If DeepSeek wiped out $1 trillion of public tech company value—Just watch what happens to private AI company valuations…

Anthony Perkins and The Cryptonite Weekly Rap Feb 1 ∙ Preview

READ IN APP

Chinese panda street art.

Who will end up holding the bag of AI stocks when this market pops? 

During the Internet bubble era (1993-2023), we created $5 trillion of new wealth and then destroyed it. Driven by 200+ Internet companies that should never have gone public (or got funding at all), NASDAQ skyrocketed to 5,048 by March 10, 2000, blew up, and then punishingly twirled down over a two-and-a-half year period and bottom-out at 1,114 on October 9, 2002.

Before the pop, the VCs who bet early on the Internet were able to push their portfolio companies' stock out into the public markets largely into the hands of overly whooped-up retail buyers. Between September 1999 and July 2000, VC firms like Kleiner Perkins and Sequoia Capital and their networks of angels cashed out approximately $43 billion to the public markets (when a $billion was BIG money💰!)—twice the return rates of the no so bad two previous years—and claimed big wins to their limited partner investors. In retrospect, the VCs won big, and the retail investors were caught holding the bag.

Nvidia's half-a-trillion-dollar market cap drop last Monday and the other BigTech stock carnage was an early warning signal that tech companies, both public and private, are overvalued. Public and private stocks and cryptos markets have been over-inflated by the Trump Pump and AI hype. It all begs the question—Who is going to be holding the stock bubble bag this time around? The VCs or the overly enthusiastic retail investors. 

Cryptonite's predictions for 2025—Short Google, Nvidia, & OpenAI?, the Elon/Zuck alliance, Robotaxis takeover, IPOs up... The Cryptonite Weekly Rap and Anthony Perkins · Jan 6 Read full story

Nvidia’s plunge

In our predictions for 2025, we stated that tech stocks, especially AI-related public and private companies, are overvalued and ripe for decompression in price. In 2025, we see a couple of dozen premium IPOs getting out the door, coupled with a brewing M&A frenzy and consolidation of first-generation AI companies. To support our prediction, we suggested that both Nvidia and OpenAI might be good 'shorts.' (We use that word figuratively as we are not stock jockeys by trade). On January 6, the date we posted our predictions, NVDA was trading at $152, and since has dropped to as low as $114 in just the last three weeks…so we are just saying. 🤷🏻‍♀️

One of the biggest differences between the Internet and AI stock bubbles at this moment in history is that 90%+ of the AI companies are still private. During the Web1 bubble, 217 'dot-com' companies had already gone public. If our overvaluation premise is true, then it will be interesting/entertaining to see how the VCs handle this predicament. In their ideal world, public tech investor jitters settle, and they start pumping their portfolio of AI companies out the public door like it was in 1999.

Tip for VCs: When you are up early doing your calisthenics and pounding your blueberry + whey protein smoothies, blast this epic Prince tune, 1999, on your headset to keep psyched up. Even Christian McCaffrey would have a tough time trying to blow through the tiny window that sits in front of you.

The book my brother Michael and I wrote, called The Internet Bubble: Inside the Overvalued World of High-Tech Stocks,' was published by HarperBusiness on November 1, 1999, when NASDAQ was at 2,970. The tech-heavy NASDAQ kept rising to its height of 5,048 on March 10, 2000, so we didn't look so bright for the six months before the pop.😅😅 The interesting thing to note is of the over 50 top VCs (e.g., Don Valentine, Tom Perkins, and Bill Gurly) and Internet entrepreneurs (e.g., Jeff Bezos, Yahoo founder Jerry Yang, and Internet browser inventor Marc Andreessen), we interviewed for the book they ALL acknowledged that Internet stock prices were out of control and not sustainable. Almost none of the Internet stocks were making money, and a high percentage didn't even have revenues.

But as one VC admitted at the time, ‘Are these valuations rational? Not at all. But we either sit on the sidelines and watch everyone else make money or jump in and make money too while the market is hot.’ A top technology investment banker shrugged at the time, saying it is their 'job to put the IPOs on the stoop and see if the dogs will eat it.’ 

The 15 most highly valued private AI companies with venture capital investors. Cryptonite's paid members can see the rest of the list and more data by clicking the link at the very bottom of this post.

What’s is in a value?

Nvidia’s total market value-to-sales ratio (with its stock at $120) is 50x, Apple’s is 8.2x, Meta’s is 11.1x, Intel’s in 1.5.x, and Qualcomm’s is 4.9x.🤔 We couldn’t find reliable sales numbers for the private AI companies in our table, but the question is—Is it possible for each venture-backed AI company to achieve the required revenues and earnings over the next couple of years to support their current VC valuation? We are not so sure.

The primary point of this post is to widely broadcast that since the DeepSeek story slapped Silicon Valley back to reality, we don't believe there is a single high-rolling, AI-betting VC on this planet that isn't pacing their luxury second home floors on Kona or in Montana somewhere feeling powerless. They have been boasting about the rise in valuation of their AI portfolio companies to their Limited Partners for the last two to three years now.

The new reality is that AI VCs may have invested too much money at too high valuations, and their minds are spinning. The only option, before the rest of the world figures out what they now know, is to sell, baby, sell! The problem is that the public markets are too overvalued to open up the IPO doors just yet, and BigTech company buyers know what the VCs know, so they ain't going to buy any GenAI startup at the top.

‘It's going to be an intense year, so buckle up. We've got a lot to do. I'm excited about it,’ says Zuck to the Meta tribe.

What Zuck thinks about DeepSeek

Question: Thoughts on DeepSeek and how Meta should respond or pivot its infrastructure spend to set itself up for success?

Zuck: DeepSeek had a few pretty novel infrastructure optimization advances, which, fortunately, they published, so we can not only observe what they did, but we can read about it and implement it, so that'll benefit us. It's always interesting when there's someone who does something better than you, and that motivates me to make sure we're on it.

—Question from an all-hands Meta employee meeting on Thursday that was leaked and pissed-off Zuck.😤 ‘Everything I say leaks. And it sucks, right?’

DeepSeek and the Open Source AI model

DeepSeek, the Chinese AI startup that has been causing all the stock market commotion, was founded in May of 2023 by hedge fund manager Liang Wenfeng. Mr. Wenfeng is also the founder of High-Flyer, one of China's top quantitative hedge funds, with $8 billion under management. The company's recent announcement of its DeepSeek-R1 model has not only knocked a trillion of market cap out of tech stocks but has also put the company's brand on the cutting edge of AI research. DeepSeek is now a shining example of China's growing capabilities in AI and quantitative finance.

DeepSeek founder and hedge fund manager Liang Wenfeng was motivated to start the AI company by regulatory crackdowns on quantitative trading in China and his desire to pursue more ambitious AI research.

The DeepSeek-V3 base language model foundation uses a mixture-of-experts architecture to handle tasks more accurately and efficiently. The recently introduced DeepSeek-R1, built upon the DeepSeek-V3 foundation, is already matching or outperforming the power of closed model AI bots, including OpenAI's ChatGPT, Meta's Llama, and Anthropic's Claude. DeepSeek-R1 is also 10 to 50 times more affordable and cheaper to run, depending on the task. These power stats and the fact that DeepSeek's model efficiency means needing fewer AI chips are what sent Nvidia's stock tumbling.

The final market shocker was DeepSeek's now debunked report of only spending a mere $5.6 million in model training costs versus an estimated '$100 million to $1 billion' for Anthropic's Claude, according to its CEO Dario Amodei, or Sam Altman's claim that it cost more than $100 million to build OpenAI's GPT-4 model.

Driven by all this excitement, DeepSeek's app rose to No. 1 on the Apple App Store, and downloads soared, reaching 2.6 million across the App Store and Google Play as of last Monday morning.

US and the Might AI Dragoon

Palmer Luckey, Founder & CEO of military AI startup Anduril shows mullets are back in fashion and says don’t listen to CCP propaganda.

"I think the problem is DeepSpeek put out a low development number [$5.6 million to train an AI model] specifically to harm US companies. I don't think that people should take what they're saying at face value. There's a reason they put out the news that way, and if the stock market is any indication, it's accomplishing exactly what they hoped to. We can recognize that Chinese AI is a real competitive threat without losing our minds over it and falling for CCP [Chinese Communist Party] propaganda.'

—Palmer Luckey, Founder & CEO of Anduril, who previously sold Oculus to Facebook for $2 billion in 2014 in an interview with FOX Business.

Editor's note: DeepSeek, like other Chinese technology companies, including TikTok, is subject to content restrictions and data-sharing requirements imposed by the Chinese government. DeepSeek's chatbot self-censors avoid answering questions on sensitive topics like Tiananmen Square, Taiwan, the maltreatment of the mostly-Muslim Uyghur population, and the 'Great Firewall'. That's why we will never use it.

In 2017, Xi Jinping also added the principle of 'civil-military fusion' to the CCP's constitution,' which mandates that all research and intellectual property in China be shared with the People's Liberation Army to use it for what they wish. In other words, all roads lead back to the Pigs running the farm. For all these reasons, we need to keep our eyes on the Mighty AI Dragon, and not use DeepSeek.

Are the VCs funding the noose with which China will hang us? Anthony Perkins · June 22, 2023 Read full story

"I've been reading about China and some of the companies in China, one in particular, DeepSeek, coming up with a faster method and less expensive way to develop AI. And that's good because you don't have to spend as much money. That is a positive, as an asset. If it's a fact, and if it's true, that is very much a positive because we won't be spending billions and billions either, and we will hopefully get the same result. The release of DeepSeek's app should be a wake-up call for our industries, and we need to be laser-focused on competing to win."

—President Donal Trump in an address to House Republicans last Monday, the day Nvidia's stock tanked 17%.

Part of the chatter on the streets of Silicon Valley is that in the long run, the AI winners will not be the companies that produce what will become remarkably fast marginal cost commodity services. The winners will be AI-powered users leveraging the new tools to spend less time producing a lot more work. 🤔

Ethernet co-inventor, company founder, Turing Award winner, Metcalfe's Law namesake, and one of our heroes, Bob Metcalfe.

Moore's and Metcalfe's Laws meet AI development

Moore's Law dictates that technology hardware gets cheaper and faster over time. Metcalfe's Law dictates clever companies can grow exponentially with a little help from their friends. In the next few year's, cheaper and faster AI chips, the use of synthetic data and more efficient training techniques, and the open-source AI model will dramatically reduce AI development and operating costs.

All hype aside, DeepSeek's much cheaper and more efficient open source AI strategy is the real story we should be dissecting. DeepSeek's open source approach is fostering a collaborative and sharing community that drives code enhancements, makes fixes and offers new solutions—all at accelerated rates and a fraction of the cost of closed models. This is where Metcalfe's Law (V=n(n−1)/2) comes in, because the value of DeepSeek's platform's value will goes up not only in relation to its number of active users, but also via the rise in the size of its core open source development community helping to power its software innovation and accelerate its decentralized growth model.

DeepSeek's open source strategy also offers several other value-building, cost-efficiency advantages over its closed source rivals, such as eliminating hefty licensing costs and and the ability to offer lower API fees. For instance, DeepSeek charges $0.55 per million input tokens and $2.19 per million output tokens, compared to OpenAI's $15 and $60 respectively. This dramatic cost reduction makes open source AI companies more accessible to high-volume users and resource-constrained organizations. Read more about the additional performance and other advantages of open source AI here.

In our view, the biggest significance of the DeepSeek splash is highlighting a trend that suggests open source AI communities will play an increasingly significant role as AI innovators, accelerators, and cost cutters, spurring open access to advanced AI capabilities for the betterment of all.

Tulipmania reached its zenith on February 5, 1637, at an auction in Alkmaar, Netherlands, where a single bulb of a rare variety sold for 5,200 guilders—five times the price of a modest urban house at that time. Street art: Kaiya with Tulips by @nilsrva

AI Bubble Epilogue

The bottom line is that the AI bot business will evolve into a highly valued, commodity-priced service delivered with significantly less development and operating costs compared to those paid by first-generation GenAI companies. Just as stocks can't escape P/E ratios, AI companies can't escape Moore's and Metcalfe's Laws.

We lean with our boy Palmer Luckey, the founder of Anduril, about DeepSeek doing a DeepFake on its cheap training numbers. For one, the DeepSeek foundation reportedly used 2,048 H800 GPUs for its training, which alone costs between $50-100 million. At best, this ploy was part of a very successful marketing stunt worth a ton of app downloads, or at worse, part of a CCP plot to control more of our personal information. We sadly think it's both.

No matter how deep the conspiracy goes, it can not be denied that the Chinese now have a highly competitive AI platform driven by open source development and a model that challenges the financial models of ALL the private AI companies listed in the table above, as well as Nvidia, Meta, and many other companies in the AI space. 

Web1