Market Overview April 21, 2026

Summary

• Price-insensitive systematic buying has been the main fuel behind the rebound of the past 3 weeks. Global investors bought $135B over the past 2 weeks, with the bid coming from CTA buying and prime-book short covering; this week, the models still have to buy under bull, bear or flat cases: $20B of U.S. S&P exposure and about $35B globally, according to Goldman’s report late on Apr 20. But the acceleration is fading: the tinder is nearly gone, and the rally powered by systematic buying ends this week.
• Next week is a “show-me” earnings week. Starting the week after next, CTA has little left to buy in the bull case—positions are largely filled—and would have to sell heavily in the bear case, which looks dangerous on the surface. But overall market positioning is not full: the prime book has only covered U.S. ETF shorts, while single-name and global shorts remain; hedge funds and long-only investors are still waiting. Good earnings do not necessarily mean the market has to fall. For those who followed the trade off the lows or added leverage, risk starts accumulating next week.
• Political headlines around the Iran talks will be largely discounted this week. The gamma book has been rebuilt and is back at very high levels; on a selloff, the market would have to absorb both $20B of forced CTA buying and dealer hedging flows, making it hard to punch a hole in the market. The base case is a modest rise or a flat week, with dealers selling into strength and capping the upside.
• The medium- and long-term story is unchanged. Wall Street’s skepticism about AI-token demand is gradually being resolved: “almost everyone listening to this call should be using Anthropic for coding.” GPU purchases pay back in 1.5 years; capex growth will slow next year, but semiconductor earnings should not face a major problem. The 2023 SVB analogy still holds—Nasdaq gained 37% that year after SVB failed, and short-term shocks do not change the underlying picture. Even during the war, S&P first-quarter earnings estimates were raised 5%, with storage companies accounting for a large share of the revisions.
• Mithius, Anthropic’s best model, is available only to enterprise customers. The official explanation is safety, but his view is that 50% of it is marketing—“Codex and 4.6 are not that different”—with compute capacity the real constraint. Models will eventually stratify: free users take GLM, paying users take 4.6, and the best model goes to enterprise customers. Model tiering and compute scarcity will drive semiconductor earnings and demand.
• Kevin Warsh will ultimately be appointed after tonight’s Senate hearing. Trump will, as usual, TACO when policy collides with personal animus, and Warsh should be more dovish than the market’s hawkish pricing implies. He has cited Greenspan’s 1990 decision not to hike during a major productivity surge, and his prepared remark is that “inflation is a choice.” He could well hike twice this year, allowing the market’s hawkish-risk pricing to unwind.
• Liquidity footnote: SOFR briefly fell below the Fed’s policy rate over the past 2 weeks. The interbank rate was below the policy rate—“pretty scary”—with cash so abundant that it flowed into the RRP. Tax-season payments will move funds into the TGA, producing seasonal tightening over the next few weeks; position accordingly.
• Intel will “definitely fail the test” on Thursday, and AMD will probably fail as well. The market may not be disappointed, since 陈立武’s standard handover playbook is to flush out the legacy business this quarter. CPU shortages will be driven by Agent calls rather than human calls. If there is a pullback, $100 by the end of 2027 is fair; the current move to $70 is excessive. And SK Hynix is the better trade than Micron, with Korean valuations still low.

Deep dive

1. The Main Fuel Behind the 3-Week Rebound: $135B of Price-Insensitive Buying

• The starting mechanism was straightforward: systematic selling of roughly $180B over 1 month drove the S&P down to 6,300, where it stopped going lower—“$180B was thrown at it and it still would not break; that level was clearly a tactical bottom, and even a little buying could ignite it.” He said on Twitter 3 weeks ago that the market had accumulated a great deal of tinder. Once the market stopped breaking, someone lit a match; even without an Iran ceasefire, it would still rebound because the certainty of the buying was high.
• CTA trend-following is not wrong, but it is not chasing a 1-2 day move. The selloff changed the moving averages, prompting CTAs to go broadly short; once prices returned to the prior highs, the models had to buy back according to their weights. “Even if it falls, it still has to buy it back.” That is the source of this week’s forced $20B purchase of U.S. S&P exposure, regardless of the bull, bear or flat case, and roughly $35B globally, according to Goldman’s report late on Apr 20.
• The rally follows the relationship between distance, speed and acceleration: the higher prices go, the more money has to be put in. The S&P rose for 12 straight days; the gain was largest 2 weeks ago, last week required more money but produced a smaller gain, and this week’s inflow is below either of the prior 2 weeks. The tinder is nearly gone, and the rally powered by systematic buying ends this week.

2. Next Week’s “Exam”: Positioning Is Not Full, and Good Earnings May Not Mean a Selloff

• Starting next week, the market returns to show-me earnings and the question of whether AI demand is real. From the week after next, CTA has little left to buy in the bull case as positioning fills out, but would have to sell heavily in the bear case. That looks dangerous, but the key caveat is that CTA positioning cannot be viewed in isolation.
• Prime-book selling after the war reached 2.5 sigma. Over the past 2 weeks, the book has covered only U.S. ETF shorts; single-name shorts have not been covered, and the global book remains at the same short exposure. In effect, ETF longs are hedging single-name shorts, leaving the book in wait-and-see mode. Market-timing and stock-picking funds and long-only investors are also waiting. “The market as a whole is not fully positioned… good earnings do not mean the market has to fall.” If positioning were full, he would expect a sizable correction next week—but it is not.
• The direct warning to anyone who went long at the lows or added leverage: “Risk starts accumulating next week. Whether the test goes well or badly, you need to make your own call.”

3. Political Headlines Discounted This Week: Gamma and CTA Provide a Double Floor

• Iran is alternately willing and unwilling to negotiate, and Vance is due to hold another round of talks on Wednesday. But the market should be desensitized to the headlines this week. The gamma book has been rebuilt and is back at very high levels: dealers buy when the market falls and sell when it rises, providing a balancing force. A market selloff would have to fight both CTA’s forced buying and dealer hedging demand, making it difficult to push prices through. On the other side, dealer selling into strength means the week is most likely to be “modestly higher or fairly flat.”
• Buying is concentrated in ETFs and e-mini futures—the speaker called them “san mi,” likely e-mini—so large caps are leading: Microsoft and Meta are both rebounding. Heavily shorted single names have not seen a broad rebound; the next quarter is still the test.

4. The AI Story Is Unchanged: Wall Street Is Coming Around to Token Demand

• The core reason risk assets were suppressed from late last year into early this year was Wall Street’s skepticism about whether token demand was real and whether capex would earn a return. “Wall Street never believed this demand existed.” That skepticism is now being worked through: almost everyone listening to the call should be using Anthropic for all kinds of coding. If you are not, you may be the one who gets eliminated.
• Mithius, Anthropic’s best model, is available only to enterprise customers. The official explanation is safety and excessive capability, but his view is that “50% of it is marketing”—Codex and 4.6 are not that different, while Anthropic’s execution is simply somewhat better. The real issue is insufficient compute. Models may stratify into free GLM for retail users, 4.6 for paying users, and the highest-compute, best model for enterprise customers.
• The economics are compelling: GPU purchases pay back in 1.5 years, while a large amount of capex is going into construction and other infrastructure amortized over 20 years. The market will worry when capex growth slows next year, but chip purchases still have to continue, and semiconductor earnings should not face a major problem; the market will digest the concern gradually by year-end. Even during the war, S&P first-quarter earnings estimates were raised 5%, with storage companies accounting for a large share of the revisions.
• The 2023 SVB analogy still applies: as long as earnings hold and the business model continues to make money, short-term shocks—war or oil at $200—do not change the fundamentals. The Nasdaq still gained 37% in the year SVB ran into trouble. “Go ahead and fight the war.” On relative value, Korean valuations are low, making SK Hynix preferable to Micron. SK Hynix has already broken to a new high and was said to be listing in June. It had previously been hit harder by oil prices, while the U.S. is not short of oil—“it threw a rock into someone else’s latrine.”

5. Kevin Warsh: More Dovish Than Priced, with 2 Hikes Possible This Year

• The Senate hearing is tonight. After the wrangling, the conclusion is that Warsh will ultimately be appointed: the DOJ is investigating a corruption case involving Powell, and the Senate wants that dropped before approving a new chair. But Warsh is a staunch Trump loyalist who would execute much of Trump’s policy agenda; when policy execution collides with a personal vendetta, Trump habitually TACO’s. If Warsh were genuinely hawkish, Trump could simply drag out the appointment.
• The dovish case starts with Warsh’s late-2025 reference to Greenspan’s decision not to hike in 1990: a major technology-driven productivity surge was deflationary, so there was no need to rush. The Anthropic founder said 50% of investment-banking jobs and programmers could disappear within 1-5 years; the host called that “too polite,” arguing that all of them could be eliminated in 1-2 years. Warsh is therefore more focused on unemployment risk. His prepared remark is that “inflation is a choice,” while “Fed independence is largely up to the Fed.”
• The conclusion is that “the market broadly thinks Kevin Warsh is extremely, extremely hawkish, but he is not.” He could well hike twice this year, causing the pricing of hawkish fears to unwind. Iran is similarly priced for the worst case: “Price it however you want.”
• A liquidity footnote: SOFR briefly traded below the Fed’s policy rate over the past 2 weeks. “The interbank rate was below the policy rate—pretty scary, right?” There was enough money to push funds into the RRP. Tax-season payments will move money from bank accounts into the TGA, equivalent to the Fed draining liquidity, bringing seasonal tightening over the next few weeks. Alongside the positioning discussed above, consider how to size your own positions.

6. Intel Will “Definitely Fail the Test” on Thursday—but $100 by End-2027 Is Fair

• He was among the earliest Intel bulls, calling it a “no-brainer buy” at $20; the stock is now at $65-$70. The standard view that data centers should be built around a 1:10 GPU-to-CPU ratio for training is wrong. At inference, “it is often not a person making the call—it is an Agent.” The machine calls the model, then calls tools and searches, all of which require CPU. CPU supply will remain structurally tight as compute and AI applications scale exponentially. The emergence of production-grade AI models at the end of last year—4.5 at the time—was the inflection point. He is his own example: “My entire deck was written by Claude, come on.”
• This quarter has 2 problems. Capacity cannot ramp—the data center wants newer chips, not the old ones—and the data center business is only 8-10% of Intel’s operations. The PC business is being hit by higher storage prices; storage is a commodity, “like trading oil,” so computers get more expensive and fewer people buy them. PC production lines have already shifted toward data center products, which helps earnings, but “it is not the explosive help everyone expected.”
• The exam will “definitely go badly,” and classmate AMD will probably have a bad quarter too, but the market may not be disappointed. “He is smart, the family is wealthy, and they have spent a lot on tutoring; it is fine if he does poorly this time.” 陈立武 has been in the seat for more than 1 year, and the standard incoming-CEO playbook is to flush out the legacy business in this quarter and the next, after which the results can be credited to his own work. Q3 and Q4, especially the forward outlook after that, could improve. A pullback would be an opportunity: $100 by the end of 2027 is fair, while the current move to $70 is excessive and earnings may not support it.