$26 billion. That is how much Nvidia has committed to multi-year cloud service agreements, according to its SEC filings - double the $12.6 billion disclosed just three months earlier. The spend is spread across six fiscal years through 2031, with $6 billion annually in the peak years of 2027 and 2028.
The filings state the compute supports "research and development efforts and DGX Cloud offerings." But read alongside Nvidia's recent moves, the picture becomes clearer: the company that sells picks and shovels to the AI gold rush is building its own mine.
From Chip Seller to Model Builder
Nvidia already ships the Nemotron 3 family of open-weight models (models whose internal parameters are publicly available for anyone to inspect, modify, and deploy). The lineup includes Nemotron 3 Nano at 30 billion parameters, Super at roughly 100 billion, and Ultra at around 500 billion. These are not side projects - they are production models designed for enterprise agentic AI (systems where AI autonomously performs multi-step tasks rather than just answering questions).
The company is also preparing NemoClaw, an open-source AI agent platform expected to debut at GTC 2026 on March 15. Wired reported that Nvidia has been pitching the platform to Salesforce, Cisco, Google, Adobe, and CrowdStrike. NemoClaw is hardware-agnostic, meaning companies can run it even without Nvidia chips - an unusual move that signals Nvidia is prioritizing platform adoption over hardware lock-in.
Why Build Models When You Sell the Hardware?
Nvidia's GPU business generated $215.9 billion in fiscal 2026 revenue, up 65% year-over-year. So why spend $26 billion on cloud compute for internal R&D?
The strategic logic: if open-weight models from DeepSeek and Alibaba's Qwen keep improving at their current pace, demand for the most expensive Nvidia hardware could plateau. DeepSeek V4 launched in March 2026 with inference costs roughly 1/20th of GPT-5. Companies that can run capable open-weight models on cheaper hardware have less reason to buy top-tier Nvidia GPUs.
By building its own competitive open-weight models and an agent platform, Nvidia creates a reason for enterprises to stay in the Nvidia ecosystem even as the cost of running AI drops. It is the classic razor-and-blades strategy inverted: give away the software, make the hardware more valuable by optimizing everything to run best on your chips.
The Competitive Math
Nvidia's $26 billion in cloud commitments dwarfs what most AI labs spend on compute. OpenAI is reportedly burning through $14 billion in 2026, and even that requires $100 billion in fundraising to sustain. Anthropic's revenue run rate is projected at $26 billion for 2026, but its compute spending is a fraction of that.
Nvidia CEO Jensen Huang recently called the company's $30 billion investment in OpenAI "might be the last" - a comment that reads differently when you realize Nvidia is simultaneously building infrastructure to compete with OpenAI's models directly.
The GTC keynote on March 16 will likely reveal how far along Nvidia's model ambitions actually are. For now, the $26 billion in committed compute tells the story: Nvidia is not content selling shovels.