news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

Text | Shidao, Author | Shidao AI Team, Editor | Rika

Nvidia has been in trouble recently, and it's a big one.

The first is "external trouble". Due to constant "small moves", the company is facing an antitrust investigation by the US Department of Justice.

The second is "internal troubles". Due to product design errors, the delivery time of the new Blackwell chip will be delayed by three months.

Last week, Nvidia's stock price went on a roller coaster ride - on Tuesday, it plummeted 7%; on Wednesday, it soared nearly 13%; on Thursday, it closed down nearly 7%; and on Friday, it fell 7% during the session.Even looking at volatility, Nvidia has surpassed Bitcoin.Its 30-day options implied volatility recently surged from 48% to 71%, while the Bitcoin DVOL index fell from 68% to 49%.

Well-known hedge fund Elliott told investors that Nvidia is in a "bubble" and AI technology is over-hyped. The agency said that Nvidia's largest customers are developing their own chips and doubted whether the giants will continue to buy Nvidia's GPUs in large quantities.When will the bubble burst? Elliott believes that if Nvidia's financial report is not good, the bursting may happen.

It is undeniable that Nvidia is entering a "tumultuous period", but it is still too early to be "besieged on all sides".

The regulators are targeting Nvidia’s weak spot, but they are steady and accurate but not harsh.

In early July, France fired the first shot in its antitrust case against Nvidia, and in the following weeks, Nvidia was also targeted by the U.S. Department of Justice.

The Information reported that investigators from the U.S. Department of Justice visited the homes of Nvidia's "rivals" and "customers" to collect their accusations.

Allegation 1: Nvidia threatens customers.When you buy an AMD chip, Nvidia may raise your price or even reduce the number of GPU shipments it originally promised.

Accusation 2: Nvidia forced the distribution of network cables and racks.The biggest victims are: Microsoft, Google, Amazon........

In order to get the priority to ship, Microsoft gritted its teeth and bought Nvidia's network cables. When it settled the accounts in early 2023, it found that one-third of the money paid to Nvidia was spent on these "edge" products. It is reported that Nvidia's network cable sales have more than tripled to US$3.2 billion, accounting for 14% of its total data center revenue.

Now, Nvidia Vice President Andrew Bell has also said: "Whoever buys the rack will get priority to get GB200."
Microsoft escaped a disaster after weeks of wrangling, while Amazon and Google, under duress, have already agreed.

Allegation three: Nvidia acquires Run:ai.The transaction was officially announced at the end of April this year, with the purchase price being US$700 million.

Run:ai is an Israeli startup focused on simplifying AI applications and workloads on GPUs, thereby increasing the efficiency of AI chips, which in turn helps reduce the number of Nvidia GPUs required to complete a task.

Nvidia integrates Run:ai's features into existing products, not only "killing" the possibility of it being used by competitors,It also increased its DGX and DGX Cloud presence.

Huang's cloud business is "a good way out for Nvidia to avoid becoming Cisco." First, the cloud and software business can generate billions of dollars each year. Second, Nvidia has "all the levers" to develop services that complement its chips.

Whether it is the above-mentioned "strict allocation of GPUs", "forced purchase and sale of goods", or "building a cloud and ecosystem", they are all "ruthless moves" made by Huang Renxun to consolidate the Nvidia empire in the long term. After all, his goal is that all companies in the world...will run on Nvidia AI Enterprise.

Therefore, this investigation by the US Department of Justice can be described as "steady and accurate", hitting Nvidia's weak spot.

But whether it is “ruthless” is another matter. After all, antitrust investigations often last for years, and the government may not ultimately file a lawsuit against the target company. Even if a lawsuit is filed, the case may take years to conclude.

For example, the 2020 Google search antitrust case was only heard this year, and the judge in charge of the case has not yet made a ruling.

In summary, although this antitrust investigation came quickly and fiercely, it does not seem to damage the foundation of the GPU empire.

Two direct impacts on Nvidia are expected.

First, it affects future acquisitions. For example, in 2021, the FTC filed a lawsuit on antitrust grounds to prevent Nvidia from spending $40 billion to acquire Arm from SoftBank. Now it seems that Masayoshi Son can be resurrected, thanks to the FTC.

Second, crack down on "tyrannical agreements." After all, Nvidia has been targeted, so its "small moves" may be restrained.

Nvidia's shock slows global AI development

Let’s talk about Nvidia’s product “internal problems” again.

According to reports,Due to design defects, NVIDIA's B100 and B200 models have been cancelled. Previous orders will be delivered at 20% of the order quantity, and will be upgraded to B200A, with deliveries expected to begin in the middle of next year.

But the progress of GB200 has not been delayed. The "strongest king" GB200 is not just a chip, but a powerful platform that combines two B200 chips. (One GB200 has one Grace CPU + two B200s).

It is reported that the market demand for B100 itself is not large, and customers have shifted from B100 to B200 (demand for more than 450,000 pieces).

According to feedback from some institutions, the current solution is to use H200 to fill the gap first.NVIDIA will mainly launch H200 in the third quarter, and will add some H200 to customers in October and November for emergency response, and everything will be normal by December at the latest.

The delay will disrupt deployment plans of major customers such as Meta, Google and Microsoft, and may affect the development progress of products and services that rely on Nvidia's AI chips, including generative AI, video processing and other AI applications.

In other words, even if analysts shout "Nvidia bubble", once something happens to Nvidia, the global AI development process will slow down.

That is, if you want to prove that the Nvidia bubble is about to burst, or to prove that the AI ​​bubble is about to burst; or to prove that the market demand for GPUs is slowing down; or to prove that Nvidia's competitors are starting to shine.

But at the moment, you can't provide proof of a hit.

Why we're bullish on Nvidia

First, is the AI ​​bubble about to burst? No.

As the most "pessimistic" boss, Daron Acemoglu, the god of international top magazines and "Professor Dalong", believes that in the next 10 years, AI will improve productivity by about 0.53% to 0.66% and increase GDP by about 0.9%.

Jim Covello, head of global equity research at Goldman Sachs, is also quite pessimistic: AI must be able to solve complex problems in order for revenue to exceed expenditure, about $1 trillion.As for the AI ​​bubble, Coverlo's view is that it may take a long time for it to burst.

At the same time, Morgan Stanley and Wall Street's "TMT King" Coatue are very optimistic about the prospects of AI.

Morgan Stanley believes that the current AI infrastructure investment boom is in its early stages and has not yet reached the level of the Internet bubble in 1999. GPU investment, in particular, has just started.

Coatue believes that AI is not hype and its golden age has not yet arrived; AI is not a valuation bubble, but is mainly a game for giants.

Secondly, are GPUs fully utilized? No.

Here is the battle between Sequoia and a16z.

Sequoia is "pessimistic" about GPUs, believing that the peak of supply shortages has passed, and calculated that the AI ​​industry's revenue gap due to purchasing GPU manufacturing is as high as $500 billion.

a16z, on the other hand, has not only invested hundreds of millions of dollars to stockpile thousands of GPUs, but has also launched the "Oxygen Plan" in a high-profile manner, with the ultimate plan to expand the size of its GPU cluster to more than 20,000.

What is the reality? Let's change our thinking.

Conventional thinking: The more powerful the model and the emergence of killer applications, the greater the demand for computing power, and the higher the demand for GPUs.

But on the other hand, if we compare the AI ​​era with the Internet era, we can conclude that the Internet is a network of computers, and AI models are computers connected to the network. Just like everyone has a PC and a smartphone, a variety of AI models will appear in the AI ​​era, and all of them will form a complete ecosystem.

Therefore, the AI ​​era requires more computing power. However, computing power is too expensive and scarce, which further affects the development of AI. This can be seen from the fact that Nvidia's failure has slowed down the development of AI worldwide.

In addition, even if we don’t look at the long term, but only at the short term, there is a huge and unmet demand for computing power from training to inference.

Quoting the calculation of New Cortex x First Financial Daily. According to Omdia statistics, by the end of 2023, the total shipments of NVIDIA H100 reached 1.2 million. In the four quarters ending at the end of April 2024, NVIDIA earned $65.8 billion through AI chip sales. Based on the price of $40,000 per H100, NVIDIA has sold approximately 1.645 million H100s in the past four quarters. This is approximately equivalent to 1/4 of the overall model training market (assuming that the world has trained 100 GPT-3.5-level, 50 GPT-4-level, 10 GPT-5-level large language models, 10 Midjourney-level image generation models, and 20 Sora-level video generation models).

This is just the market demand for model training. Inference is a market of a much larger magnitude and has not yet been truly opened. If we conservatively calculate that each person only generates one video per month, the global computing power demand for large model inference is equivalent to 10.4 million H100 chips, which is twice the demand for model training. If the video generation technology and market are more mature, the computing power consumption required for inference will soar to dozens or even hundreds of times the training demand.

Google CEO Pichai said: "The risk of underinvesting is far greater than the risk of overinvesting."

As Meta CEO Zuckerberg said: "I would rather overinvest than save money by slowing down development."

Microsoft said that the current AI computing capacity has limited its financial income, and this situation will continue until at least the first quarter of fiscal year 2025. Musk even developed his own supercomputer to challenge Nvidia.

Finally, is Nvidia's moat strong? Yes.

Regarding Nvidia's future, Coatue directly expressed his views:If there is a bubble, it is not a valuation bubble, but a bubble in which profits are pulled forward.

So, who can challenge Nvidia?

On the one hand, compared with the three major accusations mentioned above, Nvidia has a bigger crime - software and hardware bundling.

Today, more than 95% of processors in data centers use NVIDIA GPUs, and the entire cloud AI demand still relies on the CUDA ecosystem, which can only be used with NVIDIA chips. Programmers don't want to learn another language, and CUDA is even considered NVIDIA's moat.

In order to overthrow Nvidia's "tyranny", Google, Meta, and Microsoft currently want to join forces to participate in the open source language project Triton initiated by OpenAI; Intel, AMD, and Qualcomm also want to use Triton to poach Nvidia's customers.

While Triton could erode Nvidia’s market share, Citi analysts estimate that Nvidia’s share of the generative AI chip market will still be around 63% by 2030, meaning it will remain the dominant player for many years to come.

On the other hand, giants have issued challenges one after another, and Nvidia is also running. After all, Huang Renxun definitely remembers the "arrogance of his predecessor Cisco".

According to the semiconductor "Makimoto cycle" - chip types regularly alternate between general and customized - general-purpose structures are the most popular in a certain period of time, but after reaching a certain stage, specialized structures that meet specific needs will catch up.

At present, the era of general-purpose structures represented by NVIDIA is undergoing subversion.

What's more, at the beginning of this year, Nvidia established a new business unit to build custom chips for customers in the fields of cloud computing, 5G telecommunications, games, and automobiles. It is reported that the new Nintendo Switch launched this year is likely to be equipped with Nvidia's custom chips.

But is Nvidia absolutely as stable as a rock? I thought of a saying: "The process is right, but the result is wrong." To make an inappropriate analogy, if Nvidia's strong bargaining power drives the cost of AI computing power to remain high for a long time, thereby inhibiting large-scale innovation, there may be a backlash.