On November 2, 2026, Applied Compute announced $350M in new funding at a $3.25B valuation, led by Elad Gil. Participants included Kleiner Perkins, Lux Capital, NVIDIA, Blackstone, Shopify Ventures and HubSpot Ventures, and NVIDIA and Shopify also appear among its customers. The company plans to spend the money on compute capacity, hiring and enterprise deployments.

That round followed an $80M financing announced April 8, 2026 at a $1.3B post-money valuation, led by Kleiner Perkins, which brought total funding to $160M. Benchmark and Sequoia were among its backers when the company launched publicly in October 2025.

Yash Patil (CEO), Rhythm Garg (CTO) and Linden Li founded Applied Compute in 2025. The founders previously worked at OpenAI on Codex, reasoning models and the systems used to train models with reinforcement learning. They left OpenAI betting that open-weight models, whose parameters anyone can download, would become strong enough for companies to build on with their own data. The team has tripled since January 2026, and 85% of employees are researchers and engineers.

Applied Compute helps companies take an open-weight AI model, train it on their own work and run it in production as a model they own. Its buyers are AI and product teams at companies like Harvey, Clay and DoorDash. These teams make huge volumes of model calls and want better accuracy or lower cost than a general-purpose frontier model delivers.

A project starts with three inputs. The first is the customer's data, such as past documents or records of agent behavior. The second is evaluation criteria, often a program called a grader that scores each answer. The third is the customer's existing agent harness, the code that wraps a model with prompts, tools and access to internal systems. Applied Compute trains the model inside that harness and saves checkpoints, which are snapshots of the model during training. It then deploys the best checkpoint as a live endpoint and feeds production traffic back into the next training run.

As a hypothetical illustration, an insurer could supply past claims, its adjusters' final decisions and a grader that checks whether an agent's approval recommendation and cited policy clause match those decisions. This is not a named Applied Compute customer.

AC2, the Applied Compute Agent Cloud, launched in private beta on August 25, 2026. Before that, it was the platform the company's own researchers used for customer work. Its training offering has two paths. In the Managed path, a customer's ML team runs training, environments and graders itself through a command-line tool and Python SDK. In the Embedded path, Applied Compute's forward-deployed researchers write custom evals based on the customer's product metrics and hand back a model whose weights the customer keeps.

AC2 gives the harness a standard OpenAI-style model address. This lets developers reuse the code surrounding their production model while AC2 captures the model interactions needed for training and a grader scores the outcomes. A research agent called Ari watches training runs, reads outputs for failure modes, fixes configurations and relaunches jobs.

Harvey's Review Table shows what this produces. In that product, a lawyer uploads up to 10,000 files and gets back a grid of cited answers to their questions. Applied Compute said the model it trained improved answer and citation quality at less than half the inference cost of a frontier alternative.

AC Inference serves models as live endpoints. It can deploy a finished checkpoint or a Hugging Face model in minutes, by default on dedicated reserved Blackwell GPUs. It offers autoscaling, a 99.9% uptime target, single-tenant and region-locked deployments, zero data retention and the option to run in the customer's own cloud. New checkpoints swap in behind the same endpoint.

Routing applies the same loop to choosing which model handles each request. Clay's Claygent research agent previously ran every request through one user-selected model. Applied Compute post-trained an open-weight Qwen model to pick a model for each query and serves it on AC2 at hundreds of millions of queries a month. Applied Compute reported that in a live A/B test, the router cut costs 47% relative to a fixed baseline of sending every request to the frontier model Sol, while holding steady on primary quality guardrails.

Applied Compute sells research services and compute. Customers pay for researchers who build evals and training recipes, and for the GPUs that run training and serve models. Embedded work bundles dedicated capacity under service-level agreements. Dedicated training and inference capacity are sold through enterprise engagements rather than a published standard price list.

Its direct costs mirror that split: GPU capacity and researcher pay. If reserved GPUs stay busy and the same training stack, Ari automation and serving software are reused across customers, each new workload spreads those costs further. Idle capacity or bespoke research on every account would push the other way.

The economic trade-off is how much a customer can save by switching from a general-purpose model while still leaving Applied Compute a profitable price for its research and serving infrastructure. Clay's lower inference bill demonstrates customer savings; it does not establish Applied Compute's gross margin.

The main account-expansion engine is the move from training to serving. A trained checkpoint becomes an inference workload whose traces feed the next run, turning a one-time project into recurring compute spend. Hosting in a customer's own cloud may win security-sensitive deals, but it could leave less hardware revenue with Applied Compute.

Applied Compute's most direct competitors are Together AI and Fireworks AI, independent clouds that both fine-tune open-weight models and host production inference. The central contest is whether enterprises buy customization as a self-serve utility priced per token or GPU-minute, or as a research-led engagement bundled with dedicated capacity.

Together runs the full fine-tuning lifecycle, from data upload to hosting on a dedicated endpoint. It uses LoRA adapters (small add-on weights) by default, with full fine-tuning available. A separate post-training product runs supervised, DPO and GRPO reinforcement learning on GPUs customers provision, billed per GPU-minute. Fine-tuning is billed per token processed, hosting is billed by the minute, and jobs launch from a console, API or command-line tool. Against that broad self-serve catalog, Applied Compute differentiates on setup rather than unique training methods: it trains inside a customer's existing harness, with researchers writing the evals.

Fireworks offers serverless training on shared GPUs, charged per token, and dedicated capacity for larger or full-parameter training jobs. Its training API supports reinforcement learning, preference training and distillation, giving developers control over the training loop rather than restricting customization to a single managed recipe.

This makes Fireworks a direct alternative for teams capable of designing their own experiments. Applied Compute's embedded researchers can make it easier to translate an application problem into a useful training objective, but it must show that the additional engagement produces better results than a developer can achieve on another training and inference platform.

Applied Compute began by helping companies build specialized models and is extending the infrastructure behind that work into a broader platform. Its opportunity is to reach buyers who want to operate model development themselves or purchase production infrastructure without first commissioning a custom research project.

The Managed training path gives established machine-learning teams a way to use AC2 without handing every experiment to Applied Compute's researchers. Those buyers can bring their own agent code, data and scoring rules while buying the infrastructure that schedules training, records experiments and deploys checkpoints.

This could move Applied Compute into research-platform and computing budgets at companies that already employ AI researchers, expanding beyond buyers seeking an outsourced custom model. Its advantage is a platform developed through its own customer engagements; reaching this market would require onboarding, documentation and reliability that let customer teams operate independently, especially while AC2 remains in private beta.

Clay's production router provides an entry point into inference budgets even where a customer does not need a new model for every underlying task. A specialized router can direct easy requests to cheaper models and reserve more expensive models for harder requests, letting Applied Compute sell the system that manages a company's model spending.

The incremental buyer is an application or infrastructure team responsible for a large production inference bill, rather than a research team commissioning model training. Applied Compute could earn recurring serving revenue by demonstrating savings against that team's existing deployment while maintaining its quality and latency requirements, although it would compete directly with other inference providers and routing products.

Applied Compute says its customers already span financial services, legal work, life sciences and other industries, while its infrastructure supports single-tenant, region-restricted and customer-cloud deployments. Those capabilities could help it reach internal enterprise teams whose proprietary records cannot readily be moved into a shared consumer AI service.

The commercial opportunity is to sell model development and deployment into budgets for document review, operational decision support and other specialized work beyond AI-native software products. Expanding here would require proving each model against the customer's own standards, integrating with its systems and satisfying security and procurement requirements; cloud isolation alone does not establish regulatory compliance or permission to automate consequential decisions.

Specialization must keep earning its cost: Applied Compute's customer proposition depends on custom models delivering better quality or lower costs than general-purpose alternatives on specific tasks. Faster improvements or price cuts from frontier providers could make an expensive training engagement less attractive or force repeated work to preserve the advantage.

Compute costs and capacity utilization: Production inference and reserved GPU capacity introduce substantial operating costs alongside the research work needed to win deployments. Growing token volume can increase revenue without improving gross profit if capacity sits idle, serving prices fall or workloads require more computing than expected.

Research-intensive enterprise delivery: Applied Compute said 85% of its team were researchers and engineers after headcount tripled since January 2026. If each new customer continues to require extensive custom research, evaluation and integration work, hiring and delivery demands could constrain growth despite the reuse of AC2's infrastructure.