Nvidia is testing a licence fee for the software its chips use to run AI agents, and the reference price it gave at GTC Berlin is about USD 2,000 per GPU per year, roughly RMB 13,500. The figure came from chief executive Jensen Huang. It is not a published product yet, and Nvidia has signalled that pricing will vary by GPU model and by which services a customer takes.
The direction of travel has been visible for a while. On the BG2 podcast earlier this year, Huang said the day may come when Nvidia is no longer classified as a chip company. The company has been pushing into automotive, telecoms, healthcare and life sciences, and software subscriptions are the part of that push with the cleanest margin.
What the first software tier already costs
Nvidia has sold AI software before. AI Enterprise, launched in 2021, is priced at USD 4,500 per GPU per year for production deployments. That bundle covers virtual GPU compute, Run:ai cluster scheduling, Omniverse and the full RAPIDS stack, with NIM microservices, the NeMo framework and an enterprise AI platform as its core services. It exists to get pretrained models, retrieval and orchestration into production.
The new tier targets agentic AI instead. Its focus is long-running agents, tool calls and production inference, the layer that lets an agent safely call tools, read data and execute tasks on a compute cluster. In practice it supplies the software, management tooling and support that an AI factory needs to run model agents at all.
Read together, the two tiers look like a stack being priced in layers: hardware bundled software, then enterprise platform, then agent software. Each layer charges again.
The arithmetic behind the price
Semiconductor analysis firm TechInsights reported that Nvidia sold about 3.76 million GPUs into data centres in 2023 alone. With AI demand since then, the installed base has passed the ten million mark. Assume five million of those cards buy the new subscription, and the line item reaches USD 10 billion a year.
For scale, Nvidia’s fiscal 2026 revenue, covering February 2025 to January 2026, was about USD 215.9 billion. Ten billion is a meaningful slice of that, and the installed base is still growing.
The customer concentration is sharper than the average suggests. Reuters has reported that Tencent, Alibaba and ByteDance jointly ordered more than three million RTX PRO 5500 GPUs, mainly for inference. If that entire order subscribes, the annual bill is USD 6 billion, or about USD 2 billion per company, roughly RMB 13.3 billion each. That is a serious number even for a platform group spending tens of billions of yuan a year on AI.
A stronger lock than silicon
The commercial logic is that software subscriptions bind harder than chips. If a customer’s agent security, orchestration and observability all run on Nvidia’s stack, moving to AMD silicon and the ROCm ecosystem is not a code change. It is a rebuild of the deployment and operations chain.
That matters because AMD, Intel and others can ship competitive accelerators without automatically winning the account. The migration cost sits with the customer, not the vendor.
Nvidia has run this play before. CUDA is free, near-monopoly in general GPU acceleration, model training and scientific computing, and it still provoked a decade of escape attempts because it is so tightly bound to Nvidia hardware.
Those attempts are now arriving in numbers. In October, Amazon introduced TorchNeuron, which runs PyTorch workloads on its own Trainium chips with almost no code changes. Microsoft built Maia 200 around Triton, the compiler OpenAI open-sourced. Google and Meta jointly released TorchTPU, a native PyTorch path for their tensor processing units. In China, Moore Threads ships MUSIFY to translate CUDA into its own MUSA format, and eight domestic GPU vendors completed Day0 support for DeepSeek’s V4 model.
If the free layer already generates that much defection pressure, a paid layer above it will generate more.
Who actually pays
The likely split is by company size. Small teams debugging locally, running experiments and prototyping with open-source models keep CUDA, cuDNN and the base frameworks for free. A subscription only triggers when they serve external traffic, need a vendor SLA, maintain long-lived branches or deploy NIM commercially.
Large labs with big engineering teams have another option: build the deployment layer themselves. Alibaba runs its own inference engine, deployment platform and compiler stack. Groups with heavy agent plans and little appetite for rebuilding plumbing are the natural buyers of the new tier.
Mid-sized firms sit in the worst spot. They need production reliability and cannot afford to re-engineer the stack, so they either absorb the cost or hunt for substitutes.
The risk in the trade
Nvidia’s side of the bargain is that customers must feel the price is earned. Recurring revenue, stickiness and a re-rated multiple are all real prizes, and agent security and observability genuinely cost money to build.
The risk is that a paid layer accelerates the de-CUDA movement it is meant to prevent. Free and developer-friendly was CUDA’s original attraction. Charge above it, and more projects will make ROCm or open alternatives the primary target rather than the fallback.
The short-term effect is higher costs for anyone deploying AI at scale, plus a harder moat. The longer-term effect is a compute software stack that fragments along national and vendor lines, which raises the commercial value of domestic and open alternatives in China and everywhere else.
Editor’s note: This is an adapted translation of the original Sohu Tech report. It has been trimmed and restructured for readability for an international business audience. The full original (in Chinese) is at https://www.sohu.com/a/1093196205_116157.