Applied Compute raises $350M to build custom AI models
Applied Compute raised $350M at $3.25B after hitting a $100M run rate building custom models on open weights. Harvey and Clay cut inference costs. Price yours.
Applied Compute, a custom AI model startup, raised $350 million at a $3.25 billion valuation, led by Elad Gil, the company announced on November 2. Nvidia, Blackstone, Kleiner Perkins, Shopify Ventures and HubSpot Ventures are among the investors. The company says it reached a $100 million annualized revenue run rate less than 18 months after founding. Its thesis is simple: companies want to own and improve the models behind their products, built on open weights, instead of renting a frontier model per token.
What actually happened
From the company's post and SiliconANGLE:
- Customers named include Microsoft, Shopify, DoorDash, Harvey, Cognition, Clay and Decagon.
- Harvey got a model for its Review Table product with better answers and citations at "less than half the inference cost of a frontier alternative," per Applied Compute.
- Clay used a custom router that cut production inference costs "by up to 47%."
- Its AC2 platform combines training, inference and continual improvement. The company says customers have run more than 1,700 training runs since July and its inference serves more than 1 trillion tokens a day.
- SiliconANGLE describes reinforcement learning tooling and Shift-1, a router that sends each coding-agent request to the cheapest model that can handle it.
All of these numbers are company-reported. The cost claims are case studies, not audits.
Why it matters for your business
You will not train a custom model this quarter. You can steal the two ideas behind the $350 million.
Routing first. Clay's savings came from a router, not a new model. Most small-business AI stacks send every call to the most expensive model because that is the default in the SDK. Classification, extraction and summarizing product data rarely need a frontier model. A routing rule by task type is an afternoon of work.
Your data is the asset. Applied Compute's whole pitch is that the customer's data and expertise make the model. Start saving it now: prompts, outputs, and the corrections your team made. That log becomes your eval set today and your fine-tuning set later, on whichever vendor wins.
Key takeaways
- Applied Compute raised $350M at a $3.25B valuation, led by Elad Gil
- It reports a $100M run rate in under 18 months, building custom models on open weights
- Case studies: Harvey at under half the frontier inference cost; Clay cut costs up to 47% with a router
- Route cheap tasks to cheap models before you think about training
- Log prompts, outputs and corrections now; they are your future eval and training data
Find out what your AI calls actually cost. Run your volume through our ROI calculator, then talk to us about routing and logging that you own.
Sources: Applied Compute announcement, SiliconANGLE.
- #custom-models
- #open-weights
- #inference-costs
- #model-routing
- #funding
Tommy Rush — Founder, Rush Commerce
Operator turned builder. 15+ years running operations — now shipping the systems businesses run on. More
Get The Rush Report weekly — one email, zero fluff.
Keep reading
Tencent Hy4 ships with cyber-offense filtered out
Tencent released Hy4 with exploit, reverse-engineering and pentest content filtered from training data. Weights and a safety report come later. Test the capability you depend on.
Read itSoniox STT v6: voice agents stop hearing themselves
Soniox STT v6 adds agent text input and agent voice echo cancellation for voice agents. v5 is removed December 15 and prices now come back as words. Retest first.
Read it