Market Overview May 16, 2026

Summary

• The short-term pullback is “normal—and necessary.” The last semi rally ran on 3 engines: depressed valuations (Nvidia’s earnings are up more than 70%, effectively putting the stock at last year’s $90 price), FOMO around Anthropic and fantasies of “infinite tokens,” and widespread semiconductor shortages. The actual transmission mechanism, however, was options leverage: call/put skew completely reversed, dealers were loaded with calls, and negative gamma forced them to chase the underlying higher. Calls began to unwind Thursday and Friday; negative gamma had “basically disappeared,” and Friday’s late-session buying never showed up. With the reflexive buying gone, a further rise powered only by retail leverage carries greater near-term risk.
• The pullback target is somewhere before May 1, not the March low. Before May 1, CTA forced buying drove the market higher; after May 1, the move was earnings plus FOMO, leverage and options. A healthy reset only needs to flush the leverage from semiconductors. “Demand is real, token demand is real, but macro risk is real too.”
• CTA flows are the tail risk. Over the next week, CTA models imply roughly $20B of selling even if markets rise (about $10B in the US); in a decline, global selling could reach roughly $100B—“if the market falls this week, the downside tail risk is substantial”—potentially taking the market in one move to a post-May level. For reference, the May 1 monthly rebalance sold $23B and the market still squeezed higher.
• A break above 4.6% in the 10-year is a danger signal. The danger is not the level of rates itself but rates and rate vol rising together—equivalent to roughly a 3% increase in TLT vol; how high is the equity VIX, really? The Nasdaq Sharpe ratio has reached 1.6x, creating a “huge disconnect” between risk assets and macro and splitting the market into token-demand and macro-risk camps.
• Inflation is a shock, not a trend; later this year, unemployment may become the risk. CPI and PPI data spooked markets, with CPI driven in large part by lagged rents. PCE could look worse near term as AI competes with retail for memory and products, pushing up prices across software and hardware. But wage growth is only 3.6%, white-collar job openings are weak, AI itself is deflationary for labor, and China is exporting severe deflation—there is not the same tinderbox as before the Russia-Ukraine war. Later this year could bring unemployment and negative cash flow at large companies (CDS are already higher); if the market also decides capex will miss buy-side expectations, the combination could produce “a somewhat longer correction”—possibly a better opportunity.
• Oil at $105 is actually being held down. Of roughly 20 million barrels/day of Gulf capacity, 12-14 million barrels/day are missing; oil should theoretically be far above this price. The missing upside has been absorbed by Trump’s repeated “2 or 3 weeks” peace-talk expectations—a slow-boiling-frog effect—and by excess inventories held around the world. Goldman’s base case is reopening in May and June, but in reality “the reopening could take a very long time.”
• This year’s 2 gold mines are the optical-communications bottleneck and an IGV breakout. The optical-communications industry has never gone through this kind of expansion: a certain substrate is in short supply, and Japan’s 5802 calling a 30%-40% output increase large is itself telling. Packaging, optical-chip testing, and tuning/alignment are all choke points, leaving the sector with a longer runway than memory, which has expansion experience and the fastest earnings ramp. In software, “long semi, short IGV” is too simple: AI beneficiaries such as Snow (possibly Snowflake), Palantir and Oracle (Oracle’s compute-center revenue is estimated to exceed 50%) are being indiscriminately shorted and compressed on valuation; semiconductor deleveraging will come with dispersion inside IGV.

Deep dive

1. The engines of the last rally: low valuations, Anthropic FOMO, and shortages—but the mechanism was options leverage

• Jin’s recap: bought the March 31 dip and ran for the exits in early May, though the position was not fully closed. This rally had 3 fundamental drivers: semi valuations were crushed (Nvidia earnings were up more than 70%, “equivalent to a $90 share price relative to last year”), Anthropic FOMO—everyone needed to dream about “infinite tokens”—and widespread semiconductor shortages that drove memory higher.
• The real driver was hidden leverage and the options structure: after the March rally, call/put skew completely reversed, with heavy call buying and some dip buyers selling puts as well. Dealers were loaded with calls; a 50% delta meant that a rise forced them to buy the underlying, effectively levering the market into a large amount of stock and pushing the whole market higher.

2. Reflexive buying is gone; the pullback is “normal—and necessary”

• Friday’s signal: with heavy call positioning, there should have been a large late-session buy program, but none appeared. Many options began to close Thursday and Friday, calls unwound as well, and upside negative gamma had “basically disappeared,” so strength no longer triggered mindless buying. The Nasdaq Sharpe ratio has reached 1.6x; straight-line rallies require a pullback in the short term.
• The pullback anchor is not March. “A lot of what came after earnings is real”—Anthropic’s earnings growth is real, right? Anthropic’s revenue growth is real, right?—but the market should reset to some level before May 1, clearing out semiconductor leverage. The weakness in the shortage thesis is timing: “Every time earnings come around, you can say shortages are severe. Outside earnings, there is actually a window of dead time in the process.” This week’s Nvidia earnings will address whether demand can keep pushing forward; Jin says it is “hard to call.”
• CTA is a tail risk hanging overhead: CTA models imply roughly $20B of selling over the next week (about $10B in the US) even if the market rises; in a decline, they could sell $20B-$30B, and global selling could reach roughly $100B—“if the market falls this week, the tail risk is substantial”—taking the market in one move to some point after May. But the risk should not be overstated: the May 1 monthly rebalance sold $23B, and the market still squeezed higher.

3. Macro fracture: oil held down by a “slow-boiling frog,” rates and rate vol in sync

• Oil is around $105, above $100, but nowhere near where it should be. Gulf capacity is roughly 20 million barrels/day; a Saudi pipeline can potentially carry about 3 million barrels/day, leaving an actual reduction of 12-14 million barrels/day. In theory, that degree of demand disruption should have pushed oil much higher. The missing upside has been absorbed by Trump repeatedly talking up negotiations over “2 or 3 weeks,” together with large-scale inventory releases around the world—a slow-boiling-frog effect. Goldman’s base case is reopening in May and June, with China also pressuring Iran, but the actual reopening “could take a very long time.”
• The 10-year breaking 4.6% is a danger signal. The level itself may not break capital markets; the danger is that it is moving in tandem with rate vol. Translated into TLT, that is roughly a 3% rise in volatility—how high is the equity VIX, anyway? The 30-year at 5.2-5.3 seems less important; the buyers there are mainly pensions and long-term capital. The market is splitting into 2 camps: one clearly sees token demand and semiconductor productivity, while the other sees rising awareness of macro risk.

4. Inflation is a shock, not persistent; unemployment may emerge later this year

• CPI and PPI both beat expectations, spooking markets. But under the surface, a major part of the CPI move came from lagged rents, including adjustments not captured in the prior 2 readings. With heavy housing starts over the past 2 years, US rents have little room for a fundamental increase. PCE could look worse than CPI in the short term because AI is moving directly into consumer pricing: Microsoft 360 raised prices with an AI feature, while Mac Studio introduced a lower-memory version. “AI is competing with retail for prices and products: it takes your job, then your memory”（AI是在跟retail抢价格、抢产品……一方面抢你的饭碗,然后还抢你的内存）.
• The ingredients for persistent inflation are missing: annualized wage growth is only 3.6%, white-collar job openings are weak, labor-force participation is much higher than during the Russia-Ukraine war period, AI is generating labor-force deflation, and China’s severe deflation—retail, industrial production and fixed-asset investment data are “a mess”—is exporting deflationary pressure globally. Back then, fiscal transfers piled up dry tinder and oil could ignite it with the smallest spark; those foundations largely do not exist now.
• The risk sequence to watch is unemployment later this year, followed by recession fears, compounded by negative cash flow at large companies. CDS have already risen, and the market may worry that capex will fall short of buy-side expectations. If these issues resonate later this year, there could be “a somewhat longer correction”—possibly a better opportunity.

5. Goldmine one: invest in choke points—memory trades fastest, optical communications runs longest

• The framework is simple: “The market never anticipated having to accommodate this much volume; the final bottleneck is always the shortest plank.” Memory is the first wave and the fastest trade—once it chokes, prices rise and earnings improve immediately—but every memory cycle has an expansion playbook, so earnings ramp fastest and capacity comes on fastest. The next variable is ChangXin’s listing; the market will debate whether China can dramatically expand DDR capacity, although China cannot make HBM.
• Optics and optical communications are longer-duration trades. The supply chain has not been built for large-scale compute-center deployment: it must iterate generation by generation, stacking onto the CPU, while also ramping capacity. A certain substrate is severely tight in Japan; Japan’s 5802 says a 30%-40% production increase is already substantial for it. Packaging capacity is limited, optical-chip testing is much slower than electrical-chip testing, and CW-laser tuning and alignment remain problems. Every link can break, and every break becomes a choke point.

6. Goldmine two: IGV breakout—software moves out of the center, but indiscriminately shorted winners will separate

• Jin’s software framework is a move out of center stage (“C位降位”), not replacement. Software no longer has to sustain a fixed 20% annual growth rate across the board; it becomes something AI calls as needed, breaking the DCF logic. Tax software names and Salesforce are “obviously no good in the visible future,” but some companies that directly benefit from AI are growing faster instead.
• As the market has levered into semis, the popular trade has been long semi, short IGV, “but that framing is too simple.” Snow (possibly Snowflake), Palantir and Oracle are beneficiaries. Oracle’s core software revenue is about 35%, while its compute-center business contributes roughly more than 50% of revenue. The first stop for token demand is cloud—Amazon, Oracle, Microsoft and Google—which then transmits demand to semis through hyperscaler procurement. “Semis are the direct expression of tokens; tokens equal compute”（半导体直接变token,token等于算力）.
• The hedge leg against long semis should therefore not be IGV. Semi deleveraging—SMP was essentially lifted by semis—will arrive alongside dispersion within IGV: AI-winning software that has been indiscriminately shorted and compressed on valuation will emerge, “another major development this year.” The third line is to watch catch-up names and capex efficiency: products that could not previously be sold can now be sold; the question is who gets the best results from capex.