Morgan Stanley: Despite ASICs, Nvidia will continue to maintain a dominant market share. The ASIC category is neither superior nor inferior to commercial GPUs—it is simply another means to achieve the same result. Over the past six months, momentum in the AI sector has clearly shifted toward custom silicon. Nvidia has remained stagnant while AMD has performed sharply poorly. Nvidia’s $3 trillion market capitalization is supported by over $32 billion in quarterly AI revenue, whereas AVGO’s $1.1 trillion market cap is based on $3.2 billion in quarterly revenue. Clearly, the market has judged that the growth potential of ASICs will be several multiples that of commercial GPUs. Will ASICs outperform commercial solutions in the long term? While there are various possible outcomes, our view is that—unless something changes—the incumbent, namely Nvidia, will continue to hold a dominant market share. The ASIC category is neither superior nor inferior to commercial GPUs. It is merely an alternative means to achieve the same result. We would evaluate an ASIC the same way we evaluate a chip from AMD, a chip from Intel, or any chip from a startup—in the context of the price/performance of that silicon relative to the value the incumbent (primarily Nvidia) can deliver. The development budget for an ASIC is typically less than $1 billion, and in some cases, much lower. This contrasts with our assumption that Nvidia will invest approximately $16 billion in R&D this year alone. With that funding, Nvidia can maintain a 4–5‑year development cycle by running three design teams sequentially—each with an 18–24‑month architectural cadence—delivering innovation over a five‑year span. In addition, they invest billions in interconnect technologies to boost rack‑scale and cluster‑scale performance, and by being present in every cloud worldwide (subject to U.S. Department of Commerce approval), any investment in improving the Nvidia ecosystem propagates throughout the global ecosystem. We have examined several theories regarding whether ASICs can outperform commercial solutions. Could a custom chip be superior because it is better suited for a narrower set of applications? Sometimes, yes—and in such cases, it becomes one of the most compelling use cases for customization. The clearest example is Google’s successful TPU tensor processor. Google invented modern LLM transformer technology and directed Broadcom to develop a chip optimized for that technology, while at the time Nvidia was developing GPUs optimized for vision models using convolutional neural networks. To date, the TPU is the most prominent example of a cloud customer gaining a clear advantage through a customized solution, resulting in over $8 billion in TPU revenue for AVGO. However, an important point is that Nvidia is expanding its share of Google’s spending in 2025. Part of this is attributable to investments in the cloud sector, where commercial products tend to outperform ASICs. Yet the customization aspect of the TPU may be less valuable than it once was. Nvidia is also optimizing for transformer models, and in fact, the largest training and inference clusters are not currently highly customized. Instead, the greater benefits of customization going forward seem more likely to manifest in legacy workloads. While Nvidia’s focus on training multimodal AGI models may be overkill for some older applications, it will be difficult to outperform Nvidia when it comes to delivering high‑end training capabilities. In particular, for Google, while purchases of Nvidia products are expected to roughly double this year, TPU growth is anticipated to be more modest—partly due to Google’s investments in enterprise cloud, but also because Nvidia’s LLM transformer performance is exceptionally strong even for internal workloads. Is an ASIC cheaper? This is the most frequently heard argument. Yes, a given ASIC may not reach the performance levels of the H100 until three years after its debut, but it costs only $3,000, whereas the H100 is priced at $20,000—resulting in a lower total cost of ownership. But does this make sense? Is it really that easy to compete with Nvidia merely by producing a $3,000 AI chip? We have seen countless startups, priced only partially at Nvidia’s level, fail to gain a foothold in this space. Moreover, Intel has struggled for over a decade—even after acquiring companies that were already shipping—while pricing below cost. AMD’s first few generations were non-starters until they broke through with the MI300 last year. If a $3,000 chip were all that was needed, why wouldn’t every competitor simply produce a $3,000 chip? Meanwhile, Nvidia has introduced several lower-priced chips aimed at legacy inference applications. Two years ago, we reported enthusiastically on the L4, L40, and other announcements. However, Nvidia has also discovered that there is a gravitational pull toward the most expensive cards. Why is that? While the processor itself may be cheaper, system costs can be higher. The cluster cost for ASICs can be materially higher than that of Nvidia—which has built a 72‑GPU NVLINK domain using copper—because ASICs use more expensive optical technology. Other major cost components are at similar levels. High‑bandwidth memory costs are the same—and in fact, Nvidia is likely favored by its monopsony purchasing power for the latest HBM versions. The same applies to CoWoS; because many ASICs use smaller dies with larger stacks, the CoWoS cost can be higher than that of Nvidia. Of course, Nvidia’s wafer costs might be higher due to reticle‑limited dies, but overall, Nvidia delivers exceptional value. Furthermore, the software aspect also places a burden on customers, as it is a challenging and time‑consuming task. Ease of use when managing software changes and running various workloads is crucial, and minimizing “software developer hours” is often an overlooked element in total cost of ownership (TCO) calculations. For instance, Databricks—a Trainium customer—anticipates that it will take “weeks or months to get the system up and running” (link reference). Deployment delays can put customers even further behind Nvidia, which is equipped with the widely used CUDA (Compute Unified Device Architecture) SDK. Over the years, we have observed 20 to 25 alternatives to Nvidia’s products. Initially, there was enthusiasm based on price and potential performance that drove early deployments, but ultimately customers gravitated back to Nvidia, which has the most mature ecosystem. Consequently, alternative products are often put on hold or sometimes disappear from the market altogether. TPU and Trainium, as well as AMD’s MI300, are notable exceptions—but even in those cases, strong investments in Nvidia’s ecosystem continue this year. This is not to say that cheaper processors are entirely irrelevant; rather, historically, cheaper processors have often failed to garner as much market attention as initially expected. Could the ASIC relationship give ASIC vendors additional visibility that commercial products do not possess? We disagree. Every hyperscaler investing in ASIC technology intends to use it. ASIC development incurs significant costs because there is an expectation of returns. Therefore, there is a possibility that the SAM analysis is correct. However, the technology must ultimately be more cost‑effective than commercial solutions when delivered; if not, purchasing will often pivot back to the best‑in‑class commercial chips. One exasperated cloud executive recently remarked, “Every two years, our ASIC team delivers technology that is 2–3 years behind Nvidia. It is economically not that useful.” Of course, this is not everyone’s view, nor is it the objective at the outset of the design process. However, it is a fairly common complaint among cloud vendors, who—even while considering it an investment in the future—are not entirely convinced that adopting inferior ASICs will form the foundation of a differentiated long‑term strategy. Put another way—why does AVGO’s SAM analysis attract far more attention than AMD’s predictions of “tens of billions in annual revenue potential” for products like the MI400? AMD’s forecasts are also based on conversations with potential customers, yet investor responses are, “How can we be confident in AMD’s 2026 product (MI400) compared to Nvidia’s product (Rubin)?” Indeed, we cannot be sure. Nevertheless, for some reason, we expect that ASICs will be competitive with Rubin in 2026. However, AMD’s investments should not be taken lightly. The scale of AMD’s investments across the ecosystem tends to far exceed that of ASIC vendors. This year, AMD has completed two acquisitions of AI software assets. One of these—the acquisition of ZT Systems—involved acquiring a major server ODM, divesting the ODM business while retaining key engineering talent related to rack‑ and cluster‑scale computing. Once in possession of such assets, AMD can deploy them across multiple cloud environments, which drives third‑party support and accelerates ecosystem development. Can ASIC designers within the cloud replicate this? Are ASIC chip suppliers willing to replicate it? It is possible—and their networking expertise may help—but it is by no means a given. AMD has secured dominant technology in the server ecosystem through such iterative investments, with the progression from the early struggles of Naples to Rome and Genoa serving as a success story (though Intel’s struggles have clearly contributed as well). So why is it presumed that AMD will fail to meet its stretch goals while ASICs will succeed? In our view, AMD has just as much potential to achieve its SAM as ASIC competitors, and this is perceived by buyers as a “show-me” story. So why has ASIC not yet become a “show-me” story? In our perspective, ASICs should also be expected to deliver; this does not mean that either technology offers a guaranteed solution, but rather that the challenges of competing with Nvidia should not be taken lightly. It is difficult to say that the ASIC relationship provides suppliers with a more fixed market share, as nearly every ASIC we know of faces direct competition. For example, MediaTek (as reported by Charlie Chan) is expected to challenge Broadcom within the Google TPU ecosystem, and Alchip (also reported by Charlie Chan) is anticipated to challenge Marvell within the Amazon (as reported by Brian Nowak) Trainium ecosystem. Recent reports indicate that Meta (also reported by Brian Nowak) is considering the acquisition of a Korean ASIC design—suggesting that its relationship with Broadcom is not exclusive—and Meta is also investing significantly in AMD. Bytedance is Broadcom’s third ASIC customer, but there is also the possibility that they have alternative ASIC designs from Chinese vendors in the pipeline due to export control issues. In fact, export controls on AI-related shipments to China have lowered the technological threshold to 80% of the state‑of‑the‑art, creating a highly open competitive environment. OpenAI could become a major ASIC user, but reports indicate that they are already working with multiple suppliers. Nevertheless, Broadcom and Marvell possess enviable capabilities in this field that will undoubtedly drive growth. Success begets success, yet there are clearly challenges within the ASIC domain as well. We expect that in 2025, Nvidia and AMD will outperform ASIC competitors—especially in the second half of the year. For 2024, Nvidia’s processor revenue is projected to be approximately $98 billion, AMD’s around $5 billion, AVGO’s about $8 billion, and the combined revenue of Alchip/MRVL roughly $2 billion. This implies that commercial silicon accounts for roughly 90% of the market share, while ASICs constitute about 10%. We expect the 90% share for commercial products to increase slightly this year. The two major users of ASICs are likely to further increase the proportion of spending on Nvidia. AVGO’s largest customer is Google, which drives most of its 2024 revenue, and we expect that in the 2025 fiscal year, Nvidia will grow 50–100% more than TPU. Marvell is driven by Amazon, and while Amazon’s ASIC purchases are expected to double from approximately $2 billion to about $4 billion this year, Amazon’s purchases of Nvidia products are also projected to more than double beyond that level. Other ASIC customers, such as Meta and Bytedance, are still in the early stages. In our view, Nvidia’s revenue momentum in the second half will be significantly stronger than the builds from ASIC or AMD. Beyond 2026, ASICs could grow much faster. However, this largely depends on the value delivered by ASICs versus what Nvidia or AMD can offer. According to AVGO’s SAM analysis, Google, Meta, and Bytedance could each achieve between $60 billion and $90 billion in revenue in the 2027 fiscal year with clusters numbering in the millions—which is entirely plausible, though that technology has yet to be delivered. Yet, with the market discounting such significant market share gains for ASICs—for instance, Nvidia generates 10 times the AI revenue of AVGO this quarter while its market cap is only three times as high—doesn’t the onus fall on the ASIC side to prove its case? And if we consider the Blackwell ramp in the second half of 2025 as the major story, wouldn’t the market capitalization that shifted from Nvidia and AMD to AVGO/MRVL over the past six months swing back? Does this contradict our earlier work on the ASIC opportunity? In the deep‑dive report on the ASIC opportunity led by Charlie Chan last year (link here), we forecasted that the total addressable market (TAM) for AI ASICs would grow from $12 billion in 2024 to $30 billion in 2027—a figure considerably lower than what is being discounted in the market now (for example, AVGO’s SAM for 2027, based mainly on ASICs for three out of five potential customers, amounts to $60–90 billion). Our view remains unchanged, but market expectations have shifted dramatically. In particular, Charlie Chan maintains a bearish view on Asian peers such as Alchip and MediaTek, as they are securing projects at the expense of their U.S. competitors. Referring to the same report, Asian vendors like Alchip (with the AWS 3nm project) and MediaTek (with Google’s 3nm TPU v7 for training) are expected to expand their market share in 2026. So what should we do regarding U.S. AI semiconductor stocks? Simply put, buy Nvidia. The stock remains our top pick and is currently trading at a significant discount compared to AVGO/MRVL. We believe that this discount will not last long. Nvidia’s biggest short‑term risk is U.S. export controls, which are equally problematic for AVGO. In the long term, the greatest risk is not competition but a slowdown in investment—which we forecast to occur around mid‑2026, though we continue to push that timeframe out as visibility remains strong. Meanwhile, we continue to maintain a bearish view on AVGO while closely monitoring expectations. Frankly, feedback on this note will play a crucial role in shaping investor expectations for AVGO relative to the 2027 SAM analysis. This year, AVGO AI remains in a transitional phase—with one fully penetrated customer (Google) that has slowed due to a product transition and two new customers that are still small. The company’s three‑year SAM analysis was partly presented to explain the short‑term slowdown and the strong ongoing investment, but the trajectory—from $3.2 billion per quarter to $18 billion over 12 quarters—could raise short‑term expectations even further. We are sidelining AMD, MRVL, ALAB, MU, etc. We are not fundamentally negative on any AI‑related stocks, but we do have concerns about short‑term momentum shifts—particularly as the multiple premiums for MRVL and AVGO far exceed those of Nvidia. While AMD’s valuation is becoming increasingly attractive, the MI350 does not seem revolutionary, and the MI400 remains somewhat of an unknown. We have confidence in the execution of both AMD and Nvidia. ALAB stock is so volatile that it is difficult to make a persistent valuation call, but at current levels, we maintain a constructive view. Management has indicated that Nvidia will be the strongest driver of demand in 202X, yet Nvidia’s data center—which returns twice the operating margins of its peers—continues to record the lowest EV-to-AI revenue ratio in our analysis. Source: Morgan Stanley Research $NVDA $MRVL $AVGO