RTX 5090 Review 2026: 31% Faster, $3,000 Street

NVIDIA shipped the GeForce RTX 5090 on January 30, 2025, and eighteen months later the review writes itself differently than it did on launch day. In January 2025 the story was a spec sheet — 21,760 CUDA cores, 32GB of GDDR7, the fastest consumer graphics card ever built. In July 2026 the story is a receipt. The $1,999 NVIDIA printed on the box is a number that now lives mostly in press releases. The card in your cart costs three thousand dollars, often more, and it pulls 575 watts through a connector that a non-trivial number of owners have watched turn brown.

This is a review of both things: the silicon, which is genuinely the best money can buy, and the situation, which is genuinely the worst it has been for a halo GPU in living memory. We will do the numbers — real ones, from GamersNexus, Digital Foundry, TechRadar and DSOGaming, not the "up to 8x" slide from the keynote — then we will talk about why the thing costs what a used car costs, and whether any human who is not training a language model in a spare bedroom should buy one.

What NVIDIA Shipped

A launch that set the tone

The RTX 5090 is the flagship of the RTX 50-series "Blackwell" generation, and it arrived the way flagships now arrive: in a puff of smoke. Founders Edition stock at the $1,999 MSRP evaporated in minutes, board-partner cards listed hundreds of dollars higher on day one, and the restocks never really came. What looked in January 2025 like ordinary launch-day scarcity turned out to be the permanent condition of this product. If you want the longer view on how the card has aged, we wrote a standalone piece on where the 5090 sits in 2026; this review is about the launch verdict and everything that has calcified around it since.

GB202 and the Blackwell die

Under the triple-slot cooler is GB202, NVIDIA's first consumer Blackwell flagship die, fabricated on a TSMC 4nm-class process. It is a big chip — the full GB202 is larger still, and even the 5090 does not enable every SM on it. Against the RTX 4090's AD102, the generational move is more of everything: more cores, a wider memory bus, faster memory, and a meaningfully higher power budget. What it is not is a new manufacturing node. Ada was "5nm," Blackwell is "4nm," and in practice these are close cousins in the same TSMC family. Hold that thought, because it explains almost everything about the performance story.

The reviews landed split

Day-one coverage was not a coronation. Tom's Hardware headlined its Founders Edition review "Blackwell commences its reign with a few stumbles." The Verge's Tom Warren wrote that "a new king of 4K is here," and then spent the rest of the review qualifying it. GamersNexus, TechRadar and DSOGaming all landed in the same neighborhood: fastest card on the planet, yes; a generational leap on the order of the 4090, no. That split — best-in-class product, underwhelming improvement — is the review, and it has only sharpened with time.

The Spec Sheet

More cores, more memory, more bus

On paper the 5090 is a monster, and the paper is not lying about the hardware — only about the price. Here is the generational comparison that matters.

575 watts on the same node

The number that should stop you is 575W of total graphics power, up from 450W on the 4090 — a 28% jump in draw for roughly a 31% jump in frames. That ratio is the whole ballgame. Because Blackwell did not get a real node shrink, NVIDIA bought most of its performance the old-fashioned way: more silicon fed more electricity. GamersNexus measured efficiency as "about the same as the 4090," which is a polite way of saying this is a bigger engine, not a better one. Plan for a 1,000W power supply and a case that can exhaust the heat, and if your card is going to hang unsupported in a horizontal slot, read our guide on stopping GPU sag in 12 steps — this is a heavy slab of metal.

What the sheet won't tell you

Two things the spec table hides. First, the 32GB of GDDR7 is the single most consequential number on it — not for gaming, where 24GB was rarely the bottleneck, but for local AI and content work, where VRAM is destiny. Second, that same memory is why the card costs what it costs, a point we will return to with feeling. The spec sheet describes a gaming GPU. The bill of materials describes an AI part wearing a gaming cooler.

Real-World Performance: 31%

Native 4K, the number that matters

Strip out the marketing and here is the honest figure: in native 4K, the RTX 5090 is roughly 31% faster than the RTX 4090. GamersNexus reported a "27–35% uplift in 4K gaming over the RTX 4090." DSOGaming's John Papadopoulos ran twenty of the most demanding PC games at native 4K and landed on a 33% average — 108 fps on the 4090 becoming 137 fps on the 5090 — with a spread from 23% to 47% depending on the title. His summary was blunt: NVIDIA "raised the price of its XX90 GPU model by 25%, offering 30–40% better performance." TechRadar, testing a broader mix, put the average uplift closer to 21%, rising to "up to 50%" in the workloads that scale best.

Ray tracing and path tracing

The 5090 stretches its legs most in heavy ray tracing, where the extra RT and tensor hardware plus the memory bandwidth actually get to work. DSOGaming measured about +38% in Cyberpunk 2077 and +37% in Black Myth: Wukong at native 4K with ray tracing enabled. The famous "60 vs 58 fps" Cyberpunk result that made the rounds was a native path-tracing corner case where both cards choke on principle — a useful reminder that when you turn every knob to maximum, even a $2,000 GPU renders a slideshow. This is exactly the workload NVIDIA built DLSS for, which is the next section's problem.

The laptop asterisk and the synthetic trap

Two traps for the unwary. First, the laptop RTX 5090 is a completely different animal — power-constrained mobile silicon that reviewers found beating the 4090 laptop part by under 2% at 4K, and occasionally losing at 1080p. "RTX 5090" on a laptop spec sheet is marketing, not the card in this review. Second, be careful with synthetics. Some 3DMark subtests show the desktop 5090 up nearly 49%; a saturated benchmark like PassMark's G3D shows barely 2%, because it tops out at the high end and stops measuring the gap. The honest gaming truth sits in the low 30s, and anyone quoting you 2% or 49% is selling a narrative.

DLSS 4 and 'Fake Frames'

What Multi-Frame Generation actually does

DLSS 4's headline feature is Multi-Frame Generation — the ability to generate up to three AI frames for every one the GPU actually renders, quadrupling the frame-rate counter. It shipped with the cards on January 30, 2025, supported in 75 games at launch, and NVIDIA's own tally grew it past 100 titles by spring 2025, 125 at Computex, and 250-plus by CES 2026. When NVIDIA's keynote says "8x performance," this is the trick: three generated frames plus upscaling stacked on top of the raster number.

The problem with counting fake frames

PC Gamer's verdict on the brute-force raster gain was that the 5090 is "only incrementally faster," and it filed MFG under "fake frames" — the enthusiast shorthand for the fact that generated frames add smoothness but not responsiveness. Interpolated frames do not sample new input, so a 240 fps MFG figure can carry the input latency of the 60 fps it was built from. It looks fluid; it does not necessarily feel like 240 fps. For a single-player path-traced showcase, that trade is often worth it. For a competitive shooter, it is not — a distinction we get into in our piece on whether 240Hz actually beats 144Hz.

The 8K myth and DLSS 4.5

Ignore the "8K gaming without frame generation" milestone; it is the same marketing NVIDIA ran for the 3090 in 2020, and native 8K is a 33-megapixel wall that still needs upscaling to clear 60 fps. What is real is the roadmap: at CES 2026 NVIDIA announced DLSS 4.5, adding a 6x multi-frame mode and a second-generation transformer model that improves image quality at a modest compute cost, shipping spring 2026. Frame generation is now the axis NVIDIA competes on, and every review from here forward has to fight the "up to 8x" number to tell you what the card actually renders.

The Connector That Still Melts

der8auer's clamp meter

In February and March 2025, overclocker and YouTuber der8auer put a clamp meter on the 5090 Founders Edition's power cable and found current wildly unbalanced across the six 12V conductors. The target is roughly 8.3 amps per wire — 600 watts over 12 volts is 50 amps, split six ways. He measured something very different.

# der8auer -- 12V-2x6 per-wire current, RTX 5090 FE (Feb 2025)
# Target: ~8.3 A/wire  (600 W / 12 V = 50 A across 6 wires)

wire 1:   ~2 A   |##
wire 2:   ~5 A   |#####
wire 3:  ~11 A   |###########
wire 4:  ~22 A   |######################  (2.6x spec, one wire)

connector temp, PSU side : 150 C
connector temp, GPU side :  ~90 C
load balancing           : none (3090 Ti had shunts; 5090 FE, none)

The load-balancing regression

The root cause is not the connector geometry — it is that the Founders Edition has effectively no per-pin load balancing. The RTX 3090 Ti, which introduced this 16-pin connector in 2022, used three shunt resistors and could sense and spread current across pin pairs. The 4090 dropped to two. The 5090 FE collapses all six 12V pins into a single sensing domain, so if one wire hogs the load — as der8auer's did, at 22 amps — the card has no idea and no way to react. Worse, this happens on the revised 12V-2x6 connector, the one that was supposed to have fixed the 4090-era melting. It did not, because the melting was never really a connector-geometry problem.

'Not fit for purpose'

Club386's Ben Hardwidge did not hedge: "This socket isn't fit for purpose, and our industry, particularly Nvidia, needs to admit it." PC Gamer's headline was a weary "Surely not again." A 575W card running a connector rated for about 600W has almost no thermal headroom, and GamersNexus described the design as "pushing so close to the limit of the cable spec." There has been no recall. NVIDIA's guidance amounts to: seat the plug fully and use a proper cable. If you own one of these, that advice is not optional.

The $1,999 Fiction

The MSRP is a paper number

Let us be precise about the money, because it is the actual 2026 story. The $1,999 Founders Edition MSRP is real in the sense that the price exists; it is fiction in the sense that you cannot buy the card for it. FE stock sells out instantly and rarely returns, and board-partner cards have settled far above sticker. By mid-2026, industry reporting had the street price of a normal, air-cooled 5090 sitting between roughly $2,900 and $3,700, with some listings pushing past $4,300 in July 2026. NVIDIA itself is reported to have applied an approximately $300 wholesale price increase around May 2026. The polite industry term for a $1,999 MSRP nobody can buy at is a "paper" launch.

AI ate the memory

The reason is not gamer demand — it is that AI data centers have absorbed the world's GDDR7 and DRAM supply, and the 5090's 32GB is caught in the crossfire. Reporting through 2026 tracked the cost of the card's memory climbing from tens of dollars to well over $200, with VRAM alone estimated at over 80% of the card's bill of materials. When memory is most of what a product costs, and memory prices are set by a datacenter bidding war, the retail price follows the auction, not NVIDIA's press release. It is the same squeeze crushing storage — we wrote about how the AI buildout is eating consumer SSD supply too.

The Lightning Z lottery

The absurdist peak of all this arrived at CES 2026: the MSI GeForce RTX 5090 32G Lightning Z, an AIO-cooled halo card with an 8-inch telemetry screen, firmware-selectable 800W and 1,000W power limits, and a separate 2,500W "XOC" BIOS for liquid-nitrogen lunatics. Its official price is $5,090.99 — the model number is the price — and MSI sold its 1,300 units by lottery. It resold on eBay for $6,700 to $8,800, with hopeful listings near $27,000. Tom's Hardware reviewed it and asked the only sane question: "RTX 5090 Ti, anyone?" For roughly 12% more performance at stock, or 18% overclocked, over a Founders Edition, it is a monument to a market that has lost the plot.

The XX90 Price Ladder

From Titan to xx90

NVIDIA's halo tier has a lineage, and the prices tell a story. The Titan brand gave way to the RTX 3090 at $1,499 in September 2020, then the RTX 3090 Ti at $1,999 in March 2022. The RTX 4090 actually walked the price back to $1,599 in October 2022 — a genuine bargain by this tier's standards, in hindsight. The 5090's $1,999 MSRP simply returns to the 3090 Ti high-water mark. Except that the 3090 Ti's $1,999 was a price you could actually pay, and the 5090's is not.

The diminishing leap

The other pattern is generational. The 4090 over the 3090 Ti was a knockout — a real node shrink, huge efficiency gains, frame-rate leaps that occasionally hit 60–70%. The 5090 over the 4090 is a low-30s uplift bought with 28% more power on the same node family. This is what Digital Foundry meant by a modest generational jump, and it is the fingerprint of Moore's Law slowing at the leading edge. When you cannot shrink the transistors, you add more of them and turn up the voltage — and you hand the customer the power bill.

The connector's rap sheet

The melting saga is also not new. The 12VHPWR connector debuted on the 3090 Ti, went infamous on the 4090 in late 2022 when adapters browned, was revised into 12V-2x6 to "fix" it, and then melted again on the 5090. Three generations, one unsolved problem — because, as der8auer's measurements showed, the failure mode is unbalanced load on a connector with no headroom, not a bent pin. The industry keeps treating a systems-engineering problem as a user-error problem.

The Competition

AMD sat this one out

At the top of the stack, there is no competition. AMD's RDNA 4 generation targeted the mainstream and mid-high market, explicitly declining to chase the flagship crown, which leaves the 5090 without a rival at native 4K or in ray tracing. NVIDIA is competing with itself, and it prices accordingly. A monopoly at the halo tier is a large part of why $3,000 sticks — there is no red-team card forcing a correction.

The used-4090 argument

The most rational competitor to a new 5090 is a used 4090. The 4090 is end-of-life in 2026, so its price no longer reflects a $1,599 MSRP — but even elevated, it delivers something like three-quarters of the 5090's gaming performance, and at 1440p the gap narrows further. If you are not gaming at 4K or leaning on the 32GB for AI, the value math points hard at last generation. One Canadian outlet put it plainly: with an $800-plus price gap, the 4090 is the smarter 1440p buy.

Down the stack

If the flagship math is insane — and it is — the interesting cards are lower. The RTX 5080 delivers a large fraction of the experience for a fraction of the price, and even the midrange has its own version of this fight, which we cover in the RTX 5070 versus 4070 in 2026. For the overwhelming majority of players, the right Blackwell card is not the one in this review.

6-12 Month Forecast

Prices stay broken

Here is where I think this goes over the next six to twelve months, into mid-2027.

1. The $1,999 MSRP stays fictional. The GDDR7 crunch is driven by AI datacenter demand that shows no sign of easing before new memory capacity comes online in 2027. Expect street prices to hold above $3,000 and to spike higher on any supply shock. The paper launch is now the business model.
2. No recall for the connector. NVIDIA will keep pushing responsibility onto cable seating and PSU makers rather than redesign the Founders Edition. Board partners will quietly add per-pin sensing as a differentiator, and melting reports will scale with the install base — a slow-burn liability, not a headline event.
3. A refresh will not cut prices. If a 5090 "Super" or a higher-VRAM variant appears, it will use scarce 3GB GDDR7 modules and cost more, not less. There is no version of the current memory market that produces a cheaper flagship.
4. DLSS becomes the whole pitch. With DLSS 4.5 and 6x frame generation shipping in spring 2026, NVIDIA will lean even harder on "multiplied" frame numbers, and the gap between marketing performance and rendered performance will widen. Read every future benchmark for whether frame-gen is switched on.
5. Halo cards get sillier. Expect more Lightning Z-style lottery drops above $5,000 as partners chase margin in a supply-starved market. The $2,000 flagship is dead; the $3,000-to-$5,000 flagship is the new normal.

The one thing that could change it

The single variable that breaks this forecast is memory. If AI capital expenditure cools, or GDDR7 capacity expands faster than expected, the 5090's price could fall toward its MSRP quickly — because the silicon was never the expensive part. I would not bet on it inside twelve months, but it is the pressure valve to watch.

Who Should Buy One

The enthusiast with no ceiling

If you game at 4K, want maximum ray tracing with DLSS 4 headroom, and genuinely do not care whether the card costs $2,000 or $3,500, the 5090 is the fastest thing you can buy and nothing else is close. That is a real customer, and this is their card. Just buy a proper 1,000W PSU, a native 12V-2x6 cable, and seat the plug like your GPU depends on it, because it does.

The AI and creator case

The stronger argument is professional. The 32GB of GDDR7 runs local language models and large 3D and video projects that simply will not fit on the 4090's 24GB — we clocked Llama 70B at roughly 85 tokens per second versus the 4090's 52, a 63% jump that dwarfs the gaming gap. If the card earns money, $3,000 is a business expense with a payback period, and the value calculus inverts entirely.

Everyone else

For everyone else — which is almost everyone — the RTX 5090 is a bad-value trophy. The performance is real and the price is indefensible, and those two facts do not cancel out. A used 4090 or an RTX 5080 will deliver dramatically better frames per dollar, and the money you save buys a very good monitor and a chair you can survive the ten-hour sessions this card enables. The 5090 is the best graphics card in the world. It is also, at $3,000-plus for a gaming machine, the easiest flagship to talk yourself out of that NVIDIA has ever built. Fastest, yes. Worth it, almost certainly not.