Here are some of the lessons I learned from Sabaio.com and Legalcomplex.com.
What Is Sabaio? In two words: Private Intelligence. Sabaio is an AI platform that leverages open-source software such as Ollama and vLLM, running in Docker or Kubernetes and compatible with Windows, Linux, and macOS. It is a flexible framework that allows enterprises to deliver basic AI capabilities like proofreading, summarizing, and building RAG workflows. Sabaio also supported advanced legal tasks, such as generating synthetic legal data, legal reasoning, contract workflows, and compliance research. Sabaio was deployed at a 400-person European law firm with single sign-on.
What Made Sabaio Special? Sabaio runs completely on-premise and does not require internet access. It processes all data locally and does not track or share any user data, making it ideal for governments and other sensitive sectors such as healthcare, banking, universities, and legal.
When the DeepSeek app rocketed to the No.1 spot in the US iPhone App Store, Americans were quick to push for legislation banning its cloud version. Using cloud AI is not necessarily unsafe, but the key question is: Which country is hosting your cloud AI?
One would assume a Western country is safer. Yet, the UK has issued a secret order requiring Apple to install backdoors in iPhones and iCloud – effectively giving the government access to all iPhone users’ data globally, (as reported by The Washington Post). Essentially, any government can compel a cloud provider to provide access to all the data it processes. AI is the perfect choke point to tap into human activity.
Why Did Sabaio Die? Despite being open source and solving an obvious need, the economics didn’t work out. Running AI locally requires specific expertise, sufficient electricity, and expensive computing power. At a minimum, Sabaio required a single NVIDIA H100 GPU to serve 100 users with a 70B parameter large language model. Scaling hardware to accommodate more users required specialized infrastructure, known only to a small group of experts. Beyond talent and hardware, there are also cooling and energy costs.
All this requires significant capital. Running on-prem AI means recreating the same data center infrastructure that Microsoft, Amazon, Meta, and Google are frantically building. These four companies alone are projected to invest $320 billion in AI in 2025, as reported by The Wall Street Journal. Imagine how much a single enterprise must allocate to support its employees.
Why Not Raise Venture Capital? In my humble opinion, that would just add fuel to the fire. The customer acquisition and onboarding costs for Sabaio would force me onto a fundraising treadmill with no clear exit in sight. Additionally, the slow onboarding process of on-prem AI doesn’t align with the ‘grow fast’ venture capital mentality.
VCs are also more comfortable funding Software-as-a-Service (SaaS) models that charge licensing fees. An open-source business model, which doesn’t rely on licensing fees, instead requires a large user base to convert. The private intelligence market is relatively small and firmly locked into the Microsoft ecosystem.
Even though the premise of distributed, decentralized, open-source AI is perfect, there is no sustainable financial model for it. While AI costs will decrease as adoption increases, future cheap AI doesn’t help offset today’s expensive infrastructure costs. Everyone will eventually need to recover their initial investments – but when will that point arrive?
What’s Next for Raymond? I will wind down Sabaio and Legalcomplex. The last Legalcomplex analysis will be the Q1 2025 report.
This 7-year rollercoaster has been an incredible journey. I’ve had the opportunity to travel, meet amazing people who loved Legalcomplex, and develop skills in financial analysis, identity access management, LLM fine-tuning, and LLM orchestration.
I’m a relentlessly resourceful, creative introvert, and I’m honestly excited to start a new chapter in my life and embrace a new challenge.
Originally posted by Artificial Lawyer on Feb 11, 2025
Added links and minor edits