Companies are moving their AI off the commercial APIs like Anthropic, OpenAI and Google and onto open-weight models (models that can be downloaded and run on hardware the company controls) hosted on their own servers. This change has two drivers: an entry-level private LLM server now costs $8,000 to $12,000, and the models themselves have closed most of the distance to the commercial frontier models.
We saw this before 15 years ago when highly regulated sectors like banking or healthcare had to repatriate servers back on-premise against the cloud trend of the moment. Now it's happening for different reasons: economics, and data sovereignty.
Quality and Cost
Quality is still the reason why frontier models (led by Anthropic today) remain the best option available. Open-weight models, though, are considered to sit only a few months behind them, and to be closing in: about seven months on average since 2023, according to Epoch AI, and four to seven on cyber tasks in the UK AI Safety Institute's evaluation. As of May 2026, the top 15 places on the LM Arena leaderboard still belonged to proprietary models, with the strongest open-weight entries, GLM-5.1, Kimi and DeepSeek, at 16th, 26th and 32nd.
For AI agents executing business work, the quality trade-off is barely relevant, because most business tasks need consistent, fast inference at volume rather than frontier reasoning. Open-weight models cost roughly a tenth as much per token, and above 50,000 queries a month a private server pays for itself within three to six months. Lambert expects most companies to end up running both kinds: closed frontier models directing AI agents, and small open-weight models executing the repetitive steps at a fraction of the cost.
Control
Cost is however no longer the only argument engineers make for on-premise deployments. AI agents increasingly depend on fast access to company data, through retrieval systems that supply a model with internal documents and through custom models trained on proprietary data, and both are easier to run and to keep under control when the model runs on the same infrastructure as that data. Control matters more as AI agents gain decision power and access to internal tools, a trend expected to accelerate; routing those operations through cloud servers is becoming a compliance and security problem that regulated companies struggle to defend to auditors.
Geopolitics
But there is a catch: the leading open-weight models come almost entirely from Chinese labs (Z.ai, DeepSeek, Moonshot, Alibaba), which for some buyers turns a cost and security decision into a policy question. Z.ai's GLM-5.2 sits only a few months behind GPT-5.5 and Claude Opus 4.7 on cyber and bio capabilities, and refused none of the offensive tasks in SaferAI's tests. A company that leaves the cloud also leaves the provider's guardrails behind and takes on that responsibility itself.
Time will tell, but power may be moving from the American cloud providers to Chinese models. They cost less and, oddly enough, pass compliance more easily because they run locally. The question is: will this be tolerated by American regulators?
