Nvidia Isn't Selling Chips. It's Selling the AI Factory — and the Price per Gigawatt Keeps Rising

Nvidia just reported the biggest quarter in its history — $96.2 billion in revenue, with data center sales up 117% to $89 billion — and the stock has taken it largely in stride. That muted reaction is where the real debate sits. It was never really about whether demand holds; demand is the one thing nobody is questioning. The question is whether Nvidia is a chip company whose lead can be undercut, or a platform whose value runs across the whole AI factory. Those two readings justify very different multiples — and very different upside.

Here is the single number that separates them. Nvidia says each new generation raises the value it captures per gigawatt of data center capacity a customer builds: $18 billion on the older Hopper line, $25 billion on Grace Blackwell, and $40 billion on the Vera Rubin platform now ramping. That last figure is not a chip sale. It is the full stack priced together — compute, networking, software, the racks and systems that bind them — measured against the thing customers are actually buying, which is working AI capacity, not silicon.

Where the attack actually lands

The case against Nvidia has always found its target at the point where the AI cycle shifts from training to inference. Training runs on CUDA's installed base, a deep moat. Inference is contested on cost, latency, and efficiency — precisely the dimensions where AMD, custom in-house silicon, and new entrants are aiming. That is why the inference transition, not silicon share, is the load-bearing question.

Rubin is Nvidia's answer to that specific attack. The company claims the new platform delivers 30x higher throughput per megawatt and 35x lower token cost than the prior Grace Blackwell Ultra generation — moving the fight onto the cost axes that inference rewards. Its stated reasoning is that one fungible system spans the whole lifecycle, from data preparation through agentic inference, which it says can require 15 to 100 times more compute per interaction than a human user. The thesis is stated, not yet delivered: management expects Vera Rubin to reach roughly 20% of data center revenue by the third quarter of fiscal 2028 as supply grows. The per-gigawatt ladder, by contrast, is already priced and shipping.

Beyond the GPU

The platform reading also shows up in the pieces that are no longer chips. Networking was already a $31 billion business in the fiscal year that ended in January, and Spectrum-X Ethernet grew 2.6x year over year — enough for Nvidia to describe itself as the largest and fastest-growing network company in the world. By the first quarter of 2026, per IDC, it had become the No. 1 vendor in data center Ethernet switching, taking 21.5% of a market it held under 4% of two years earlier.

Software is the other leg, and it is where the multiple historically gets set. On this call Nvidia detailed the revenue share behind its NeoCloud partners: a take-or-pay floor that lets them borrow, plus a share of their rental revenue above it. Jensen Huang's phrase was that Nvidia now gets paid twice — once on the hardware sale, again through usage — which he calls a recurring stream layered on top of a one-time equipment purchase. It is early, framed as "billions over the medium to long term," but the direction is the argument.

The diversification extends to who is buying. Data center revenue split into $49 billion of hyperscale and $40 billion from everything else — AI clouds, enterprise, industrial, sovereign — the latter up 138% year over year, faster than the giants. Huang described these customers as the half of the business that is "invisible to everybody," because they need a whole factory, not parts. Sovereign AI alone roughly tripled.

What the platform is worth — and what it costs

Here is where the reframing earns its keep, and where it stops. At a market cap near $5.5 trillion, Nvidia trades at roughly 18x trailing sales and 27x EBITDA on the trailing year — expensive on absolute terms, but the platform framing is why the market keeps the multiple propped up. It is also why the multiples alone no longer argue for a penalty: this is being priced as an infrastructure business, not a chip cyclical.

Yet calling it a platform does not remove the risk; it moves it. Supply commitments more than doubled, to $279 billion, up 135% sequentially, and management has been explicit that most of it is memory. That is the footprint of a company that now buys the bottleneck rather than just the parts — a wager on demand that is also a fixed-cost obligation if memory pricing or demand turns. It fits a company guiding to roughly 70% growth in fiscal 2028 and describing itself as supply-constrained, not demand-limited.

My judgment: Nvidia was always going to be judged on whether it can migrate value from hardware into a software-and-recurring layer as the workload shifts toward inference. The per-gigawatt ladder says it is doing exactly that, generation after generation. The open question is not the direction — it is the cost, timing, and delivery risk loaded into a ramp this big. A company at this size priced for platform economics is a different risk than a chip stock, and it is not priced like one. It is priced to keep delivering the migration every quarter.