

Business model

Forward-deployed work

Much of Applied Compute's customer work has been done through engagements in which its own staff work inside the customer's organization.[2][3] In an April 2026 post the company said it works in customers' offices, sorting through unstructured data and turning research into production systems. It described two customer-facing roles: forward deployed engineers, who build evaluation frameworks, agent environments and systems for ingesting a customer's data, and applied research engineers, who fine-tune models on proprietary data and tune agent performance. The post presents this as a newer form of the forward-deployed model that Palantir pioneered.[18] The company's website offers an "embedded" mode, in which its researchers work beside the customer's engineers from the first evaluation to deployment, and a "managed" mode, in which the customer runs its own experiments while Applied Compute handles training, serving and infrastructure.[1]

Forbes described the company as aiming to be a one-stop shop for training, deploying and running custom open-source agents, and described AC2 as a cheaper option than its high-touch consulting services.[2] Forbes presented open models as a much cheaper alternative to relying on OpenAI or Anthropic, and Patil told the magazine that in some cases the company's models can be 10 times cheaper than frontier models.[2] In a September 2026 post the company said it had seen open-weight model deployments across more than twenty enterprise applications and five model families move from early ideation to production, and argued that models of both U.S. and Chinese origin can be deployed safely in high-stakes applications if they are properly sandboxed and evaluated.[19]

AC2

AC2 is the platform Applied Compute's researchers use to build custom models for customers; the company opened it to other teams in private beta on August 25, 2026.[4] According to the launch post, AC2 supports recent open models as starting points, lets customers train inside their own agent harness, and includes a console for inspecting rollouts, comparing runs and iterating on graders.[4] A separate post explains the harness integration: AC2 replaces the model endpoint that an existing agent calls, and its trainer captures the tokens and metadata needed for gradient updates while the agent keeps running in its own environment.[20] Trained checkpoints can be served on dedicated inference capacity using the same sampling configuration, numerical precision and kernels used in training, and production traces can be fed into later training runs, including through on-policy self-distillation.[4] The company's website says it is SOC 2 and ISO 42001 certified and offers serverless or virtual private cloud deployment.[1]

In August 2026 Applied Compute added full fine-tuning of Kimi K3, which it described as having just under 3 trillion parameters. It said it had to re-architect parts of its training stack, including memory management, weight transfer and checkpointing, and that two memory-streaming optimizations cut the number of GPUs needed per training replica by about 40 percent.[21]

Ari

Ari (Applied Research Intelligence) is an in-house AI research agent that Applied Compute described on August 12, 2026. The company said it began as a shared set of Claude and Codex skills and now runs inside AC2 sandboxes, where it reads training logs and agent traces, looks for problems such as reward hacking, keeps a shared project wiki of past findings, and produces plots and reports. Customers can use it as well.[22]

Shift-1 and AC2 Routers

On October 21, 2026, Applied Compute released Shift-1, a model router for coding and knowledge work, together with AC2 Routers, a service for customers to train and serve their own routers.[23] Shift-1 chooses a model for each request rather than once per task, and it takes into account the cache cost of switching models; its routing policy is a custom decision model with 4 billion parameters.[23] The company said that on its own internal coding use, Shift-1 cut Codex spending by 55 percent compared with using GPT-6 Astra alone, by routing between Astra and GPT-6.1 Sol, and cut Claude Code spending by 24 percent compared with Claude Opus 5.5 alone, by routing between Opus 5.5 and Sonnet 5.5.[23][24] Its research write-up says the Codex configuration preserved SWE-bench Verified accuracy and user satisfaction, while the Claude Code configuration had a small drop in SWE-bench Verified accuracy and no significant drop in user satisfaction.[24] These are the company's own measurements. Applied Compute said its router research was done in collaboration with Nvidia's NeMo Switchyard team, and Shift-1 is offered through the site shift.ac.[23]

Customer projects

Applied Compute has published case studies of several customer projects. The results below are as reported by Applied Compute and its customers.

Published	Customer	Project	Reported result
February 17, 2026	DoorDash	Menu error-correction model for merchant onboarding, trained with reinforcement learning against an automated grader calibrated to DoorDash's human reviewers	In an A/B test the share of low-quality menus fell by about 30% relative to the baseline; the model was rolled out to all U.S. menu traffic[25]
February 24, 2026	Mercor	Long-horizon RL on Mercor's APEX-Agents professional tasks, first on GLM-4.6 and then on GLM-4.7	The resulting model, "Applied Compute: Small", ranked first in corporate law and fourth overall on the APEX-Agents leaderboard at publication; corporate-law Pass@1 rose from 4.7% to 26.6%[26]
May 11, 2026	Cognition AI	SWE-check, a bug-detection model trained inside a replica of the Windsurf environment using code from public repositories, served on Cerebras hardware	About 10 times faster bug detection than the frontier alternative; the model powers Quick Review in Windsurf[27]
June 22, 2026	Harvey	Full-parameter RL on GLM-5.1 for Harvey's Legal Agent Benchmark (LAB)	Rubric pass rate rose from 0.853 to 0.913, above GPT-5.5 xhigh and Opus 4.8 Max in Applied Compute's evaluation[28]
August 14, 2026	Harvey	Model for Harvey's Review Table document-analysis product, post-trained from GLM-5.2 on a synthetic dataset built from open-source legal documents	Answer score of 0.903, against 0.867 for Fable 5 and 0.857 for GPT-5.6-Sol on the project's benchmark[29]
September 4, 2026	Turbopuffer	Code-search specialist trained from Qwen3.6-35B-A3B with GRPO, searching a turbopuffer index	10-100 times lower cost and 2-10 times lower latency than most frontier models on the tested tasks[30]

In the Cognition project, the case study says Cognition had found that Claude Opus 4.6 was the only model good enough to ship in the feature but was too slow and expensive for on-demand use in the editor.[27] Niko Grupen, Harvey's head of applied research, told Forbes that building the document-review agent with Applied Compute took less than two months, work he said would previously have taken several months to years along with specialist researchers, infrastructure engineers and large amounts of compute.[2]

Applied Compute also works with Nvidia on its Nemotron open models. In an August 2026 case study the company said it deploys custom Nemotron models for enterprise customers in software engineering, financial services, logistics and AI-native companies, runs RL experiments on targeted capabilities, and shares the results with Nvidia, including which reward designs hold up and which get hacked.[31] Forbes reported that Nvidia used the startup to help post-train Nemotron models for building agents.[2] The company's website also carries endorsements from people at Handshake, Decagon, Semgrep and LatchBio, among them Handshake's chief executive and the chief technology officers of Semgrep and LatchBio.[1]

Research

Applied Compute publishes technical write-ups on its website. Examples include:

Date	Title	Authors	Subject
March 24, 2026	Speeding up RL with high-leverage samples	Agastya Goel, Linden Li	Defines "sample leverage", a measure of how much training signal a rollout carries under binary rewards, and a thresholding method that trains only on high-leverage rollouts to save compute[32]
April 22, 2026	Benchmarking Inference Engines on Agentic Workloads	Oam Patel, Linden Li	Releases three multi-turn workload profiles drawn from production post-training runs (agentic coding, code QA and office work) and an open-source harness for replaying them; compares vLLM and SGLang serving DeepSeek-R1[33]
June 4, 2026	Training an Agentic Router for Optimal Cost-Performance on SWE Tasks	Nic Becker, Rhythm Garg	Trains Qwen3.6-35B-A3B to route SWE-bench Verified tasks among Nemotron 3 Ultra, GPT-5.5 and Claude Opus 4.7[34]
June 16, 2026	Continued Training with Entropy Preserving RL	Nic Becker, Jacob Philips, Rhythm Garg	Examines entropy collapse during continued RL training and tests an entropy-preserving controller against GRPO[35]
July 3, 2026	Predicting and Controlling Staleness in Fully Asynchronous RL Training	Chengyu Dong, Agastya Goel, Dylan Yu	Derives closed-form formulas for the mean staleness of rollouts in fully asynchronous RL and its trade-off with system performance[36]
August 28, 2026	Frontier-scale RL with Kimi K3	Raymond Feng, Lucca Bertoncini	Memory streaming, low-precision (MXFP4) rollouts, weight transfer and checkpointing changes for full fine-tuning of Kimi K3[21]
October 26, 2026	Shift-1: Research Overview	Alex Noviello, Sam Denton	Evaluation and training of the Shift-1 router[24]

Partnerships

Nvidia. Nvidia's August 2026 announcement of Nemotron 3.5 Lightning listed Applied Compute among its post-training partners, alongside companies such as Thinking Machines Lab and Prime Intellect.[37]
Reflection AI. When Reflection AI released the open weights of its Beam model in October 2026, it listed Applied Compute among the platforms offering the model at launch.[38]
Modal. Patil was quoted in Modal's May 2026 Series C announcement, saying that sandboxes "are one of the most important building blocks for Reinforcement Learning" and that Modal was structured in a way that let the company build complex environments.[39]

Leadership

Name	Role	Background
Yash Patil	Co-founder and chief executive officer	Studied computer science at Stanford and left without a degree; joined OpenAI at 20 and worked on post-training infrastructure and Codex[2][6]
Rhythm Garg	Co-founder and chief technology officer	Stanford; reasoning research at OpenAI, listed as a core contributor to o1[2][7][13]
Linden Li	Co-founder; listed as chief architect by AI Engineer	Stanford; machine learning systems at OpenAI, listed as an o1 contributor; earlier at Databricks Mosaic AI and Nvidia, according to his X profile[2][7][8][14]

Forbes gave the founders' ages in September 2026 as 23 (Patil), 24 (Garg) and 25 (Li).[2]

Reception

In June 2026 Patil interviewed Microsoft chief executive Satya Nadella about enterprise AI and model ownership. Business Insider reported that in the interview, published on June 26, Nadella said every company should build AI models suited to its own business, adding that he did not want to be locked into one model.[40] Applied Compute features his remark "There should be as many models in the world as there are firms" on its homepage.[1] The same month Patil appeared on The Generalist podcast to explain why he thinks companies need models of their own; the episode is headlined with his remark "No one is data ready. Absolutely no one is data ready."[6]

Elad Gil, the lead investor in the 2026 talks, told Forbes, "I always believed open source would play an important role in things," and said that because the founders were "the former OpenAI research team that worked on these types of problems, they're really at the cutting edge."[2]

Debarghya Das's "Ultimate Neolab List", a list of recently founded AI labs, includes Applied Compute but rates it "Borderline". The list's note credits the company's former OpenAI researchers and heavy compute use, but says that its roughly $50 million in annualized revenue and its services model fail the list's low-revenue test.[41]

See also

References

Applied Compute. "The Best AI is Built, Not Bought." Company website, accessed October 27, 2026.
Richard Nieva, Forbes. "Meet The Ex-OpenAI Guys Building A Cheaper Open Source Alternative." September 2, 2026.
Yash Patil, Rhythm Garg and Linden Li, Applied Compute. "It's time to get specific." October 29, 2025.
Applied Compute. "Applied Compute Agent Cloud." August 25, 2026.
Yash Patil, Rhythm Garg and Linden Li, Applied Compute. "The Advantage You Own: Applied Compute Raises $80M to Help Enterprises Advance from Generalized to Specific Intelligence." April 8, 2026.
Mario Gabriele, The Generalist. "Own or Be Owned: Why Every Company Needs Its Own AI Model (Yash Patil, Co-Founder & CEO of Applied Compute)." June 23, 2026.
OpenAI. "OpenAI o1 Contributions." Accessed October 27, 2026.
AI Engineer. "Applied Compute." Organization page, accessed October 27, 2026.
TechStartups. "Ex-OpenAI researchers' startup Applied Compute targets $500M valuation in new funding round." September 26, 2025 (quoting The Information).
Alex Konrad, Upstarts Media. "Scoop: 3 Ex-OpenAI Staffers Raise $20M For New Startup Applied Compute." June 27, 2025.
The Information (@theinformation). "A startup founded by three former OpenAI staffers, Applied Compute, is in talks to raise new funding at a $500 million valuation." X, September 27, 2025.
TechStartups. "Applied Compute, AI startup founded by ex-OpenAI researchers, in talks to raise funds at $1.3B valuation just months after $500M round." January 20, 2026 (quoting The Information).
Rhythm Garg (@rhythmrg). "Profile." X, accessed October 27, 2026.
Linden Li (@lindensli). "Profile." X, accessed October 27, 2026.
Latham & Watkins. "Latham & Watkins Represents Applied Compute in US$80 Million Fundraise." April 15, 2026.
Stephanie Palazzolo (@steph_palazzolo), The Information. "New w/ @julia_hornstein: Applied Compute is raising a new round at a ~$3B valuation." X, August 10, 2026. Links to "Applied Compute in Talks to Double Valuation to $3 Billion on Open-Source Demand."
Dealroom News. "Applied Compute in talks to double valuation to $3B on open-source demand." August 12, 2026.
Bryan Lee, Applied Compute. "Unlocking the AI Overhang." April 10, 2026.
Dan DiPietro and Michael Chen, Applied Compute. "Securely Deploying Open-Weight Models." September 24, 2026.
Vinjai Vale, Applied Compute. "Bring Your Own Harness to AC2." August 6, 2026.
Raymond Feng and Lucca Bertoncini, Applied Compute. "Frontier-scale RL with Kimi K3." August 28, 2026.
Aniruddh Sriram, Applied Compute. "Ari, Applied Compute's in-house AI research agent." August 12, 2026.
Applied Compute. "Shift-1, a SOTA coding router." October 21, 2026.
Alex Noviello and Sam Denton, Applied Compute. "Shift-1: Research Overview." October 26, 2026.
Applied Compute. "Automating Merchant Onboarding at DoorDash." February 17, 2026.
Applied Compute. "Building State-of-the-Art Agents with Mercor." February 24, 2026.
Applied Compute. "Unlocking Real-Time Bug Detection at Cognition." May 11, 2026.
Applied Compute. "Training a State-of-the-Art Legal Agent with Harvey." June 22, 2026.
Niko Grupen et al. (Harvey and Applied Compute). "Training Frontier Review Table Agents with Harvey." August 14, 2026.
Applied Compute. "Training a Specialist Code Search Agent with turbopuffer." September 4, 2026.
Applied Compute. "Post-Training Collaboration with NVIDIA." August 5, 2026.
Agastya Goel and Linden Li, Applied Compute. "Speeding up RL with high-leverage samples." March 24, 2026.
Oam Patel and Linden Li, Applied Compute. "Benchmarking Inference Engines on Agentic Workloads." April 22, 2026.
Nic Becker and Rhythm Garg, Applied Compute. "Training an Agentic Router for Optimal Cost-Performance on SWE Tasks." June 4, 2026.
Nic Becker, Jacob Philips and Rhythm Garg, Applied Compute. "Continued Training with Entropy Preserving RL." June 16, 2026.
Chengyu Dong, Agastya Goel and Dylan Yu, Applied Compute. "Predicting and Controlling Staleness in Fully Asynchronous RL Training." July 3, 2026.
Chris Alexiuk and Chintan Patel, NVIDIA Technical Blog. "NVIDIA Nemotron 3.5 Lightning Delivers Fast, Accurate Specialized Task Execution for Long-Running Agents." August 11, 2026.
Reflection. "Beam is now available." October 16, 2026.
Modal. "Modal's Series C: Raising $355M at a $4.65B valuation." May 21, 2026.
Katherine Li, Business Insider Japan. "マイクロソフトのサティア・ナデラ、「すべての企業が独自のAIモデルを構築すべき」だと語る" Japanese translation of a Business Insider article. July 1, 2026.
Debarghya Das. "The Ultimate Neolab List." Accessed October 27, 2026.

Editing: Applied Compute