How DeepSeek Nuked Nvidia Again?

Liang Wenfeng, the founding CEO of DeepSeek

On February 21, Nvidia’s stock—after a brief rebound—nosedived straight to hell again. In just six days, $300 billion vanished from its market cap. That’s because DeepSeek announced its Open Source Week on the 21st, and investors around the world knew Nvidia was about to face another round of nuclear strikes. Unsurprisingly, a revolutionary technology that blew up one of Nvidia’s key AI training moats is published on the first day and is free for all.

For years, Nvidia convinced the entire AI industry that smarter AI required ever-increasing computing power—and that Nvidia’s newest GPUs would always crush previous generations. This created a vicious cycle: Every product launch forced AI companies to empty their wallets to upgrade hardware, terrified of falling behind rivals. This “forced obsolescence” strategy fueled Nvidia’s meteoric rise to a $3 trillion valuation.

Then DeepSeek changed the game.

Most AI firms use two types of Nvidia chips: the budget H800 and the premium H100 (twice as powerful but $10k pricier per unit). But DeepSeek’s new FlashMLA technology works like magic—it turbocharges the “outdated” H800 to handle long-text processing 10x faster, while even outperforming H100 chips in reasoning tasks by 2x.

Here’s the kicker: FlashMLA doesn’t require new hardware. It’s like discovering untapped superpowers in existing GPUs. Imagine computing power as water supply: Traditional algorithms are like delivering water with fixed-capacity buckets—you waste tons of resources even for tiny tasks. Worse, simple requests hog all the bandwidth while complex jobs wait forever.

FlashMLA flips this with smart “computing faucets.” It breaks down text tasks, dynamically allocating resources based on complexity—a trickle for simple sentences, a firehose for brain-melting logic. This “precision watering” system saves resources and speeds up workflows dramatically.

The fallout? Two existential threats to Nvidia:

First, AI firms are cancelling their GPU shopping sprees. Reuters reported on the Feb 25 that Microsoft already started axing data centre leases. Even utility companies powering these centres saw stock prices tank.

Second, Nobody’s buying Nvidia’s incremental upgrades anymore. Unless Huang can deliver a 10x performance leap (physically impossible under Moore’s Law), the upgrade treadmill is broken.

Some users of NVIDIA’s high-performance RTX 50 series GPUs report that the power connector has melted

Meanwhile, Nvidia’s consumer GPUs are becoming a meme. Their flagship RTX 50 “Blackwall” series keeps failing spectacularly—melting power cables, stability glitches, and missing hardware features. Since February, driver updates haven’t fixed anything. The official forum’s top “solution”? Cap your screen refresh rate at 60Hz. Imagine buying a Lamborghini only to be told “Don’t drive over 50 mph—the engine might explode.” Early adopters are fuming.

Here’s the irony: On January 13, former US president Joe Biden tried to cripple China’s AI growth by blocking H100 exports. He successfully forced Chinese AI companies like DeepSeek to squeeze every drop of performance from older chips—ultimately undermining Nvidia’s entire business model. Talk about poetic justice.

And this is just Day 1 of DeepSeek’s “Open Source Week.” More disruptive announcements are coming. Let’s see how many walls of America’s AI fortress can survive this siege.

Editor: Charriot Zhai