EP693 | 🍖

The market quickly recovered after the sharp July sell-off

The market plunged about 30% from its high in July. At the time, the host had expected it to take until the end of the year to recover the lost ground, but the rebound since August has been surprisingly fast. Although the broader market is still volatile and remains some distance from the peak reached during July’s most euphoric period, it has generally moved close to a new high.

In the absence of any major surprises, the host believes it could reach another high within the next one to two months. This kind of movement can be viewed as the market gradually recovering lost ground while taking a breather, rather than rising in a straight line every day.

The host’s core approach to this market phase is to hold positions while reducing frequent trading. Momentum investors may feel that the market lacks explosive strength, but many sectors have continued advancing since August, and optical communications-related stocks have remained strong. Market events often trigger dramatic reactions at first, while the fundamentals ultimately may not have changed much.

Short-term traders therefore need to accept that volatility can create opportunities as well as inflict losses. The host notes that the AI industry is still in its early stages, and the market may become large enough for multiple participants to benefit at the same time. The difficulty lies in predicting the ups and downs in their share prices along the way.

MediaTek’s $3.5 billion convertible bonds deepen the partnership

MediaTek2454.TW issued approximately $3.5 billion worth of convertible bonds. NVIDIA$NVDA took the vast majority of the offering, while Google$GOOGL also participated. This marks NVIDIA’s first investment in a publicly listed Taiwanese company and is one of the largest convertible bond investments in history. The bonds carry a 0% coupon, and investors paid a premium of more than 10%. This shows that subscribers were not interested in the interest payments, but in the option to convert the bonds into shares if the stock price rises in the future.

The host believes the transaction should not be simplified as NVIDIA financing MediaTek or helping it deal with inventory. MediaTek is financially sound and highly profitable. Unlike companies that depend on renting out computing capacity or surviving on a single customer, it is in a much stronger position. The deal therefore looks more like a long-term alignment between strategic partners. MediaTek has already worked with NVIDIA on laptop chips and automotive cockpits, and the relationship may now deepen further by combining NVIDIA’s AI platform with MediaTek’s custom XPUs.

Google is a MediaTek customer, and MediaTek is also helping design Google’s TPU for the first time. The product roadmap already has a basic shape. However, the host believes that Google and NVIDIA subscribing at the same time does not mean the three companies are deliberately forming an alliance. Google continues to develop its own TPU, ICI interconnect, and OCS optical switching solutions. These products are even important competitors to NVIDIA’s interconnect offerings. The two subscriptions are therefore more likely to mean that both companies see value in MediaTek’s capabilities, rather than that they are jointly building a single ecosystem.

NVIDIA wants to sell an entire AI factory, not just GPUs

Customers developing their own XPUs could indeed take away some demand for GPGPUs, directly affecting NVIDIA and AMD$AMD. But the host believes it is not enough to focus on the market share of any single type of chip, because the total addressable market for AI infrastructure is also expanding. As applications continue to grow, an increase in custom chips does not necessarily mean NVIDIA’s revenue will decline accordingly.

A more likely scenario is that the market gradually shifts from domination by a few companies toward greater division of labor among multiple players, while the existing leader continues to earn revenue from other parts of the system.

NVIDIA is therefore gradually shifting from “selling GPUs” to “selling complete solutions.” In addition to the main NVL72 Compute Rack, it offers MGX, ETL, Vera CPU Rack, storage racks, and Spectrum networking solutions. These products can work alongside XPU designs developed by customers themselves. In other words, if customers want control over their own XPUs, NVIDIA does not necessarily have to stop them. It can continue selling its own interconnect, rack, storage, and networking equipment outside the core chip. This is like moving from selling only an engine to supplying the chassis, drivetrain, and control systems for an entire vehicle.

This model may be less straightforward for Google, which already has a relatively complete interconnect solution. But it could be attractive to other customers that have not built a complete data center ecosystem. NVIDIA is betting that after MediaTek helps a customer develop an XPU, NVIDIA can still capture value through the surrounding solutions. The host also notes that NVIDIA appears to be trying to bring all the major players into its partnership network, potentially turning companies that might otherwise compete with it into customers or supply partners at different levels.

NV-HBM is still being validated, and TSMC may capture more value

NVIDIA says its NV-HBM offers 30% more bandwidth and 15% lower power consumption than HBM4, while moving the controller from the accelerator to the HBM base die. Its difference from the current JEDEC standard is that it appears to be a NVIDIA-specific specification. This could allow NVIDIA to capture additional value in the memory supply chain. The host sees it as NVIDIA bringing hardware specifications further under the control of its own ecosystem, thereby increasing product lock-in.