Eighteen months ago, breaking news about DeepSeek hurt tech company valuations. DeepSeek is an open source Chinese AI model that promised to do the same thing as foundational AI models from companies like OpenAI and Perplexity, but at far less cost. DeepSeek was seen as a direct threat to the foundation model companies, and an indirect threat to investors and stakeholders such as the large public cloud providers.
Now Moonshot AI, another Chinese AI company, has opened up access to Kimi K3. K3 is an open source, open-weights model that’s said to be competitive with Anthropic's Fable 5, according to benchmarks published by Moonshot. Fable 5 is arguably the most powerful publicly available model today.
Fortune's Nicholas Gordon delivered the story in Fortune last Friday, quoting Larridin co-founder and CTO Ameya Kanitcar. Key points:
- The performance gap between K3 and Fable is quite narrow.
- The price gap is brutal for Anthropic, with K3 charging $15 per million output tokens, vs. $50 for the same number of tokens from Fable, a savings of more than 3x. Third party providers will be able to host K3 in the near future, then compete on price, likely driving prices even lower. And companies will be able to host K3 themselves, or use other providers.
- Export bans kept Chinese companies from accessing the latest and greatest AI processors from US companies. This may have forced the Chinese companies to be so efficient that they can match cutting-edge performance at far lower costs.
Larridin research shows that companies are using these products as a cost strategy. Often, however, the usage is not through official IT projects. Instead, it’s in the form of skunkworks projects. This AI usage is internal, but still unsanctioned—happening out of view of upper management and outside of governance and reporting frameworks.
Because Chinese labs release their weights for free, third-party providers—including U.S. ones—can host these models at cost. As Larridin CTO Karnitkar told Fortune, the only costs that matter in that scenario are "GPUs and energy."
Microsoft has even discussed using Kimi K3 or a similar model as an offering in Azure, and incorporating it within Copilot to save more than $600M a year. However, such moves by US companies to offshore AI technology may run afoul of the US government.
Chinese companies can be expected to compete with Western companies on performance and far undercut them on price, along with offering open source and open-weights models for anyone who wants to run them.
The bottom line: Open source / open weight usage has started, inside companies and from suppliers. It can only be expected to increase, unless prevented by strong protective action. The conversation hasn't started, even within enterprise companies, and management can't see this usage without help. Companies need measurement and the ability to optimize usage, just what Larridin provides.