Reliable AI for the enterprise, from the lab that built Jamba
HQ: Tel Aviv, Israel | Founded: 2017 | Employees: ~200 | Stage: Series D — $300M from Google and NVIDIA (May 2025); $636M raised total | Website: https://www.ai21.com
AI21 Labs builds AI systems for enterprises that need results they can trust. The company emerged from Tel Aviv's deep-tech scene in 2017, founded by Amnon Shashua — who also founded and runs Mobileye — alongside Stanford professor and AI textbook author Yoav Shoham and Ori Goshen. Its early Jurassic models were among the first serious alternatives to OpenAI's, and its Wordtune writing assistant showed early product instincts.
The company's lasting technical contribution is Jamba: a hybrid architecture combining state-space models (like Mamba) with traditional attention, published as open weights for private deployment. Jamba handles very long contexts efficiently — the exact trade-off enterprises face when they want AI over large document sets without hyperscaler costs. March 2025's Jamba 1.6 targeted private enterprise deployment specifically.
On top of the models sits Maestro, AI21's model-agnostic orchestration platform for agentic AI. Maestro's pitch is reliability: getting agentic systems to deliver accurate, controllable results in production, across whatever models a customer chooses. That 'trust layer' positioning — combined with deep research pedigree — persuaded Google and NVIDIA to invest $300M in May 2025, bringing total funding to $636M.
AI21 Labs has been doing generative AI since before it had a name. Founded in 2017 — the same year as the Transformer paper — by Amnon Shashua (the Mobileye founder), Stanford professor Yoav Shoham and Ori Goshen, it shipped the Jurassic models years before ChatGPT and survived the frontier-model land grab that buried most of its 2020-era peers.
Survival came from pivoting. Rather than chase OpenAI head-on, AI21 found its wedge in two technical contributions: Jamba, a hybrid architecture mixing state-space models with attention for long context at lower cost, released openly for enterprise deployment; and Maestro, an orchestration platform that makes agentic AI reliable — model-agnostic, with the accuracy and control enterprises need before they let agents near real work. The positioning is deliberate: the trust layer above whatever models you run.
The strategic validation is unusual for the category: both Google and NVIDIA invested — $300M in May 2025, taking total funding to $636M — even as their own teams compete in models. That reflects AI21's role as a neutral enterprise supplier with genuinely deep research credentials (its founders literally write AI textbooks).
The hard truth remains scale. Frontier training runs now cost more than AI21 has raised in total, and the enterprise AI market is crowded with Microsoft, Salesforce and every model vendor selling 'reliable agents.' AI21's bet is that reliability is a distinct discipline from capability — that enterprises will pay for systems that fail gracefully and stay auditable, and that Jamba-style efficient architectures keep the cost profile sane. With ~200 people and $636M behind it, that bet has runway.
Browse: Home | Blog | About | All Startups A-Z | View full profile on StartupWiki