AI Storm: NVIDIA's Twilight

Today, let's start with NVIDIA.

After NVIDIA's stock fell for two consecutive days, with a drop of 11% and a loss of $222 billion in market value, one question has become particularly eye-catching.

Will it continue to fall?

There are two possibilities: the decline this week is just a routine pullback after the split.

Moreover, 20% of those who hold this view provide a famous contrarian indicator, which is that "Wood Sister" bought a large amount of NVIDIA stocks on Friday.

In fact, this game of guessing the rise and fall of stock prices is meaningless.

What's important is whether NVIDIA will still have value in the future.

Advertisement

After all, it is currently a very important "high ground," a symbol of the United States' ability to maintain a leading edge in the field of high technology.

Firstly, as the AI revolution led by the United States shines like the sunshine of a new world, its shadow is particularly clear.

The development of AI has not brought a huge leap in world technology and economy, but first brought a huge burden.

Bloomberg published an article last week titled "AI Has Caused Serious Damage to the Global Energy System," which said: "There are more than 7,000 data centers around the world that have been built or are in different stages of development, higher than the 3,600 in 2015.

If these data centers continue to operate, they can consume a total of 508 terawatt-hours of electricity per year.

This is higher than the annual power generation of Italy or Australia.

By 2034, the energy consumption of global data centers is expected to exceed 1,580 terawatt-hours, equivalent to the entire power consumption of India."

In the United States, large technology companies in Silicon Valley are developing artificial intelligence at all costs, and the result is that the sharp increase in electricity demand is seriously threatening the United States' energy transition plan and clean energy goals.

At the same time, due to the sudden increase in power supply capacity, it is also encroaching and threatening the electricity used by residents in the areas where data centers are located, as well as some startups or small and medium-sized enterprises that are in operation.

Their electricity demand applications often need to go through a very long waiting period.

On June 20, SoftBank's Masayoshi Son announced that he was preparing to make the next major investment after a break, and this time he chose the direction of investment, which is a race that is so hot that it can burn any investor - artificial intelligence.

After suffering heavy losses of billions of dollars in the investment of WeWork, Masayoshi Son has been focusing on the chip department Arm Holdings Plc and the investment strategy around artificial intelligence.

At the annual shareholders' meeting of the wireless operator on June 20, he said to the gathered SoftBank shareholders: "We need to look for our next big move, not afraid of whether it will succeed or fail.

SoftBank's main direction currently looks like two: one is the start-up of artificial intelligence.

The second is the power demand of artificial intelligence, and this part of the project will focus on the United States.

It is said that Masayoshi Son is looking for up to $100 billion in funds to finance a chip enterprise (possibly ARM), ready to compete with NVIDIA.

When most of the funds all flow to artificial intelligence, how will other industries develop?

AI is like a giant baby that absorbs nutrients, it devours a lot of things, but it still grows slowly.

As I shared before, the improvement of productivity by AI is a one-time pass, and currently, at least in the United States, there is no successful case of generative AI generating positive cash flow income.

Apple is preparing to combine OpenAI into the mobile phone for sale, but it was immediately cooled down by the European Union, and the European Union's information bill does not support the Apple mobile phone with AI.

So what about the other big market in China?

Will it allow OpenAI, with the former director of the National Security Agency of the United States as a director, to enter China directly with Apple?

I am skeptical.

Secondly, back to NVIDIA, the biggest problem of this company is "irregularity".

1.

On June 21, the U.S. Supreme Court announced that it would hear the case of shareholders suing NVIDIA for defrauding investors.

This case occurred in May 2022, NVIDIA agreed to pay the U.S. Securities and Exchange Commission (SEC) $5.5 million to resolve the civil charges brought by the agency: the company did not properly disclose the important impact of "cryptocurrency mining" in the growth of its game business chip sales.

In the two quarters of the fiscal year 2018, the company did not disclose that cryptocurrency mining was an "important factor" in its revenue growth.

This is just a small detail in NVIDIA's many "revenue control methods" and is also the only small problem that has been "caught out" and dealt with.

As NVIDIA's importance to the United States increases day by day, the government's supervision of it becomes more and more relaxed.

2.

What is the big problem of NVIDIA?

Massive inflation of revenue through "revenue control methods".

I once saw an internal research report that analyzed in detail the problem of inflated revenue in NVIDIA's 10-Q report (equivalent to the quarterly report).

According to the 10-Q report: "Two indirect customers will each account for 10% or more of total revenue in the first quarter of the fiscal year 2025; one indirect customer mainly purchases our products through direct customer B.

Both belong to the computing and networking department."

"A direct customer A accounts for 13% of the total revenue in the first quarter of the fiscal year 2025; another direct customer B accounts for 11% of the total revenue, both belong to the computing and networking department.

See this chain?

Customer A, accounting for 13%, this is no problem; customer B, accounting for 11%, one of the indirect customers mainly purchases our products through direct customer B.

It means that all of customer B's inventory is completely sold to an anonymous "indirect" customer C. Why can't NVIDIA sell directly to customer C?

There are two options: However, NVIDIA's disclosure behind this is very strange, "If the final demand increases or our finished product supply is concentrated near the end of the quarter, system integrators, distributors, and channel partners have limited ability to increase their credit limits, which may affect the timing and amount of our revenue.

Haven't they all been sold to B and confirmed revenue?

What's the significance of this disclosure?

Then look down: Q-10 says, "Most of our sales are based on purchase orders, and our customers can usually cancel, change, or delay product purchase commitments without prior notice to us, and without paying a fine."

Seeing this, readers familiar with financial fraud may feel the "familiar taste", the so-called customer B-C may just be a system of inflated revenue.

Or simply label the operation mode of this system - "brushing orders".

This may also be the most important reason why NVIDIA relies heavily on system integrators, distributors, and channel partners.

Currently, NVIDIA sells billions of GPUs to Taiwan every quarter, but Taiwan's data center revenue is expected to be only about $220 million per year by 2028, and the total investment in the region is expected to be $3.2 billion by 2028.

At the same time, there is also such a description in the report: "For example, most of the goods related to Singapore's revenue in the first quarter of the fiscal year 2025 were shipped to the United States or Taiwan."

3.

Although the research report is extremely obscure, it is easy to clarify the context under reasonable suspicion.

The first category, normal sales, such as customer A; the second category, brushing order system, such as customer B-C; the third category, gray income, such as TW.

(The part that is transshipped back to the United States may involve the B-C system) The question is, NVIDIA's method, Enron also used it before.

But how long can it be used?

4.

In addition to these, NVIDIA also has another widely criticized revenue control system: NVIDIA - equity investment start-up company - start-up company raises funds on the grounds of purchasing GPUs - uses GPUs as collateral.

There are dozens of start-up companies invested by NVIDIA.

However, the above is currently still covered under the "compliant" and "exciting" financial statements, what is the problem when and who will lift it?

But this is a "soft rib", when the capital wants to short NVIDIA, it will be very easy.

However, these are not the fatal threat to NVIDIA.

Thirdly, the only thing that can subvert NVIDIA is that the world no longer needs "so many GPUs".

According to the convention, the moat built by NVIDIA through CUDA is almost unbreakable, but there is a very important premise - as long as the GPU is still key in the creation of large-scale language models (LLM).

However, the change has arrived.

Recently, two papers have attracted widespread attention: "Accelerating Pretrained LLMs via Post-Training Multiplication-Less Reparameterization" (Accelerating Pretrained Large-Scale Language Models through Post-Training Multiplication-Less Reparameterization).

This paper proposes a method called ShiftAddLLM, which eliminates the dependence on matrix multiplication through post-training shift and addition reparameterization.

This method quantifies each weight matrix into a binary matrix and pairs it with a group scaling factor, reparameterizing the multiplication operation between the activation and scaling factor, and performing query and addition operations according to the binary matrix.

This method has shown significant advantages in reducing memory usage and improving energy efficiency.

This paper demonstrates a language model architecture that completely eliminates matrix multiplication.

In the experiment, this matrix-free model shows comparable performance to the current most advanced Transformer model when the parameter scale reaches 2.7 billion, and shows significant optimization in memory usage.

This architecture achieves very low power consumption in the inference process through customized hardware solutions, approaching brain-like efficiency.Here is the translation of the provided content into English: These two papers jointly demonstrate the feasibility of no longer relying on GPUs in the training of large-scale language models, and according to experts, it has strong operability.

This disruptive change is enough to make Nvidia the next Cisco.

I don't like Nvidia because it is a kingdom built on sand dunes, which could be swept away by the wind at any time.

Like its CEO, it is filled with illusory passion but lacks down-to-earth qualities.