Morgan Stanley semiconductor analyst Joe Moore's latest discussion: What's the current view on Nvidia, Broadcom, Micron, AMD, Intel, Marvell Technology, and Astera Labs?

Today's content comes from the latest episode of Morgan Stanley Tech Talk, hosted by Tom Wade of the Tech Specialty Sales Department, with guest semiconductor analyst Joe Moore, who just returned from the Nvidia GTC conference.

This episode covers a broad spectrum—Nvidia, Broadcom, AMD, Intel,Marvell Technology,Astera Labs, and ends with Micron's quarterly results. The information density is very high, and I will do my best to fully restore their core views.

Nvidia: Why Isn't the Stock Moving?

Tom began with a question everyone is asking: Nvidia has beaten expectations for two consecutive quarters, the new products at GTC are impressive, valuations are still cheap, so why isn’t the stock price rising?

Joe’s answer was direct—there are concerns about the sustainability of AI trades and market share, which are the key suppressing factors. However, he believes Nvidia is doing everything right: the data is excellent, and management’s comments are very measured. He specifically points out that Nvidia, unlike other AI companies, did not give hard performance guidance for next year, which means their numbers are much more conservative than their peers.

In the long run, Joe is very optimistic about the technological capabilities and market share growth brought by Vera Rubin. He admits that around this time every year the market gets anxious about the following year, but emphasizes that short-term data points are more supportive than the same periods of the past two years. All semiconductor companies currently state that prospects look good up to 2027, with Oracle publicly saying its capital expenditures will continue to rise into 2027.

As for the issue of limited cash flow among hyperscale companies, Joe’s view is: these are indeed long-cycle investments with complex cash flow characteristics, but the return speed of chip investments is very fast. Once a data center is built, racks start generating returns immediately—Nvidia's share in the entire financing process has no risk at all.

Joe named Nvidia as the top pick for Morgan Stanley’s TMT Conference, as it was in the first half of last year as well. He says he’s not sure exactly when market confidence will return, but is convinced the direction is right.

One Trillion Dollar Revenue Target: What's Included?

At GTC, Nvidia offered a Wall Street-facing figure—the accumulated revenue target for Rubin and Blackwell-related businesses from 2025 to 2027 is one trillion dollars. This figure is the focus of much debate in the market.

Joe broke down this number from the most conservative perspective. According to the CFO, the one trillion dollars includes networking business and Viva Rubin. The market consensus for data center business revenues in the same period is about $950 billion, leaving $50 billion growth under a conservative framework.

But even under this conservative framework, the Hopper series has not been included—about $25 billion. There are also CPU, software, RTX, and other businesses totaling about $25 billion per year. Adding it all up, the total reaches $1.1 trillion, far exceeding the consensus expectation of $950 billion.

However, it should be noted that 5 out of the 12 quarters have already passed—no room for growth in those. But looking at the remaining 7 quarters, there’s about 20% revenue upside potential compared to market expectations. Furthermore, new demand keeps pouring in, token growth is surging, Agoric is just starting—Joe finds it hard to imagine growth slowing down.

Joe also mentioned an interesting contrast: Wall Street’s current concern over “sustainability” is at a three-year peak, but his own visibility on sustainability is at an all-time high. The reason is that semiconductor shortages have led to prepayments and long-term contracts. The general industry response is—what exactly is Wall Street worried about? There’s a clear shortage of computing resources.

He also notes that CNBC keeps discussing whether “AI is a bubble.” People consider this a contrarian view, but the reality is the opposite—the adoption rate of models is very strong.

Nemo Claw: Ecosystem Play or Revenue Driver?

Nvidia launched Nemo Claw at GTC, and Tom asked whether it’s a potential source of revenue or an ecosystem play.

Joe believes this will directly generate revenue. He was surprised that this open-source project had only been around for weeks, but Nvidia has already built a complete business model around it. The core issue is: how to run AI agents safely in enterprise environments? After all, these agents access sensitive enterprise data. Nvidia is developing software around this pain point and has even done benchmarking against other similar software companies.

Joe sees this as part of Nvidia’s integrated hardware and software strategy, providing business optionality. More importantly, the rapid adoption of such technologies will drive token growth and computing demand. Unlike the past three years, however, semiconductor capacity is already maxed out, with very limited ability to respond to extra demand.

AI Capital Expenditure: Still Accelerating

Based on his personal channels, Joe observes that AI capex is not only strong but accelerating further.

In the short term, token growth is explosive. Companies originally planned to use half their processors for training and half for inference, but actual demand has far exceeded expectations, and more training hardware is being shifted to inference. Hock and Broadcom mentioned that some clients saw a 10x increase in token volume, yet their investment plans were for just 3x—that already seems aggressive, yet capacity still can’t keep up.

Joe says if you look at the S1 filings of upcoming IPOs, you’ll see clues explaining the current spending levels. The enthusiasm across the ecosystem is very high. A multi-trillion dollar five-year spend sounds aggressive, but capacity expansion is linear while demand growth is super-linear. Moreover, the bottleneck is not just in processors—it’s memory, optical components, hard drives—every component faces supply constraints.

Broadcom: Attractive, but Not as Good as Nvidia

Tom brought up Broadcom, and Joe’s attitude was clear: Nvidia has a lower valuation and he’s more optimistic about Nvidia, with more upside as well. But Broadcom’s position is also attractive.

Broadcom expects AI revenue to exceed $100 billion next year, and Morgan Stanley raised its forecast from $85 billion to $115 billion. Yet Joe thinks Broadcom’s upside is not as big as what Nvidia’s model offers.

For specific business segments: Joe has confidence in Google’s TPU and the networking business (the fastest-growing part of Broadcom’s AI business). But he is cautious about Meta’s MTA—scale is small, teams are mostly from Google and Apple, and Meta people are using TPU through Google Cloud. As for OpenAI's custom chips, as new team products, they still need time to prove themselves.

Looking at the remaining businesses (software and RF), Broadcom’s valuation is about twice Nvidia’s. Although there are more options, Nvidia’s risk/reward at current price levels is clearly better.

CPU Recovery: Real but Hard to Participate Through AMD and Intel

On the topic of CPU recovery, Joe confirmed that demand for traditional CPUs is indeed strong. On one hand, Intel’s capacity has long been limited, supporting supply; on the other, smart agents are just emerging, and traditional compute platforms are seeing a large bump in workload.

However, Joe is cautious about participating in this round of recovery through AMD and Intel. AMD’s CPU business is doing okay, but with the market focused on AI, he is in wait-and-see mode, wanting to first see the results of the 455 and related data validation. Intel’s main debate lies in its foundry business, about which Joe is skeptical. Market improvements for CPUs might boost Intel's profitability, but current profits are poor—a small improvement won't excite the market.

He prefers to allocate to the server market through memory—demand for DRAM and NAND is currently very strong. The core controversies for AMD and Intel are not in their CPU segments.

Marvell Technology and Astera Labs: Both Worth Watching

Joe is positive on both companies.

On Marvell Technology, the market had been pessimistic, but last quarter the company increased its focus on non-ASIC businesses. Joe believes competing directly with Nvidia is very tough; positioning the ASIC business in niche markets is the smarter move. In optical communications, 800G continues to grow, and 1.6T DSP products are ramping up quickly. After last year’s negative incidents, Marvell now looks undervalued and more like a value play.

As for Astera Labs, although small-cap stocks usually have higher valuations, the growth drivers are clear. The lateral expansion market for copper cables (scale-up) is still at an early stage, and there are more opportunities ahead for rack-level full-switching architectures. The Scorpio X switch’s primary customer is Amazon Trion 3, with 20 more potential clients. Scorpio Op has also added several large customers.

Micron: Big Beat on Earnings, So Why Did the Stock Drop?

This is both the final and the most in-depth part of the conversation. Joe has always been sharp in storage—last year he picked SanDisk first, then switched to Micron, both proven right.

Micron’s results this quarter significantly exceeded expectations, but the stock price fell. Joe attributes this to market worries over rising capex and the sustainability of high-expected profits.

On pricing, Joe provided concrete numbers: first-quarter general storage prices rose 90%, Q2 is expected to rise over 50% (“over”), and Q3 will continue rising. DRAM supply shortages have been an ongoing issue, and if AI maintains a growth path above 50%, demand will only increase year by year.

Micron’s gross margin reached 81% this quarter—the highest Joe has ever seen as a semiconductor analyst. Traditionally, such high gross margins, alongside a flat spot market for a month, slower Q2 price increases versus Q1, and rising capex, do trigger profit-taking. But he thinks, if AI holds steady, memory will remain the spending bottleneck, and an 81% gross margin will be hard to bring down.

Joe did the math: per performance guidance, Micron’s current annualized earnings are $76. If they can maintain over $80 in the next two years, despite high capex, $150 billion in cash flow could be generated. Chip Act restrictions prevent large returns to shareholders, so most of it will accumulate on the balance sheet. All debt will be cleared, and the cash surplus can be used for buybacks, providing downside protection. The current valuation is about 5x PE—if profits hold for two years, value will be unlocked.

Regarding long-term contracts, Joe admits Micron disclosed little this quarter. One five-year agreement is set, others are being negotiated. But the deal structures he’s learned about—prepayments, large upfront cash—will still show up on the balance sheet even if the company doesn’t disclose: rising cash, increased deferred income. This should relieve concerns over sustainability over time.

Lastly, Joe circled back to Nvidia: if Nvidia can relieve market doubts about sustainability, the whole sector benefits. Micron’s optimistic price target is about $600, and if profits last for years, that’s entirely reasonable.

My Understanding:

This episode of Morgan Stanley’s Tech Talk is packed with information, but the core logic boils down to one line: the sustainability of AI demand.

One contrast Joe repeatedly highlighted is worth noting—while Wall Street’s worries over sustainability are at a three-year high, industry-side visibility is as high as it’s ever been. Prepayments, long-term contracts, explosive token growth, and semiconductor capacity at its limits—these are hard data, not emotions.

For ordinary investors, a key takeaway from this conversation is not to focus solely on single-quarter financials or stock reactions, but to understand where the core constraints of this cycle are. Joe’s answer is clear—memory is the current bottleneck for AI spending, processor capacity is expanding linearly while demand grows super-linearly. Within this framework, many “too high” margins and capital expenditures may just be structural, not cyclical.

Of course, Joe doesn’t ignore risks: hyperscale company cash flow pressure is real, and returns on long-cycle investments are indeed complex. But he distinguishes between the long cycle of construction/infrastructure and the short cycle of chips generating returns immediately after deployment. This distinction is important.