Morgan Stanley internal disagreements? The Greater China region has just lowered its expectations, while North America is loudly calling to bottom fish NVIDIA

Morgan Stanley's internal analysts seem to have differing views on NVIDIA: analysts in Greater China have lowered their shipment expectations for the GB200 chip this year, predicting a worst-case scenario of less than 20,000 units, and expect the growth cycle of the cloud computing market may peak this year, with year-on-year growth in the fourth quarter potentially dropping to single digits. However, analysts in North America reaffirm that NVIDIA remains their top choice, claiming that the sell-off triggered by DeepSeek is a buying opportunity.

Analysts in the semiconductor sector at Morgan Stanley, including Joseph Moore, released a report on Thursday, February 6, giving NVIDIA a target price of $152, which represents an expected increase of about 21.8% based on the stock price at the close on Wednesday. They believe that investor sentiment around NVIDIA's potential long-term risks has deteriorated, but NVIDIA's short-term business remains robust, with increasing visibility on the supply of Blackwell chips and a clear willingness to spend from customers, still considering NVIDIA as the preferred choice.

The aforementioned analysts in North America first acknowledge in their report that DeepSeek has caused some headwinds in export controls and long-term investments. In the worst-case scenario, they anticipate three areas of disruption:

• First, further export controls. Although leading hardware companies in the AI field are based in the U.S., that does not mean the U.S. can control global software development. If the U.S. had imposed export controls on microprocessors 25 years ago, its market share in the microprocessor field today might be significantly lower, and the situation in the software industry could be similar. However, analysts are almost certain that there will be more restrictions. Any limitations imposed by the U.S. government could undermine NVIDIA's profit margins.
• Second, the financing environment for spenders in the AI field. The massive sell-off of AI concept stocks may create a different financing environment than in the past. Over the past 12 months, any growth in AI was viewed as a positive factor. Now, considering some views that there are limitations on scaling, financing may face more scrutiny.
• Third, investor sentiment has become quite negative, seemingly unrelated to recent results. While there are several reasons to believe that large cluster construction will continue, primarily because the architects of these clusters are striving to persist. Morgan Stanley's analysts believe that accelerated revenue growth can alleviate these concerns, but this view is contentious, and whether it can truly be so remains to be seen.

Considering the above negative impacts, Morgan Stanley's analysts remain very optimistic about NVIDIA's balanced development prospects this year. They emphasize the following four supporting factors:

• First, it is still a transitional environment, but confidence in Hopper and Blackwell chips is continuously strengthening, and these analysts expect positive evaluations to re-emerge after this quarter.

Regarding the Hopper chip, analysts recently found that demand for Hopper has weakened, but industry discussions suggest that demand for this chip may strengthen, with export control risks possibly being a reason, as regions potentially affected by such controls are taking action to prepare in advanceThis is a certain driving force, considering the transitional nature of this environment, confidence is still being established.

For the Blackwell chip, analysts are still focusing on NVIDIA's information regarding the "unprecedented complexity" of various Blackwell solutions, and they feel these issues are being resolved quickly, including the final form of the GB200. Limitations still exist, and confidence is being enhanced. Previously considered stopgap measures by NVIDIA, such as hastily pushing the GB300 to market without alleviating adjustments to the GB200, are no longer seen as necessary. Demand for all forms of Blackwell is strong, initially still likely leaning more towards non-rack scale solutions, but confidence in all Blackwell is certainly increasing, believing that the availability of GB200 NVL72 from computing power rental companies, the so-called "GPU scalper" Coreweave, is a very promising signal.

All of this may indicate that NVIDIA will provide guidance for the current fiscal quarter in April that meets expectations, but analysts expect that comments from NVIDIA's management regarding the second half of the year will significantly decrease, and comments about particularly high demand for Blackwell may reappear.

• Second, although investor sentiment towards large training clusters is under pressure, analysts say there are signs that large clusters are still being built.

Comments on capital expenditures from NVIDIA's major customers reaffirm the investment trajectory and emphasize that the supply-demand mismatch will continue in the short term. For NVIDIA's cloud customers, the revenue algorithm remains clear—"the more GPUs I buy, the more money I make." And for investments that are currently not generating revenue, there is still a commitment to advancing cutting-edge technology. Many of the architects of the largest AGI clusters have reiterated their commitment to scaling large training clusters, with no signs that DeepSeek will change this momentum.

• Third, inference seems poised to drive growth for years, and NVIDIA's position in the inference field remains solid.

The inference market is clearly more deflationary than the training market, but analysts also see a market tendency towards the highest performance solutions. Even NVIDIA's lower-priced inference solutions have largely given way to the comprehensive training versions of Hopper and now Blackwell. They remain confident that NVIDIA is the biggest beneficiary of long-term inference workloads.

• Fourth, the stock market has recently favored ASICs over GPUs, and analysts expect this preference to reverse in the second half of the year. They continue to expect revenue from GPUs to accelerate in the second half, while expecting ASICs will not see this. Both markets are promising in the long run, but they believe there is more room for sentiment to shift towards NVIDIA in the second half.

Analysts believe that the strong performance of ASICs after the rise of DeepSeek is due to an increasing emphasis on service costs, and they see ASICs as more suitable to become low-cost inference platforms for many NVIDIA customers. They confirm that optimized products are more suitable for specific workloads, which makes sense. However, based on their observations, it is Blackwell, not ASICs, that will change the cost curve in 2025Analysts have believed in recent years that cheaper inference-specific products, such as T4 and L40, would be more popular. However, as these workloads become increasingly complex, they continue to view high-end GPUs as the primary source for inference computing. The extension of compute testing times further increases the computational demands for inference, which bodes well for top-performance solutions, and NVIDIA provides such solutions.

According to ML performance benchmarks, the B200 delivers 2.5 times the number of single-chip tokens per second on LLama2 70b compared to the H200, and the multi-chip performance should be even higher, as it has already shifted to using Blackwell's 5th generation nv-link, which has doubled the bandwidth compared to the 4th generation. In other words, for every dollar spent, performance could potentially improve by at least 2 times, which is in reference to products that only started shipping a few quarters ago. Analysts believe this performance has set a high comparative standard for ASICs, and NVIDIA is expected to release the Rubin chip in about a month, which may further raise the target that ASICs need to surpass.

Another viewpoint suggests that ASIC projects are more durable compared to spending on NVIDIA products. Analysts believe that to some extent, this viewpoint is correct, as there are indeed commitments to enhance ASIC capabilities in areas with currently low penetration rates. However, if training demand slows down, capital commitments related to training and inference may waver. Analysts point out that Broadcom has indicated that the company's serviceable available market (SAM) of $60 to $90 billion is primarily related to training.

Analysts believe that when the training market growth slows, NVIDIA's existing advantages in training are a strength, as once training runs are completed, it automatically releases GPUs that are typically used for inference. Without initial purchases, the competitive barrier may be lower, but they consider this a long-term risk. In the short term, these GPUs may replace inference ASIC projects, as NVIDIA's flexibility allows for a higher degree of overall unification across cloud, on-premises, inference, and training workloads. NVIDIA is also continuously adjusting its software stack and providing these updates to customers for free, some of which have significantly improved performance in the past. Even if overall AI capital expenditures shrink, it would not surprise analysts if NVIDIA gains market share, especially compared to first, second, or third-generation ASIC products