Open source is going to win it all: Harvey proves it | E2328

Harvey's release of its first proprietary model, Harvey Tenant, marks a significant strategic shift in legal AI. Built on top of Kimi K3 and developed with Fireworks Research, the post-trained open weight model is designed to handle long-horizon legal work without relying on OpenAI or Anthropic. Jason Calacanis estimated Harvey was spending roughly 10 million dollars per month with OpenAI before this move, with 99 percent of that spend on frontier models. The departure is partly competitive: OpenAI led Harvey's original five million dollar seed round but now wants to build the final product in the legal AI space itself, creating a direct conflict of interest. A second conflict drives Harvey's architecture further, as competing law firms sharing a single model creates a data trust problem, which Harvey is solving by building separate isolated model instances for each client. Harvey raised 200 million dollars at an 11 billion dollar valuation in March 2026, co-led by GIC and Sequoia Capital.

Calacanis argued that frontier model providers study token usage patterns from application companies and will eventually replicate the most successful use cases as free platform features, making it strategically dangerous for companies to keep feeding data to OpenAI or Anthropic. He predicted that by 2027 the dominant trend will be startups moving off frontier models to run open source models and retain ownership of their own training data, a call he said he made in 2023. His thesis is that language models hit parity quickly, become commodities, and that real value accumulates with whoever controls proprietary training data. He compared the coming commoditization to how WordPress drove margins out of web publishing. He estimated companies like Lovable, ElevenLabs, and Cursor were each spending between 10 and 30 million dollars per month with frontier providers, and predicted that as those companies build their own models, the revenue loss will represent a meaningful headwind for OpenAI and Anthropic heading into their IPOs. One speaker noted that approximately 99 percent of enterprise AI spend is reportedly shifting toward open source or internal deployments, with the remaining roughly 5 percent used for distillation from frontier models at an estimated cost of around one million dollars per month.

OpenAI confidentially filed its IPO prospectus in June and raised 122 billion dollars in cash in March. Calacanis identified churn rates and true cost to serve as the critical unknowns for AI companies going public, noting that churn data will not be disclosed in the filing and that AI companies will likely avoid being required to report it the way mature public companies must. He drew a parallel to Uber and Lyft price wars, during which Uber had one billion rides in a quarter while losing roughly two dollars per ride on rides priced well below their true cost, with venture capital funding those subsidies as a substitute for advertising. He argued AI companies are likely running a similar playbook and will present a reasonable profitability story ahead of their IPOs while a real gap between usage and spending persists underneath. A counterargument was raised that AI differs from ride-share because users can switch freely between Claude, Grok, and other models without meaningful lock-in, making the habituation strategy harder to execute.

On the question of which AI tools are most useful in practice, Grokbot was described as more reliable than Claude for tasks requiring access to X and LinkedIn, less brittle, and capable of running continuously in the cloud without requiring an app to remain open. After a single login, Grokbot can cross-reference news with X posts automatically and send scheduled reports to Slack or integrate with Google Sheets. Calacanis said he built recurring Grokbot agent systems for newsletter competitive intelligence, speaker sourcing, and co-host candidate research, estimating that AI agents could save five to ten hours of research per week, with guest booking alone requiring roughly three hours of effort per guest before an appearance.

Alan Guo, founder of Willow Voice, described his app as a free AI-powered voice dictation tool offering under 250 milliseconds latency and claiming superior performance over competitors in formatting, correction, speed, and accuracy. His paid product, Scribe, functions as a writing assistant generating full responses using stored context and knowledge bases, deployed across airlines, customer support, and sales teams. Guo's strategic decision was to commoditize dictation first by making the core product free rather than waiting to be undercut by larger platforms. He predicts Apple will reach 70 to 80 percent of dedicated dictation tool quality within two to three years but will not close the gap fully because Willow offers fine-tuned models, corrections, and personalization that platform tools will not prioritize.

A first-time hardware builder named Ethan built a self-driving golf cart retrofit completed in two to three weeks, using three front cameras, three back cameras, an Nvidia Jetson for onboard compute, and a Starlink Mini for cloud inference and live code deployment while in motion. Golf carts cost ten to twenty thousand dollars, and Ethan estimated a retrofit autonomous kit could be sold for under a couple thousand dollars. The project has drawn interest from the golf industry and from Disney. Calacanis revised an earlier prediction that 20 companies would figure out self-driving between 2026 and 2028, now suggesting it could be 2,000 people solving the problem within a 36-month window after seeing the project.