Hims & Hers CEO: Open-Weight AI Models Cut Costs 70-80%

Telehealth executive makes case for custom AI training

Companies sitting on large proprietary datasets should consider training their own open-weight AI models rather than relying on commercial services, according to Hims & Hers CEO Andrew Dudum. Speaking with CNBC on Tuesday, Dudum said the approach can slash AI costs by 70 to 80 percent while delivering superior performance for specific use cases.

Dudum pointed to his own company's experience as evidence. Hims & Hers, a telehealth provider offering prescription drugs and personal care products through subscriptions, has accumulated what he described as a "real asset": a closed-loop dataset of patient interactions that can be used to train specialized models.

"I think for companies that have the resources and scale, if they have an independent dataset, that is a path that they will go, no question," Dudum told CNBC host Andrew Ross Sorkin.

Why it matters

As AI spending comes under scrutiny across the business world, Dudum's comments highlight a strategic fork in the road. Companies with substantial proprietary data may find better economics and performance by building their own models rather than paying per-token fees to OpenAI, Anthropic, or other providers. This shift could reshape competitive dynamics in AI adoption, favoring organizations that have been collecting high-quality operational data for years.

Performance gains beyond cost savings

While the cost reduction is significant, Dudum emphasized that performance improvements may be even more valuable. Models trained on a company's actual workflows and customer interactions can outperform general-purpose alternatives from day one.

"What we launched, our first version, immediately outsurpassed what we could get in market," Dudum said. He added that continuous learning from new patient data creates a compounding advantage over time, with transformative improvements possible within just six months.

Open-weight models allow organizations to access and modify trained parameters, offering flexibility that closed commercial models cannot match. This customization enables fine-tuning for specific domains and use cases.

Broader trend toward AI cost optimization

Dudum's strategy reflects wider industry efforts to extract more value from AI investments. Companies have moved away from "tokenmaxxing"—encouraging employees to use AI tools liberally without regard to cost—and toward more disciplined approaches.

Model routing, which matches tasks to appropriately sized AI systems based on complexity, has become common practice. Some executives, including Coinbase CEO Brian Armstrong, have adopted lower-cost Chinese models like Kimi K2.7 and GLM 5.2 as defaults for routine tasks.

Kimi K3, an open-weight model from Chinese lab Moonshot AI, generated significant attention in Silicon Valley last month by promising capabilities comparable to OpenAI and Anthropic offerings at reduced costs.

Data emerges as critical competitive asset

The emphasis on proprietary datasets for model training has intensified competition for high-quality data. Google recently paid $10 million for internal data from defunct carrier Spirit Airlines, including documents, workflows, emails, and millions of Teams messages. AI training firm Handshake AI has offered to pay $6 per page for quality work documents, with payments capped at $30,000 per contributor.

These details were first reported by Business Insider.