Mark Zuckerberg has a big idea: put superintelligence in the hands of billions of people. It's a sweeping vision, laid out in an essay Monday, where the Meta Platforms, Inc. (NASDAQ: META) CEO argued that powerful AI shouldn't be hoarded by a few companies or institutions. Meta plans to offer free versions of its AI tools to the masses, while letting users pay up for extra computing power.
But is that vision actually practical? Paddy Srinivasan, CEO of DigitalOcean Holdings, Inc (NYSE: DOCN), thinks the economics of AI might finally be moving in that direction. He points to open-weight models as the key shift—models that deliver increasingly strong performance at a fraction of the cost.
"Open weight models are balancing out the intelligence per dollar equation and driving even more usage of AI," Srinivasan told MarketDash in an exclusive email interview.
That could be the missing piece for Zuckerberg's grand plan.
Open AI Models Are Changing the Math
The appeal of open-weight models isn't just that they're cheaper. Businesses can customize them with their own data, and sometimes run them on their own infrastructure. Srinivasan says that helps companies keep their intellectual property in-house while tailoring models to proprietary information and real-world use.
This changes how companies think about AI spending. Instead of paying top dollar for every task, they can pick models based on what they actually need. Srinivasan's goal: the "right model, right cost, for every task."
DigitalOcean is already building around that idea, offering access to multiple models through its inference infrastructure. The platform supports models from providers like OpenAI, Anthropic, and open-weight developers, letting customers choose based on cost and performance.
The Frontier Model May Not Be Needed Every Time
Srinivasan says frontier models—the most capable and typically most expensive—are needed for only about 25% of the work he sees among customers. The other 75% can often be handled by open-weight models, creating what he calls a typical 25/75 split between closed and open-weight models.
That doesn't mean OpenAI or Anthropic suddenly lose most of their business. The hardest reasoning and specialized tasks still justify the cost of frontier models. But it does suggest that AI adoption doesn't have to hinge on the economics of the most expensive systems.
And that's where Zuckerberg's vision meets Srinivasan's thesis. Meta is betting that powerful AI should reach billions. Open-weight models could make that access cheaper, more customizable, and less dependent on a handful of providers. Meta itself is leaning into the approach, with Zuckerberg saying the company plans to resume releasing some open-source models, while also launching the smaller open-weight Muse Glimmer model for personal computers.
The bigger AI race, then, may not simply be about building the smartest model. It could be about making useful intelligence cheap enough, flexible enough, and accessible enough that far more people can actually use it.
For Zuckerberg, that's the vision. For Srinivasan, "intelligence per dollar" may be what makes it economically possible.
