At GTC 2026, Charlie Boyle stood on stage and walked through Vera Rubin, the platform NVIDIA announced earlier in the year, framing it around three words the company now repeats like a mantra: productive, durable, fungible. That’s the pitch. Not FLOPS, not transistor counts. Three adjectives aimed squarely at whoever signs the purchase order.
As someone who spends most of his time testing tools and then writing down what actually happened, I find this interesting for a reason that has nothing to do with silicon. NVIDIA has stopped selling speed and started selling accounting.
What the three words actually mean
Strip the marketing away and NVIDIA’s argument is that an AI factory earns its keep on three axes. Productive means more tokens per megawatt, since power is the real constraint in a modern data center, not rack space. Durable means the hardware stays commercially useful for years rather than becoming a tax write-off. Fungible means one cluster can serve training, inference, and whatever workload shows up next quarter, instead of being purpose-built for a model architecture that’s already going stale.
It’s a tidy framework. It’s also a framework NVIDIA gets to grade itself on, which is where my reviewer instincts kick in.
The numbers, and what I’d want to see before trusting them
NVIDIA says Vera Rubin NVL72 delivers more than 30x higher throughput per megawatt than GB300 NVL72, and up to 45x lower cost per million tokens on DeepSeek V4 Pro. Those are enormous multiples. They’re also vendor-measured, on a vendor-chosen model, with “up to” doing the heavy lifting in the second figure.
I’m not calling them wrong. Generational jumps in this category have been real. But I review things for a living, and the pattern with numbers like these is consistent: the headline multiple comes from a configuration tuned specifically to produce that multiple. If you’re evaluating this for your own stack, the questions I’d ask are:
- What batch size and sequence length produced the tokens-per-megawatt figure?
- Is the 45x cost comparison against GB300 at list price, or against a street price someone could actually negotiate?
- Does the throughput hold on your models, or only on the one named in the benchmark?
- What’s the power envelope you need to even reach the comparison baseline?
None of that invalidates the claim. It just means the claim is a ceiling, not a floor, and the gap between those two is where most buyer disappointment lives.
The durability argument is the strongest part
Here’s where I think NVIDIA has the better case, and it’s the part getting the least attention. The A100 shipped in 2020 and remains in commercial service. CoreWeave has extended bookings on its fleet through 2029. Every major operator has pushed out the depreciation schedule on its servers, which is a quiet but meaningful signal — depreciation schedules are set by finance teams who get in trouble for being wrong, not by marketing teams who don’t.
That matters more than any throughput multiple. If a GPU bought in 2020 is still earning revenue in 2026, the total cost of ownership math changes completely. You’re not amortizing over three years, you’re amortizing over five or six, and the hardware’s value curve looks more like commercial real estate than like a laptop.
This is also the hardest thing for a competitor to replicate. Raw performance can be matched by whoever ships the next good chip. Six years of proven software support, driver continuity, and a framework ecosystem that still targets your older parts cannot be matched on a product cycle. The A100 staying useful is a CUDA story more than a transistor story.
Fungible is the word I’d watch
Of the three, fungibility is the one that’s hardest to verify and most valuable if true. The risk in buying AI infrastructure right now isn’t that it’ll be slow. It’s that model architectures shift, inference patterns change, and the cluster you bought for one shape of workload turns out to be awkward for the next one. A platform that handles training and inference and whatever comes after is insurance against a bet you can’t make confidently.
But fungibility is a claim about the future, which means nobody can test it today. You can measure tokens per megawatt in an afternoon. You can confirm durability by looking at what’s still racked and billing. You find out whether your infrastructure was fungible about two years after you needed to know.
My read
NVIDIA has built a genuinely useful framework for thinking about this spend, and then filled it with numbers only NVIDIA can produce. Take the durability evidence seriously — it’s external, it’s financial, and it comes from operators with no incentive to flatter anyone. Treat the 30x and 45x figures as the best possible outcome rather than the expected one. And ask hard questions about fungibility, because it’s the claim carrying the most weight and the least proof.
Good framework. Verify the inputs yourself.
đź•’ Published: