Original Report: Morgan Stanley Research "NVDA NDR Highlights Diversified Growth Opportunities," July 10, 2026

Compiled & Organized: DaiDai

Editor: Frank

Key Points Overview:

• NVIDIA believes growth is still accelerating, but the drivers are changing: Demand is no longer solely reliant on leading AI labs and hyper-scale cloud providers; networks, CPUs, Neocloud, sovereign AI, industrial and enterprise customers are becoming new incremental sources;
• ASICs and GPUs are not a zero-sum competition: For a previously ASIC-dependent leading edge model client, NVIDIA's computing power share has risen to nearly 50%. In many practical workloads, customers ultimately compare not just chip prices, but the comprehensive computing costs and deployment efficiency per token;
• The Vera Rubin roadmap has not been delayed: NVIDIA denies that Rubin Ultra has been pushed to 2028 and confirms it still plans to ship in 2027. Although some rack solutions will be adjusted, key directions such as the 800V power architecture and inter-rack optical interconnect remain unchanged;
• Storage and power remain supply-side constraints, not signals of demand weakness: NVIDIA expects storage shortages may last for several years and will improve system delivery efficiency under limited supply by adjusting memory configurations, networks, and system architectures;
• Morgan Stanley continues to regard NVIDIA as the preferred semiconductor: The report maintains an "Overweight" rating and a $288 target price, believing that as the company increases cash return rates, NVIDIA is transitioning from a purely high-growth asset to gradually embodying value attributes of large tech stocks;

When a company's quarterly revenue approaches $100 billion, the market usually becomes less concerned about how much it can grow, and more focused on how long that growth can be sustained.

NVIDIA is currently facing such scrutiny.

Recently, the controversies surrounding NVIDIA have primarily centered on three issues: Whether global AI infrastructure investment is approaching a peak; whether cloud providers like Google, Amazon, and Meta pushing for self-developed ASICs will gradually erode the market share of GPUs; and whether the next generation Vera Rubin product roadmap will see delays or adjustments.

In early July, Morgan Stanley organized multiple NVIDIA non-deal roadshows (NDR) in California, where NVIDIA CEO Jensen Huang, CFO Colette Kress, and the management and investor relations teams directly faced institutional investors to respond to the market's most concerned issues regarding growth, competition, and products, hoping to explain the changes occurring in its sources of growth for the next phase.

Morgan Stanley summarized the tone of the meeting in its post-meeting report as "positive," while NVIDIA's management gave a more aggressive assessment: Even with quarterly revenues approaching $100 billion, the company's current growth rate is still expected to accelerate.

The core supporting this judgment lies not in a sudden increase in purchases from a single customer, but in the fact that NVIDIA is transitioning from a company primarily reliant on GPU cycles and a few leading customers to a full-stack AI platform covering computing, networking, CPUs, models, and cloud infrastructure.

1. NVIDIA is Redefining Its Sources of Growth

According to NVIDIA management's classification during the meeting, the future demand for the company primarily comes from three types of markets: AI labs, traditional hyperscale cloud providers, and emerging customers including AI cloud, industrial, enterprise, and sovereign AI.

This means that the narrative familiar to the market in the past of "AI labs training large models, cloud providers purchasing GPUs" can no longer fully explain the next phase of NVIDIA's growth.

1. AI Labs: ASIC Customers are Also Increasing GPU Usage

Currently, AI labs account for about 20% of NVIDIA's total demand.

During this roadshow, one particularly noteworthy detail was that a leading edge model customer that previously relied predominantly on ASICs had a low initial usage ratio of the NVIDIA platform, but now NVIDIA's share in its relevant computing power has risen to nearly 50%.

This indicates, at the very least, that ASIC's substitution of GPU is not a one-way path.

As model size, inference load, and system complexity continue to rise, customers need to compare not just the price of a single chip, but also the overall cost of generating tokens across the entire system. In numerous practical workloads, NVIDIA believes that the lowest unit token cost still comes from its complete platform.

Other leading edge models are still primarily operating on NVIDIA's platform. For NVIDIA, the opportunity with AI labs is no longer just about following the growth of model training scales, but also about regaining share from customers that originally favored self-developed chips.

2. Hyperscale Cloud Providers: Expanding from GPUs to Networking and CPUs

According to new business metrics, traditional hyperscale cloud providers remain one of the most important sources of demand for NVIDIA, accounting for about half of the company's revenue.

However, the constraints currently faced by the cloud provider market have gradually shifted from chip supply to land, power, and data center space. In other words, customers are not unwilling to continue investing, but are increasingly finding it difficult to quickly expand data centers in the original way.

In this context, NVIDIA is increasing its value share in each data center—beyond GPUs, the company is also expanding its serviceable market with networking equipment, CPUs, interconnects, and rack-level system solutions.

The management reiterated its goal of achieving $20 billion in CPU-related revenue this year and hinted that a significant portion of this demand does not come from the Head Node responsible for cluster control and scheduling, but rather from independent computer racks equipped with Vera CPUs.

Vera is not just a general-purpose CPU pursuing more cores; it is optimized for single-thread workloads and memory access within AI data centers. It represents a direction where NVIDIA isn't just selling accelerator cards but is further entering into CPUs, networking, interconnects, and complete system solutions.

This also explains why Morgan Stanley believes cloud providers developing their own ASICs and NVIDIA's continued growth can coexist. Large cloud providers will still continue to develop and deploy custom chips, and ASIC suppliers like Broadcom may also maintain rapid growth, but as long as the entire AI computing market expands quickly enough, GPUs and ASICs do not need to engage in zero-sum competition.

Morgan Stanley expects that both NVIDIA's and Broadcom's AI-related businesses may maintain extremely high growth rates over the next year, and there may not be significant short-term shifts in market share.

3. Sovereign AI, Industrial, and Neocloud: More Dispersed, and Possibly Larger

Compared to AI labs and hyperscale cloud providers, the demand from sovereign AI, industrial enterprises, and Neocloud is more dispersed, and project initiation speeds are often slower, but this might be the most noteworthy incremental market for NVIDIA in the next phase.

Geopolitics, data sovereignty, and supply chain localization are driving more countries and regions to build independent AI infrastructure. Meanwhile, industries such as finance, retail, biotechnology, and manufacturing are starting to transition AI from a generic tool to an integral part of their internal production systems.

Such projects typically require longer project initiation, approval, and deployment cycles, but once they enter the implementation phase, the scale of individual projects can be substantial, and they often face less direct chip competition than in the leading cloud market.

Neocloud offers another growth model.

These new types of cloud providers, centered around GPU computing power, are fulfilling the computing demand that traditional cloud providers cannot meet in a timely manner. NVIDIA's collaborations with some Neoclouds may involve joint investments and credit support, and could also introduce revenue sharing mechanisms.

The market is currently more concerned with whether NVIDIA is taking on credit risks for its customers, but from another angle, revenue sharing also means NVIDIA may share the long-term benefits of downstream GPU cloud services. Based on this, Morgan Stanley believes that NVIDIA may not only sell hardware to Neoclouds in the future, but will gradually become a stakeholder in large GPU cloud networks.

2. ASICs, Rubin, and Storage Constraints—What Truly Impacts Growth?

To be pragmatic, the largest controversy currently faced by NVIDIA is not whether AI will continue to grow, but whether chip competition, product roadmaps, and supply constraints will weaken the company's market share in this current growth wave.

From the information released during this roadshow, NVIDIA's response is that these issues certainly exist but have not changed the company’s core growth direction.

1. ASICs will grow, but that does not mean GPUs will necessarily lose share

For stable, mature, and sufficiently large workloads, ASICs may indeed provide lower chip costs and higher customization efficiency. However, model architectures, inference methods, and development tools are still rapidly evolving, and customers need both flexibility, software ecosystems, and system delivery capabilities.

Thus, what truly determines customer choices is not the price of a single chip, but the comprehensive costs after considering model training, inference, networking, memory, and software coordination.

NVIDIA's moat is no longer just GPU performance but the platform capabilities collectively formed by CUDA, networking, CPUs, interconnects, complete systems, and model tools.

2. Rubin Roadmap has not seen substantive delays

This roadshow also addressed recent controversies surrounding the Vera Rubin product roadmap.

NVIDIA denied market rumors that Rubin Ultra has been delayed to 2028, confirming that the product is still planned for 2027, but it also acknowledged that some rack forms will undergo adjustments; the original Kyber rack solution will be replaced by another design referred to by management as "better."

This change may support larger single-rack scale-up domains, but does not mean a change in core technical direction; key technologies including 800V power architecture and inter-rack optical interconnect are still progressing according to the original plan.

For NVIDIA, attempting more aggressive system designs with each generation of products is part of maintaining its leading position. What truly needs to be observed is not whether the roadmap has undergone adjustments, but whether the company can timely correct itself before mass production and manage large-scale delivery risks.

Morgan Stanley believes that Vera Rubin will remain an important product cycle driving NVIDIA's growth over the next 12 months.

3. Storage shortages are another issue easily misunderstood by the market

Management anticipates that storage supply constraints may last for several years.

If the AI industry aims for significant annual token growth, and storage supply cannot expand accordingly, simply increasing the number of GPUs will not solve the problem; the configuration relationships between computing, networking, and memory must also be redesigned.

One potential adjustment could be to reduce the LPDDR5 configuration amount per rack so that limited storage supplies can support more rack deliveries; another direction could be to reduce the reliance on traditional DRAM through networking, caching, and faster on-chip storage in certain scenarios.

Management also mentioned technologies with SRAM as the main storage architecture; while the cost of SRAM is generally higher than that of DRAM, its low latency and high bandwidth could provide system-level value under certain inference loads.

These statements may temporarily affect market sentiment toward certain storage companies, but their logical starting point is not a weakening of storage demand. On the contrary, NVIDIA needs to continuously adjust architectures because the company expects long-term storage supply to struggle to meet the growth of AI computing.

Therefore, for the storage industry, what truly warrants attention is not whether NVIDIA reduces a certain memory configuration per machine, but how the value of memory within the AI system will be redistributed among HBM, LPDDR, SRAM, caching, and networks.

4. The Importance of Nemotron and Open Models

In addition to hardware, NVIDIA is also strengthening the role of open models and enterprise AI software.

Management cited professional scenarios like circuit design, indicating that generic closed-source models may not meet enterprises' needs for specialized knowledge, data security, and workflows; what enterprises truly require is often not just to directly call an external model, but to establish a controllable and adjustable AI system around their own data and experience.

The significance of open models like Nemotron lies here.

They allow enterprises to train, fine-tune, and deploy around their own business while retaining control over models, data, and infrastructure. For NVIDIA, open models are not merely a software product; they are an important entry point that connects GPU, networking, inference frameworks, models, and enterprise applications.

NVIDIA hopes to construct not just a high-performance computing hardware suite but an AI technology stack that enterprises can fully grasp and deploy.

3. How Does Morgan Stanley Value NVIDIA?

Another important purpose of this roadshow is to expand NVIDIA's coverage among value investors.

As is well-known, over the past few years, NVIDIA has primarily been viewed as a typical high-growth asset, but with the company's market capitalization and institutional holdings continuously expanding, many growth funds have widely held shares of NVIDIA, with some institutions even approaching their internal single holding limits. Therefore, for NVIDIA, relying solely on growth funds to continue increasing allocations has become limited.

NVIDIA thus needs to provide the market with another valuation logic.

Management anticipates that from this stage onward, the company will allocate 50% or more of its cash flow to shareholder returns. In this context, the continuously increasing stock buybacks and cash returns enable NVIDIA to maintain high growth attributes while gradually acquiring the value characteristics of a large mature tech company.

In other words, NVIDIA hopes that the market will use two frameworks to value it: on one hand, it remains a company capable of sustaining high growth in AI infrastructure; on the other hand, it is becoming a super-large tech platform capable of continuously generating and returning significant cash flow.

Based on this judgement, Morgan Stanley continues to list NVIDIA as the preferred choice in the semiconductor sector, maintaining an "Overweight" rating and a target price of $288, while providing three risk-return scenarios:

• In the base scenario, the target price is $288, applying about 22 times the projected earnings per share (EPS) for calendar year 2027, corresponding to EPS of $13.08: the report predicts that NVIDIA's revenue will grow by 82.0% in 2026 and by 52.4% in 2027;
• In the optimistic scenario, the target price is $330, corresponding to about 23 times the calendar year 2027 EPS of $14: its core assumption is that data center business continues to grow rapidly, and networking, Vera Rubin systems and software revenues drive NVIDIA to obtain a higher premium for the full-stack AI computing platform. Further upside potential may come from scaling AI PCs, autonomous driving, and robotics businesses, as well as increased proportions of high-margin software and AI services revenue.
• In the pessimistic scenario, the target price is $160, corresponding to about 16 times the calendar year 2027 EPS of $10: major risks include data center supply catching up with demand more quickly, a significant slowdown in the growth of AI infrastructure; customers reducing their dependency on NVIDIA through self-developed ASICs; AMD or other competitors regaining market share; and tariffs and export restrictions having a greater than expected impact on revenue;

From the profitability forecast perspective, Morgan Stanley estimates that NVIDIA’s GAAP revenue will respectively reach $215.938 billion, $393.005 billion, $598.809 billion, and $783.877 billion from 2026 to 2029; gross margins will be 71.3%, 74.4%, 72.5%, and 72.0%, with EPS of 4.61, 8.96, 13.08, and 17.63 dollars respectively.

The core judgment behind these forecasts is that even though NVIDIA's revenue base is already enormous, the expansion of AI infrastructure is still expected to drive its revenue and profit to maintain high growth rates over the next several years; networking, CPUs, software, and system-level solutions will contribute more increments beyond GPUs.

Of course, Morgan Stanley also acknowledges that storage, networking, and other supply chain segments may have higher profit elasticity at certain stages. In fact, in past semiconductor cycles, Morgan Stanley has rotated its sector preference from NVIDIA to companies like SanDisk and Micron.

However, from a risk-return perspective, NVIDIA remains one of the most attractive core assets in the semiconductor sector.

On one hand, the company’s current valuation multiples may be constrained by its large market capitalization and index weight; on the other hand, compared to computing semiconductor companies like AMD and Broadcom, NVIDIA does not enjoy a particularly notable valuation premium. Morgan Stanley believes that as the sources of growth diversify and cash returns increase, the valuation differences between NVIDIA and some of its peers are expected to converge.

Nevertheless, market positions do not exhibit completely uniform optimistic expectations. The report shows that the ownership ratio of active institutional investors has reached 50.9%, indicating that the holding base of growth funds is quite high; at the same time, Morgan Stanley's quantitative models for NVIDIA's future three months and 24 months are both in the 4th quintile, not yet entering the most favored range by quantitative funds (the 1st quintile represents the most favored, and the 5th quintile represents the least favored).

This also constitutes a rather subtle aspect of NVIDIA's current valuation: while fundamental expectations remain strong, the high institutional holdings, large market capitalization, and market expectations also mean that the company needs to continuously deliver growth to further drive its valuation upwards.

Final Thoughts

The past discussions about NVIDIA often revolved around a simple question: How long can AI capital expenditures continue to grow?

However, this roadshow offered another perspective.

NVIDIA's next phase of growth does not entirely depend on whether a few large cloud providers continue to increase GPU purchases, but rather on its ability to further expand its platform capabilities to networking, CPUs, models, enterprise software, sovereign AI, and new cloud infrastructure.

ASICs will continue to grow, storage and power constraints will not disappear, and product roadmaps may still adjust; what truly determines NVIDIA's long-term value is not whether it can prevent these changes, but whether it can continue to translate each industry change into a larger serviceable market.

For a company with quarterly revenue approaching $100 billion, "accelerating growth" may sound counterintuitive.

But what NVIDIA is trying to prove to the market is that as AI infrastructure transitions from chip procurement to system construction, and expands from a few cloud providers to nations, enterprises, and new cloud platforms, it still has the capability to secure a larger value share at each layer of the technology stack.

This also marks a longer growth curve after transitioning from an AI chip supplier to a full-stack infrastructure platform.

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。