Stream thérapeutique

Summary

The stream focuses on a sharp market sell-off, particularly in AI and semiconductor stocks. The presenter sees the pullback as a possible buying opportunity, arguing that prices—especially for memory-chip makers—have fallen much further than available company commentary and demand indicators seem to justify. He also emphasizes that the investment thesis could be wrong: memory demand, AI spending, and future earnings remain uncertain.

Market and portfolio context

• The presenter describes a severe drawdown in his portfolio, with gains returning to roughly €27,000—a level he recalls from an earlier period. He says he will keep about €30,000 in cash in the portfolio and does not plan to add more overall.
• He characterizes the broader move as a market panic and a flight to perceived safety. He cites Walmart at roughly 47× earnings as an example of a stock being valued richly amid that rotation.
• He says the Nasdaq had fallen about 10% and argues that historically, buying the index on a pullback has generally been more rewarding than buying after a rise, unless the market is in a lasting bubble.
• He points to daily declines of about 15% for AMD and 11% for Tesla. Samsung was about 16% below his purchase price, after having traded higher.

Main investment thesis: AI infrastructure and memory

The presenter’s central argument is that the market may be treating memory-chip companies as if another severe boom-and-bust cycle is imminent, while the economics of the market may have changed:

• Past memory downturns were driven partly by Samsung and SK Hynix expanding production, followed by phone and PC customers cutting orders after accumulating large inventories.
• He argues that AI and data-center customers differ from those earlier buyers: they may have little inventory, make multi-year investment plans, and sign longer-term supply agreements.
• He says Samsung and SK Hynix have become more cautious after repeated past cycles. According to his account, they are signing five-year, and sometimes ten-year, agreements, with a significant share of production reportedly committed through 2029. Some new factory capacity is being built in stages rather than brought fully online immediately.
• He says the market is pricing in an exceptionally severe downturn—at one point describing a potential 90% earnings contraction, and elsewhere a 90–95% margin contraction. He argues that if earnings do not collapse by anything close to that amount, current valuations could offer substantial upside.
• He acknowledges the opposite possibility: AI-related memory demand could fade, production could exceed demand, and the traditional memory cycle could return. At one point he suggests hyperscalers might buy memory for only two or three years before needing less; elsewhere he argues that AI, robotics, and wider technology adoption could extend demand. The outlook is therefore presented as uncertain rather than settled.

AI capex and data-center debate

• The presenter disputes the idea that AI-related capital expenditure has clearly peaked. He says estimates of the peak have repeatedly moved outward—from 2026 to 2028, and potentially 2030 or later if robotics and other uses of AI expand.
• His view is that capex in 2035 need not be lower than in 2026, although this is his speculation, not a stated company forecast.
• He discusses political opposition to data-center construction, particularly proposals associated with Democrats. He argues that restrictions in one state may not halt investment because companies can relocate projects and states compete for jobs and tax revenue. A nationwide ban, he says, would be much more consequential but politically difficult.
• He also notes that even a halt to new construction could increase the value of already-built or nearly completed data-center infrastructure, though he observes that related companies have also sold off.

Historical and analytical framework

The presenter’s approach to judging a sell-off is to distinguish price movements from changes in business fundamentals:

• Do not treat a falling share price alone as proof that the investment thesis is wrong. He compares current concerns with past drawdowns in Netflix, Nvidia, Amazon, and Apple, arguing that large declines did not by themselves establish that the businesses were fundamentally impaired.
• Look for concrete evidence that would invalidate the thesis, such as sharply reduced hyperscaler capex, weakening GPU-rental demand, memory customers reporting lower demand, or material earnings and margin deterioration.
• Compare market expectations with company statements and commitments. He sees a large gap between pessimistic market pricing and what he says Samsung and SK Hynix are communicating about demand and orders.
• Account for historical industry cycles. He describes earlier memory cycles in which production expanded, customers built inventories, orders were cut, and prices and profits collapsed. His thesis depends on believing that today’s customer mix and contractual arrangements reduce that risk.
• Treat low valuation as a potential opportunity, not a guarantee. He argues that very low multiples can create upside if earnings hold up, but recognizes that the market may be anticipating a genuine collapse.

Stocks, assets, and instruments mentioned

• Semiconductors and AI infrastructure: Samsung Electronics, SK Hynix, Micron (MU), Nvidia (NVDA), AMD (AMD), TSMC (TSM), and Celestica (CLS).
• Other public companies: Netflix (NFLX), Meta Platforms (META), Lyft (LYFT), Walmart (WMT), Berkshire Hathaway, Tesla (TSLA), Oracle (ORCL), and Palo Alto Networks (PANW).
• Index and sectors: Nasdaq; memory chips, semiconductors, AI, cloud computing, data centers, cybersecurity, and oil.
• Private AI companies and services: OpenAI and Anthropic.
• He mentions Samsung GDRs and says the investor is generally exposed to the Korean won through Korean holdings.
• A company rendered as “Advent” in the captions may refer to Advantest, but the identification is uncertain.

Valuation, financial, and other figures cited

The figures below are the presenter’s estimates or claims during the stream; the captions are sometimes unclear:

• Samsung is described as trading at around 3× forward earnings, or roughly 5× using a different earnings basis.
• He says memory stocks’ valuations imply an exceptionally large earnings downturn—at points described as around 90%.
• Nvidia is discussed around $180–183 per share, with the presenter citing approximately 20× earnings and 10× forward earnings. Elsewhere, the captions give different forward-multiple comments, so these figures are not fully consistent.
• Micron is cited as having generated about $37 billion in revenue in the latest quarter and having a roughly $700 billion market capitalization. The presenter says Micron’s share-buyback restrictions were expected to ease in December 2026, and speculates that buybacks could become substantial; the captions’ estimate of “5% of free float per quarter” is especially uncertain.
• The presenter discusses a California proposal to impose a one-time 5% wealth tax on billionaires, with a five-year grace period—described as equivalent to about 1% annually over five years—and says voters rejected it.
• In discussing Anthropic’s spending commitments, he refers to a model in which revenue might need to grow by around 60% annually through 2028. The specific commitment and revenue figures in the captions are difficult to reconcile and should be treated cautiously.

Trading views and cautions

• He says he would consider adding to Samsung, SK Hynix, and possibly Micron. He expresses a preference for Samsung over TSMC, while also listing TSMC among possible purchases.
• He says he might wait for Lyft’s results before selling part of that holding and using the proceeds to buy more memory stocks.
• He is less enthusiastic about adding Nvidia, citing uncertainty over AI’s future economics and the possibility that AI companies could alter the semiconductor competitive landscape.
• He is concerned that AI services could disrupt parts of the software and cybersecurity industries, while also suggesting that AI could create new cybersecurity capabilities. These comments are speculative.
• He describes an anecdote about an AI startup that moved away from self-hosted, open-source models running on rented Nvidia GPU capacity and back to a commercial model, which the presenter says had become cheaper and more capable. This supports his view that owning or renting compute does not automatically make open-source AI economical.
• He warns that his memory-stock thesis fails if demand weakens, companies overproduce, margins collapse, or technological changes reduce memory requirements. He also acknowledges broader unforeseen risks.

Disclosures: No explicit “not financial advice” disclaimer is audible in the subtitles. The presenter discusses his own portfolio and positions.

Presenters and sources: The presenter is the Crésus channel host, who refers to himself as Crésus. Sources discussed include company statements and market pricing, historical share-price and memory-cycle data, political proposals, and an anecdote from a friend working at an AI startup.