The Economics of Open Intelligence
Generative AI is spreading unevenly across the economy. Software developers have seen their work change thanks to AI coding assistants, but most companies report little gain from early investments. David Rotman and Richard Waters lay this out in a recent MIT Technology Review piece, calling it the “economic singularity.” Businesses face a probabilistic tool that hallucinates, so impacts vary wildly.
Open Models Close the Gap—But Money Stays Closed
A report by Frank Nagle from Harvard and the Linux Foundation, titled “The Latent Role of Open Models in the AI Economy,” crunches data from OpenRouter. Open models hit 90% or better of closed model performance while costing about one-sixth as much to run. Nagle’s team calculates enterprises leave $24.8 billion on the table each year by sticking with pricier options like GPT-4 instead of Llama 3.
Simon Willison, writing in InfoWorld, says don’t call it waste. Companies pay for more than tokens—they want service-level agreements, safety filters, and someone to sue if things go wrong. You can’t take a GitHub repo to court. Cloud computing showed the same pattern: free open source software existed, but people paid AWS for the hassle-free setup.
Why Open Source Wins Don’t Mean Open Everything
Open models like Meta’s Llama or DeepSeek’s releases shine as base infrastructure, much like Linux did for servers. No one competes on kernels anymore; value moved to apps. But AI differs. Fixing bugs in a 70-billion-parameter model needs massive compute and original training data—out of reach for most.
Top talent clusters at places like OpenAI and Google. Releases from Meta or Mistral act more like “source available” than true collaborative open source. Companies share to undercut rivals, then sell high-margin layers on top.
- Base models: Open and commoditized.
- Data for fine-tuning: Locked down tight.
- Agents and reasoning tools: Proprietary, with integrations and liability protection.
- Guardrails for safety and compliance: Paid services.
The future splits the market. Raw intelligence gets cheap and open; the rest stays behind closed doors where real revenue hides.
