Summary of Key Points
In the past two months, the global AI model market has experienced a revolutionary shift: the share of token calls for open-source models has soared from 28% to 62%, surpassing that of closed-source models (such as OpenAI’s GPT and Anthropic’s Opus) for the first time. Meanwhile, AI hardware giant NVIDIA has invested $7 billion (including $6 billion in technology licenses and $1 billion in investment) to support the open-source team Poolside, with the goal of creating an open-source model with over a trillion parameters that can compete with closed-source models. Despite closed-source giants frequently launching new models and cutting prices as a defense strategy, users are still shifting in large numbers to open-source models. This shift reflects the evolution of open-source models from being “usable” to being “worth large-scale commercial use,” while the closed-source model’s business model, which relies on an “intelligence premium,” is becoming ineffective.
1. Open-Source Revolt: Token Shares Surpass Closed-Source Models – The Trends Behind the Data
Tokens can be understood as the “information units” processed by AI models (for example, the length of an AI model’s response to a question corresponds to the number of tokens used). Data from the global mainstream AI deployment platform Vercel shows:
- In April, open-source models accounted for only 11% of token shares;
- By June, this share had risen to 28% (at which point open-source models were using less than 4% of the API costs but handling nearly one-third of the total calls);
- On August 22, open-source models surpassed closed-source models for the first time, with a share of 62%.
Chinese open-source models have performed particularly well: DeepSeek alone accounted for 22.6% of the token shares, and the GLM-5.2 model saw its token usage increase by 50 times within two weeks of its release. More importantly, the total number of calls for closed-source models is still growing (for example, OpenAI and Anthropic saw an acceleration in calls in July), but open-source models are growing even faster. This indicates that users are not rejecting closed-source models due to their poor performance but are choosing them for their better cost-effectiveness.
2. NVIDIA’s “Clever Strategy”: Investing Money Not to Develop Models, but to Sell More GPUs
NVIDIA’s approach is quite strategic: it did not directly acquire Poolside but instead purchased technology licenses for $6 billion and invested $1 billion (with a pre-investment valuation of $12 billion), while also incorporating over 100 engineers into its open-source product line, Nemotron. This move is intended to circumvent anti-monopoly regulations, and its real goal is to acquire Poolside’s reinforcement learning technology and top talent.
NVIDIA’s ultimate plan is to make open-source models extremely affordable and user-friendly, leading to a surge in token usage worldwide. All of this token processing will rely on NVIDIA’s GPUs (such as the A100 and H100). Whether an open-source model emerges as the winner, the profits will flow back to NVIDIA’s hardware business. NVIDIA’s previous efforts to establish open-source alliances and release its own open-source models were all part of this strategy.
3. The Anxiety of Closed-Source Giants: The Collapse of the “Intelligence Premium” Model Business Model
The traditional approach of closed-source models was to develop the most intelligent models (such as GPT-5) and then sell them at high prices, using the profits to develop the next generation. However, this cycle has been disrupted:
- Faster Catch-Up: It used to take a long time for Llama2 to catch up with GPT3.5; now, Kimi K2.6 has caught up with Anthropic Opus4.5 in just 4.8 months, reducing the lead time from one year to a few months.
- Companies Calculating Cost-Effectiveness: Is it worth paying ten times more for a 5% improvement in performance? For example, Anthropic’s flagship model Fable5, although the most powerful, is also the most expensive and only accounted for 11.4% of corporate spending two months after its release, while the cheaper Opus5 is more popular. Companies are no longer automatically choosing the “most intelligent” models but rather those that are “usable and affordable.”
4. The Evolution of Open-Source Models: From “Toys” to “Production Tools”
Open-source models have evolved from being merely “usable” (for simple tasks like answering questions) to being capable of handling production-level tasks:
- Focus on Practicality: Models like DeepSeek, Qwen, and Kimi are focused on coding and intelligent assistance;
- Commercial Capability: GLM-5.3 can handle long texts and other practical tasks, and Alibaba has made its 2.4-trillion-parameter model available;
- Integration into Workflows: Open-source models are now being used directly in everyday tasks such as email summarization, information extraction, and code writing, which is the main reason for the surge in token usage.
5. A Changing Industry Landscape: An Open-Source-Dominated “Inclusive AI” Future
This reversal is not accidental but represents a trend:
- Open-source models will penetrate most industries; small and medium-sized companies, as well as general applications, will prefer open-source solutions due to their affordability and customizability;
- Closed-source models will retreat to niche markets, where only the most demanding scenarios (such as high-end research and complex decision-making) will still rely on them;
- NVIDIA will be the biggest winner: Both open-source and closed-source models will need NVIDIA’s GPUs.
In the future, AI will no longer be monopolized by a few giants; more companies and developers will have access to it, marking the beginning of an era of “inclusive AI.”
In a nutshell: Open-source models have defeated closed-source models with their cost-effectiveness, with NVIDIA playing a key role in driving this change. The rules of the AI industry are shifting from “who is the most intelligent” to “who is the most affordable and user-friendly.”