SILICON VALLEY, California (August 14, 2026) — Cheaper, more customizable and able to generate code nearly as well as the best from Anthropic and OpenAI, open artificial intelligence models from China have taken Silicon Valley by storm this year. American tech firms are seeking to fight back, countering a rival they can no longer ignore. Meta said this week it would resume releasing open models, including a version of its most powerful.
Chip giant Nvidia, also among the most prominent U.S. makers of open models, released a small system and is developing a larger one to rival leading open-weight models, The Information reported. Taken together, the moves reflect a growing recognition among America’s tech leaders that open-weight models will likely play a bigger role in making the technology more widely available, especially as companies shift from spending heavily on AI — a trend known as tokenmaxxing — to scrutinizing that same spend for efficiency.
They also highlight the growing threat that open models, particularly those from Chinese firms such as Moonshot and Z.ai, pose to IPO-bound OpenAI and Anthropic. Both U.S. AI giants have to convince investors their technological edge can justify the staggering cost of developing and operating their models.
“For so many of the basic operational tasks that are being done, you don’t need the most cutting (edge) frontier model. So, it becomes a return on investment question for these companies,” said Marc Bhargava, managing director at General Catalyst, a venture capital firm that has invested in Anthropic.
Bhargava said hurdles remain because adapting open models to match the best systems can be difficult and costly, especially for smaller teams. For the most demanding tasks, such as coding, he expects models from OpenAI and Anthropic to retain an edge.
META’S RETURN
Meta’s return on Monday to open models, alongside a 6,500-word essay from CEO Mark Zuckerberg advocating AI for everyone, drew skepticism from some critics who saw it as a self-serving bet by a company without a top-tier model. But the move also points to a growing space many believe American firms can fill. Despite their rising popularity, Chinese models are still viewed with suspicion by many American companies that are worried about data security and the biases the models may have absorbed in training, potentially skewing their responses.
“People don’t want to use Chinese models. That’s their default position. But they are so cheap that the economics drive them to use it,” said Ameya Kanitkar, co-founder and chief technology officer at Larridin, a San Francisco-based startup that helps businesses measure, manage, and optimize AI adoption. “So the natural default position to win here is an American open-weight model. People would just gladly use it, without the hesitation about the regulatory and geopolitical environment.”
Running Chinese open-weight models on American platforms largely addresses data-security concerns by keeping customer prompts and information within the cloud instead of sending them to the models’ Chinese developers. But there is still unease from the fact that the models are not fully open source. Like most open-model makers, the Chinese companies release the weights, the trained settings that shape a model’s answers, for download while keeping the training data and code private, so no one can fully inspect what went into them.
That differs from open-source software such as Linux, which underpins much of the internet and makes its code available for anyone to inspect and modify, giving users greater control. Demand for American alternatives is already visible on AI developer platforms such as OpenRouter, where Nvidia’s Nemotron 3 Ultra ranks above Chinese models including MiniMax’s M3 despite trailing them on widely used performance benchmarks. Meta and U.S. open-model makers have other advantages, too.
The $1.5 trillion social media company plans to spend up to $145 billion on AI infrastructure this year, giving it a scale and computing capacity few open-weight companies can match. Most Chinese open-weight models, meanwhile, are hosted by their developers, China-based cloud giants that many U.S. and European companies are unlikely to use, or smaller American AI cloud providers such as CoreWeave.
“Our biggest bottleneck right now is scale,” said Kai Mak, chief revenue officer at Together AI, one of the U.S. cloud companies that hosts Chinese models. He told Reuters that a $240 million Nvidia-powered AI cluster it is building on IBM Cloud under a multi-year deal is expected to sell out two to three months before it launches, because of strong demand. Still, Meta’s success is far from guaranteed, and the extent of its commitment to open models remains unclear.
The company plans to release the weights of Muse Spark 1.2, its most advanced model, but has not said whether it will do the same for Watermelon, a more powerful system reportedly under development for release in the coming months. That decision could show whether Meta is merely testing the waters or, as Zuckerberg put it in his manifesto, working to make “American open source models to be the best globally.”
Data from fintech provider Ramp released on Wednesday showed that more businesses are using platforms such as OpenRouter that offer open-weight and Chinese-developed AI models. About 6.1% of businesses spending on AI used these platforms in July, up from 4.5% a year earlier. The data covers only a slice of AI spending, but shows open-weight models gaining ground in Silicon Valley.
Ramp lead economist Ara Kharazian said the trend has yet to dent spending on OpenAI and Anthropic, but their adoption is slowing, particularly for the ChatGPT maker. That, he said, leaves the top AI companies more reliant on growth from existing customers, especially the biggest spenders, who are increasingly putting money into open-weight models.
With information from Reuters
