Applied Compute Raises $350M for Enterprise AI Models

A useful answer in Harvey's legal-document software sometimes begins with teaching a model when to abstain. Applied Compute and Harvey built that judgment into training for Review Table, where correctness, formatting and citations all influence the reward a model receives. The exercise gives enterprise AI a rather demanding customer: a team that needs to explain why an answer deserves to be used.

Applied Compute is financing a larger role in that work. The San Francisco company announced $350M in new funding at a $3.25B valuation on November 2, 2026, led by Elad Gil. It also reported a $100M annualized revenue run rate less than 18 months after founding, as its business expands across training and running specialized models.

The financing supports compute capacity, hiring and enterprise deployments. For technology buyers, the interesting development is the supplier's expanding assignment: help define a model around a company's work, operate it, and keep improving it as that work changes. Those responsibilities can continue long after the first successful training run.

The founders are bringing research into a continuing service

Applied Compute's co-founders are CEO Yash Patil, CTO Rhythm Garg and Chief Architect Linden Li, roles identified by Comcast NBCUniversal LIFT Labs. The company says their OpenAI experience covered Codex, the o1 reasoning model and reinforcement-learning infrastructure, respectively. That combination matters to the commercial proposition because an enterprise deployment asks several research and engineering disciplines to work together.

In its October 2025 introduction, Applied Compute described engineers embedding with customer teams and disclosed $80M in backing. Kleiner Perkins subsequently described its Series B investment in May 2026. The latest company announcement leaves the new round's letter and valuation basis unstated; its dollar figures also do not explicitly name a currency denomination.

The November financing includes Kleiner Perkins, Lux Capital, Hanabi Capital, NVIDIA, Blackstone, Shopify Ventures and HubSpot Ventures among the participants. Applied Compute reports a revenue pace rather than a completed year's revenue, and the announcement supplies no profitability measure. The useful reading of those disclosures is a rapidly expanding infrastructure business with substantial resources to deploy, while its financial economics remain largely private.

Harvey makes quality an engineering assignment

Harvey's Review Table lets legal teams analyze collections of documents through a grid of questions and answers. Applied Compute's joint case study with Harvey explains why a single table can generate many model calls, each carrying its own requirements for extraction, reasoning and evidence. At that volume, small changes in the cost and quality of a cell can accumulate across the product.

The teams built a synthetic training dataset from public legal material, with expert quality checks. The case study explicitly says customer data is not used for training, an important boundary in a product serving legal work. The resulting model was trained around the production workflow, including the need to return an appropriate answer or abstain and to cite supporting text.

This is where a customer's expertise becomes a purchasing specification. A legal team cares about an answer's evidentiary footing; an infrastructure supplier needs a way to make that preference measurable during training. The published results are specific to the evaluated workload, but the method explains what a specialized-model engagement actually buys: sustained engineering around the standards a product must meet when people depend on it.

Clay puts model spending inside the query

Clay offers a different example of the same operational depth. Its Claygent router case study describes moving from manually chosen models to a trained routing policy that evaluates which model should handle a particular request. Contact lookups and difficult company research place different demands on the system, even when users encounter them through the same product.

Applied Compute and Clay report inference-cost reductions of up to 47% while preserving quality. Their technical account describes a cost-quality trade-off constrained by guardrails for individual query classes. Those constraints matter because an attractive average can conceal weak performance on a category of requests that a customer considers important.

The commercial insight sits inside that design choice. A buyer can specify how spending should relate to useful output, then ask the infrastructure to maintain that relationship as traffic changes. Savings figures will vary across workloads; the more transferable question is who owns the evaluation criteria, who inspects the failures, and who keeps the routing policy aligned with the product people are actually using.

AC2 gives the relationship somewhere to continue

Applied Compute introduced AC2, its Agent Cloud, in August 2026 to combine training, serving and improvement. Customers can bring their own harnesses and graders, inspect training behavior and deploy checkpoints. The company's inference offering extends that work into production traces and subsequent model updates, giving the training process information from the environment where the model earns its place.

Putting those tasks together changes the scope of the supplier relationship. The customer still needs to decide which errors matter, what successful work looks like and when a revision is ready. Applied Compute can provide researchers and infrastructure around that judgment, then serve the model once it is deployed. This is an interpretation of the product architecture, rather than a claim that customers must buy every part of it.

The company's careers page describes about 70 people in San Francisco, with researchers and engineers making up 85% of the team. Its new financing is intended to expand the people and compute available for enterprise work. As another Harvey table fills with answers or another Clay research list runs, the work keeps producing decisions about quality, cost and the next model revision that those teams will have to make together.