Alright, so this China-only RTX 4080M thing, it’s a real Frankenstein card. It’s using laptop GPU silicon, specifically the AD103 chip, on a desktop PCIe board. This is happening because of US export restrictions on high-end NVIDIA cards like the RTX 4090, which pushed Chinese manufacturers to get creative, you know, find workarounds. The card is called the RTX 4080M, not an official NVIDIA product, so no official driver support or warranties, obviously. But it’s out there, and it’s being tested.
Bilibili user 杰某 (Jie Mou) did some benchmarks. The 4080M scored 18,614 in 3DMark Time Spy. That’s not bad, but the big deal is the power draw. A regular desktop RTX 4080 can pull up to 320 W.
An official NVIDIA RTX 4080 mobile GPU can draw a maximum of 175 W TGP. This unofficial 4080M, though, it only used 100 W during testing. That’s a significant power reduction. The price point for this 4080M in China is around 2,700 to 2,800 RMB, which is roughly $400 or £300, due to component shortages. That makes it comparable to NVIDIA’s RTX 5060 Ti or AMD’s RX 9070 GRE in that market.
In gaming tests, it actually held its own against the RX 9070 GRE, even getting 10 more frames per second in Delta Force at 4K Ultra settings, and 100 more frames in PUBG at 2K Ultra. Both cards, the 4080M and the RX 9070 GRE, have 12GB of VRAM.This whole situation, it just highlights the impact of those export controls. It forces markets to adapt, to innovate in unexpected ways. We saw similar things with modified RTX 4090D and RTX 4080 Super cards appearing in China with double the video memory, up to 48GB and 32GB respectively, for AI training models.
These modifications likely use workstation-series PCBs, because a standard 4090 only supports 12 memory modules. It’s a painstaking process, usually, for a single card, so how cloud computing companies are doing this at scale, that’s the question. And speaking of NVIDIA, the company is making moves, big moves, in AI. On July 6, 2026, NVIDIA confirmed its product roadmap, including the Kyber rack-scale architecture for the upcoming Rubin Ultra chips, is on track for a 2027 release. Wait, no, actually, the Kyber rack-scale architecture, designed for the 2027 Rubin Ultra chips, has been delayed to 2028 due to manufacturing challenges with a printed circuit board.
Research firm SemiAnalysis said this PCB midplane is causing the holdup. This delay, it could create an opening for rivals like AMD and Google in the high-end AI market. NVIDIA even scrapped a fallback design that would have bolted together two current-generation racks because cloud service providers objected to the operational burden. Also on July 6, 2026, NVIDIA announced the BioNeMo Agent Toolkit. This toolkit, it’s designed to transform general AI into specialized, autonomous research assistants for life sciences, like biology and chemistry and genomics.
It leverages over a decade of NVIDIA’s accelerated computing libraries and allows developers to turn large language models into scientific research partners. Nearly 50 partners are already adopting this platform, including Eli Lilly and Snowflake. And there’s more, NVIDIA also announced the NVIDIA Vera Rubin platform. This platform, it’s ramping into full production, with Taiwan’s top server makers manufacturing Vera Rubin-based systems at scale. It’s designed to power agentic AI factories worldwide, delivering 10 times agent throughput at scale compared to the previous Grace Blackwell platform.
The Vera Rubin platform combines Rubin GPUs, Vera CPUs, DPUs, and advanced NVLink interconnects. It’s targeting partner availability and volume shipments in the second half of 2026. And a record 35 new NVIDIA AI HPC supercomputers in Europe, that was announced on June 22, 2026, not July 6, 2026, my bad, so many announcements. These supercomputers are being developed across 23 countries, representing about 800 AI exaflops of deployed or announced capacity. This is Europe’s largest one-year expansion of supercomputers, aiming to equip over 3 million researchers. Man, the market is just, it’s a lot.
I bought NVDA back on January 15, 2024, at $56.30 a share. I’m holding until it hits $300, or if the AI boom shows any real signs of slowing down, then I’m out. Today, July 6, 2026, NVDA closed at $196.33. It’s been a wild ride.