Nvidia Pitches a $2,000-per-GPU Software Fee While Racing to Turn Orders Into Hardware

Nvidia CEO Jensen Huang used a Berlin stage on 22 October to float a pricing idea that would attach a recurring charge to every accelerator a customer runs: roughly $2,000 per GPU per year to license the company's agent software suite. Media reports framed it as a pricing approach rather than a signed deal, and that distinction matters — no customer contract and no booked software revenue sit behind the figure.

The proposal lands alongside a separate, already-shipping product. Nvidia's OpenShell Reference Architecture, unveiled on 21 October, is designed to deploy and manage AI agents across enterprise infrastructure and edge systems. The company called it immediately available and named partners spanning cloud, software, cybersecurity and systems. One is an offer customers can buy today; the other is a possible way to monetize it. Availability on its own says nothing about commercial traction.

Demand Is Not the Problem — Fulfillment Is

Supply, not appetite, keeps surfacing as the binding constraint. Reuters reported on 21 October 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 requests in full, according to the report, because demand outran planned production. Those orders therefore do not translate into realized revenue, and the gap between order interest and shipped hardware is where investors are being asked to do their own math.

The company is trying to close that gap on several fronts at once. Working with manufacturing partners, it announced on 21 October that substrate capacity would rise 2.5-fold within three years. It also committed a $1.1 billion capacity-reservation advance to KYEC for new testing facilities in the United States. Amkor, for its part, is targeting an increase of more than 300% in interposer wafer capacity by the end of 2028. Announced capacity is not yet available production, but the direction of travel is toward more components and more test throughput.

Should investors sell immediately? Or is it worth buying Nvidia?

A Modular Blueprint Aimed at the Deployment Gap

Speed of installation has become its own test of the growth story. The DSX POD reference design presented on 21 October is a modular blueprint for Vera Rubin AI factories, built by partners, which Nvidia says could cut deployment times from months to weeks. The advantage would sit less in any single component than in a better-prepared build-out of entire facilities. That remains a company claim about the concept's potential, not a demonstrated result — and shorter timelines would not by themselves prove that every announced project gets delivered on schedule.

The expansion of the Oracle partnership disclosed a day earlier shows why the question is live. Scaling OCI superclusters to multi-gigawatt levels would require not just more systems but their delivery at scale. Vera Rubin NVL72 and Vera CPUs are now generally available on OCI, and Oracle plans to introduce HGX Rubin NVL8 systems — giving the collaboration both a completed availability milestone and further product plans. Meta, meanwhile, intends a broader rollout of Vera Rubin, Vera CPUs and Nvidia networking and interconnect technology, a plan that would raise the stakes on a more efficient supply and deployment chain.

The Financing Question Hangs Over the Order Book

There is a counterweight to all this expansion: who pays for it. Reuters reported on 21 October that Meta raised its planned equity offerings to $97.5 billion, with Nvidia committing to take a $3 billion stake. The report flagged concerns about circular financing when chip suppliers help fund a major customer's AI build-out. A larger planned installation cannot be assessed in isolation from its funding — the stake can support the build, but it also raises the question of how durable demand would be without financial help from the supplier side.

Execution risk compounds the issue. The capacity build-out stretches over several years, and DSX POD promises faster deployment without guaranteeing it at any individual site. If financing or construction stalls, bigger expansion plans alone would offer thinner support for the valuation.

What to Watch Next

Two checkpoints will carry the most signal. The first is Oracle's planned rollout of HGX Rubin NVL8 systems on OCI, which would show how the widened partnership moves beyond hardware that is already available. The second is Amkor's capacity target through the end of 2028, which bears on longer-term supply capability.

For investors, the question is less about the size of any single expansion promise than about its quality. Plans that turn into available systems and operating infrastructure give the growth case substance. If delivery lags the announcements, the financing concerns carry more weight — and a per-GPU software fee, however intriguing, will not settle that argument on its own.

Ad

Nvidia Stock: New Analysis - 28 October

Fresh Nvidia information released. What's the impact for investors? Our latest independent report examines recent figures and market trends.

Read our updated Nvidia analysis...