This last week started with three clowns announcing a $500B “Stargate” project that they not only lacked funding for, but that structurally made no sense whatsoever. It ended with half of Silicon Valley in panic mode after suddenly waking up from their comfortable slumber to discover that their DeepSeek nightmare was actually a brutal reality.
Let me begin by explaining why the era of AI charlatans ends by looking at what DeepSeek needed to develop a superior model to the most recent OpenAI one:
- Training Costs: The official training of DeepSeek-V3 reportedly cost around $5.57 million to $6 million. This includes the use of approximately 2.78 million GPU hours on NVIDIA H800 GPUs, which took about two months on a 2048 GPU cluster.
- Additional Research and Experimentation: Beyond the final training run, there were costs associated with prior research, and ablation experiments on architectures, algorithms, and data. Estimates suggest these additional efforts could have cost between $10 to $15 million.
- Team Costs: The development team, which included 139 authors as noted in the technical paper, would have incurred considerable personnel expenses. Annual team costs could exceed $15 million, depending on the duration of the project and salary scales.
- Data and Model Distillation: There were also costs related to using outputs from other models for training purposes. For instance, spending on OpenAI model inference for distillation was speculated to be around $5 million or more, although this is less certain due to the nature of the model’s training data.
Now, let’s look at what the major (hyper-inefficient) hyperscalers spent instead in 2024 ALONE to build massive “AI data centers” stuffed with GPUs that the whole world now knows will be completely unnecessary:
- Amazon:
- 2024: Bought 196,000 GPUs.
- Estimated Total Spending: Given the lack of specific pricing, but knowing that high-end GPUs like the H100 cost around $30,000 to $40,000, Amazon’s spending would be in the range of billions of dollars. However, exact figures are not detailed, but considering their purchase volume, let’s estimate around 5-10 billion USD.
- Microsoft:
- 2024: Bought 485,000 GPUs.
- Estimated Total Spending: Microsoft’s spending would be significantly higher due to the sheer number of GPUs. With the same pricing estimate, their spending could be around 15-20 billion USD.
- Google:
- 2024: Bought 169,000 GPUs.
- Estimated Total Spending: Google’s expenditure would be in the billions, likely around 5-8 billion USD based on the number of GPUs and assuming similar pricing.
- Meta:
- 2024: Bought 224,000 GPUs.
- Estimated Total Spending: Meta’s investment would also be in the range of 7-10 billion USD.
- Tencent:
- 2024: Bought approximately 230,000 GPUs.
- Estimated Total Spending: This would place Tencent’s spending at around 7-9 billion USD.
- Baidu:
- 2024: Bought approximately 230,000 GPUs.
- Estimated Total Spending: Similar to Tencent, Baidu’s investment would be around 7-9 billion USD.
How much will these thousands of GPUs be worth once the whole world realizes only a tiny fraction of them will be used? A figure not far from zero.
Hold on a second, weren’t these GPUs also used as collateral to borrow money to buy more GPUs by companies like CoreWeave? Yes, they were: insert article.
Imagine the huge financial hole about to blow open on this front too when these companies find themselves: IS MAGNETAR CAPITAL BLOWING AIR INTO THE NVIDIA (PONZI) SCHEME SO THEY CAN BET ON ITS IMPLOSION?
- With no revenues because their data center business (that so far only existed on papers full of unrealistic projections) will never begin operations
- Carrying billions in debt borrowed from Private Equity firms and Banks to buy thousands of overpriced GPUs posted as collateral now worth pennies on the dollar
Last but not least, how much will those tens of billions of USD investments made in startups, mostly spent to buy GPUs, be worth now that whatever they were developing can be effectively replicated in the blink of an eye, for a fraction of the cost and then made open source for everyone to use?
Summing it all up, we are looking at hundreds of billions of dollars in CAPEX, and Goodwill has just gone puff. Yes, I know the hyperscalers never made money with their AI businesses even before DeepSeek surfaced, but at least they could pretend they would make it in the future. How will this impact their overblown valuations when the rest of their businesses haven’t meaningfully grown above the level of inflation for several years? Assuming the P/E ratios go back to their long-term average, we are talking about a 50% loss of value, believe it or not.
Of course, the biggest loser among all will be Nvidia, which spared no effort for years to convince everyone they had “overwhelming” demand for their GPUs for many years to come – an effort both in terms of PR and in terms of inflating its revenues with the biggest revenue round-tripping scheme in history. I’m not going to spend more ink today on this matter since I’m excited to talk about what comes next; you can find an abundance of details in my archive here.
Alright, why then am I happy to say the real era of AI finally begins? This situation isn’t very different from what happened more than 20 years ago when the DotCom bubble burst. Similar to today, at that time the biggest companies in the space built up a massive and eventually unnecessary infrastructure to build and run Internet businesses.
As a consequence, the vast amount of cheap computing power allowed very small companies to enter the space and build sustainable and scalable business models. Some of these companies later became the trillion-dollar titans we are so familiar with today: Amazon, Google, and Meta. However, twenty years later, the leaders in the space eventually ended up making the exact same mistakes that allowed them to become what they did, and now rest assured there will be more and more nimble and hungry companies like DeepSeek starting to pop up like mushrooms, thanks to the massive and unnecessary AI data center capacity built.
This will result in a race to the bottom in price cuts to ensure they attract companies willing to use it rather than leaving it idle (something that will compound the write-off losses in data center providers’ financial statements). Overall, this is a great win for society, while the likes of Sam Altman, Masa Son, or Jensen Huang effectively tried to build a tight oligopoly from which they would have immensely profited at the expense of governments and their citizens, who would have instead been scammed out of hundreds of billions to invest in totally useless data centers and completely illogical AI projects from a business perspective. Thank God there will no longer be any credibility for the likes of Sam Altman running around trying to fundraise 7 TRILLION USD to splurge in building GPU data centers: Sam Altman Seeks Trillions of Dollars to Reshape Business of Chips and AI
JustDario on X | JustDario on Instagram | JustDario on YouTube