If you've been shopping for a flagship GPU lately, I'm sorry. The market is doing that thing again where it stops making sense, and this time even Nvidia's own launch price looks like a distant memory.
The RTX 5090 situation is officially out of hand
TechRadar's latest look at the RTX 5090 market shows the cheapest card available in the US now sits above $6,000, with third-party listings wandering past $9,000. Launch MSRP was $1,999. Let that sink in for a second.
The driver? AI server builders reportedly buying the flagship by the pallet — a practice Nvidia's own driver license has prohibited since 2017, funnily enough. Before you panic, the photos backing that story may be AI-generated, so treat the "pallet" part as a rumor with a side of skepticism. Unverified industry rumor, for reference only.
Meanwhile, the 12V-2x6 power connector saga refuses to die. XDA points out that RTX 5090 cables are still melting under the near-600W TDP, and Nvidia apparently had a fix years ago that it didn't roll out. I've built with these cards; I triple-check the connector click every time. Honestly, waiting for that little plastic latch to seat properly should not feel like defusing a bomb.
Workstation memory gets weird (in a good way)
On the pro side, Nvidia quietly dropped the RTX PRO 5500 with 84GB of GDDR7. The trick: 28 memory chips split 14 above and 14 below the PCB — a "memory sandwich" with a 416-bit bus. It looks bizarre, but cramming that much VRAM on one board is exactly what local LLM folks have been begging for. To be fair, the 416-bit bus leaves me a bit puzzled about the memory bandwidth math, but for AI workloads that just need capacity, this thing talks.
Games are actually getting kinder to your GPU
Two game stories this week, and both are good news for normal people. Gears of War: E-Day system requirements are out, and 4K@60FPS with upscaling lands on an RTX 5070 Ti — not the flagship everyone expected. Unreal Engine 5 games have been a GPU tax for years, so this is a nice reversal.
Tom's Hardware also put Control Resonant through its paces across 28 GPUs. Path-traced Manhattan is brutal, and it leans hard on DLSS 4.5. Keep this in mind: the gap between "runs on paper" and "actually smooth with ray tracing" is wider than ever.
One more: RPCS3, the PS3 emulator, just got an up to 25% FPS boost on NVIDIA GPUs thanks to a driver bug fix. Emulation nerds, go update your drivers.
Small machines, big numbers
Honor revealed the Tiangong AXB35, a roughly 2-liter workstation with an Nvidia RTX Spark, 128GB of unified memory and a 2TB SSD at 1.5kg. It's a Mac Studio rival for people who want local LLMs without the Apple tax. Intel and AMD versions are coming too. Microsoft is also teasing an October 7 event covering Windows, Surface, and RTX Spark — local AI computing is clearly the theme of the season.
The industry is shifting under our feet
Piper Sandler claims Amazon, Apple, AMD, Google, Tesla, Microsoft, NVIDIA, and Qualcomm are all evaluating Intel's 14A process. That's a heavy list, and it's analyst chatter rather than confirmed contracts — Unverified industry rumor, for reference only. But the logic is sound: TSMC capacity is reportedly choked until at least 2028, and Intel is the only onshore alternative that could scale.
Also worth a shrug: San Francisco is banning new data centers while Santa Clara cashes in $29.5M a year from them. And AMD +6% / Intel +8% after the Fed's rate cut shows investors are finally looking past NVIDIA.
DIY corner: ROG goes full tower
Asus ROG announced the Cronox ARGB full-tower case and the Eurux GR120 fans. Big EATX chassis, ARGB everything — if you're planning a build around one of those giant GPUs, a full tower suddenly makes sense again. Those 5090s need airflow, and this is a case that takes airflow seriously.
By the way, if you are budgeting a build around these absurd GPU prices — Decision Calculator is handy for weighing the monthly cost against performance. No sign-up, just the numbers.
