Microsoft Upgraded, Broadcom Slides, Korean Chips Rattle

The central tension in AI investing right now is visible in a single day’s news: fundamentals keep printing strong, but markets keep selling the stocks anyway. Tuesday’s session crystallized that gap across several names.

Key points

• A new Seeking Alpha analysis upgraded Microsoft (MSFT), arguing that pressure on the stock reflects margin anxiety around AI capital spending rather than any weakness in actual demand, and that the multiple looks cheap relative to Azure’s trajectory.
• Broadcom’s AI revenue grew 143%, yet the stock has dropped roughly 22% over the past month, a split that underscores how elevated expectations can overwhelm even exceptional results.
• South Korea’s Kospi fell 6%, with Samsung and SK Hynix each sliding about 8%, as investors questioned whether AI-driven memory demand can justify current valuations.
• DeepSeek is reportedly building a custom AI inference chip to cut reliance on Nvidia, joining Zhipu in a pattern of Chinese AI firms pursuing in-house silicon.
• The Nasdaq-100 fell 525 points in the session, with chip selling, Microsoft job cuts, the DeepSeek chip report, and geopolitical tensions (US-Iran airstrikes lifting oil and the dollar) all cited as contributing factors.

What is actually driving the Microsoft bear case?

The Seeking Alpha upgrade lays out the argument clearly: MSFT’s stock is not under pressure because Azure is struggling. The concern is that the massive wave of AI infrastructure spending will compress margins before revenue scales to match it. The analysis points to three specific signals investors should watch, centered on how Azure and cloud cost-of-goods trends compare to revenue growth over coming quarters. The core thesis is that if those two lines begin to converge favorably, the current valuation looks genuinely cheap.

For investors, this framing matters because it reframes the risk. Microsoft is not a demand story to doubt; it is a timing story about when capex translates into operating leverage. That is a meaningfully different bet, and one with a clearer set of observable checkpoints.

Why is Broadcom falling despite triple-digit AI revenue growth?

A 143% jump in AI revenue is an extraordinary result by any normal measure. The problem for Broadcom’s stock is that expectations had already priced in an aggressive growth trajectory, and a 22% decline over the past month suggests the market is now recalibrating what the next leg of growth looks like. This is a recurring dynamic in the current AI cycle: the question is rarely whether a company is growing fast, but whether the growth rate itself is accelerating or beginning to plateau.

Broadcom’s custom chip business for hyperscaler customers remains a genuine competitive advantage. But investors appear to be asking whether the concentration of that business among a small number of large clients creates its own risk, particularly as those same clients explore more in-house silicon options.

Korean memory stocks: a demand reality check or an overreaction?

The Kospi’s 6% single-day drop and the 8% declines in Samsung and SK Hynix look jarring alongside Samsung’s own record $61 billion profit, which was itself driven by AI memory demand. Foreign investors were net sellers throughout the session, and the concern articulated across coverage was not that AI demand has collapsed but that record earnings may already be fully reflected in share prices.

That distinction is important for anyone holding memory-adjacent positions. A company can print record profits and still see its stock fall if the market decides the peak is visible. HBM demand from Nvidia and the hyperscalers remains structurally strong, but the Korean market appears to be pricing in a more cautious view of how long the current cycle sustains those margins.

Chinese firms keep pushing toward chip independence

Both DeepSeek and Zhipu are now reported to be pursuing custom silicon, with DeepSeek focusing specifically on inference rather than training workloads. This is a logical split given where costs accumulate once a model is deployed at scale. Separately, a Chinese research team unveiled a memristor chip described as capable of brain-structure mapping up to 478 times faster than Nvidia’s A100 in specific tasks. That claim is specific to neural simulation workloads and does not represent a general-purpose GPU challenge, but it reflects the breadth of silicon innovation now targeting Nvidia’s dominance from multiple angles.

None of these efforts pose a near-term threat to Nvidia’s data center revenue. Over a multi-year horizon, the cumulative effect of hyperscalers, sovereign AI programs, and well-funded startups all reducing their Nvidia dependence is a risk worth tracking.

---

This update is for informational purposes only and does not constitute investment advice.