Open-weight models now carry most gateway traffic. The hosting bill lands on buyers

The traffic has moved

Open-weight models have crossed from an option to the default route for AI traffic. io.net reports that open-weight models processed more than 71.5% of AI gateway tokens as of October 2026, with their spend share up more than fourfold since June 2026. The same report puts open-weight models at roughly 23% of total AI spend. The gap between those two numbers is the point: the traffic is open, but much of the money still flows through frontier providers. m:…1476

Databricks tells a similar story from inside the enterprise. The company says open-weight models now account for about 40% of tokens used on its Unity Gateway, and the average customer there uses 7.1 unique models, up from 4.2 since January 2026. This suggests buyers are not picking one open model and stopping. They are assembling a bench, and the bench keeps changing: Databricks notes the top intelligence ranking for AI models has changed hands more than 20 times since 2023. m:…6685

Why enterprises moved now

Two barriers fell in the autumn, according to io.net. A wave of Western open-weight releases in September 2026 gave enterprises alternatives to Chinese-developed frontier models, and Microsoft and NVIDIA published joint guidance on managing open-weight model risk in October 2026. That guidance matters because the objection to open weights was rarely capability. It was accountability: who owns the risk when the model is yours to run. m:…1476

Databricks frames the shift as a response to single-provider pressure. The company says enterprises are re-evaluating reliance on one frontier provider because of seven pressures, including rising costs, model volatility and restrictive guardrails. Its own example is pointed: Hugging Face responders used GLM 5.2 to analyse attack logs after frontier models refused the task, because the guardrails blocked legitimate work such as malware analysis. m:…6685

The serving layer is the new product

The supply side is racing to meet the demand. Chutes added seven new models to its catalog from Z.ai, DeepSeek, Alibaba and Moonshot AI, all reachable through its existing OpenAI-compatible API. The cheapest listed token rate in the expansion is $0.30 per million regular input tokens for DeepSeek V4.1 Flash 1017, while GLM-5.3 is priced at $1.225 per million regular input tokens and Qwen3.8 Max 0902 at $2.00. The company says the new models do not run inside a trusted execution environment, a detail buyers who care about confidentiality should note. m:…8097

The cost argument for self-hosting is stark in io.net's numbers. The company puts running a 70B parameter model on its spot instances at about $856.80 a month, against about $2,829.60 a month on AWS on-demand instances. Databricks, meanwhile, says it raised $3 billion in July 2026 to secure GPU capacity for hosting open models, and its Unity Gateway now offers smart routing to shift routine tasks to cheaper models and fallback chains to reroute requests during rate limits or outages. The implication is that the cheapest model is not the point. The routing is. m:…1476 m:…6685

The smallest end of the market shows how far open weights now reach. Preferred Networks released PLaMo 3 Translate, a suite of translation models in four sizes from 0.8B to 31B parameters, covering 53 languages including English and Japanese. The company says the models run locally on edge devices without an internet connection, and organisations with annual revenue of 1 billion Japanese yen or less can use them free under the PLaMo Community License. Translation that once needed a cloud call now fits on the device. m:…0117

The catch: the risk moves with the model

Hosting the model moves the risk, and Airia is blunt about it. The company says hosting an open-weight model shifts responsibility for monitoring outputs and patching failure modes from the vendor to the enterprise, and that when an enterprise hosts its own model it assumes full ownership of the incident response process. Fine-tuning or substantially modifying a model can also change an organisation's legal role under the EU AI Act, according to Airia. Enterprises are evaluating open-weight models because of cost pressure and the need to fine-tune on proprietary data, the company says, but the governance work does not disappear. It changes hands. m:…7452

The practical question for buyers is therefore not whether open weights are good enough. Databricks says open-weight models now benchmark at levels comparable to top frontier models at lower cost, and io.net notes that Muse Open and Beam models support vLLM and SGLang on release day. The question is whether the buyer has the team to own serving, version upgrades, hardware costs and incident response. io.net says infrastructure teams are now responsible for exactly that. An enterprise that treats open weights as a cheaper API call, rather than a system to operate, will learn the difference in production. m:…6685 m:…1476