Newsletter commentary July 2026

Shift in Value Distribution

In July, the AI supply chain underwent a sharp correction, while most other sectors with somewhat better fundamentals experienced localized performance opportunities. We recognized early on that issues were emerging, driven primarily by overly rich valuations: market expectations were built on multiplying projected 2028 or 2030 market shares by profit margins, and then applying a 30x P/E multiple. In reality, the probability of these expectations materializing is extremely low. Even if they are realized, if the market still applies a 30x P/E multiple years down the road, where is the time value of money?

From a relative standpoint, however, we retained a degree of attachment toward the domestic supply chain, which was on the verge of large-scale replication and capacity expansion, as well as optical module companies, whose growth relied not on price hikes, but on tech upgrades driving ASP expansion and higher interconnect density driving volume growth. Yet, the market’s response was clear: grander overarching narratives take precedence.

What Happened

Following a surge in both token volume and pricing during the first half of the year—driven by breakthroughs in AI coding capabilities, which marked the most favorable period for equity investments, the AI supply chain faced renewed market skepticism regarding high token costs and Cloud Service Provider (CSP) ROI, returning capital expenditure sustainability to the center of debate.

Furthermore, competitive dynamics shifted across multiple sub-segments. For instance, after Broadcom partnered with Google to challenge Nvidia's GPUs with TPUs, Broadcom's own market share faced diversion from MediaTek and Marvell. Meanwhile, Chinese open-source large language models (LLMs) continued to narrow the gap with leading frontier closed-source models. These shifting dynamics have introduced substantial uncertainty to corporate earnings outlooks. Domestic hardware firms, while riding the tailwinds of overseas AI hardware upgrades, were also disrupted by engineering complexity and technology roadmap uncertainties. Many hardware innovations naturally experience normal delays from design to mass production, a reality the market consistently underappreciated.

Some shifts in competitive dynamics are clearly visible, such as capacity expansions in optical fiber and preforms. Historically, demand has rarely withstood capacity expansions that lack high entry barriers. As for the AI narrative, we believe AI’s transformation of production and lifestyle remains in its early stages; however, this alone does not justify continuous stock price appreciation. The new energy sector has been well off its peak equity valuations for years, yet the underlying industry has made tremendous progress since then.

Ultimately, any industrial development must integrate into the existing price framework. For instance, wind and solar power tariffs eventually had to benchmark against and align with existing coal and gas power prices before ultimately replacing them. As AI moves from the laboratory into diverse enterprise applications, it is similarly constrained by societal input-output dynamics. Even if it ultimately restructures industries, it remains bound by the incumbent framework. Demand assumptions that ignore pricing are meaningless—much like everyone wanting to live in a large house without considering affordability: desire is not effective demand.

Moreover, genuine end-user demand differs fundamentally from demand supported by vendor financing. If GPUs were truly in short supply, vendors would hold back inventory for higher prices rather than extending the massive vendor financing we currently observe. The real scenario may be that management teams find it acceptable to extend vendor credit based on future demand assumptions. Even so, this differs markedly from demand backed purely by end-user capital without vendor financing, carrying entirely different risk profiles.

Trading Structures and Industry Dynamics

Extreme positioning also surfaced within market trading structures. Public data indicates that domestic institutional allocations to the electronics and communication sectors at one point reached 60%, against an overall equity exposure of just over 80%. Such a concentration level is unprecedented and unlikely to be repeated anytime soon. A similar phenomenon unfolded in South Korea, where circuit breakers were triggered with unimaginable frequency, leverage ratios ran extremely high, and regulators have stepped in to curb speculative trading. In the U.S., a benchmark leveraged AI portfolio was taken over by Citadel. In the short term, a peak at the market microstructure level has materialized.

At the industry level, ChangXin Technology's public listing implies that memory supply chains in regions like South Korea will have no choice but to accelerate capital expenditures. Even if a technology gap remains between domestic memory producers and global leaders, the catch-up ratchet has been engaged irreversibly. Excessively high profit margins are the root cause driving this structural shift. South Korea's GDP stands at approximately USD 1.8 trillion, yet its top two memory giants are projected to generate over USD 100 billion in profits—a level that is fundamentally unsustainable. Looking back in a few years, the industry may realize that while these firms earned unprecedented profits, the incumbent oligopolistic market structure was dismantled in the process, making it a net loss over the long run. That said, even without the surge in memory prices, the breaking of the incumbent structure was inevitable; current market conditions merely accelerated it.

Viewed through this lens, low P/E multiples lose their relevance. Even if a company earns back its market capitalization within a few years, if the bulk of subsequent earnings must be converted into continuous capex while facing depressed long-term margins, those earned profits are purely nominal accounting values. Unless demand expands rapidly enough to absorb capacity additions of this scale—which appears unlikely—this dynamic will persist. By 2027, memory is projected to account for over 30% of total server rack value, while consumer electronics demand has largely stabilized. Over the long horizon, as SK Hynix Chairman Chey noted, this will cease to be merely an economic issue.

Progress in Chinese open-source models also presents a challenge to closed frontier models. If massive capital investments fail to maintain a technological moat, where lies the economic justification? The mobility of talent and cross-learning cannot be halted, a reality that will prove difficult to reverse.

Current equity investment logic relies on the premise of running a marathon at a 100-meter sprint pace. Without denying the long-term structural value of the "marathon," if market participants revert to a normal pace, much of the upward price pressure will ease—a transition with a high probability of occurring. Today's "sprint" is fueled by credit-driven incentives demanding rapid victories; however, Artificial General Intelligence (AGI) does not appear imminent. Credit expansion will inevitably face constraints imposed by financial market mechanics, and "bond vigilantes" will emerge periodically. Once CSPs realize that moderating their pace satisfies shareholders, preserves their market standing, lowers component prices, and keeps the AI narrative intact, they will adjust their cadence. Indeed, during recent earnings calls, several CSPs emphasized capital expenditure flexibility, particularly regarding short-duration assets.

Macroeconomic Environment and FX Dynamics

Additional uncertainty stems from the Federal Reserve. The new Fed Chair's reforms urge market participants to focus on the "game itself" rather than over-analyzing the "referee's expressions." Concurrently, the Fed plans to reduce market communications, decrease the frequency of FOMC meetings, and reform inflation measurement methodologies. Overall, the new Chair exhibits a strong reluctance to raise interest rates while harboring dissatisfaction with the central bank's traditional operating framework. However, implementing these reforms requires theoretical backing and empirical evidence. While old frameworks are being dismantled, new ones have yet to be established. Although the old system was far from flawless, the market's current inability to anchor expectations creates a challenging environment for asset pricing.

The Fed remains in a wait-and-see posture. While maintaining a observation period amidst long-term unfulfilled inflation targets carries logical justification, it has sparked controversy: an overly prolonged wait risks eroding central bank credibility. Recently, U.S. long-term Treasury yields approached their highest levels since the second half of 2007. Should inflation trends prove disappointing, neither financial markets nor the public will grant the Fed much leeway. Furthermore, inflation dynamics remain linked to the trajectory of the U.S.-Iran conflict, where mutually acceptable resolutions appear difficult to reach under current conditions. This represents a significant risk for markets, rendering August data critical.

Severe volatility in foreign exchange markets signals that long-standing equilibria are breaking down. Global asset allocation relies on relative currency stability: gold previously experienced major volatility, and more recently, the Japanese Yen followed suit. Key structural anchors of the incumbent system—including low Japanese interest rates, carry-trade flows, and Japan's position as a primary holder of U.S. Treasuries—are loosening under persistent global inflation. Historically, global dollar recycling flowed primarily into U.S. Treasuries; today, a new cycle centered around AI investments has emerged. As inflation clouds the outlook for U.S. Treasuries, trade-surplus nations display diminishing appetite for U.S. debt. Whether volatile AI investments can absorb massive long-term capital flows remains to be seen.

Systemic stability generally resists trend-level shifts in major variables. Prolonged inflation, China's economic ascent, and global capital markets' underallocation to Chinese assets will periodically breed asset bubbles. Surging gold prices and sustained Yen depreciation serve as warning signs. Left unaddressed, these factors could accumulate into a "gray rhino" event that becomes self-fulfilling.

Shift in Value Distribution

At present, AI models remain in a stage where they cannot handle everything, yet can perform a wide range of tasks. If models fail to rapidly acquire end-to-end problem-solving capabilities over the coming period, Forward-Deployed Engineers (FDEs) will command significant value. Value distribution along the AI value chain, across Hardware, Models, and Applications, may gradually shift from current levels.

Hardware: During the arms-race phase, hardware commanded peak profitability as industry spending proceeded with little regard for cost. However, as the market transitions into large-scale inference deployment, cost becomes a paramount consideration, and hardware profit margins will likely normalize.

Models: It is difficult for any single company's model to maintain a long-term lead, and switching costs between models remain low. Consequently, standalone model economics appear mediocre—especially given the steady performance gains of Chinese open-source models, which cast uncertainty over the commercial viability of closed-source models requiring massive capital investments. Current token consumption remains concentrated in AI labs, a handful of breakout generative AI applications, and enterprise clients seeking cost avoidance and productivity gains (e.g., customer service and business process automation).

Applications & Engineering: The future of AI applications lies in enterprise production environment workloads. We observe major model developers aggressively hiring FDE talent. By developing robust Harnesses to decouple models from execution systems, product competitiveness will depend less on the underlying model itself and more on the engineering system encapsulating it. Given current model capabilities, expanding AI applications beyond coding into broader enterprise workflows requires deep engineering execution to adapt to diverse industries.

Overseas model giants are actively recruiting talent to perform Palantir-like work, while domestically, Tencent, Alibaba, and ByteDance are consolidating their productivity suites. The next frontier for AI centers on enterprise production workloads, where engineering capability and distribution channels form the primary competitive moats—creating significant opportunities for platform operators. Over the past week, Alphabet, Microsoft, and Amazon gained USD 1.5 trillion in combined market capitalization, while hardware-centric peers faced downward pressure.

The AI transformation of the global economy continues to unfold, but the value distribution within AI investments is undergoing a structural shift: moving from a winner-take-all hardware narrative to a broader set of high-quality enterprises across the ecosystem. This forms our core investment thesis today.