The Second Cold War—or the First AI War?

18-11-2026 | Insights

for GPU

Nvidia RTX 5090 cards are sold out on the official website, while prices elsewhere have skyrocketed to absurd levels. Another example: the Apple Mac Studio with 512 GB of RAM (for a cool $25,000) has become virtually impossible to order. Customers trying to buy one are met with a “page not found” error. And with flagship hardware vanishing from shelves, even budget components are seeing price hikes.

When Jensen Huang was signing autographs for fans, he knew exactly where the market was heading. And he has every reason to make gamers less of a priority.

Jensen was signing autographs for fans

The entire semiconductor industry is reshaping itself around AI, leaving the consumer market as an afterthought. Think of it like a quiet town where a wealthy developer just moved in. Nobody is outright kicking the locals out—familiar trails simply get gated off, the woods are suddenly declared private property, and a members-only club pops up where the public beach used to be.

Why all the fuss, and where did all the graphics cards go?

Someone else needed those chips, and they were willing to pay top dollar. What does a business choose when faced with that kind of dilemma? Money, of course. Apparently, the offers were so lucrative that literal spies were sent to approach Nvidia’s CEO to negotiate supply deals.

When analyzing a giant like Nvidia, you shouldn’t be counting mistresses or chasing tabloid headlines—you need to look at the financials. A P/E of 19.41, a forward P/E of 9.75, and profits that have more than doubled year-over-year make Nvidia an attractive play no matter what.

Of course, Jensen’s departure will impact the company’s future, though not overnight. He is irreplaceable when it comes to executing product launches and creating entirely new markets. Long-term investors and the board were foolish to accept his resignation. It seems far more likely that Washington politics is the real driver behind his exit.

real driver behind his exit

Hundreds of thousands in the tech community have rallied behind Jensen, including Elon Musk and Tim Sweeney. There are even reports that Jensen himself might be open to returning. This is precisely why you shouldn’t trade on the news: much of it is just noise designed to push a stock price one way or the other.

According to The Information, around 240,000 export-banned Nvidia accelerators—worth roughly $7.5 billion—made their way into China through firms tied to Alice Huang (the alleged spy). Counting other channels alongside the Huang case, an estimated 1.5 million GPUs were smuggled in.

To put that in perspective, legal shipments to OpenAI alone in 2025 are estimated at 4 to 5 million GPUs. Gamers, meanwhile, got roughly 44 million GPUs that same year. Sure, that sounds like a lot of cards. But in terms of revenue for fiscal 2026 (Feb 2025 – Jan 2026), Nvidia’s gaming division brought in $16 billion, while data centers dwarfed it at $193.7 billion.

AI Is Hungrier Than We Think

To make AI cheaper, developers are trying to cut down on the amount of compute required. They rely on adapters, quantization, distillation, more efficient architectures, and pretrained models.

When building their own models in Russia, teams followed a pragmatic principle: replicating what works is cheaper than inventing from scratch—and requires fewer Nvidia chips. Instead of hunting for novel architectures, they build on proven foundations. For instance, training Alice AI LLM, initialized with Qwen3-235B weights, took far less time (and fewer GPUs) than training from random weights. Similarly, the first version of GigaChat was essentially a fine-tuned version of GPT-3.

This is a textbook example of a well-known economic phenomenon (often called the Jevons paradox): making a resource more efficient to use can actually drive up its overall consumption.

AI experimentation has hooked far more than just big corporations. Open-source projects are racking up hundreds of thousands of stars on GitHub, maintained and pushed forward by thousands of independent developers. And they all need hardware.

While most gamers get by just fine with 32GB of RAM, running AI locally makes you want 128GB—and even 256GB is hardly the limit. That’s why memory is so expensive right now. Everyone is buying up everything in sight.

What Can We Do?

Almost nothing. Just enjoy the ride.

The worst part is that consumers have virtually no way to opt out of this arms race. You can’t simply refuse to use AI—you’ll immediately fall behind at work, in school, and in almost every other field. What we can do, however, is refuse to accept or pay for low-quality AI content:

  • Don’t buy games where the core creative elements—artwork, music, and story—are AI-generated.
  • Don’t buy products that slap on AI features purely as a marketing gimmick.
  • Don’t support companies that trade real quality for automated generation.

We can only dream that future generations of home PCs will gradually edge closer to 512 GB of RAM, paired with graphics cards offering well over 8 GB of VRAM. That would unlock entirely new horizons for both gaming and running AI locally.

Right now, even 64 GB of RAM in a standard consumer PC is out of reach for many people—and when it comes to GPUs, the picture is even bleaker.

Still, there’s comfort in knowing that everything can flip overnight. For instance, if local AI models become so cheap to run that paying for cloud subscriptions stops making sense—and the bubble finally bursts.