# Applied Compute raises $350 million for custom AI infrastructure

> The startup reports a $100 million annualized revenue run rate and builds models and routing systems for business workflows.

By BIGCHANGE.AI Editorial

Published: 2026-11-04T04:16:38.192Z
Updated: 2026-11-04T04:16:38.192Z
Canonical: https://bigchange.ai/blog/applied-compute-funding-custom-model-routing

![Conceptual illustration of a generic server cabinet in a data-centre aisle, with cables routed along overhead trays.](https://bigchange.ai/api/media/file/cuda-shared-compute-hero-v1.png)
AI-generated conceptual illustration by BIG CHANGE.

Applied Compute says it has raised $350 million at a $3.25 billion valuation as it sells companies infrastructure to train and run specialized AI models. In its [November 2 announcement](https://www.appliedcompute.com/company/100m-run-rate-fundraise), the startup reported a $100 million annualized revenue run rate, a pace rather than revenue already earned over a year.

## The big change

- Applied Compute plans to expand compute capacity and enterprise deployments with the new funding.
- Its cloud combines model training, inference and ongoing improvement, the company says.
- Two company-reported cases show workload-specific routing and training can change cost and quality, within their tests.

## What the funding supports

Applied Compute says the round was led by Elad Gil, with NVIDIA and Blackstone among participants. It plans to expand compute capacity, grow its team and accelerate enterprise deployments.

## Choosing a model for each request

In its [October Clay case study](https://www.appliedcompute.com/case-studies/clay), Applied Compute describes an automated router for Claygent, a research agent used across lists of companies and people. The team generated representative synthetic tasks and scored candidate-model answers with an AI judge. The trained router receives a query and candidate configurations, predicts a quality score for each, and selects a model using a cost-quality trade-off. That lets the team choose where a more expensive model is worth using rather than applying one model to every request.

Applied Compute reports a 47% cost reduction in a live A/B test against routing all requests to GPT-5.6 Sol, while maintaining its primary guardrail metrics. The result concerns this workload, not every customer or query.

## Training for document review

A separate [joint Harvey case study](https://www.appliedcompute.com/case-studies/harvey-review-table) describes lawyers posing questions across document collections, with each table cell requiring an answer and citations. The teams say they built synthetic training tasks from public legal material and evaluated both response correctness and supporting citations. Their reward design treats incorrect answers as failures even if they carry citations, while rewarding relevant evidence.

The authors report improved answer and citation quality versus the base GLM 5.2 checkpoint, evaluating models at minimal reasoning effort. Average cost per cell was 54.8% lower than Claude Sonnet 5 on the same tasks. They separately compared iterative search with fixed retrieval. These company-reported results are task-specific, not an independent guarantee of legal accuracy.

Both results are task-specific.

## Sources

- [Applied Compute funding announcement](https://www.appliedcompute.com/company/100m-run-rate-fundraise) — Company funding and run-rate claims.
- [Clay router case study](https://www.appliedcompute.com/case-studies/clay) — Company account of a customer routing system.
- [Harvey Review Table case study](https://www.appliedcompute.com/case-studies/harvey-review-table) — Joint report on legal review models.