The Treasury Doubles Long-Dated Buybacks and the 30-Year Eases. Asia Breaks Anyway as the KOSPI Loses 5.8%.

Wednesday, August 19, 2026, written at 9:15am Eastern, fifteen minutes before the bell, with the July FOMC minutes at 2:00pm and three retailers already out. Two things happened overnight and they point in opposite directions. The US Treasury said it will at least double the size of its long-dated buyback operations, and the 30-year fell nine basis points off Tuesday’s nineteen-year high. Then Asia’s semiconductor complex fell apart anyway, with Korea’s Kospi down 5.80% and an exchange halt tripped inside the first hour. Reconciling those two facts is the whole job this morning.

Top Call: The Bond Problem Got a Backstop and the AI Problem Did Not

Treasury announced it will raise the maximum size of liquidity support buybacks in the 10-to-20-year and 20-to-30-year nominal sectors from $2 billion per operation to at least $4 billion, effective September 9 through November 4. The stated rationale is liquidity support in sectors where sponsorship is strong. The market read it as a bid, and the response was immediate. The 30-year went from above 5.33% Tuesday, its highest since 2007, to 5.196% this morning, and the 10-year gave up six basis points to 4.647%. The 2-year did not move, sitting at 4.175%. S&P 500 futures are up about 0.5%, Dow futures are up 305 points, and the Russell leads at roughly 0.7%.

Our take: this is a real intervention and it should be read as one rather than as a technical adjustment. Doubling operation size in the two longest nominal buckets is Treasury telling the market it will lean against disorderly moves at the back of the curve, and it worked within hours. But notice what did not happen. The Nasdaq 100 is up the same 0.5% as the S&P and less than the Russell. Yesterday this note argued that the growth complex now trades as a rates product, and if that were the entire story then a nine basis point relief rally at the long end should have produced obvious Nasdaq outperformance. It did not, and that is a genuine partial falsification of the duration thesis in its strong form. The rest of the explanation sits in a Wall Street Journal analysis published Tuesday that counted roughly $3 trillion of AI-related commitments living outside nine tech balance sheets. Financing got cheaper this morning. The scale of what is being financed did not change.

The Tape

The semiconductor ETF losing 5.02% in a session while the S&P gave up 0.69% is the number that frames everything else. Micron fell about 7%, Western Digital 7.4% and SanDisk 9.1%, and none of it came from memory-specific news. SanDisk entered Tuesday up 653% year to date and Micron up 255%. When positions that size begin unwinding, the trigger matters less than the crowding, and Monday’s Bank of America survey had already flagged long global semiconductors as the most crowded trade at 53%, down from 82% a month earlier.

Asia: The Memory Complex Sells Its Own Buyback

Korea took the worst of it by a wide margin. The Kospi closed down 5.80% at 6,471.17 after falling more than 6% within minutes of the open, which triggered the Korea Exchange sidecar and halted program sell orders for five minutes. Samsung Electronics fell 7.82% to 247,500 won. SK Hynix fell 9.75% to 1,500,000 won. In Japan the Nikkei lost 3.16% to 65,326.42, its lowest close in almost two weeks and a second consecutive down day, with Kioxia off 12.60% and SoftBank Group off 10.34% following its recent bond issuance to fund AI investment. Furukawa Electric, an optical and connectivity name that had run hard on the datacentre theme, fell 13.69%.

The detail that matters most is the timing. SK Hynix announced this morning that it will buy back and cancel 40 trillion won, roughly $29 billion, of its own stock. That is up to 24 million shares, about 3.3% of the 730.49 million outstanding, purchased between August 20 and November 19 and then retired. It is the largest treasury share cancellation in the history of Korean listed companies. Management simultaneously raised its 2025 to 2027 shareholder return commitment to more than half of cumulative free cash flow, backed by roughly 69 trillion won of net cash at the end of the second quarter, and said plainly that intrinsic value is not reflected in the share price.

Our take: a company announcing the largest buyback in its market’s history and then falling almost 10% is not a fundamental verdict, it is a positioning verdict. The buyback is well constructed and the balance sheet behind it is real. What the tape is saying is that the marginal holder of Korean memory in August 2026 is not a value investor who needed a yield signal, it is a momentum holder who needed the trend intact, and the trend broke. That distinction matters because the same announcement will read very differently in three weeks once forced selling clears. It also means the size of this drawdown understates nothing and overstates a great deal about end demand, which has not changed since Micron cleared $1,000 on Monday.

China went the other way on the surface and the same way underneath. The CSI 300 was down 2.4% at the lunch break with the semiconductor complex off 7% and the CSI Robot Index down more than 6%, while Hong Kong’s Hang Seng edged up 0.2% on energy strength. The divergence between a mainland tech rout and a flat Hong Kong index is mostly composition, and it is the third session running where Chinese energy names have offset Chinese AI names.

The $3 Trillion Sitting Off Nine Balance Sheets

The Wall Street Journal analysis that helped light Tuesday’s fuse deserves precision, because the headline number invites sloppy use. Across Alphabet, Amazon, Meta, Microsoft, Oracle, Nvidia, Broadcom, SpaceX and AMD, the paper counted roughly $3 trillion of AI-related commitments that do not appear as balance sheet liabilities, against roughly $600 billion of reported trailing twelve-month capital expenditure. The composition splits into about $904 billion to $1.2 trillion of leases that have been signed but not commenced, which under current accounting create no liability until they do, and roughly $1.52 trillion to $1.9 trillion of purchase commitments covering chips, construction and power. Alphabet alone carried $811 billion of purchase and contractual commitments as of June 30. Meta carried $347 billion of leases not yet commenced.

Our take: these are not hidden debts and calling them that is wrong. They are contractual obligations disclosed in the footnotes, and the accounting treatment is entirely conventional. What is genuinely new is the ratio. Five times reported capex in forward commitments means the spending path is far more rigid than a capex line implies, because a company can slow purchases but cannot easily walk away from a signed lease on a datacentre shell. That rigidity is the real risk transfer inside the AI buildout, and it is why the long end and the AI complex stay linked even after Treasury steps in. A cheaper 30-year makes financing the commitments easier. It does not make the commitments smaller, and if demand disappoints in 2028 the leases still commence. That asymmetry is what the market started pricing on Tuesday and did not fully unprice this morning.

Analog Devices: The Cleanest Print of the Morning

Analog Devices grew 40% year over year to a record $4.02 billion and guided the current quarter above consensus, with management pointing to data centre and industrial as the leaders. Shares rose. In a week where the semiconductor tape has been a positioning story rather than a fundamentals story, this is the counterweight, and it arrives from the least speculative corner of the sector.

Our take: analog is the part of semis that does not participate in narrative inflation, because it sells power management, signal chain and interconnect content into systems other people design. Forty percent growth off a base that spent 2024 and most of 2025 in an inventory correction says the buildout is still pulling content through the unglamorous parts of the bill of materials, and industrial recovering alongside data centre says the cycle is broadening rather than narrowing. If you wanted evidence that Tuesday’s 5% move in the semiconductor index was about crowding rather than demand, this print is the strongest single piece of it. We would treat the ADI guide as more informative about 2027 semiconductor revenue than anything the Kospi did overnight.

Target and Lowe’s: One Consumer, Two Very Different Reads

Target’s comparable sales grew 3.8% against a 2.5% consensus, driven by a 3.6% rise in comparable traffic, with store comps up 2.7% and digital comps up 8.7% on more than 25% growth in same-day delivery. Earnings per share doubled to $4.11. Of that, $1.65 came from $994 million of pretax tariff refunds flowing through gross margin, worth $752 million of net earnings. Excluding the refund, earnings grew 20%. Management raised the full-year net sales growth outlook to around 5%. The stock is down about 3%.

Lowe’s told the opposite story. Net income of $2.4 billion and $4.27 per share were flat against a year ago, and the company cut full-year guidance to the bottom of its prior ranges: sales of $92 billion rather than $92 billion to $94 billion, comparable sales flat rather than flat to up 2%, and adjusted earnings of $12.25 rather than $12.25 to $12.75, citing persistent softness in do-it-yourself spending. TJX raised its full-year earnings forecast to $5.31 to $5.36 but guided the current quarter to $1.30 to $1.32 excluding tariff benefits, below the $1.35 consensus, and fell about 4.6% premarket.

Our take: the tariff refund is the story inside the story and it applies well beyond Target. Roughly 40% of Target’s quarterly earnings per share came from recovering duties already paid, which means a large slice of this retail earnings season will be flattered by a legal outcome rather than by operations. The market is treating it correctly by selling a genuine operational beat, because 20% underlying earnings growth on 3.8% comps is good and the printed 100% is not repeatable. The Lowe’s cut alongside Home Depot’s reaffirmation yesterday is the cleaner signal: the professional and homeowner-with-equity customer is holding, the discretionary do-it-yourself customer is not, and that split has been widening for three quarters. Neither company can fix it while the 30-year sits near 5.2% and mortgage rates follow it.

Unitree and Lyntris: Two IPO Windows, Two Temperatures

Unitree Robotics, formally Yushu Technology, debuted on Shanghai’s STAR Market and opened 629% above its 150.8 yuan offer price at 1,100 yuan before closing at 845 yuan, up about 460%. The company raised 6.1 billion yuan, roughly $904 million, in an offering more than 8,000 times oversubscribed, a record for the exchange. Peak valuation touched about $66 billion. DeepSeek put in around 140.8 million yuan and Tencent was already on the register. Founder Wang Xingxing, who is 36, holds 121.4 million shares worth roughly 103 billion yuan at the close.

Lyntris, a defence technology roll-up assembled by Trive Capital from Accelint and Vitesse, priced 17 million shares at $17.50 against a marketed range of $19 to $22 and an original plan for 24 million shares. The company and selling holders raised $297.5 million rather than the $528 million initially sought. It carries a $923.9 million backlog and begins trading on the NYSE today.

Our take: put those two side by side and you get a decent map of where risk appetite actually lives. Chinese retail investors will pay $66 billion for a humanoid robot maker with modest revenue on the same day the mainland semiconductor index falls 7%, which is thematic speculation operating independently of the broader tape. US institutional investors just took a defence contractor with a real backlog at a 20% discount to its marketed range in the middle of an active shooting war in the Gulf. That is not a risk-on market and a risk-off market. It is one market pricing stories and another pricing cash flows, and the gap between them is as wide as we have seen this year.

Oil and Iran: The Escalation That Has Not Been Priced

Brent traded toward $92 and sits around $91.28, with WTI at $85.31, a fourth consecutive session of gains and the highest level since July 24. The war is in its twenty-fifth week. President Trump said no talks are underway or scheduled with Iran, days after saying a back channel had been established with the Revolutionary Guard, and posted an image showing the Strait of Hormuz labelled as New US Territory. A senior Iranian official told Reuters that Tehran will shift to a fully offensive military posture because negotiations have stalled, and threatened a timely and precise attack to break the American naval blockade. Eight vessel attacks in the strait have been reported this month, including ships linked to the UAE and Saudi Arabia. Turkey’s President Erdogan called Trump to urge de-escalation. Iran and Oman continue to discuss a fee-based passage arrangement run through a newly created corporation.

Our take: the crude price is not reflecting the change in posture. A stated shift to fully offensive operations combined with an explicit threat against the blockade is a different risk than the attrition of the past two months, and Brent moving a dollar is not the response that language deserves. Two explanations are available. Either the market has habituated to Iranian rhetoric after twenty-five weeks, which is plausible and has been correct so far, or the physical shortage is already so priced into refined products that the crude benchmark has stopped being the marginal signal. Yesterday’s record $102 diesel crack argues for the second. We would rather own this risk in product spreads than in flat crude, and we would treat any Hormuz incident involving a US naval asset as a step change rather than another headline.

Europe and the Rest of the Overnight

The Stoxx 600 was flat at 652.36 in the morning session, holding just above Tuesday’s close of 651.90, which was a two-week low. The eurozone government bond selloff that pushed long-dated yields above 3.2% on Tuesday, a fifteen-year high, paused as the US Treasury market stabilised, which is the buyback announcement transmitting across the Atlantic within hours. Energy is the only European sector working consistently. The DAX and CAC are little changed and both sit well below the records they set in early August.

Elsewhere, Japan’s 10-year government bond is still near 2.95%, its highest since 1996, with the market pricing a possible Bank of Japan hike in September, which is the domestic half of why Japanese equities fell harder than the global move alone would explain. Madison Air agreed to buy Airflow Technologies for $5.4 billion yesterday, and Alibaba agreed to sell its videogame business for at least $1.5 billion earlier this week. Neither is large enough to change a sector view, but both confirm that strategic acquirers are still transacting at scale while the IPO window prices at discounts.

Key Events Today

• Overnight: Kospi -5.80% with a sidecar halt, Nikkei -3.16%, SK Hynix -9.75% on a $29 billion buyback
• This morning: Treasury doubles long-end buyback operations, 30-year to 5.196% from above 5.33%
• Before the open: Analog Devices $4.02 billion and +40%, Target comps +3.8%, Lowe’s cuts guidance, TJX guides light
• 6:00am: MBA mortgage applications, refinance index prior 744.4
• 9:30am: EIA crude, gasoline and distillate inventories, prior crude build 17.4 million barrels
• Today: Lyntris begins trading on the NYSE after a downsized $297.5 million IPO
• 2:00pm: July 28 to 29 FOMC minutes, from a hold at 3.50% to 3.75% with Logan, Hammack and Kashkari dissenting for a hike
• Tomorrow: Walmart fiscal Q2, jobless claims, Philadelphia Fed, flash PMIs
• August 26: Nvidia fiscal Q2, consensus roughly $92 billion
• Later this month: Jackson Hole, where Chair Warsh frames September

What We Are Actually Watching

Three things, and the first is a correction to ourselves.

Yesterday this note argued that the AI complex now trades as a pure rates product. This morning gave that thesis a clean test and it only half passed. Long rates fell nine basis points at the 30-year and the Nasdaq did not lead the bounce. The honest revision is that the long end sets the financing cost and therefore the timing of stress, while the magnitude of stress is being set by something else, and the Journal’s $3 trillion count is the best available candidate for what that something is. Duration explains why the selling started Tuesday. Commitment scale explains why a rate reprieve did not stop it.

The second is the FOMC minutes at 2:00pm, which have become more interesting than they looked a week ago. Three regional presidents dissented in favour of a hike in July, September is priced at roughly a 65% hold, and JP Morgan has pushed its first hike call to December. The minutes will show how close the committee came to moving, and they were written before the last two weeks of oil. Any language suggesting the hawkish bloc is broader than three votes lands into a market that just received a Treasury-engineered rally in the long end, and an interventionist Treasury alongside a hawkish Fed is an awkward policy mix to trade.

The third is whether Korea’s break spreads or stays local. The Kospi is down roughly 7.3% in two sessions and it tripped an exchange halt this morning, which is mechanically different from a bad day. The case for containment is that the fundamentals underneath have not deteriorated at all, from Micron above $1,000 to Analog Devices at 40% growth to high bandwidth memory sold out through 2027. The case for worry is that Korean memory has been the highest-beta expression of the global AI trade all year, and high-beta expressions usually break first rather than break alone. Watch whether US memory names open green on ADI or red on Asia, because that answer is worth more than anything else the tape produces today.

All marks above are pre-market and will move by the open.