Sora Ignites a New AI Arms Race
Lately, semiconductor and AI momentum has driven a broad rally across the major stock markets.
What's fueling this surge isn't just Nvidia's blockbuster earnings — it's the arrival of a genuinely killer application for all that compute: OpenAI's Sora, an image-and-video generation AI. So far we've only seen demo clips, since the technology's potential social impact is too large to open up to the public just yet. But the demo alone was enough to leave jaws on the floor. It really is remarkable.
Blazing-Fast Video AI Technology Sparks a New Market Frenzy
There's already no shortage of articles dissecting how Sora works, so I'll keep this brief. In essence, OpenAI combined the Transformer architecture — the technology that has proven itself the backbone of this entire AI revolution — with diffusion models, another AI workhorse. Transformers handle understanding content; diffusion is the basic tool for generating images. The real breakthrough, though, is something called spacetime patches: a technique that chops three-dimensional space and time into small chunks. By working with these chunks — each carrying both spatial and temporal information — Sora can generate what used to be extraordinarily complex video content in a coherent, logical sequence.
Why does this matter so much? If you've played with AI image generation, you know how fast the technology has advanced — today's models handle still images quite well. So why not just generate video frame by frame, like a flipbook?
In practice, that's exactly the problem: a still image never has to reckon with how objects relate to each other in space over time. Ignore those relationships and the temporal dimension, and video generation quickly falls apart — the result looks more incoherent than a Dalí painting. Spacetime patches solve that problem. Of course, having a solution on paper and actually being able to execute it are two very different things — and in the end, it still comes down to OpenAI's staggering compute budget.
Before Sora, two video-AI startups, Runway and Pika, were doing quite well and had raised plenty of venture funding. Now OpenAI has essentially delivered them a devastating blow.
Runway has been in the AI video business for a while now; Pika is a buzzy Stanford spinout. Both companies' generated clips have been hits on Twitter/X, and both ship usable products. But Sora's demo reel is operating on an entirely different level from anything these smaller players have shown — a brutal illustration of what it means to get outclassed by an order of magnitude.
Video AI is likely to be one of the biggest consumers of compute over the next two years. At some point, video's share of total compute demand should ease off — but generalized AI (AGI) is still waiting in the wings. If generating a single video already eats up an almost unimaginable number of chips, imagine what AGI will require; the chip count becomes essentially incalculable. Some might say: why not wait until the best models get open-sourced, then buy the chips and leapfrog everyone else? It's a tempting, seemingly clever strategy — but apparently no one in Silicon Valley's top ranks is actually betting on it.
Semiconductor Chips: The Key to Who Actually Wins Big
Big Tech isn't foolish. At today's pace of technological turnover, falling behind once means falling off the ride entirely — you never even get to the day the good models go open source. So that "wait and leapfrog" strategy simply isn't realistic.
The recent video-AI battle already gave us the answer: once a stable, deep-pocketed giant offers something as powerful as Sora, who's going to bother with a small startup's free version that maxes out at ten-odd seconds of footage? Of course you wait for the big player's reliable, polished $20-a-month subscription instead. Which means any tech giant that doesn't keep buying chips gets bounced out of the AI game immediately.
So will Nvidia and AI-related chipmakers just keep climbing forever? One thing we know for certain: no stock on Earth rises forever, so ups and downs are inevitable. But we're still in the very early innings of the AI revolution, and the core of that revolution is the compute that semiconductor chips provide. Sure, Nvidia could someday fade away too — but short of someone finding a path completely detached from today's technical trajectory, an insatiable appetite for compute is simply unavoidable.
The world's major tech giants have already seen this coming. The mainstream approach is still to go to Nvidia and buy up whatever chip supply you can get — it's become such a seller's market that customers are practically lining up to beg Jensen Huang for allocation. The other approach is to build your own chips; Google is still pursuing that path, and OpenAI's Sam Altman has floated a trillion-dollar chip-manufacturing plan of his own. But that's still at the planning stage — building fabs on the back of trillions of dollars in spending takes a long time to pay off even if it succeeds.
Video AI's impact has already hit film and TV investment in its first wave: Hollywood director Tyler Perry halted his reported $800 million studio expansion project.
In an interview, Perry said the best thing about generative video is that you no longer have to travel to shoot on location — a computer can conjure up a mountain cabin or a lunar landscape just as easily. At the same time, he's deeply worried about the jobs this will wipe out. The film and TV industry hasn't settled on how to respond yet, but since Sora hasn't been released to the public and still has some technical kinks to work out, there should be a bit of time left to figure out a response.
Yizaihui's New President: Kuo Su-hui, the Iron Woman Who Revived Weizhi Twice