In today’s special edition, I’m answering your questions on DeepSeek—the Chinese artificial intelligence (AI) startup that shocked the world.
Over the weekend, news broke that DeepSeek’s AI model—a groundbreaking rival to OpenAI’s ChatGPT—was reportedly developed in under two months for less than $6 million. That’s 90%+ cheaper than similar models developed by major US tech companies.
(Technically, the news broke in December—we’ve known about DeepSeek since last year. But over the weekend, it became the #1 downloaded app, alerting the investing community at large.)
Parts of this narrative are true. Other parts are greatly exaggerated. But it sent the markets—specifically AI stocks—into a tailspin. Nvidia (NVDA) shed nearly $600 billion.
Let’s get to your questions. I’ll share my thoughts on Nvidia at the end.
Q: Why did investors panic about DeepSeek? Did it really only cost $6 million?
The US is not as far ahead of China in AI as many thought. That’s one reason for the market reaction.
The more important reason is DeepSeek was supposedly trained at a fraction of the cost of leading US AI models, like ChatGPT and Claude.
It is definitely true that DeepSeek is built on innovative training techniques, plus efficient use of hardware. Specifically, it employs parts of the model called “experts” that only wake up when they’re needed, which saves time, efficiency, and money.
BUT—the $6 million number is extremely misleading. If you read the DeepSeek paper, it’s clear these costs are only for the official final training run. They exclude all other expenses, like the costs associated with prior research, experiments, architecture, algorithms, and data.
So while it cost just $6 million to train the model, they probably spent hundreds of millions of dollars on it prior.
Q: Does this mean big tech’s infrastructure spending spree is misguided?
Investors are worried the DeepSeek news means we need less chips and less AI infrastructure. If its model is really that efficient, do the big tech companies really need to spend historic sums of money on data centers, chips, and associated infrastructure for AI?
I do not believe this is a legitimate worry. Keep in mind, AI models were already getting more efficient at a rapid pace. The cost for GPT-4o level intelligence, which is the latest GPT model, has dropped 1,000X in the last 18 months.
The key difference here is that these efficiency gains came from China. A big surprise, for sure—but it’s really just a continuation of what’s been happening already.
|
Most new tech goes through this process. Early versions are expensive. Then costs drop rapidly, which increases demand. As AI gets cheaper to develop and run, more organizations will use it. We’re still going to need A LOT more AI infrastructure.
Q: Is it bad for the US that China has emerged as a top AI powerhouse?
It depends on how we react. Competition is great for technological advancement. We’re nearly neck and neck with China, and we’re going to have to keep innovating and pushing the boundaries to win this race.
It’s all hands on deck. It means even faster innovation and more spending. Probably with renewed support from the government.
Q: Does DeepSeek mark the bursting of the AI bubble?
No. There’s no AI bubble. Did you know Nvidia’s stock is trading below its five-year average P/E ratio?
But I do think investors will become more discerning with their AI investments, which is a good thing. It introduced some uncertainty and doubt around the AI theme, which was needed. AI stocks had been going up in a straight line.
This is an opportunity to scale into some of the world’s best companies at a better price. We’re still very early in this AI story. The selloff will look like a blip a year from now.
Q: Regarding Nvidia... it essentially has a monopoly on top-level GPU chips. The thesis was that demand for these chips was unlimited. Has that changed?
A mysterious part of this story is that DeepSeek claims it achieved its breakthroughs using older, slower chips. It has to say that, because it’s not supposed to have Nvidia’s latest and greatest chips, as exports to China are banned.
I put a 95% probability on the fact that China is not telling the truth about this. All signs point toward DeepSeek having access to a lot more powerful GPUs than it is letting on. My contacts tell me DeepSeek likely had access to 50,000+ cutting-edge Nvidia chips.
Q: Is Nvidia a buy today?
If you’ve been on the sidelines wanting to invest in NVDA, this is a good opportunity.
Of course, it is not the same kind of opportunity as it was in 2018, when we first recommended it. Nvidia is a giant tech company now, which means there are limits on how fast it can grow. It’s not going to gain 1,000% again.
For that, seek smaller, disruptive AI stocks—of which there are plenty. Get my new 10-bagger research here for how to find them.
But Nvidia is still the clear leader of the AI boom. It still has the best chips on the market.
Q: Where will the most rapidly growing AI stocks come from?
If DeepSeek showed us anything, it’s that a small upstart can disrupt the plans of much bigger companies. That’s been an accelerating trend for a while. But with AI rapidly improving, there are many, many opportunities to buy stock in smaller companies that grow quickly as they disrupt, transform, or create new industries.
That’s why my team and I have been scrambling to create something new to help you take advantage of this opportunity.
As Step 1, we used AI to study every stock that went up 1,000%+ in the last 20 years. We will use our findings to identify today’s small, fast-growing stocks that are set to appreciate by 1,000% or more.
Get our research here. By claiming this free report, which I wrote with RiskHedge business partner Dan Steinhart, you’ll be in the front of the line to hear more about this project very soon.
Exciting times.
Stephen McBride
Chief Analyst, RiskHedge
