Is an oversold bounce coming? South Korean stocks have turned positive, and U.S. equities are extending their gains in overnight trading. The AI trade is finally getting some breathing room. How far this rebound can go will depend on three major events this week: earnings, the U.S. midterms, and nonfarm payrolls. 1. AI earnings: Can applications make money, and can demand for computing power hold up? Watch Palantir early Tuesday morning and AMD early Wednesday morning, Beijing time. Palantir needs to show that businesses will keep paying for AI, and that orders can translate into revenue and profit. AMD needs to show that demand for data centers and AI chips can continue to support growth. Investors are increasingly focused on one question: When will Big Tech’s massive spending start paying off? Earnings beats and management’s willingness to raise guidance could determine the AI trade’s next move. 2. The midterms: Will political uncertainty clear? The U.S. votes on November 3. From Wednesday, Beijing time, watch the vote count and the battle for control of Congress. The outcome will shape the scope for tax, fiscal, and regulatory policy. For the AI industry, policy expectations around data center construction and power supply also deserve attention. But the end of the election does not guarantee a stock market rally. What matters is how much the results differ from what markets have priced in, and whether investors are willing to put money back to work once the picture becomes clearer. 3. Friday’s payrolls: Will interest rates keep weighing on valuations? The October U.S. nonfarm payrolls report is due Friday at 9:30 p.m. Beijing time. Strong employment and wage growth could reinforce expectations that rates will stay higher for longer. Weak numbers could revive concerns about an economic slowdown. For AI stocks, whose valuations depend heavily on future earnings, interest rates remain a critical hurdle. My view: Oversold conditions create room for a bounce. The next question is whether this week’s events can improve expectations for earnings and interest rates. Overnight gains are only the first step. More meaningful signals would be sustained buying after the regular session opens, leading stocks holding their gains after earnings, and the rally spreading across the broader AI supply chain. Watch buying support in the short term and the major catalysts this week. A lasting recovery will still depend on earnings delivery.

Trump says the war is almost over while threatening another strike. That’s a weak basis for calling a bottom in US stocks. The longer oil stays elevated, the harder the earnings math gets: • Higher transport and production costs squeeze margins. • Sticky inflation leaves the Fed less room to cut. • Falling Treasury yields may reflect demand for safety. They don’t automatically mean cheaper corporate funding. AI stocks have another bill coming due: massive spending, with growing pressure to show when revenue and cash flow will catch up. Add energy costs and geopolitical uncertainty, and rich valuations get harder to defend. A threat is still a threat, not a new strike. Credible progress in talks and a sustained oil pullback would change the outlook. Renewed attacks would make it worse. For now, there’s too little evidence to say the correction has run its course.

U.S. memory, optical networking and cloud stocks are broadly lower again in premarket trading. SK hynix’s ADR is down more than 4%, with SanDisk, Micron and others following. Over the next two days, SpaceX and Meta earnings will test an increasingly urgent question: Can the money AI brings in keep pace with the money going out? SpaceX today Its AI business will be one of my main focuses. Following the xAI merger, SpaceX’s results now span space, Starlink and AI. Last quarter, the company disclosed $14.1 billion in signed cloud services contracts. But signed contracts still need to turn into recognized revenue and cash receipts. Three things to watch: ① Where is AI revenue growth coming from? How much comes from cloud compute, Grok subscriptions and enterprise customers? How much of the signed contract value has converted into revenue? ② What does that growth cost? Can data center investment, capex and free cash flow support expansion? Even if adjusted earnings improve, depreciation, interest expense and cash spending still matter. ③ How much cash support can Starlink provide? Beyond subscriber growth, watch revenue per user, margins and cash generation, alongside further investment in Starship. Meta tomorrow Meta offers a more direct test of whether AI can improve an existing business. Its much-discussed AI agent project, Muse, deserves particular attention: how much is it actually contributing to results? Last quarter, Meta guided to full-year capex of $130 billion–$145 billion. This time, watch whether ad growth, conversion efficiency and margins can support continued spending. Next-quarter revenue guidance and the spending outlook will be especially important. More compute spending signals demand for chip suppliers. For Meta shareholders, it raises a return-on-investment question. The same dollar of capex means different things on opposite sides of the supply chain. My view: these two reports could drive further divergence within the AI trade. If capex guidance simply rises again, equipment suppliers may benefit while valuation pressure on the companies footing the bill persists. Faster revenue conversion and a narrowing cash burn would give sentiment a basis to recover. But if spending keeps rising while returns move further into the future, one company beating expectations won’t be enough to confirm a sector-wide reversal. The market will increasingly care about who can turn AI demand into returns shareholders actually receive.

South Korea’s Kospi plunged nearly 4% today to its lowest close in two months. For retail investors who rode the AI boom, the days of seemingly effortless gains may be over. More worrying than the selloff itself: fewer buyers are stepping in. After leading global markets in the first half, the Kospi has fallen 22% in the second half, becoming the world’s worst-performing major stock index. Trading turnover has shrunk 70% from its late-May peak. Foreign investors are pulling out, and local retail investors are retreating too. Foreign investors have withdrawn a net $131 billion from Korean equities this year—the largest outflow in Asia. They sold another roughly ₩2 trillion today, as Samsung Electronics and SK Hynix both fell more than 4%. Taiwan has emerged as an alternative for institutional investors. The Taiex is up 70% this year and outperformed the Kospi by about 23 percentage points last quarter—the widest quarterly gap this century. Attacks in the Middle East, risks to energy supplies and expectations of a Bank of Korea rate hike help explain today’s risk aversion. But the persistent decline in turnover suggests this trade had been losing its appeal for some time. Buying Korean stocks had become a shortcut to betting on memory chips: AI demand expectations lifted Samsung and SK Hynix, rising prices drew in retail investors, and leverage added fuel. Now, as foreign investors take profits and fresh buying dries up, that same concentration and leverage can accelerate the decline. The apology over single-stock leveraged ETFs captures the problem. Products introduced just months ago in hopes of keeping money in the market have exposed their other side: leverage can amplify returns, but it cannot create lasting demand. AI demand may keep growing. Chipmakers may keep delivering strong results. Neither guarantees a profit for investors who chased the rally. If memory-chip earnings keep beating expectations and foreign inflows return sustainably, Korean stocks could regain momentum. Until then, my view is that the easy-money phase—when relentless inflows kept pushing valuations higher—is behind us.

⚠️ Bought a Ledger from CryptoBilis? Pay attention. Onchain investigator @SpecterAnalyst estimates more than $86 million in suspected thefts across Bitcoin, Ethereum and Tron, tracing funds from hundreds of reported victim wallets. Ledger is investigating losses involving customers of the reseller. The total and its connection to CryptoBilis remain unconfirmed. The reported individual losses are devastating: one buyer allegedly lost 80 BTC just a week after purchasing a device; another reportedly lost 7 million USDT after buying one three weeks earlier. For anyone who bought from CryptoBilis in the past 90 days, Ledger’s precautionary guidance is: • Haven’t set it up? Don’t initialize it. • Already using it? Consider moving your assets to a new Ledger device with a fresh recovery phrase. Ledger’s advisory The cause is still under investigation. A Ledger-wide vulnerability has not been established. A hardware wallet is meant to protect your savings. These reports are a painful reminder that how a device reaches you—and how its recovery phrase is generated—matters too.

OpenAI has dropped a “bombshell” on crypto, potentially threatening the security of Bitcoin and Ethereum. Vitalik Buterin @VitalikButerin and Ethereum researcher Justin Drake @drakefjustin have issued security warnings. The key concern: AI could deliver more mathematical progress in a few weeks than humanity achieved over centuries. AI might break elliptic-curve cryptography (ECC) before quantum computers do. (Note: Bitcoin, Ethereum, and many other crypto projects rely on ECC and its derivative, the Elliptic Curve Digital Signature Algorithm (ECDSA), to generate key pairs and sign transactions.) OpenAI has organized 722 mathematical preprints into 372 groups of related results, generated by an unreleased internal model. Around 160 papers include formal proofs in Lean. OpenAI also acknowledges that results without formal verification may contain errors. Even with those caveats, this leap in AI capabilities has already raised concerns among prominent technical figures such as Vitalik. After reviewing OpenAI’s mathematical results, Justin Drake urged the crypto community to enter “bunker mode.” He recommended moving assets to fresh addresses that have never signed a transaction. Any address whose public key has been exposed, he argued, should move its funds to a new address. Vitalik subsequently issued a similar warning: while quantum computing has long been the concern, AI could pose a threat sooner. If AI delivers roughly 50 years’ worth of mathematical progress over the next two years, the concrete security of lattice-based schemes such as ML-DSA and fully homomorphic encryption (FHE) could decline significantly. ECDSA’s remaining security window could also be shorter than expected. Justin Drake believes mathematical superintelligence has already arrived and that ECC is particularly vulnerable to it. In the worst case, he suggests, ECC could be broken within months—for example, a large GPU cluster could derive a private key from its corresponding public key in just one week.

AMD just agreed to spend $8.2B in stock on Fei-Fei Li’s World Labs. This is not mainly a bet on 3D content. It is AMD buying the layer it still lacks against Nvidia: a model lab capable of turning competitive silicon into a sticky developer ecosystem. Look at what Nvidia actually sells. Cosmos gives robotics and autonomous-driving teams open world models plus training, data and evaluation tools. Once a team builds its pipeline around that stack, the next GPU order naturally leans Nvidia. That is the flywheel. AMD can compete on hardware. What it has lacked is a comparable flagship world-model layer. Selling accelerators without the model ecosystem is like selling an engine without the transmission: powerful, but harder for developers to turn into a complete product. The deal structure matters: all stock, about $8.2B. Li will become AMD’s EVP and chief scientist, reporting to Lisa Su. World Labs brings Atlas, its omni world model, and Marble, which creates persistent 3D worlds from images, video or text. This looks less like buying one product than installing a model research division. The battleground is shifting from peak compute to developer defaults. If your models, tools and SDK become the default workflow, the next hardware order gets much easier to win. The risk is equally clear. “World model” remains an overloaded category, while Marble’s first clear use cases are creative 3D workflows. Robot training and simulation are still far from mass deployment. AMD is paying $8.2B for a company founded in 2024. That is a bet on the ecosystem three years out—not next year’s revenue. This is not a cheap way to fill a product gap. It is an expensive option on where the AI stack may be heading. ImageNet helped ignite the deep-learning boom that made GPUs AI’s default engine. Nineteen years later, Fei-Fei Li is joining Nvidia’s biggest challenger. The next AI war may be decided less by transistor counts than by whose SDK developers reach for first. $AMD closed at $607.87, down 3.61%, at a ~$992B market cap. The wider AI selloff mattered, but investors may also see the price as steep. The real question: if world models become the default layer for robotics and AVs in three years, will $8.2B look expensive—or cheap?

Four shocks hit markets at once last night: hot September PMI, a weak Treasury auction, a hawkish warning from Fed Governor Barr that more hikes may be needed, and a 3.9% jump in Brent. Stocks, gold and crypto all sold off. The 10-year yield jumped about 14 bps to 5.113%, its first move above 5% since 2007. Markets now put the odds of another 25 bp hike in October at 68%, up from 49% last week. But this wasn’t mainly an inflation scare. Bloomberg data show real yields drove roughly 80–85% of the move: the 10-year nominal yield rose 15 bps, 10-year TIPS yields rose 12.5 bps, while breakeven inflation added just 2 bps. Investors are repricing a hotter economy, a tougher Fed and a return of term premium—and raising the discount rate on long-duration assets. That’s why the AI trade split. Alphabet fell 3.8%, AMD 1.5% and Nvidia 1.4%, while Meta rose 1.0% amid its developer conference and Charm hardware launch, and Microsoft gained 0.5%. The pressure isn’t on the AI story as a whole; it’s on AI valuations whose cash flows are still far out. Bitcoin slid from above $87K to $83,744, down 2.2% in 24 hours. Yet spot BTC ETFs still took in $999M Monday and $715M Tuesday. That suggests institutional flows held up even as leverage got flushed. The next week hinges on which story the data confirm. If PCE and October payrolls keep pointing to an overheated economy, 5% may be a floor, not a ceiling—and leveraged positions in long-duration AI and BTC could face another washout. If monthly PCE cools, the 2007 comparison comes back into play: the day nominal yields peak could mark the starting line for Nasdaq.

1/ The Fed hiked last week. US equities rallied anyway. Memory stocks are extending the move pre-market: $SNDK briefly traded above $1,800, $MU pushed toward $1,030, and $WDC/$STX followed. Everyone is calling it a supercycle. But the real signal is in the contracts—not the stocks. 2/ DRAM prices are expected to rise another 20–30% QoQ in Q3, with NAND up 15%+. More importantly, hyperscalers have already agreed to pay higher DRAM prices in Q1 2027 than in Q4 2026. That changes the cycle. Historically: suppliers raised prices, buyers hoarded inventory, capacity expanded, then prices collapsed. Now Google/AWS/Meta/Microsoft-tier buyers are locking higher prices into multi-year contracts. 3/ The shortage makes this possible. Estimated 2026 supply coverage is only 79% for DRAM and 80% for NAND, with balance unlikely before 2027. Each bit of HBM consumes roughly 3× the wafer capacity of conventional DRAM. The more capacity the Big Three shift toward HBM, the tighter commodity memory becomes. 4/ BofA raised its 2030 memory TAM estimate from $1.8T to $2.0T and expects DRAM/NAND revenue to compound at 21% through 2030. Its memory-cycle index reached 180 in July versus a historical median of 100. The 2017–18 supercycle peaked at only 120. This cycle is already running 50% hotter than the last peak. 5/ The consensus trade is: “The memory cycle is back, so buy memory stocks.” My view: half of the cycle is being contracted away. For maximum short-term beta, $SNDK/$WDC/$STX offer the cleanest exposure. But after rising from roughly $93.50 to $1,791 in a year, $SNDK leaves little room for disappointment. For contract protection, $MU is stronger. Micron has signed 16 strategic agreements, 14 of which represent roughly $100B in minimum contracted revenue. 6/ The setup is simple: If you are trading momentum, $SNDK has the most torque—but needs a hard exit level. If you believe contract pricing is turning memory into a structurally inflationary asset, $MU is the cleaner bet. At roughly 23× trailing earnings, with a $1,547 consensus target, $MU gives the thesis more time to work. Watch its Sept. 30 earnings. 7/ Two risks: ① CXMT’s fifth-generation DRAM platform claims at least 50% more dies per wafer. If volume ramps, commodity DRAM contract floors could weaken. ② Samsung’s HBM4 yield has reportedly improved from below 60% to around 80%. If Samsung and Micron both scale HBM4, the HBM premium could compress and capacity may flow back toward conventional DRAM. Memory cyclicality has not disappeared. Hyperscalers are simply paying to absorb part of it.

“A 25bp hike is priced in” is a dangerous shortcut to “the bad news is over.” AI stocks are bouncing ahead of the Fed, but whether that rally lasts depends on the rate path and long-term yields. Reuters’ Sept. 16 Asia-session report put the odds of a 25bp hike at roughly 93%, with the 10-year Treasury yield near 5% and Brent above $108. Sentiment is improving. Financing costs and inflation pressure remain elevated. Three risks to watch: The expected hike, followed by a more hawkish path. A dot plot pointing to more hikes or later cuts could force another valuation reset, especially for expensive tech and companies dependent on external funding. A 2 p.m. rally that fades after the 2:30 p.m. press conference. Warsh doesn’t need to sound aggressively hawkish to disappoint. If traders are betting on “one and done,” simply refusing to endorse a pause could trigger profit-taking. Stocks rally, but bonds refuse to confirm. A 10-year yield near 5% keeps pressure on financing costs and equity valuations. Digesting the hike doesn’t remove that pressure. Even a surprise hold may fail to deliver lasting relief. If investors question the Fed’s inflation-fighting resolve, long-term yields could rise and erase the initial equity bounce. Before calling this a durable relief rally, I want three confirmations: • The 2-year yield stops climbing after the press conference, with no further hawkish repricing. • The 10-year and 30-year yields fall. • Stocks hold their gains into the close, with participation broadening beyond the megacaps. Premarket green shows traders are willing to bet. Gains that survive the press conference, backed by falling yields, would make that bet more convincing.