Surprise! Nvidia’s GeForce RTX 50-series GPUs cost less than you thought

It’s official: Nvidia’s next generation of gaming graphics is here, and friends, the GeForce RTX 50-series looks pretty compellingly priced on paper. Nvidia CEO Jensen Huang announced eight different GeForce RTX 50-series graphics cards during his CES 2025 keynote – four for desktop, and four for laptops, all compatible with a new DLSS 4 generation.

The same GPUs were announced for both form factors: The GeForce RTX 5090, 5080, 5070 Ti, and 5070. It was a big difference from AMD’s keynote, where RDNA 4 and the Radeon RX 9070 weren’t even mentioned despite press receiving a high-level briefing.

Let’s start with Nvidia’s hotly anticipated desktop graphics cards, powered by Nvidia’s new Blackwell architecture. They’re actually cheaper than expected, with lower prices than their predecessors…except for the RTX 5090. But there’s a reason for that.

Meet Nvidia’s RTX 50-series desktop graphics cards

I was convinced that the GeForce RTX 5090 would cost $2,500 or more based off the leaked specs. Well, Jensen didn’t really get into product level specs, but at $1,999, it’s a relative bargain for AI researchers, if not necessarily gamers. With 32GB of VRAM and a beefy 512-bit bus, paired with Nvidia’s flagship Blackwell GPU, this thing will be ferocious in gaming, but unparalleled in machine learning tasks if you can’t afford a pricey Nvidia enterprise-class card.

The theme goes down the line: At $999, the RTX 5080 costs $200 less than the 4080 did at launch (and the same as the RTX 4080 Super). At $749 and $549, the RTX 5070 Ti and RTX 5070 cost $50 less than their predecessors (though to be fair, I called the RTX 4070 Ti “hobbled and wildly overpriced”). Better yet, Jensen said that the RTX 5070 delivers RTX 4090-class performance for roughly a third of the price.

If that’s true in games and not just AI workloads or benchmarks, it’s an exciting start to the new generation – but be warned that these starting prices for Nvidia’s overhauled Founders Edition versions may not reflect the same price you’ll see on custom third-party cards. Nvidia’s announcement post for the 50-series, published after the keynote, includes some high level performance details for each desktop GPU, but only in full ray tracing and DLSS-enabled situations. Since we have no tangible idea what sort of performance benefits DLSS 4 brings, take these with a bunch punch of salt for now. It shows what’s possible though.

Nvidia says the RTX 5090 is “2X Faster Than The GeForce RTX 4090;” the RTX 5080 is “Twice As Fast As The GeForce RTX 4080;” the RTX 5070 is “2X Faster Than The GeForce RTX 4070 Ti;” and the RTX 5070 is as fast as the RTX 4090. Again, only expect to reach those high-water marks in specific games and scenarios that support the full suite of Nvidia ray tracing and DLSS features. The RTX 5090 offers wayyyyyy more CUDA cores than the 4090, so it should be a beast no matter what, while the lower-tier RTX 50-series cards have more modest CUDA upgrades, so Blackwell’s architectural changes (and DLSS 4) may need to do some heavy lifting if Nvidia plans on offering big performance uplifts. We’ll see!

Look for RTX 50-series graphics cards to start hitting the market later this January in some form. Nvidia didn’t say which cards are launching when.

Nvidia GeForce RTX 50-series laptops

Usually, Nvidia launches desktop GPUs well before their laptop counterparts, but not with the RTX 50-series! Well, kinda. They’re launching in March, while the desktop cards launch this month, but Jensen revealed both onstage.

Details were scant beyond what you see in the screenshot above. The GeForce RTX 5090 will be available in laptops starting at $2,899; the RTX 5080 in laptops starting at $2,199; the RTX 5070 Ti in laptops starting at $1,599; and the RTX 5070 in laptops starting at $1,299.

Pay close attention to the cited AI TOPs under the model numbers for both the laptop and desktop GPUs, though. Laptop graphics tend to be cut-down from their desktop cousins, and the AI TOPS suggests that’s the case with the RTX 50-series too. The RTX 5090 desktop card offers 3,400 AI TOPS while the laptop version hits 1,850 – about the same level as the desktop RTX 5080. (No surprise there, as the laptop 4090 is basically a desktop 4080 stuffed into a notebook.) The same trend continues down the laptop 50-series line.

Update: Nvidia’s GeForce website provided more specs after the keynote and yep, my initial suspicions were correct. Be sure to also pay attention to memory capacity (especially for the RTX 5070) and bus width size, in addition to CUDA core counts. That laptop 5090 is nowhere near the firepower of the desktop 5090 (though Nvidia wisely increased the memory capacity of the laptop 5090 over the desktop 5080), and the laptop 5070 offers just 8GB of VRAM over a 128-bit bus — a skimpy total that frankly doesn’t cut it at even 1080p resolution in 2025’s latest games, though Nvidia may argue that DLSS 4’s AI benefits reduce memory demands.

What’s inside Blackwell?

That’s all well and good, but how do these puppies perform? Jensen didn’t get into details on-stream, instead relaying some high-level Blackwell specs that matter more for AI researchers than gamers – AI TOPS, RT TFLOPS, “AI management processor,” and the like. You can see it all in the screenshot above.

On the plus side, at least some of Nvidia’s RTX 50-series will include bleeding-edge GDDR7 memory, helping to deliver up to 1.8TB/s of memory bandwidth, or twice what was possible with its predecessor. It will also offer dual shaders for INT and floating point calculations (the big two for traditional gaming graphics), the ability to intermix GPU and AI workloads as needed, and programmable GPU shaders that can process neural network tasks.

If it sounds like Nvidia is going all-in on AI, well, they have been for years. Haven’t you seen DLSS and DLSS 3 Frame Gen? On that note, Jensen also teased DLSS 4, complete with “neural texture compression” and “neural materials” that reduce the need for traditional GPU rendering even further.

Nvidia says that while DLSS 3 could inject AI-generated frames between every GPU-rendered frame, DLSS 4 can infer three full frames off of a single traditional frame. Hey, AI upscaling is killing native graphics, after all. Jensen said that a total of about 33 million pixels are generated for four frames of 4K imagery; with DLSS 4, that’s still true, but the GPU is only rendering about two million of those, with AI doing the rest of the heavy lifting. You can read much more about DLSS 4 and multi-frame generation on Nvidia’s website if you’re interested.

As DLSS 3 users who forgot to turn on Nvidia Reflex can tell you, injecting AI-generated frames also injects latency — though Reflex can claw it back. That’s a much, much more pronounced problem when you’re using AI to create multiple frames. Enter Nvidia Reflex 2, which uses “frame warp technology” to reduce latency by up to 75 percent. It’s complicated stuff, so watch the video and hit that link for an in-depth explanation of the technology if you’re interested in more details.

Bonkers – and potentially very, very game changing if it works as advertised. That’s a big if, though, and there are plenty of questions still swirling around the RTX 50-series’ capabilities after Nvidia’s detail-light keynote. Stay tuned to PCWorld (and PCWorld’s YouTube channel!) for all the latest updates and news coming from CES 2025.