Nvidia and Alibaba Push Free Models to Sell GPUs and Cloud

The article frames the local LLM boom as the intersection of three trends that accelerated and crossed by 2026: the maturation of models, hardware, and software; real anxiety that cloud AI providers hold life-or-death control over enterprise AI usage and could cut off access; and the strategic ambitions of both vendors and nations that want open models to win.

On the vendor side, Nvidia's position is structural. Its revenue currently rests on a handful of hyperscalers — Microsoft, Google, Amazon, and Meta — and its AI Factory concept depends on ordinary enterprises buying their own GPU servers to 'manufacture' AI tokens with their own data. That only works if enterprises have strong models they can run themselves. Hence Nvidia develops the Nemotron series, releases not just weights but datasets and training recipes, and in March 2026 formed the Nemotron Coalition with 8 founding members including Mistral AI, Perplexity, Cursor, and Thinking Machines Lab. Alibaba Cloud takes a different route: it keeps top-tier 'Max' models closed for API revenue while freely licensing lower tiers for commercial use, aiming to make Alibaba Cloud the default ecosystem and capture the computing demand that follows. The Qwen download total — over 1 billion cumulative by April 2026 — is the evidence that this funnel is filling.

The geopolitical layer makes the picture sharper. China, restricted by US export controls on leading-edge GPUs, cannot out-compute America's closed labs on raw hardware. Distributing weights freely worldwide is therefore a way to win ecosystem and mindshare, since published weights cannot be recalled. The US has moved in parallel: the July 2025 AI Action Plan acknowledged the geopolitical value of open-weight AI, a July 2026 public letter signed by Nvidia, Microsoft, Meta, IBM, Hugging Face, Mistral AI, and Palantir argued for spreading rather than restricting open weights (OpenAI and Google signed later), and in August 2026 the government told companies that US open-weight models would be exempt from pre-deployment safety review.

The practical read for businesses is that this is not a story about replacing cloud AI. The article's own guidance is that closed frontier models still win on performance and cost-effectiveness, and the realistic answer is task-by-task selection: frontier cloud models for the hardest tasks, local models for data that cannot leave or workflows that must not stop, and cheaper clouds hosting large open models in between. Holding the option to run AI yourself is the leverage — and a business continuity plan for a moment when a model provider might go dark.