Companies with huge datasets can save up to 80% in AI costs by using open-weight models: Hims CEO

One executive is touting open-weight models, particularly if your institution produces tons of data.

In an question and reply pinch CNBC Squawkbox, released connected Tuesday, Hims & Hers CEO Andrew Dudum said companies that already person ample datasets should displacement distant from utilizing large AI models and toward utilizing open-weight models.

He cited his company, a telehealth supplier that delivers medicine narcotics and individual attraction products done a subscription-based service. He said the institution has a ample closed-loop dataset of patients it has worked with, which he called a "real asset" because it tin beryllium utilized to train AI models.

"I deliberation for companies that person the resources and scale, if they person an independent dataset, that is simply a way that they will go, nary question," he said to CNBC big Andrew Ross Sorkin. "The costs tin beryllium upward of 70, 80% less."

Open-weight models fto users entree and customize a model's trained parameters, and tin beryllium cheaper to tally astatine scale. More importantly, Dudum said models trained connected a company's existent usage cases would execute better.

"And truthful what we launched, our first version, instantly outsurpassed what we could get successful market," he said. "But the trajectory of what's imaginable successful conscionable six months is transformative because it's continuously learning connected each caller diligent that comes into Hims & Hers."

Dudum's comments travel arsenic redeeming AI costs is 1 of the hottest topics successful the tech and business spheres. Companies are phasing retired tokenmaxxing, successful which labor were urged to pain arsenic galore AI tokens arsenic possible, and are now focusing connected getting the astir worth from their spending.

Many are model routing, referring to matching tasks to AI models based connected their complexity. Some executives, for illustration Coinbase's Brian Armstrong, person implemented the usage of Chinese models for illustration Kimi K2.7 and GLM 5.2 arsenic default models wrong their companies.

Kimi K3, an open-weight exemplary developed by the Chinese AI laboratory Moonshot AI, made waves successful Silicon Valley past period for promising to rival the capabilities of models from OpenAI and Anthropic astatine little costs.

Meanwhile, information has go 1 of the astir sought-after resources successful the AI space, arsenic it is important to exemplary training. Earlier this month, Google paid $10 cardinal for soul information from the shuttered bearer Spirit Airlines, which included soul documents, workflows, emails, and millions of Teams messages.

AI training patient Handshake AI has promised to salary $6 per page of "high-quality" activity documents, capped astatine $30,000, successful a bid to summation entree to much data.