Market Overview: March 17, 2026

Summary

• The central market event over the next month is the US-Iran conflict. Normal traffic through the Strait of Hormuz is 19-21M barrels/day; it is now effectively shut. The IEA has agreed to release 400M barrels from strategic reserves (2.5-4M/day, depending on how the situation develops), while a Russian easing of restrictions could add another 0.3-0.5M; the net shortfall would still be about 16M/day—“the scale is terrifying.” Goldman Sachs’ base case is a 21-day closure, implying a roughly 330M-barrel shortfall and a halving of global commercial inventories. Crude is already pricing in no passage this month, but “definitely not 2-3 months” of disruption.
• Stagflation is unlikely; once the market has over-priced it, Treasuries become a buy. UK front-end rates moved 10 standard deviations in 2 weeks, with large positions betting on stagnation, but the speaker disagrees: oil’s pass-through to core inflation is weak—every 10% rise lifts PCE by just 0.2 and core PCE by 0.04—and worsening unemployment would offset part of the increase in core inflation. When forced to choose between employment and inflation, the Fed usually chooses employment. “This is the year for bonds”—if oil triggers a capital-markets selloff, Treasuries are the best dip-buy.
• This is different from the 2022 Russia-Ukraine shock. Rates were exceptionally low then, central banks were printing money and fiscal policy was still expanding; “the kindling for inflation was already stacked high, so the flames took off.” Rates are high now, employment data are very weak—the NFP print the week before last was extremely poor—and JPMorgan has even shifted toward expecting a rate hike this year, while the speaker believes the Fed “doesn’t have the capacity to hike.” The 10-year yield is already at 4.22-4.25, high but “not yet a level to go outright long duration”—the market is like a drawn bow, and “when it is pulled all the way back, the rebound will be very sharp.”
• CTA selling peaks this week. CTAs were expected to sell $21B last week and actually sold $20.5B, including more than $20B of the S&P, broadly in line with expectations. CTA positioning is now flat: short- and medium-term momentum have flipped short, while long-term momentum in the 6300s has not flipped yet. This week, the S&P still has $28B to sell, US equities more than $30B and global CTAs roughly $70B, after global CTAs sold $80B last month. Once the liquidation is done, the pace of selling should slow—“a glimpse of light”—and once all CTAs are short, only long exposure remains, “a bit like last April.” About $2B of negative gamma sits 5% lower; the book is thin, so a drop there would accelerate instantly.
• Nvidia’s $1T revenue call at GTC failed to excite the market—the real issue is private credit. Hyperscalers ultimately fund capex through private credit. A sharp rise in oil widens CDS spreads, weakens financing capacity, suppresses AI capex and damages market sentiment. Private credit has been falling since year-end, led lower by Oracle and other AI-capex names; software exposure may be roughly 25-30%. As big-tech free cash flow declines, funds such as Apollo and Blue Owl must take assets off balance sheet and add leverage, but those funds are being slowed by credit stress. AI capex could reach $800B next year and more than $2T over the next 2-3 years.
• A capex slowdown is not necessarily bad news for semiconductors. Investment divides into construction, depreciated over 20 years, and buying GPUs, with H100s potentially paying back in about 1.5 years. As capex slows, direct GPU purchases make up a larger share and ROE rises; completed data centers could still be short more than 3M GPUs this year and more than 4M next year. Memory offers a precedent: after Huang bought up most of the GPUs, HBM cards and capacity, Xiaomi and Huawei (Honor) could not source memory for their phones, and prices surged when demand was cut. Oil works the same way in a panic—using half of commercial inventories is not the same as using 80%.
• Geopolitically, both sides are preserving their final negotiating chips. The US killed Khamenei but did not bomb Iran’s largest crude-export port or blockade Iranian vessels. Iran, which exports about 1.5M/day, may lack the ability to mount a large-scale counterattack: any coastal firing position would be destroyed by US forces within 5 minutes of opening fire. Tehran may instead quietly lay mines to harass shipping—“like scattering a bag of sand into a bowl of rice; picking it all out would be extremely difficult.” If the closure persists, oil-short Europe, China, Japan and South Korea would be hurt far more than the US, which has 4M/day of its own exports versus 20M/day for the Persian Gulf as a whole.

Deep dive

1. The Hormuz math: a net 16M/day shortfall burns through half of commercial inventories in 21 days

• Goldman Sachs research gives the pass-through coefficients: a 10% rise in oil lifts PCE by 0.2 and core PCE by 0.04. Normal traffic through the Strait is 19-21M barrels/day; now “roughly less than 1M” is getting through. Some Iranian operators may know where the mines are, so Iranian ships may still quietly make it to China to sell oil.
• On the supply side, the IEA has agreed to release 400M barrels from strategic reserves over roughly 120 days, or 2.5-4M/day—“a little less at first, and more if conditions worsen.” A Russian easing of restrictions could add another 0.3-0.5M. Total replacement supply stays below 3M/day, leaving a net shortfall of about 16M/day. Maintaining the closure for 21 days through month-end implies 330M barrels; another 3 weeks would reduce global commercial oil inventories by roughly 50%.
• The key mechanism is that panic buying emerges only once demand is cut. The memory-market analogy is straightforward: after Huang bought up HBM cards and capacity, Chinese EV makers, Xiaomi and Huawei (Honor) could not source memory, hurting shipments; “once demand is cut, prices explode.” A 21-day closure is Goldman Sachs’ base case, after which continued IEA releases could allow the market to recover. “If it runs beyond 21 days and turns into 3 or 4 months, this oil price will be very hard to recover.”

2. Stagflation is unlikely: the pass-through chain is weak, and the Fed usually favors employment

• UK front-end rates saw a 10-standard-deviation move at year-end. Rates had barely moved before that, implying that large positions were betting on stagflation. The speaker’s rebuttal is that oil first feeds headline inflation and only later core inflation; by the time the core effect arrives, unemployment has already worsened, and “that itself offsets part of the increase in core inflation.” Between employment and inflation, the Fed’s “usual choice” is employment. “Personally, I don’t think it will.”
• This is not the 2022 Russia-Ukraine setup. Rates were exceptionally low, central banks were printing money and fiscal policy was still spending; “the kindling for inflation was already stacked high, so the flames took off.” Rates are high now, and the US macro direction is toward rate cuts.
• Politics is the variable. The Fed chair and Trump have “basically fallen out”; whether Powell leaves and what action he takes beforehand is difficult to call. The left-right fight is so intense that even the Save America Act, which would require voter ID, cannot pass. Political infighting inside the Fed is also severe, so the final policy response remains uncertain.

3. This is the year for bonds: the bowstring is not fully drawn

• The speaker’s call at the start of the year was that, absent war, “this is the year for bonds” (“今年是债的大年”). Long-end yields should fall given the Treasury-supply backdrop and a macro deflationary setup. If the market prices in too much stagflation, “you should buy Treasuries; I think the market is underestimating them far too much.”
• This oil shock is “cut from the same cloth” as Trump’s tariffs. In the worst case—oil above $150 and staying there—core inflation would be 0.5 higher by year-end while the economy weakened further. But the Iran problem cannot remain unresolved for 1, 2 or 3 years; “it will be either weeks or months.” Once the crisis is resolved, the market should rebound sharply, much as it did when tariffs were lifted. Unlike the tariff episode, however, the economy’s fundamentals will already be damaged, so yields should fall in addition to the capital-markets rebound.
• Timing still argues for restraint. The 10-year at 4.22-4.25 is “a relatively high level, but not yet a level to go outright long duration.” The market is like a drawn bow: “pull the bowstring back—it definitely isn’t all the way back yet, but when it is, the rebound will be very sharp.”

4. Positioning: CTA selling peaks this week, with light at the end

• CTAs were expected to sell $21B last week and sold $20.5B, including more than $20B of the S&P, broadly in line. CTA positioning is now flat: short-term and medium-term momentum have flipped short, while long-term momentum in the 6300s has not flipped that quickly. About $2B of negative gamma sits roughly 5% lower—“not a large amount, but once the market gets there, it detonates instantly”—and displayed liquidity remains thin.
• This week, the S&P is expected to sell $28B, expanding to more than $30B across US equities and about $70B from global CTAs, with the US accounting for roughly half. Global CTAs sold $80B last month. The market’s sideways action over the past month was driven by “CTA buying and buying,” gamma overhead preventing a breakout, and retail dip-buying that kept the tape flat.
• Last week’s decline was far smaller than the speaker expected. Under negative gamma, $20.5B of selling should have produced a much larger drop, suggesting retail’s appetite to buy was strong. Once this week’s selling is finished, the monthly pace of liquidation should slow—“a glimpse of light.” After CTAs are fully pushed short, only long exposure remains, and “the rebound will be just as sharp, a bit like last April.”

5. Private credit is pulling AI capex around

• At GTC, Nvidia talked about $1T in revenue, but “the market did not push the stock up very much”; the concern underneath is what happens to capex. Hyperscaler funding ultimately comes from private credit. Oil up sharply leads to wider CDS spreads, weaker access to financing, lower AI capex and weaker market sentiment: “Nvidia’s earnings are not a problem, but the whole market’s sentiment is affected, and stocks fall again.”
• Private credit has been falling and CDS spreads widening since year-end, initially led lower by Oracle and other AI-capex names. Software exposure may be roughly 25-30%—lower at Goldman Sachs, higher at some funds. As big-tech free cash flow declines, capex requires financial sponsors to take assets off balance sheet and add leverage, relying on private funds such as Apollo and Blue Owl. Credit stress is slowing the buildout of those off-balance-sheet businesses. On the demand side, AI capex could reach $800B next year and more than $2T over the next 2-3 years.
• The funding structure is changing. A new generation of investors does not buy funds but invests directly; funds typically allocate 20-30% to credit bonds, while retail investors allocate only 1%, so capital is continually being redeemed from private markets. Many funds have imposed a 5% monthly redemption cap. Being unable to redeem makes clients want to redeem even more, while some of the cash that does come out is recycled into equities—one explanation for retail’s persistent dip-buying. Capex growth will eventually slow to a critical point; “it could happen this year, or it could happen next year.”

6. Semiconductors are not in as much trouble; AI’s next step is revenue and social impact

• A capex slowdown is “not that negative” for semiconductor companies. Investment splits between construction, depreciated over 20 years, and GPU purchases, whose payback period is fast: an H100 could pay back in about 1.5 years, while B-series cards could take 1.5-2 years, depending on the rental price per kilowatt-hour. Oracle-like capex spenders have low ROE because they are “playing Monopoly”—buying land and waiting 1.5 years before deploying cards. As capex slows, direct GPU purchases make up a larger share and ROE rises. Completed data centers could still be short 3M-plus GPUs this year and 4M-plus next year.
• AI’s next proof point is revenue, while token demand has been rising exponentially over the past few months. The speaker’s own example: an insurance comparison company he had invested in can now have a tool—possibly Claude—run the comparison in a few minutes. Websites requiring front-end and back-end work have gone from taking several weeks and several people to being built in 30 minutes to 1 hour once the requirements are clearly written. This deck itself may have been written by Claude.
• The political implication is mass unemployment among cognitive workers, “most of whom are Democratic supporters,” reshaping the entire voter base. “America looks very right today—is it about to correct course and swing back left? The shift right may be a much longer cycle.” The same dynamic will play out in China.
• The closing geopolitical view comes with an explicit hedge: “I’m really not very knowledgeable about military affairs, but my sense is that Iran may not have a large-scale retaliatory capability.” Neither side has exhausted its negotiating chips: the US has not bombed Iran’s export port, while Iran has not gone all-in on mining the Strait and truly sealing it. Captains also do not want to risk being bombed: “I’m just doing a job—why would I get myself blown up?”