The AI cost war just flipped from "who has the biggest model" to "who has the best data to train their own."
The Summary
- Hims & Hers CEO says companies with proprietary datasets can cut AI costs by 70-80% using open-weight models instead of renting access to frontier models
- China's aggressive open-weight releases are forcing US players like Nvidia and Meta to reconsider their closed-model strategy, a tacit admission the moat isn't holding
- The real competitive advantage isn't model size anymore. It's having a closed-loop dataset worth training on
- If you're sitting on years of customer interactions, transactions, or operational data, you just became an AI infrastructure company whether you planned to or not
The Signal
Andrew Dudum runs a telehealth company that ships prescription drugs and personal care products. Not exactly the first place you'd look for AI infrastructure insights. But Hims & Hers has a "large closed-loop dataset" of patient interactions that Dudum now calls "a real asset." He told CNBC that companies with scale and proprietary data should train their own models using open-weight foundations, not keep paying OpenAI or Anthropic by the token.
The math is stark. Training on your own data with open-weight models costs 70-80% less than renting inference from the big labs. More importantly, models trained on actual use cases perform better than general-purpose behemoths trying to be good at everything.
"For companies that have the resources and scale, if they have an independent dataset, that is a path they will go, no question."
This isn't just one CEO's opinion. China's flood of open-weight AI models is forcing Nvidia and Meta to reconsider their closed strategies. Bloomberg frames it as a tacit admission: the US moat around frontier AI isn't holding. When your geopolitical competitor is giving away comparable capability for free, the rent-seeking business model gets harder to defend.
The shift changes who wins in AI. It's not the companies with the most PhDs or the biggest GPU clusters anymore. It's the companies with:
- Years of structured customer interaction data
- Operational datasets from repetitive business processes
- Transaction histories that reveal human behavior patterns
- Proprietary knowledge bases that can't be scraped from the internet
Every SaaS company, every logistics operation, every financial services firm has been collecting this exhaust for a decade. Most treated it like a compliance burden. Now it's the most valuable thing they own. Open-weight models let you access and customize trained parameters, which means you can fine-tune foundation models on your specific domain without sharing your data with a third party or paying inference costs forever.
The geopolitical angle matters too. If China keeps releasing competitive open-weight models, US companies paying premium prices for closed API access start to look like they're choosing ideology over economics. That's a hard sell to a board watching quarterly cloud bills. The Nvidia and Meta strategy shift Bloomberg reports isn't charity. It's recognizing that the closed-model game only works if you maintain a significant capability gap. When that gap narrows, openness becomes the competitive move.
The Implication
If you run a company with proprietary data, your AI strategy just changed. Waiting for ChatGPT to get better is the wrong bet. Your edge is training models that know your specific domain, your customer patterns, your operational quirks. Open-weight foundations make that possible without building a research lab from scratch.
The companies that figure this out first will run AI infrastructure at a fraction of what their competitors pay. The ones that keep renting general-purpose intelligence by the API call will wonder why their margins keep shrinking while their cloud bills keep growing. The data you've been sitting on just became your moat. Start treating it like one.