How did the AI bubble form?
Inexplicably soared
In the past six months or so, countless listed companies have nothing to do with AI at all, but only issued a press release stating that they have established commercial cooperation with the popular Nvidia, Microsoft, or OpenAI, and the investors are not clear. In this case, the company’s stock price immediately rose sharply the next day, ranging from single digits to dozens of percentages.
What about Taiwan?
The same situation also happened in Taiwan stocks. Jensen Huang of Nvidia and Lisa Su of AMD visited Taiwan separately in the past few weeks. As long as the companies visited by the two of them, the stock prices immediately rose the next day regardless of the reason. The media even Keen to report on the itinerary and tidbits of the two in detail and on the fly.
But the two CEO’s big conversations about Taiwan’s role in industry and the loss of IT competitiveness, or the fact that Taiwan’s role in the global supply chain has been under intense competition and threats from China, Japan, and South Korea, are rarely used to it. The Taiwanese media, which report good news but not bad news, are willing to report on the icing on the cake.
History is not far away
Dotcom bubble
History is not far away. It is reminiscent of the scene before the dot-com bubble burst only 23 years ago and the current AI craze is highly similar. Among the countless losses 23 years ago, even investors did not even know the business content The stock prices of online listed companies can rise several times or even dozens of times in a few months. Listed companies are also bragging and exaggerating the extent to which the company embraces the current popular technological trends, but the Internet 23 years ago is replaced by the current one. AI nothing more.
Buffett’s wisdom
23 years ago, Buffett also warned against the stock market phenomenon at that time. Instead, the mainstream Barron’s Weekly published “What’s wrong with Buffett?” at the peak of the bubble in 1999. Ridiculed him with “Buffett lost his mojo”.
In Taiwan, Wang Yongqing, the leader of Taiwan’s industrial and commercial circles at that time, also openly and loudly expressed his disapproval of the phenomenon that Taiwan-listed companies speculated on the Internet and did not focus on their own business. views of the past.
Wall Street and regulators are warning
Now even Morgan Stanley, which is regarded as the mainstream of Wall Street, has rarely publicly and loudly warned that the current AI frenzy in the stock market has reached the peak of the bubble. This is a rare speech on Wall Street that always encourages investors to enter the market.
It doesn’t matter, even Gensler, chairman of the US Securities and Exchange Commission (SEC), said in an interview with the “New York Times” in August 2023 that the United States is likely to end up with only 2 to 3 large language models (LLB) , the result will deepen the economic system connection and increase the incidence of financial crises, because when a certain model or database becomes mainstream, herding behavior will become more common, and investors will follow the same information and have similar reactions, causing the stock market to fluctuate Happens faster and more often. The point is that Gensler believes that AI will thus be at the center of the next financial crisis.
There will be only a few winners from AI
Why are there only a few winners?
AI is real, but it cannot support the current bubble market. The rapid development of AI in the past few years has indeed far exceeded the sum of the past half century, and it has also achieved great breakthroughs in certain theories, application scenarios, and infrastructure fields. This is an undeniable fact.
But the problem is that the current development of AI has not yet reached the stage where everyone can afford it. Any product or technology that cannot be popularized will not have a market. This is the law and common sense of the capital market. However, fans of the authorities are always easily persuaded by a few successful cases. They are afraid that they will miss the train to get rich in the stock market, causing investors to rush to chase the popular AI targets, causing the rapid formation of this bubble.
AIGC requires large language models
In terms of the development of the most popular field of generative AI (AIGC) since 2023, it is impressive. AIGC can indeed provide breakthrough application assistance for the development of many application fields. But investors forget that AIGC is based on the foundation of AI called LLM. The problem is that the cost of investing in a large language model is too high, and it is still ordinary.
How much does AIGC cost?
Here are some data for your reference:
On a global scale, the only countries with such financial resources and engineering capabilities are the two superpowers China and the United States.
On a global scale, there are only a handful of companies with financial and human resources that can invest for a long time; only Microsoft, Alphabet, Meta, Amazon, and Apple in the United States, and Baidu, Alibaba, Huawei, TikTok, and Tencent in China.
OpenAI, which developed ChatGPT, has announced that the necessary LLM and cloud facilities require at least $4 million in fixed operating costs per day; this does not include high hardware and personnel costs, which is why it needs Microsoft’s $11 billion stake. reason.
The price of Nvidia’s AI chip H100 will be about 20,000 US dollars in May 2023, and the price is increasing every day. The price on eBay in August was as high as 45,000 US dollars. Coweaver even took the company’s existing H100 bank mortgage loan of 2.3 billion US dollars.
For the AI cost, please see my post of “How Much Does Generative AI Cost?“
Nvidia collects AI tax
At present, more than 80% of almost all chips used for AI are manufactured by Nvidia. The advanced chips needed for the new AIGC are rare, and the chip supply has far failed to keep up with the demand. Musk even said bluntly: “GPUs are far more difficult to obtain than drugs at the moment.” Server manufacturers have to wait half a year at the fastest to get Nvidia’s latest chips.
Notable AI startup unicorns
| Company name | Valuation (in $1B) | Famous investors | Main business and product |
| OpenAI | 29 | Microsoft | ChatGPT，Dall-E |
| Scale.ai | 7.3 | Tiger Global、Coatue Management、Founders Fund | Data Annotation |
| Anthropic | 4.4 | Alphabet | Chatbot claude |
| Abnormal Security | 4 | CrowdStrike Falcon Fund、Menlo Ventures | email security |
| Cohere | 2 | SAP, Nvidia | NLP |
| CoreWeave | 2 | Nvidia | GPU cloud computing service provider |
| Hugging Face | 1.8 | Thirty Five Ventures、Sequoia Capital | Open source Chatbot |
| Lightricks | 1.8 | Insight Partners, Viola Group, Hanaco Ventures | Image and video editing |
| Runway | 1.5 | Alphabet | AI generated video Gen-2 |
| Jasper | 1.5 | Founders Circle Capital、Insight Partners | AI generated marketing document |
| Midjourney | 1.5 | No external investors | AI image generator Midjourney |
| Replit | 1.2 | Coatue, SVA | AI generated development tools |
| Inflection | 1.2 | Microsoft, Reid Hoffman, Bill Gates, Eric Schmidt, Nvidia | AI personalized chatbot |
| ADEPT | 1 | General Catalyst, SVA | AI trained to use software and API |
| Character.ai | 1 | Andreessen Horowitz, SVA | AI personalized chatbot |
| Stability.ai | 1 | Sound Ventures、Coatue | AI geneated image, Stable Diffusion |
Table 1: Famous AI start-up unicorns (data from Bloomberg, Pitchbook, Crunchbase)
Only few participants
Microsoft
Even companies with money such as Microsoft and Alphabet-level companies will all complain in the second quarter of 2023, saying that before the business brought by AI is converted into profit (Microsoft’s AI product Copilot will be the fastest). It will take until the second half of fiscal year 2024 to start bringing in a lot of revenue).
Microsoft’s capital expenditure in the second quarter was US$10.7 billion, an annual increase of 37.18%, setting a new high in a single quarter since 2016. Even Microsoft pointed out in its latest annual report that if the GPU in the data center cannot be obtained, Azure may experience cloud service interruption.
And it will also be forced to reduce the company’s future stock repurchase due to the huge capital expenditure of AI. They are all showing to their shareholders that AI will erode the company’s profits and affect the interests of shareholders.
Alphabet
Alphabet’s investment in this area in the past 10 years is estimated to have exceeded 200 billion US dollars, but so far it has not received much appreciation and affirmation from users and investors. Alphabet expects to increase investment in AI technology infrastructure in the second half of 2023, and capital expenditures will be higher than in 2022. Chairman Hannes said on February 23, 2023 that the cost of talking to an AI robot like LLM is 10 times that of traditional search.
Meta
Meta’s total costs and expenses will increase by 10% annually in the second quarter of 2023. Zuckerberg said that capital expenditures will continue to increase as the company will invest in data centers and AI. It already has 760 DGX servers, equipped with 6,080 A100 GPUs and 1,520 AMD EPYC CPUs, and equipped with Nvidia’s Quantum InfiniBand network, which supports Meta with a bandwidth of up to 200Gb/s.
In June 2023, it announced that it plans to build another 16,000 GPUs before July, which will be the world’s largest AI super server after completion. The price of a Nvidia AI chip is tens of thousands of dollars. After all construction costs continue to rise, additional budgets will be required, and this is only the GPU chip in this plan, plus communication chips and related infrastructure. Equipment, we can estimate how much capital expenditure will be required.
Nvidia and Broadcom
However, the AI money pit is as deep as the sea, and it is not something that any company can afford. Most of these necessary capital expenditures flow into the pockets of AI chip leader Nvidia. There will only be a few beneficiaries of AI, such as Nvidia and Broadcom whose main business is network communication infrastructure.
| ticker | 8/14/2023 share price | YTD (8/14/2023) stock return | P/E | P/S |
| NVDA | 437.53 | 205.64% | 227.39 | 43.44 |
| AVGO | 854.1 | 54.31% | 26.72 | 10.22 |
| AMD | 111.98 | 74.91% | 622.11 | 8.47 |
| ANET | 178.12 | 47.32% | 32.88 | 10.56 |
| AI | 33.9 | 206.23% | N/A | 14.77 |
| PLTR | 15.72 | 146.01% | N/A | 16.28 |
| GOOGL | 131.33 | 47.36% | 28.76 | 5.74 |
| META | 306.19 | 145.46% | 36.66 | 6.63 |
| MSFT | 324.04 | 35.25% | 33.46 | 11.43 |
| AMZN | 140.57 | 63.80% | 111.76 | 2.7 |
| BIDU | 137.59 | 15.50% | 25.77 | 2.86 |
| BABA | 93.46 | 1.61% | 20.83 | 1.98 |
| TCEHY | 42.24 | -5.25% | 15.47 | 5.19 |
| 標普500 | 4,489.72 | 17.40% | 23.46 | 2.46 |
Table 2: The stock price performance of AI-related companies in the US stock market from 2023 to August 9 (figures from Google Finance and Yahoo Finance)
There will be fewer than five LLMs in the future
Just like the views of SEC Chairman Gensler, plus the above-mentioned necessary expenditures, various huge resources required, and actual market demand; the mainstream views in the technology industry almost all agree that the result of future competition and elimination , there will be less than five LLMs left in the world.
If there is no surprise, Alphabet, OpenAI, Meta, Microsoft, and Amazon, which have developed and launched their own LLM, have already achieved first-mover advantages and occupy a favorable position. Especially the following two:
- Meta’s Llama2 is considered to be OpenAI’s biggest rival, and the latest second version has been announced as commercial open source, which is free for anyone to use.
- Alphabet is currently devoting all its efforts to develop Gemini, which has been announced by the famous SemiAnalysis to have a computing power five times that of OpenAI’s GPT-4! The main reason is that the number of TPU v5 in Alphabet’s hands is more than the sum of GPUs owned by OpenAI, Meta, CoreWeave, Oracle and Amazon.
I am the author of the original text, and the condensed version of this article was originally published in Smart magazine.
Related articles
- “Economic crisis caused by overinvestment
- “Geoffrey Hinton, 2024 Nobel Physics winner, inadvertently helped Nvida transform to AI overlord“
- “2024 Nobel Chemistry Prize awarded to 3 AI experts, accurately predict the 3D structure of proteins“
- “The artificial intelligence bubble in the capital market is forming“
- “How Much Does Generative AI Cost?“
- “Why software underperforming amid the AI craze?“
- “The Importance of Small Language Models (SLM) in AI, how is Apple AI different from others?“
- “Artificial intelligence benefits industries“
- “Supermicro, a repeat offender of scandals, valuation is not justified and unsustanable, no worth for long-term holding“
- “Top five lucrative artificial lucrative intelligence listed companies“
- “Artificial intelligence investment trap“
- “How does the all-powerful Huawei make money?“
- “Major artificial intelligence companies in US stocks market“
- “OpenAI, the Generative Artificial Intelligence rising star and ChatGPT“
- “Data center, a rapidly growing semiconductor field“
- “Consider Alphabet if you can buy one stock only to retire from it? “
- “How Microsoft makes money? What’s the next?“
- “How does nVidia make money? nVidia is changing the gaming rules“
- “Revisiting Nvidia: The Absolute Leader in Artificial Intelligence, Data Center, and Graphics“
- “How AMD makes money? A rare case of turning defeat into victory“
- “Why is AMD’s performance so jaw-dropping?“
- “How does Broadcom make money? Significant changes in business approach“
Disclaimer
- The content of this site is the author’s personal opinions and is for reference only. I am not responsible for the correctness, opinions, and immediacy of the content and information of the article. Readers must make their own judgments.
- I shall not be liable for any damages or other legal liabilities for the direct or indirect losses caused by the readers’ direct or indirect reliance on and reference to the information on this site, or all the responsibilities arising therefrom, as a result of any investment behavior.
