Picture a procurement meeting somewhere in Northern Virginia. The CFO has a spreadsheet open. One column is the capital cost of a rack. Another is how many months it stays useful. A third, the one everybody argues about, is what happens to that rack when the model everyone wants to serve changes shape in eighteen months. Nobody in the room is asking which GPU wins a benchmark. They’re asking whether the thing they’re about to buy will still be earning money in 2029.
That meeting is the audience for NVIDIA’s latest pitch, and the company has compressed it into three words: productive, durable, fungible. Higher tokens per megawatt. Longer useful life. Broader workload support. It’s a tidy framing, and it’s aimed squarely at the anxiety that any depreciation-heavy business feels when it buys silicon that might be a museum piece before the lease ends.
The numbers, and what the numbers aren’t
The headline claim is Vera Rubin NVL72 delivering more than 30x higher throughput per megawatt than GB300 NVL72, plus up to 45x lower cost per million tokens on DeepSeek V4 Pro. Those are big multiples. They are also NVIDIA comparing NVIDIA to NVIDIA, on a workload NVIDIA selected, with “up to” doing load-bearing work in the second figure.
I’m not calling it false. I’m saying that a vendor-run comparison against its own previous generation is a marketing artifact until someone else reproduces it. Ask the questions a buyer should ask:
- What precision, batch size, and sequence length produced the throughput figure
- Whether the cost-per-token math includes power, cooling, and interconnect, or just the compute
- How the 45x figure degrades on a workload that isn’t the one chosen for the slide
- What the same comparison looks like against a competitor’s current silicon rather than NVIDIA’s last one
None of that is in the public framing I’ve seen. If you’re signing a nine-figure order, get it in writing from your own test use.
Durability is the more interesting argument
Here’s where the pitch gets genuinely strong, and it has nothing to do with new hardware. The A100 shipped in 2020 and is still in commercial service. CoreWeave extended bookings through 2029. Every major operator has pushed out the depreciation schedule on its servers.
That last point is the one worth sitting with. When operators extend depreciation, they’re making a public statement that the economic life of this equipment is longer than they originally modeled. That flatters quarterly earnings, which is the obvious cynical read, and I’ll take the cynical read seriously. But it also creates real accounting exposure if the assets turn out to be junk sooner than claimed. Nobody extends a schedule casually when auditors are watching.
So the durability story has independent corroboration in a way the throughput story does not. A six-year-old GPU still earning revenue is a verifiable fact about the market, not a benchmark. That’s the kind of evidence I weigh more heavily.
Fungible is doing the heaviest lifting
Of the three words, fungibility is the one that actually describes NVIDIA’s moat, and it’s the one the company talks about in the vaguest terms. “Broader workload support” means a rack bought for training can serve inference, can run recommenders, can handle whatever the next architecture turns out to be. In practice that flexibility lives in CUDA and the surrounding software stack more than in any particular chip.
That’s a real advantage. It’s also the part of the argument that is hardest to audit and most convenient to assert. A competitor can match a throughput number on a specific workload. Matching a decade of library compatibility is a different problem.
What I’d actually tell a buyer
Treat the three-word framing as a decent checklist and a bad substitute for diligence. Productivity claims from a vendor comparing against its own prior generation need your own validation. Durability has outside support, and the A100 still being in service plus operators extending depreciation schedules is a stronger signal than any slide. Fungibility is probably NVIDIA’s most defensible claim and the one with the least published detail behind it.
The strategic read is that NVIDIA has stopped selling peak performance and started selling the shape of a financial model. That’s a sign of a maturing market, not a weakening one. When the pitch moves from FLOPS to return on invested capital, the customer has changed from an engineer to a finance committee.
The part that should keep buyers honest is simple. Every one of these claims is easier to make than to check, and the only numbers NVIDIA is obligated to stand behind are the ones in your contract. Put the throughput targets there, or don’t bother citing them.
đź•’ Published: