Foreign media report that open-weight AI companies are becoming the center of Silicon Valley’s latest wave of mergers and acquisitions. Market attention is focused on NVIDIA’s reported plan to acquire Hugging Face for $13 billion. The company, which provides open-weight models and evaluation platforms, is seen as a key distribution channel for AI developers.
Mergers and acquisitions are occurring in succession.
Prior to this rumored deal, NVIDIA had already reached a $6 billion agreement with the open-weight model company Poolside, whose majority of employees will join NVIDIA. Two weeks earlier, Stripe acquired OpenRouter for over $7 billion; OpenRouter primarily provides enterprises with access to open-weight models.
The article suggests that the surge in such transactions reflects major tech companies' efforts to secure a position along another segment of the AI supply chain beyond frontier models. Although open-weight models are still not mainstream, they have become an important choice for enterprises seeking low-cost, customizable AI solutions.
NVIDIA wants to fill the model entry gap
For NVIDIA, one reason for driving mergers and acquisitions is to reduce dependence on hyperscale cloud providers and leading model companies. As model developers like OpenAI and Google advance their own custom inference chips, NVIDIA aims not only to sell chips but also to gain greater control over model distribution and development channels.
The article notes that NVIDIA already has its own Nemotron open-weight model, but its market adoption has been limited. If NVIDIA acquires Hugging Face, it would gain access to one of the largest open-model developer communities in the U.S. and have the opportunity to direct more users toward its own chips, toolchains, and technical standards.
Businesses first consider costs, then control.
Another context behind the growing interest in open-weight models is that enterprises continue to scrutinize AI inference costs. Some companies are beginning to test lower-cost models from Chinese vendors such as Moonshot AI, DeepSeek, and Alibaba. According to the data cited in the article, the adoption rate of open-weight models among enterprises is still low but showing a growing trend.
- Ramp's survey shows that approximately 6% of businesses are using open-weight models.
- Jellyfish调查显示，仅有约2%的软件工程师在使用此类模型。
Nik Albarran, Product Lead at Jellyfish AI, said that such models are currently better suited for high-frequency, repetitive reasoning tasks, such as customer service conversations, because their patterns are relatively fixed, allowing businesses to reduce per-call costs through model fine-tuning.
However, in programming and agent tasks, leading proprietary models still hold an advantage, due to stronger reasoning capabilities, easier integration, and token subsidies offered by some service providers. The article suggests that as companies streamline their AI workflows, the barriers to adopting open models may continue to decline.
The next step may be for each company to have its own model.
Fireworks, an enterprise-grade model routing and hosting platform, is also listed in the article as a potential acquisition target. Its CEO, Lin Qiao, stated that the company currently processes 40 trillion tokens per day. She believes that enterprises will increasingly prioritize model diversity and train models tailored to their specific business scenarios.
The article argues that OpenAI and Anthropic’s current leadership is not unshakable. As tech giants seek to reduce their reliance on leading labs, the strategic value of open-weight technologies and related platforms is rapidly rising.