EN
13 November 2026 · Jakub M.
Nvidia without Jensen Huang: what it means for local AI and hardware prices
Nvidia founder Jensen Huang has stepped down as CEO. For anyone running AI on their own hardware, something else matters more: GPUs and AI workstations are getting more expensive mainly because of a memory shortage, not because of a change at the top.
What happened
On Sunday, 8 November, Nvidia’s board announced an independent investigation into its chief executive. In an internal message to employees, Jensen Huang admitted to an extramarital affair. According to US media, the case also concerns his relationship with an executive at Megaspeed, a company based in Singapore. According to these reports, US authorities suspect this executive of ties to Chinese intelligence; China rejects the allegations. Megaspeed was one of Nvidia’s large foreign customers and bought nearly 5 billion dollars’ worth of chips and other hardware from it.
The board accepted Huang’s resignation on 11 November, effective immediately, and the company announced it a day later. CFO Colette Kress is serving as interim CEO, Harvey C. Jones has become chair of the board, and the search for a permanent successor is under way. Nvidia’s shares lost about 17% over five days and are at their lowest level in more than a year.
What stays the same for now
The CEO’s departure changes nothing about how today’s graphics cards, drivers and software tools, including CUDA, work. Anyone running AI models on Nvidia hardware today has no immediate reason to change anything.
Analysts agree that the business itself is not changing in the short term either. Cody Acree of StoneX says Nvidia’s product roadmap is laid out through the Feynman generation in 2028, and he sees no impact on revenue until then. Raymond James calls the market reaction overdone. Goldman Sachs expects Nvidia to honour its existing commitments, but says new large partnerships will be harder to negotiate without Huang. According to Digitimes, Taiwanese server makers expect near-term demand to hold steady.
Jensen Huang’s departure is the biggest leadership change at Nvidia in more than three decades. For people choosing hardware to run their own AI models, however, it comes at a time when the market is difficult enough already. Memory prices are rising, powerful graphics cards are hard to find, and more and more users are looking for alternatives to classic GPU builds.
The bigger issue: memory shortages
The current rise in the price of local AI hardware is driven mainly by a shortage of memory chips. According to the Wall Street Journal, AI data centres will use about 70% of high-end DRAM production this year. Memory prices rose by about 50% at the end of 2025, and TrendForce analysts expected a further increase of around 70% this year. Manufacturers are already selling capacity for 2027 and 2028 and do not expect noticeable relief before 2028. Other factors, such as demand and manufacturers’ pricing, also feed into the final prices.
Memory is crucial for local AI. Its capacity decides how large a model fits on a machine, and its bandwidth how fast the model responds.
What local AI hardware costs today
| Device | Memory for AI | Price | Type of price |
|---|---|---|---|
| GeForce RTX 5090 | 32 GB (graphics) | $1,999 | recommended price |
| GeForce RTX 5090 | 32 GB (graphics) | ~$6,000 | online marketplaces, November 2026 |
| Nvidia DGX Spark | 128 GB (shared) | $6,950 | Nvidia list price since October 2026 (at launch $3,999) |
| Nvidia DGX Spark | 64 GB (shared) | from $4,999 | Nvidia partners |
| AMD Ryzen AI Halo | 128 GB (shared) | from $3,999 | AMD, June 2026 |
| Mac Studio M5 Ultra | 256 GB (unified) | €12,429 | Apple price list, Germany |
Graphics memory on a card is much faster, but small: a model that does not fit has to be split or slows down sharply. Shared and unified memory is slower, but holds far larger models. The two figures are therefore not directly comparable.
Graphics cards have been hit particularly hard. The RTX 5090 disappeared from major US retailers back in January. According to GPU Poet, it sold for around $6,000 on online marketplaces in November, roughly three times its recommended price. According to Chinese sources, card makers are raising prices again in November, by almost $500 for the RTX 5090 and 5080.
Compact AI workstations with shared memory are the most affordable route to large models, but they are getting more expensive too. Nvidia’s 128 GB DGX Spark has risen by 75% since launch. AMD sells its own Ryzen AI Halo computer with 128 GB from $3,999. This spring, Apple quoted delivery times of up to four to five months for Mac Studio models with more memory.
What our own test showed
We tested for ourselves how far shared memory can take you. A Mac Studio M5 Ultra with 256 GB of unified memory ran the 4-bit version of Qwen3.5 with 397 billion parameters. It used about 224 GB of memory and generated around 62 tokens per second, without any change to macOS settings. With a 128,000-token context, speed dropped to 44 tokens per second. A model like this would not come close to fitting on a single graphics card. The details are in our Qwen3.5 397B test.
What could change in the coming years
Huang’s departure adds uncertainty mainly in the longer term. Much depends on who will lead Nvidia: an engineer would probably keep the current course, while an outside manager might push harder for short-term results. The Megaspeed case has also raised the question of whether Nvidia chips reach China through other countries. Stricter export controls would mainly affect large customers and do not concern home users for now. Meanwhile, AMD and Apple keep strengthening support for local AI tools, and any stumble by Nvidia gives them an opening.
The first big test comes on 17 November, when Nvidia reports its quarterly results. The focus will be on orders and on what the company says about supply.
What it means for local AI
There is no reason to panic. The hardware you have keeps working, and Nvidia’s new products remain on the roadmap. Buying under pressure does not pay off, though: an RTX 5090 costs about three times its recommended price today, and the memory shortage will not go away soon.
It also pays not to bet everything on one platform. llama.cpp runs on Nvidia and AMD graphics, on Apple Silicon and on the CPU alone. MLX was built for Apple Silicon but now supports CUDA as well. vLLM needs a different backend for each platform. Switching platforms is therefore not free, but it is possible.
Try it before you buy
With hardware costing thousands of euros, it makes sense to check first whether your model and your workload run fast enough. You can rent the Mac Studio M5 Ultra with 256 GB that we tested on for as little as one day.
Sources: NVIDIA Form 8-K (SEC), ABC News, CBS News, TipRanks, Investing.com, Digitimes, Windows Central (WSJ, TrendForce), GPU Poet, Dataconomy, TechRadar, Implicator, Digital Citizen, BGR. Prices as of 13 November 2026.