Rising AI Costs Drive Software Developers to Open-Weight Models

Numerous software firms, including Chinese entities, are gravitating towards open-weight artificial intelligence models due to the escalating expenses associated with proprietary models from U.S. AI developers, Bloomberg reported on September 21st. Open-weight models are generally more affordable and offer companies the ability to construct their own models using their proprietary data, thus mitigating the risk of entrusting their technology to another firm, according to the report.

Nonetheless, some companies have discovered that leveraging open-weight models to construct their own models proves ineffective due to elevated initial expenditures and the necessity for specialized expertise, computer infrastructure, and data, unless they possess ample data of their own. In the case of Chinese open-weight models, certain companies' clientele have expressed apprehensions about data privacy, the report noted.

Certain firms that have developed their own models persist in utilizing models from leading AI research institutions when there is a demand for the highest levels of performance, as per the report. According to PYMNTS, chief financial officers of mid-sized companies are contemplating whether the cost savings, flexibility, and control afforded by open models are ample to justify assuming additional responsibility for the underlying infrastructure.

Self-hosting necessitates computing power, storage, cybersecurity measures, monitoring tools, and skilled personnel, whereas a proprietary, closed platform typically incorporates many of those obligations into its pricing. It was disclosed in August that AT&T achieved a reduction in coding and other advanced AI tasks costs by up to 56% by employing tools that direct employees' inquiries to more cost-effective models as required.

AT&T observed a mere 2% decline in AI performance when implementing this strategy. AT&T plans to increase the proportion of employees' queries processed by open-source models from the present 40% to a range of 60% to 70% in the forthcoming years, as per the report. In June, it was reported that the surge in model costs from prominent AI labs can be attributed to the transition from chatbots to agents, which require more computational resources, as well as AI labs' shift from flat subscriptions to token-based billing.

Chinese labs are able to charge less than their U.S. counterparts owing to their more efficient models and China's comparatively lower energy costs, the report stated.