Jensen Huang’s keynote at CES 2026 carried a lot of information, but what’s actually worth chewing over isn’t some flashy compute-performance slide — it’s that NVIDIA is systematically reshaping the economics of AI.
If you only take away “Blackwell is 10x stronger than Hopper, and Rubin is another 10x on top of that,” that’s just the surface. The real core is that AI is shifting from a compute race to a cost race.
1. The shift from “performance” to “token cost” is a deliberate change in narrative#
What Huang kept coming back to in this keynote wasn’t FLOPS — it was a metric that’s rarely been discussed publicly over the past year: the cost of generating a single token.
Under the Rubin architecture, token cost gets compressed to one-tenth of Blackwell’s.
That matters far more than “another generation of chips got stronger,” because better performance doesn’t equal commercial success, while lower cost equals expanded use cases. Only when token cost keeps falling can AI actually move from being a tool for a handful of big model companies to things like enterprise-wide deployment, AI agents running as a matter of course, long-context and real-time inference, and running multiple models in parallel or personalized to the user. In other words, this isn’t “stronger AI” — it’s cheaper, scalable AI.
2. NVIDIA isn’t just “selling chips” anymore — it’s selling a complete system#
Huang said it outright: NVIDIA is now building the entire system — AI is a full stack. That statement carries at least three layers of meaning.
NVIDIA is compressing the middle of the supply chain. The old path used to be: GPU → OEM → cloud provider → developer/enterprise. It’s becoming: GPU + networking + storage + software + scheduling + reference architecture → customer directly. NVIDIA isn’t content staying in the middle of the chain anymore — it’s raising ecosystem lock-in and switching costs.
NVIDIA’s real competitors have changed. It’s no longer just facing traditional chip rivals — it’s up against cloud providers’ own in-house compute stacks, model companies’ hardware-software co-design, and national-level, self-sufficient compute programs. As NVIDIA goes full-stack, the entire industry is forced to ask itself: should we reduce how systemically dependent we are on NVIDIA?
Full-stack is a moat, but it’s also an amplifier of systemic risk. The upside is obvious — stronger ecosystem stickiness, deeper customer lock-in, bigger scale effects. But it also means heavier capital expenditure, stronger cyclical exposure, and a much bigger dependence on energy, policy, and the macro environment.
3. The conditions for “10x times 10x” to hold up are actually quite demanding#
For this logic to hold, several hidden assumptions all need to be true at once.
AI demand has to keep expanding exponentially. Rubin’s whole premise rests on the assumption that AI training and inference demand will keep scaling up over the next 2–3 years. If enterprise ROI falls short of expectations, application-layer innovation slows down, or agents don’t take off the way people expect, then “10x more compute” could temporarily turn into structural oversupply.
Energy and data centers aren’t a side issue — they’re a core constraint. Huang repeatedly mentioning “energy efficiency” wasn’t just politeness. In the real world, the actual bottlenecks are power supply, data center approvals, energy prices, and policy constraints. If energy costs can’t fall in step, the whole token-cost argument weakens noticeably.
Success itself is fueling a move away from NVIDIA. Once NVIDIA’s capabilities get strong enough, the industry’s natural response is for cloud providers to push harder on in-house development, for national systems to emphasize independence and self-sufficiency, and for model companies to optimize their own hardware-software co-design. That’s not a short-term negative — it’s a long-term, structural tug-of-war.
4. Looking at the “material downgrade” rumors through this same lens#
If you put CES’s messaging alongside the recent rumors about high-end PCB materials possibly getting “downgraded,” the two turn out to be logically consistent. If what NVIDIA keeps emphasizing is cost rather than the absolute limits of performance, then it’s not surprising if some high-end materials that overshoot on performance but underdeliver on value get systematically swapped out. That’s not necessarily bad news — it looks more like the natural result of system-level cost optimization.
5. My take: this is a long-term logic, not a short-cycle safe bet#
Putting it all together, my read is this: it’s a logically complete, clearly targeted, genuinely persuasive long-term industry narrative — one that’s trying to push AI from a capital-intensive race toward becoming universal infrastructure.
But it’s just as important to stay clear-eyed: this logic depends heavily on demand, energy, and policy all lining up together, and if any one key variable gets out of sync, the correction will be systemic too.
NVIDIA is attempting something extremely difficult — and if it succeeds, something that will genuinely matter historically. This isn’t a short-term safe bet. It’s a long-term call that requires time, patience, and the ability to stomach volatility.
