To The Moon

Speaker: Ross Dong @Ross_Macro_Trading (Founder of Gongxing Academy and Partner at Morning Cloud Asset Management)

Live Date: July 23, 2026 (Review Link>>)

In this livestream, Ross analysed the AI memory cycle through two frameworks: the industrial-technological cycle and the monetary-financial cycle, combining top-down macro analysis with bottom-up industry research.

Want a deeper dive? We broke this session down into 4 full recap articles, each covering a different piece of the puzzle>

• Live Recap 1: Why the Memory Selloff Is a Buying Opportunity, Not the End of the Cycle

• Live Recap 2: Korea's Macro Backdrop — GDP, Valuations, and Capital Flows Behind the KOSPI Story

• Live Recap 3: Inside the Memory Supercycle — NAND, DRAM, and the AI Compute Shift

• Live Recap 4: SK Hynix in Focus — Q&A Highlights on Timing, Risk, and Embracing AI

Prefer to watch the highlights? Catch these key moments from the live session in short clip form>

• Livestream Clip 1｜Ross Dong: AI Memory Down—Is the Thesis Broken?

• Livestream Clip 2｜4 Signals Behind the Pullback

• Livestream Clip 3｜HBM Shortage Until 2028?

• Livestream Clip 4｜Can AI Really Do 90% of Research?

🐯💬 Join the discussion: Share your market view or questions below. Every useful and thoughtful comment will receive Tiger Coins!

🎯 5 Key Takeaways

• The late-June memory-stock sell-off was driven by four specific market concerns—not one clear collapse in AI demand.

• Unlike previous consumer-driven memory cycles, the current upcycle is supported by AI infrastructure, HBM and DDR5 adoption.

• Ross estimated memory supply growth at only 20%–30%, compared with 200%–300% growth in AI-related demand.

• Memory undersupply may continue until 2028–2029 because advanced capacity takes years to build.

• The most important demand indicator is the quarterly capital expenditure of major cloud hyperscalers.

📉 Four Drivers Behind the Recent Sell-Off

Ross identified four reasons for the correction:

• Excess compute leasing: Some cloud hyperscalers are leasing unused GPU capacity to third parties, reducing the immediate urgency to purchase more GPUs.

• Closed-source AI commercialisation concerns: Delays surrounding OpenAI’s IPO raised questions about the profitability of closed-source frontier models.

• Shift toward the inventory cycle: Investors are moving their focus from AI order visibility toward memory inventories, capacity expansion, and future supply.

• Profit-taking from crowded trades: Capital has moved out of heavily crowded US AI and semiconductor positions following strong gains.

Ross described the pullback as a short-term valuation reset within a longer AI infrastructure cycle, rather than proof that the entire trend has ended.

Discussion: Which of these four factors do you think presents the biggest risk to the AI sector?

🧠 Why This Memory Cycle Is Different

Traditional memory cycles were largely driven by consumer products such as smartphones and personal computers.

Ross argued that the current cycle is increasingly driven by structural AI demand:

• HBM: High-bandwidth memory used to support AI computing workloads

• DRAM: Working memory required by servers and computing systems

• NAND: Storage memory used to retain large amounts of data

• DDR5: A newer server-memory standard being adopted by hyperscale data centres

• Memory content per server: AI servers require significantly more memory than traditional servers

This means memory demand is no longer dependent only on consumer-device replacement cycles. It is also linked to data-centre construction, AI model deployment and growing compute requirements.

🏭 The 20%–30% Supply vs 200%–300% Demand Gap

According to Ross, annual memory supply may grow by approximately 20%–30%, while demand for advanced AI memory products could rise by 200%–300%.

The gap cannot be closed immediately because:

• Leading-edge HBM capacity requires major capital expenditure.

• New infrastructure may take two to three years to construct and qualify.

• Memory manufacturers previously maintained conservative expansion plans due to oversupply concerns.

• Newly announced capacity will take time to enter production.

Ross therefore expects memory undersupply to continue until approximately late 2028 or 2029. In his view, additional capacity may initially bring the market closer to equilibrium rather than immediately creating oversupply.

Discussion: Do you think manufacturers can expand capacity fast enough—or could aggressive expansion eventually create another downcycle?

🇰🇷 South Korea’s Macro and Market Backdrop

Ross also connected the memory cycle with South Korea’s economic outlook:

• 2026 GDP growth forecast: 3.7%

• 2027 GDP growth forecast: 3%

• Semiconductor contribution: Approximately 2.6 percentage points of 2026 GDP growth

• KOSPI 12-month forward P/E: Around 6 times as of July 16

• Passive fund inflows: More than $3 billion reportedly entered South Korea-related ETFs after July 9

The presentation described the KOSPI’s forward valuation as being near the lowest range seen over the previous 20 years.

Ross also highlighted two sources of market volatility:

• Foreign active-investor rebalancing and profit-taking

• Rapid growth in single-stock leveraged ETFs, whose assets under management had nearly quadrupled since May 2026

Korean regulators have responded by tightening retail access to these leveraged products, with the aim of reducing excessive market volatility.

🌐 Open-Weight AI and Memory Demand

Ross discussed the growing adoption of Chinese open-weight and open-source AI models.

Compared with closed-source frontier models, these models may offer:

• Lower deployment costs

• Greater control over private data

• Private training and customisation

• Wider adoption by enterprises

However, Ross noted that open-weight models may also be more demanding in terms of HBM, storage and total memory capacity.

This creates an important industry implication: cheaper AI models may not reduce infrastructure demand. Wider adoption could instead increase the number of companies deploying AI and raise total memory requirements.

At the same time, increased adoption of open-weight models may create commercialisation pressure for closed-source developers that rely on paid access and subscription-based models.

🔗 Optical Components: The “Nervous System” of AI Clusters

Ross described optical components as the “nervous system” of an AI cluster.

As GPU clusters expand, large volumes of data must move rapidly between processors, servers and data centres. This shifts part of the infrastructure bottleneck toward high-end optical interconnects.

Key observations from the session included:

• Optical-sector valuations reached high levels in June and early July.

• Premium optical components continue to face structural shortages.

• Major manufacturers are planning multi-year capacity expansion.

• Aggressive expansion plans suggest longer-term demand visibility.

• Supply-chain access remains important because advanced components cannot be expanded quickly.

The demand outlook may be strong, but elevated valuations mean investors must distinguish between industry growth and the price already reflected in the market.

⚠️ How to Test Whether AI Demand Is Sustainable

Ross identified hyperscaler capital expenditure as the most important confirmation signal.

Investors can monitor whether major cloud companies continue spending on:

• Data centres

• AI infrastructure

• GPUs and compute capacity

• HBM and server memory

• Storage systems

• High-speed optical interconnects

Other indicators mentioned during the session include:

• NAND and DRAM pricing

• Memory inventory levels

• New capacity announcements

• Excess GPU capacity is being leased to third parties

• Changes in AI order visibility

• Capital flows into or out of semiconductor trades

A sustained reduction in hyperscaler capital expenditure would be a more serious warning than short-term share-price volatility alone.

🤖 How Ross Uses AI in Investment Research

Ross shared that AI can now handle approximately 85%–90% of the repetitive work previously completed by a five-person team of researchers or quantitative analysts.

AI tools such as Kimi, GPT and DeepSeek can assist with:

• Collecting and organising market data

• Generating initial trading ideas

• Summarising daily market developments

• Comparing companies and industries

• Supporting repetitive research tasks

However, Ross emphasised that the final 10% still requires human judgment, particularly for:

• Assessing commercial business models

• Identifying high-conviction ideas

• Determining whether an investment thesis is realistic

• Making final positioning and risk decisions

The key lesson is that AI can improve research efficiency, but it cannot take responsibility for the final decision.

🔍 Q&A Highlights

• Ross expects the wider AI infrastructure cycle to last approximately five to ten years, although it will include periods of correction and volatility.

• The memory shortage may continue until 2028–2029 because demand growth remains much faster than supply growth.

• Hyperscaler capital expenditure should be reviewed every quarter to test whether underlying demand remains intact.

• Ross expects CXMT to pursue a listing in the second half of 2026, which could temporarily draw liquidity away from other semiconductor shares, particularly in Hong Kong.

• Open-weight AI adoption may strengthen demand for HBM and storage while challenging the commercial models of closed-source AI providers.

💬 Words from Ross

“Major trends do not end in a day or a week.”

“Looking at the next 100 years, the mastery of compute will replace the mastery of resources as the primary driver of wealth.”

“The last 10% still depends on yourself, because AI cannot produce the most valuable commercial idea"

Closing Takeaway

The AI memory story isn't broken — it's maturing. What started as a pure demand narrative is now a test of supply discipline, capital allocation, and how long hyperscalers keep spending. The next 6-12 months of capex data will matter more than any single week of price action.

🐯 Your Turn: Join the Discussion

Share your view on one of these questions:

• Is the recent AI memory correction temporary or the beginning of a deeper cycle change?

• Which signal matters most: hyperscaler capex, memory prices, inventory or capacity expansion?

• Will open-weight AI models increase overall demand for HBM and storage?

• When do you expect the memory supply-demand gap to begin narrowing?

🎁 Every useful, thoughtful, and well-explained comment will receive Tiger Coins.

Let’s compare different views and learn from one another.