RTX 50 Super and RDNA 5: What the GPU Rumor Mill Actually Knows in July 2026

Rumor coverage is where hardware reporting goes to get sloppy, so this roundup follows one rule: every claim below is labeled with who said it, when, and how much weight we think it deserves. Nothing in this article is confirmed by NVIDIA or AMD unless explicitly marked as such. As of publication, neither company has announced any of these products.

With that said, there is a lot moving under the surface of a GPU market that looks frozen from the outside, and the two stories that matter are NVIDIA's on-again, off-again RTX 50 Super refresh and AMD's next architecture, which its board partners increasingly treat as a 2027 event.

NVIDIA: the Super refresh saga

The idea of a Blackwell mid-cycle refresh has been circulating for about a year, long enough to have died and revived at least twice. The sequence, with dates, tells the story better than any single leak:

In November 2025, reports circulated that the Super cards might be cancelled outright because the 3 GB GDDR7 modules they depend on had become too valuable during the memory shortage, as Tom's Hardware recapped in June.
Earlier in 2026, NVIDIA itself shot down rumors of a CES 2026 announcement, and TechPowerUp reported that board partners still had no final specifications, with NVIDIA telling them the plan was not cancelled but postponed.
On June 5, 2026, the leaker MEGAsizeGPU claimed the refresh is "back on track" and added a new SKU to the list: a 12 GB version of the RTX 5060, possibly branded RTX 5060 Super, per Tom's Hardware's coverage. The same leaker still expects a 2026 launch.
Barely a week later, on June 14, TechTimes reported the series slipping to 2027, and VideoCardz's sources now point at CES 2027. TechPowerUp has separately reported no Super cards this year, with the RTX 60 series also pushed back.

When leaks contradict each other inside ten days, the honest reading is that NVIDIA itself has not committed to a date, and the date will be decided by memory economics rather than by engineering readiness.

The specs, at least, have been stable across leak waves. The table below is the rumored lineup as compiled by Tom's Hardware; treat every cell as unconfirmed.

Rumored card	Die	VRAM	Bus	CUDA cores	TGP
RTX 5080 Super	GB203	24 GB GDDR7	256-bit	10,752	415 W
RTX 5070 Ti Super	GB203	24 GB GDDR7	256-bit	8,960	350 W
RTX 5070 Super	GB205	18 GB GDDR7	192-bit	6,400	275 W
RTX 5060 12GB (Super)	GB206	12 GB GDDR7	128-bit	unknown	unknown

The pattern is plain: core counts barely move against the existing 5080 and 5070 Ti; this is a VRAM refresh built on 3 GB GDDR7 modules. Which is also exactly why it keeps slipping: TechTimes reports those modules cost about three times as much as standard 2 GB chips, in a market where memory prices are already the story of the year, as we covered in the DDR5 price surge. A refresh whose only selling point is more of the most supply-constrained component in the industry is a refresh with a scheduling problem.

AMD: quiet 2026, real architecture in 2027

The AMD side produces fewer headlines but currently reads as more coherent. The strongest signal came out of Computex, where Tom's Hardware spoke to board partners who expect desktop cards on the next architecture, called RDNA 5 by some and UDNA by others, in the second half of 2027. Partner estimates ranged from Q2-Q3 2027 to as late as early 2028, and the same conversations pointed to no new AMD consumer GPUs in 2026 at all.

The architecture rumors, older and thinner but consistent, sketch a return to the high end that RDNA 4 skipped:

TechPowerUp reported a top part with 96 compute units on a 384-bit bus, a mid-range 64 CU part on 256-bit, and an entry 32 CU part on 128-bit.
Later reports put the flagship at 12,288 cores with 36 GB of GDDR7, aimed at whatever NVIDIA's next flagship turns out to be, and Notebookcheck relayed claims that the silicon has taped out on TSMC's N3P node with a mid-2027 launch hinted.
The same architecture is rumored to power the PlayStation 6 in 2027, which matters because a console contract makes the 2027 silicon timeline more believable than a purely desktop roadmap would be.

If those threads hold, AMD is skipping the refresh games entirely: nothing new for a year and a half, then a genuine node-jump generation (N3P) with a flagship die for the first time since RDNA 3. That is also the first rumored product on either side that could produce the kind of generational uplift this market has stopped delivering.

How much to believe, claim by claim

Our confidence ratings, which are editorial judgment and nothing more:

Claim	Source and date	Our confidence
A Super refresh exists in NVIDIA's plans	Multiple independent reports over a year, incl. partner-facing statements	High
Super cards are a VRAM refresh (3 GB GDDR7, similar cores)	Consistent across all leak waves since 2025	High
Super launches in 2026	MEGAsizeGPU, June 5; contradicted by multiple reports within days	Coin flip at best
RTX 60 series in 2027	TechPowerUp and TweakTown, thin sourcing	Low to medium
No new AMD consumer GPUs in 2026	Board partners at Computex, via Tom's Hardware	Medium-high
RDNA 5 / UDNA desktop cards in H2 2027	Same partner conversations, multiple partners	Medium-high
96 CU / 36 GB GDDR7 flagship on N3P	Single leak lineage, repeated across outlets	Medium
PS6 on the same architecture in 2027	Notebookcheck-relayed claims	Medium

What this means for anyone holding money

Three practical conclusions, assuming the rumor consensus above is roughly right.

First, nothing coming in the next twelve months changes the performance ceiling. The Super refresh, whenever it lands, moves VRAM rather than compute; a 5080 Super with 24 GB is a better-balanced card than the 16 GB original, but it will not be meaningfully faster. Anyone waiting for the next real jump in GPU performance is waiting for 2027 on both vendors' rumored timelines.

Second, VRAM-limited buyers are the one group with a genuine reason to wait. If the thing holding a purchase back is 16 GB on the 5080 or 12 GB on the 5070, the rumored refresh addresses exactly that, and buying the outgoing card at today's inflated street prices right before a memory-doubled replacement would sting twice. The risk is the schedule: that fix is somewhere between six and eighteen months away depending on which leak proves right.

Third, do not price in rumored cards. Memory costs pushed this refresh around the calendar for a year, and the same force sets its street price if it ships. A 24 GB card built from the most contested modules in the industry, in the market conditions our Market Watch tracks weekly, is unlikely to sell anywhere near whatever MSRP gets announced. Plan purchases around cards that exist.

The honest caveat

Everything in this article is rumor except where a company statement is explicitly cited, and rumor coverage ages badly by design. The claims carry their source dates for exactly that reason; the newest information here is from mid-June 2026. If NVIDIA or AMD announces any of these products, the announcement supersedes every table above, and this article will get a dated correction note rather than a silent rewrite.

Sources: