Nvidia Puts a Price Tag on Its Software While Still Chasing the Hardware Orders

Nvidia has spent the past two years selling investors on data-center capacity. This week it began sketching what the software running on that capacity might be worth — and, at the same time, laid out how much cash it is willing to commit before a single additional GPU ships.

At the GTC Berlin event on October 22, CEO Jensen Huang floated licensing the company's agent software suite at roughly $2,000 per GPU per year, according to media reports. The figure is a pricing concept rather than booked revenue: no customer contract has been signed and no software sales have been reported. What it does signal is a possible recurring revenue stream tied directly to the installed base of hardware.

That pitch lands alongside the OpenShell Reference Architecture, unveiled on October 21 and described by Nvidia as immediately available. The framework is designed to deploy and govern AI agents across enterprise infrastructure and edge systems, with partners drawn from cloud, software, cybersecurity and systems providers. The distinction matters for anyone building a valuation: the architecture is a shipping product, while Huang's licensing remark is a monetization idea. Availability alone does not establish commercial traction.

Demand Outruns the Production Line

Order flow tells a different part of the story. Reuters reported on October 21 that ByteDance, Alibaba and Tencent placed orders for more than 3 million RTX PRO 5500 graphics cards, with a combined list value above $35 billion. Nvidia could not guarantee it would fill the orders in full, according to the report, because demand exceeded planned output. Those bookings are not evidence of realized revenue — the gap between order interest and delivered hardware sits at the heart of the investment case.

The company moved to widen the funnel the same day, announcing a manufacturing capacity expansion with several suppliers. Substrate agreements are intended to lift output 2.5-fold within three years. Announced capacity, however, is not yet available production, and the expansion addresses a different leg of the growth question than software licensing does.

Two Speeds of Progress at Oracle

The Oracle relationship captures the same tension in miniature. Vera Rubin NVL72 systems and Vera CPUs are already generally available on Oracle Cloud Infrastructure, while the larger gigawatt-scale supercluster build-out remains a plan rather than a completed fact. Investors should resist treating the expansion headline as fully realized usage — and equally should not overlook that customers can already access the technology today.

Nvidia is also trying to close the gap between installation and productive operation. On October 22 it introduced the Minerva platform, which combines telemetry from infrastructure and AI workloads to spot problems and keep AI factories running. OpenShell, meanwhile, connects Vera Rubin infrastructure with other Nvidia components so enterprises can build, monitor and enforce controls on AI agents. Both target the post-installation phase. Their launch demonstrates the offering, not its effect inside customer operations.

The expanded Muse collaboration with Meta fits the same pattern. Meta plans broader use of Vera Rubin, Vera CPUs, Spectrum-X and BlueField-4 — not a confirmed full-scale rollout, but evidence that the partnership spans multiple Nvidia products. Should such plans convert into productive workloads, they would support the bullish reading of the build-out.

Capital Committed Ahead of Proof

The counterargument is one of timing. Nvidia intends to invest $3 billion in Meta's equity offering as part of the expanded engagement, and the supply-chain expansion includes a $1.1 billion capacity-reservation prepayment to KYEC, earmarked for new test facilities in the United States. The commitment therefore extends well beyond product announcements.

If customers scale their deployments more slowly than planned, those outlays could weigh more heavily. That would not prove demand is fading, but it would put the synchronization of capacity build-out and actual usage at risk. The software tools do not eliminate that risk on their own: Minerva is meant to surface problems earlier and OpenShell to enable monitoring and controls, but neither guarantees uninterrupted operations or a faster adoption curve.

What would strengthen the thesis is concrete customer progress — the planned expansion of Nvidia's stake in Meta, advances on the Oracle build-out, and above all confirmed implementation steps that go beyond letters of intent. The debate for investors is not growth versus stagnation. It is the quality of that growth: a broader offering creates options, and only reliable use of it will show how durable the next stage of the build-out really is.