Nvidia stood on a stage in Las Vegas last week and told everyone with a straight face that a $549 graphics card is going to perform like the $1,599 card from two years ago. I've had the CES keynote replaying in my head for days now, mostly because I can't decide if I'm annoyed or impressed, and honestly it might be both.
For anyone who skipped the livestream (I watched it at 1am with a bowl of cereal, which is not a flex, just what my sleep schedule looks like during CES week): Jensen Huang came out in the leather jacket, same as always, and ran through the whole RTX 50 series lineup. RTX 5090 at $1,999 with 32GB of GDDR7. RTX 5080 at $999. RTX 5070 Ti at $749. And then the RTX 5070 at $549, which is the one everyone's actually talking about, because Nvidia is claiming it matches an RTX 4090 in performance.
Let that sit for a second. The 4090 launched at $1,599 and street prices during the worst of the scalping were way north of that. Now Nvidia says a card less than half the price does the same job. If that's true without any asterisks, it's a genuinely wild jump in two years.
Except there's an asterisk, and it's a big one. That performance claim leans entirely on DLSS 4 and something they're calling Multi Frame Generation, which is the GPU generating extra frames using AI rather than actually rendering them the traditional way. So it's not raw horsepower matching a 4090, it's a software trick making the output look and feel roughly comparable in supported games. That's a meaningfully different claim than what the headline number implies, and I think a lot of outlets ran with the "4090 performance for $549" line without sitting on that distinction long enough. I get why Nvidia wants the narrative to be simple. I just don't think it's a straightforward apples to apples comparison, and I wish the marketing said so a little more plainly instead of burying it in a footnote font size.
The dual flow-through cooler on the 5090 Founders Edition is legitimately clever looking though, I'll give them that. Air pulls straight through the middle of the card instead of just across the fins, which on paper should mean better thermals in a smaller footprint. Whether it actually holds up under a sustained load in a normal case with normal airflow, as opposed to a case a hardware reviewer built specifically to make it look good, is a different question and one I don't think we'll have a real answer to until reviewers get units in hand ahead of the January 30th release date.
There was also Project DIGITS, the $3,000 personal AI box built around their Grace Blackwell GB10 chip, small enough to sit on a desk, apparently able to run models up to 200 billion parameters locally. That's the part of the keynote I found more interesting than the gaming cards if I'm honest, mostly because it's aimed at a completely different kind of buyer, researchers and tinkerers who don't want to rent cloud GPU time by the hour. I don't have $3,000 lying around for that, obviously, but the idea of a shrunk-down local AI workstation you just plug in is a bigger shift than another gaming GPU refresh, even if it'll get a fraction of the coverage.
I'm still running an RTX 3070 in the desktop under my desk, bought used off someone in 2022 after getting completely priced out during the actual 3000-series launch chaos, and I remember writing about that mess on here at the time, refreshing Best Buy's stock page at 6am like it was a part time job. So I'm not exactly the target market for a $1,999 flagship. But the 5070 at $549 is the first RTX 50 card that's made me actually open a new tab and start doing math on whether I can justify it, DLSS asterisks and all. Ask me again in February once actual independent benchmarks are out and not just Nvidia's own slide deck, and I suspect the real story will be more interesting, and more mixed, than the keynote let on.