Has NVIDIA done well already?

TL;DR

Nvidia's financial report has been released, and there seem to be several completely different opinions. Ultimately, it closed down 3% amid market pricing discrepancies, while in recent weeks, several NeoCloud companies have generally seen declines of over 20-30%, with CRWV nearly halving in the last month. Some common analyses are as follows:

"Left hand to right hand, 'The $610 billion AI Ponzi scheme collapses'"[1]

"NVIDIA's 53-day magic: A 'circular financing game' hidden in the financial report"[2]

At the same time, Bridgewater's holdings have decreased by 65.3% compared to 7.23 million shares in the second quarter, and SoftBank has also completely liquidated its NV stocks. Of course, "big short" Michael Burry has started shorting NV and has questioned the depreciation algorithm for GPUs.

On the other hand, Jensen seems to be complaining about why the stock is falling despite exceeding performance expectations. He repeatedly emphasizes NV's value. To be honest, NV is indeed a fantastic company. From a technical perspective, why hasn't there been a mature challenger since the emergence of ChatGPT in recent years? Especially in a sufficiently open competitive market like the United States, is it really the barrier of the CUDA ecosystem? Perhaps we should look into many details. NV is a very strong system solution company, not just a simple integrated circuit company. Although many other GPU-related competitors are filling their gaps in various fields, there are many flaws in its own GPU microarchitecture and performance collapse under various workloads... Some things are better left unsaid...

1. The essence of cloud computing is liquidity management

However, from the market perspective, NV has not been performing well in recent months, especially those NeoCloud little brothers... This essentially exposes a critical issue: NV basically does not understand what cloud computing is. At least I haven't seen Jensen elaborate on the difference between AI Cloud and AI Factory? Nor have I experienced any cloud-designed features in their products. For NV, this is also a problem that many equipment manufacturers struggle to overcome. For example, equipment providers like Cisco completely missed the opportunity to transform into CSPs, largely due to their deeply ingrained mindset of selling hardware boxes.

Around this time last year, I jokingly wrote an article titled "How to Leverage GPUs as a Financial Product?" I didn't expect this year to see it play out in earnest... From NV's perspective on cloud or its own interests, it would be best for each model manufacturer to build a huge AI Factory. Is there really a bubble? Essentially, it is still a liquidity management issue, rather than a simple logic of selling hardware In fact, the essence of cloud computing is similar to that of banking. Bill Gates once said, "Banking is necessary, Banks are not." Dr. Wang Jian mentioned, "Computing for Value Beyond Computation." Essentially, the doctor is also explaining the value of computing, so one could write: "Computing is necessary, Computers are not."

In other words, computing power is essential, and the composition of computing power is diverse and can be allocated as needed. Of course, for the provision of computing power, there are many constructed private cloud clusters or other proprietary cluster leasing models. But is the computing power leasing of such IDC data centers considered cloud? This is a significant misconception for many people.

The essence of cloud computing is to leverage computing power, but it pays more attention to liquidity risk. However, the current NeoClouds still largely rely on long-term orders from single customers, and any default could lead to significant operational difficulties. In other words, these NeoClouds are essentially completely ignorant of liquidity risk management. During the H card phase, there was a supply-demand imbalance, so complex liquidity management was not particularly necessary. However, in the B card era, the balance of supply and demand is undergoing subtle changes.

"Critique of a Competitor's Claim that Traditional Cloud is Still Selling Iron: Discussing Cloud Computing and Its Liquidity Management from a Financial Perspective"

When have you ever heard a major bank tell investors that all my loans are long-term stable orders from a major client? And that client is a major one?

2. Analyzing from Revenue

In fact, in July of this year, I discussed the operational risks of GPU cloud from a risk control perspective of FRM in detail in "Discussing the Operational Risks and Liquidity Management of GPU Cloud," which basically covered some recent controversies, including depreciation and the risk of default from a single customer.

The fundamental issue is that Nvidia and its NeoCloud little brothers have not figured out what cloud computing really is. In fact, liquidity risk is the key focus of cloud operational management. In 2009, Berkeley discussed cloud computing in a famous paper titled "Above the Clouds: A Berkeley View of Cloud Computing," which outlined six points:

• Unlimited computing resources can be used on demand.
• Eliminate pre-commitments for cloud users.
• Pay for short-term use of computing resources based on actual needs.
• Significantly reduce costs through economies of scale in ultra-large data centers.
• Simplify operations and improve resource utilization through resource virtualization technology.
• Improve hardware resource utilization by running loads from different organizations through multiplexing From a financial perspective, the first two points discuss the need for rigid repayment of computing power, the third point addresses the leasing relationship of computing power, and points four to six discuss the operational management of cloud computing power institutions similar to financial institutions. Essentially, these points explain how cloud computing leverages computing power and provides liquidity.

In recent months, we have already identified some signals of liquidity risk, one being Nv's accounts receivable, and another being liquidity injected through some revolving financing and committed consumption. First, the accounts receivable for FY26Q3 amounted to $33.4 billion, with quarterly revenue of $57 billion. Compared to last year's actual billing period of 46 days, this year's data calculates to 53 days. Did the current quarter use the next quarter's deposit revenue to offset and reduce DSO (Days Sales Outstanding)? Will it continue to rise in the coming quarters? Moreover, customer concentration is also very high.

On the other hand, there is an increase in inventory, with total inventory at the end of FY26Q3 reaching $19.8 billion, a quarter-on-quarter increase of 32%, and inventory turnover days (DIO) currently reaching 117 days. In fact, the increase in inventory is very contradictory to the current situation of supply not meeting demand.

Next, there are some issues with revolving investment transactions, particularly the multi-year cloud service agreements that have skyrocketed from $12.6 billion to $26 billion. This means that NV has committed to purchasing $26 billion worth of cloud services from cloud vendors in the future. Simply put, it treats GPUs as financial assets, selling them and then leasing them back at a high price while providing corresponding returns to cloud vendors. In essence, how is this different from those wealth management products that promise returns?

Additionally, the issue of depreciation rates seems to be a constant point of contention. On one hand, it is said that the A100 is still being used and generating value, while on the other hand, Microsoft's Satya is saying that the H200 cards are gathering dust. In fact, when there is no relatively certain return, it becomes meaningless to argue about depreciation. It is merely an attempt to push the losses of the previous generation cards further out by applying longer discounts, and then using the larger scale and receivables of the new generation cards to cover the losses of the previous generation. This cycle continues to grow, and the scale of H cards must be significantly larger than that of A cards to break even, followed by the scale of B cards needing to be significantly larger than H cards to keep this game of hot potato going...

In terms of revenue, through repeated investments in companies like OAI/Anthropic, it has transformed into orders, causing a significant cycle of investment. I won't elaborate on these details.

Regarding Depreciation

In fact, the depreciation of each generation of cards is different. For the A100, which is already at the end of its life cycle, although there are some search and promotion uses, especially in so-called generative recommendation (GR) systems, the H cards are already showing signs of idleness. Regardless of computing power or cost-effectiveness, phasing out the A100 and handling its residual value is undoubtedly the best choice Its lifecycle is reasonably estimated at 5 years. As for the H card, many have actually been in operation for two to three years, and will continue to be used as a mature platform for another two years, until models with FP4 precision gradually mature. The situation is different for B200/GB200, which is likely to be a very short-lived platform. Its depreciation cycle may only be around 3 years or even shorter. On one hand, the stability issues with GB200 delayed delivery, and by the time it was really deployed on a large scale, GB300 had already come out. On the other hand, there are some issues with the chip architecture; in terms of training and inference with the same FP8 models, there is not a significant advantage over the H card, and in some cases, it may even perform worse... Furthermore, it is estimated that the application and ecosystem of FP4 will take another 1 to 2 years to mature, so the leasing premium for B200 and GB200 during this time is not high. The large-scale deployment of these two cards in the cloud will face significant operational pressure. Therefore, I have always felt that B200/GB200 is not a product worth investing in; perhaps B300 will be slightly more mature, but its value will only be recognized after the industry fully realizes the benefits of FP4.

Additionally, in Satya's recent interview, he mentioned targeting some long-tail customers, which also revealed some interesting signals. In fact, it reflects a desire to gradually diversify the risk of relying on a single customer. Overall, established cloud computing vendors like AWS/Azure/Google seem to be much more mature in their focus on liquidity compared to those Neocloud small brothers...

3. Technical Analysis

In fact, there is already a detailed analysis on the new Blackwell architecture titled "Inside Nvidia GPU: Discussing the Shortcomings of Blackwell and Predicting the Microarchitecture of Rubin," which clearly states that neither B200 nor GB200 has achieved a cost-performance advantage over the H card, and in some workloads, performance may even be lower.

The advantages of FP4 may take more than a year to materialize, so I have always held the judgment that B200/GB200 is a product destined to be short-lived, and investment in it should be reduced, especially considering the interconnect issues faced by GB200, which have been discussed in detail in the previous article "The Flaw of 30 Trillion."

Essentially, GB200, as a system solution, has certain issues regarding ROI for both training and inference, and its lifecycle has been significantly compressed due to delays caused by reliability issues in the early stages.

In essence, it can be summarized in a few points:

1. GPUs have entered a new capital cycle, gradually moving towards DSA, with TensorCore becoming larger and SM becoming fewer. The evolution of programming interfaces is a challenge, especially for later models like Rubin Ultra and CuTile; any failure in this ecosystem could open up a significant gap In fact, what we are more pleased to see is that TileLang and some other hardware have achieved certain success in this market cycle.
2. In terms of interconnection architecture, ScaleUP's NVLink is likely to hold its ground, especially after Intel and ARM join the NVLink Fusion ecosystem, which will bring some changes. There is even a viewpoint that seems somewhat out of line with today's aesthetics: Is ScaleUP's solution for small/mainframe really useful? In fact, cloud service providers tend to prefer scalable ScaleOut technology, but the fundamental reason for ScaleOut is that the RDMA Verbs interface is not GPU-friendly at all. There may be some variables here, but getting Nvidia to abandon RDMA clearly faces greater resistance, and there is likely still significant controversy within Nvidia regarding its internal factions.

Of course, we must admit that Nvidia is currently an irreplaceable company, excelling in many details. For example, load balancing for CTA/CGA, warp scheduling, compiler optimization, on-chip interconnect networks, and many other fundamental and detailed aspects are done very well. These subtle differences are the true source of the gap between Nvidia and other GPU manufacturers. Other companies (such as AMD and a number of domestic card manufacturers) seem to not pay much attention to these details...

At least in the next two to three years, there are relatively few competing manufacturers that can match Nvidia in various practical scenarios. As for the domestic situation, I won't comment much; I just hope that the technology stack covers a wider range to avoid falling into the dead end of local optimization...

Finally, as usual: The content of this article is merely personal analysis and does not constitute any investment advice, nor should it be used as a basis for any legal regulations or regulatory policies. Investors should not rely on this information as a basis for decision-making or legal actions, and any consequences arising therefrom shall be borne by the investors themselves.

Risk Warning and Disclaimer

The market has risks, and investment requires caution. This article does not constitute personal investment advice and does not take into account the specific investment goals, financial conditions, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article align with their specific circumstances. Investing based on this is at one's own risk