Nvidia Underscores Support for Open-Source AI, a Boon for Hardware Spending

The emergence of frontier artificial intelligence is the new front in Silicon Valley’s open- vs. closed-model Forever Wars, and the battle lines are now clear.

On one side: OpenAI and Anthropic, championing a closed-model ecosystem in the name of safety. On the other side: just about everyone else in the tech world, including Mark Zuckerberg’s Meta, and perhaps most importantly, Nvidia. In March, Nvidia CEO Jensen Huang spent much of Nvidia’s developers conference positioning Nvidia’s chips as the perfect infrastructure for open-source tools, and last month, Huang created an X-née-Twitter account seemingly for the sole purpose of publicly praising the democratizing power of open-source AI. On Tuesday, the semiconductor king made another stand, announcing its new open-source Nemotron 3.5 Lightning model.

With Open Arms

Still, Nvidia didn’t become the world’s biggest company by practicing pure digital altruism, and the company has plenty to gain from a more open AI ecosystem. “Software traditionally is where the value is, and that is also where a lot of the costs are,” Bill Wong, AI research fellow at Info-Tech Research Group, told The Daily Upside. “What Nvidia is doing is removing that cost” by offering and supporting open-source models, “which means [enterprise clients] have more money for hardware.” 

In other words, greater demand for and access to open models will push value back down the tech stack to where the scarcity exists: hardware and infrastructure, Nvidia’s bread and butter. The flow of money away from powerful models also drains potential capital that closed-model makers like OpenAI, Google and Anthropic could use (and are already using) to craft in-house chips and reduce their reliance on Nvidia. In turn, model right-sizing only increases compute demand, which is also good for Nvidia:

• Global spending on AI inference (or the actual usage of AI models) is expected to reach $23.3 billion this year, according to a Gartner report published Monday. That would surpass spending on AI training (expected to reach $19.9 billion this year) for the first time ever.
• Nvidia, which holds an ironclad 90% market share for the training side of AI by most estimates, is increasingly eating up the inference market, too. According to a report by The Information in June, Nvidia’s inference market share is now at 74%, compared with 66% a year ago.

Fastest Route: In addition to the Nemotron 3.5 Lightning, Nvidia on Tuesday also announced the NeMo Switchyard, an open-source library for routing agent tools to the best-suited model for any given task. Nicolas Sauvage, president of venture capital firm TDK Ventures, told The Daily Upside that may just be the most important news of the day. “Open models do not eliminate value. They change where value is captured,” Sauvage said. “The future may therefore be less about one model winning and more about systems dynamically choosing the best model, or combination of models, for each task.”