HSINCHU / SEOUL — Taiwan Semiconductor Manufacturing Company reported Thursday that its third-quarter net income rose 68.7 percent year over year to NT$763 billion, on revenue of NT$1.49 trillion. Diluted earnings per share came in at NT$29.43 — roughly $4.60 per American depositary receipt. The gross margin was 66.4 percent. Management guided fourth-quarter revenue to $52.3–54.3 billion, with margins of 65 to 67 percent.

Those are not merely good numbers. They are the numbers of a company operating at the physical center of a global capital-expenditure supercycle, where every hyperscaler, every sovereign AI program, and every defense contractor on earth is bidding for the same advanced wafers. High-performance computing drove 68 percent of total revenue. The 3-nanometer node alone produced a third of wafer sales. The 2-nanometer node, barely into commercial ramp, contributed 8 percent. Chips built on 7nm and below accounted for 79 percent of all output.

Hours after TSMC's report, the mood reversed. South Korea's KOSPI index fell more than 7.5 percent in a single session, closing at 5,554, and triggered a market-wide circuit breaker — the tenth of the year. The KOSDAQ plunged 7.6 percent. Samsung Electronics dropped over 9 percent. SK hynix fell past 11 percent.

The catalysts were stacked. China's CXMT announced plans to commercialize a server-grade DDR5 module built on a 4F² DRAM architecture by year-end — a design that shrinks cell area by roughly a third and, if it works at scale, could undercut Korean memory pricing by offering comparable density from older lithography tools. Samsung and SK hynix have researched the same architecture but placed mass production closer to 2028. CXMT claims it can beat them by two years, even without access to EUV equipment.

Analysts are deeply skeptical. CXMT has disclosed no yield data, no speed ratings, no volume commitments. Building a laboratory sample on deep-ultraviolet multi-patterning is one thing; running enterprise-grade server memory at scale is another. But the market did not wait for verification. It sold first and will ask questions later.

The inflation data compounded the damage. The September U.S. Consumer Price Index came in at 3.7 percent year over year, above the 3.6 percent consensus, with core CPI at 2.6 percent versus the expected 2.5 percent. Gasoline prices surged 9.1 percent for the month — more than half the headline increase — driven by the same Gulf disruptions that have kept Brent crude above $100. Treasury yields spiked. The rate-cut thesis that had buoyed technology valuations for months was, for practical purposes, declared dead on arrival.

And then there was the debt. Oracle fell roughly 9 percent as investors re-examined the balance sheet of every company that has financed AI infrastructure on borrowed money. CoreWeave's unsecured bonds broke through 15 percent yields — the kind of number that credit markets associate with distressed issuers, not growth stories. Elon Musk responded to the yield chart with a single word: *yikes.* Gartner released a forecast estimating that more than 80 percent of neocloud providers will cease independent operations by 2030, projecting that AI data-center operators need $8.5 trillion in revenue by decade's end to cover capital and financing costs, against a projected $3.6 trillion in actual AI revenue.

The divergence between TSMC's record and Korea's rout is not a contradiction. It is the same story read from two ends. TSMC is the toll bridge — it collects margin on every chip regardless of whether the buyer's business model works. The memory makers and the neoclouds are the ones who have to earn back what they borrowed. When interest rates rise and Chinese competition materializes simultaneously, the toll bridge keeps collecting while the travelers start drowning.

The AI physical supply chain has never been more profitable at the top or more leveraged at the bottom. That is the kind of structure that works beautifully until it doesn't.

CoreWeave, the GPU-cloud company that emerged from a cryptocurrency mining operation to become the most prominent independent provider of AI computing infrastructure, saw its unsecured bonds surge past 15 percent yield this week — a threshold that credit analysts conventionally classify as distressed.

The company carries approximately $35 billion in total debt, with total obligations including leases reported near $52 billion. Quarterly net interest expense has climbed to roughly $640 million, keeping the company cash-flow negative even as revenue expands. CoreWeave touts a revenue backlog of around $104 billion, but backlog is a contractual promise, not liquid capital, and servicing debt requires the latter.

Credit default swaps on AI-infrastructure issuers have widened sharply. Chamath Palihapitiya flagged CoreWeave's bond yields on social media. Musk's single-word reply — "yikes" — amplified the anxiety. CoreWeave shares fell roughly 15 percent in a single session, dragging neocloud peers Nebius, Iren, and Hut 8 down with them.

The pressure is structural, not idiosyncratic. A hotter-than-expected September CPI print — 3.7 percent headline, 2.6 percent core — has revived rate-hike expectations, raising the cost of the very debt that finances every GPU rack in every neocloud data center. Oracle, which carries its own enormous AI infrastructure liabilities, fell 9 percent on the same day.

Gartner's latest forecast is the sharpest institutional warning yet: more than 80 percent of neocloud providers will cease independent operations by 2030. The research firm estimates that AI data-center operators collectively need $8.5 trillion in revenue by decade's end to cover planned capital expenditures and financing costs. Projected actual AI revenue over the same period: $3.6 trillion.

The arithmetic is not subtle. The companies building AI's physical layer have borrowed against a future that may not arrive on schedule. The hyperscalers — Microsoft, Google, Meta — can fund infrastructure from cash flow. The neoclouds cannot. When rates rise, the margin between solvency and distress narrows faster than revenue can grow.

CoreWeave's backlog may eventually convert to cash. Its GPU contracts may hold. But bond markets do not trade on eventually. They trade on whether the next coupon gets paid.

ChangXin Memory Technologies, the Hefei-based DRAM manufacturer known as CXMT, announced plans to commercialize a server-grade DDR5 memory module built on a 4F² cell architecture before year-end — a timeline that, if real, would place a Chinese chipmaker ahead of Samsung and SK hynix by roughly two years on one of the most significant architectural transitions in memory-chip history.

The 4F² layout shrinks the physical footprint of each memory cell by about a third compared to the standard 6F² design used across all major commercial DRAM today. It achieves this through a vertical surrounded-gate transistor that stands the channel upright and wraps the gate around it, reducing leakage and enabling tighter packing. The density gains — 30 to 50 percent more storage on the same silicon area — translate directly into cost advantage.

The announcement hammered Korean chipmakers. Samsung Electronics fell over 9 percent. SK hynix dropped past 11 percent. The KOSPI triggered its tenth circuit breaker of the year.

But the details that would convert an announcement into a threat are conspicuously absent. CXMT has disclosed no yield rates, no operating speeds, no volume production commitments, and no reliability data for enterprise server environments. Building a working sample in a laboratory and running high-volume mass production at enterprise reliability are separated by years of engineering, not months.

More fundamentally, CXMT operates under U.S. and allied export controls that deny it access to extreme ultraviolet lithography equipment. It must achieve the 4F² architecture using older deep-ultraviolet tools and multi-patterning — a process that multiplies manufacturing steps, reduces throughput, and compresses the yield margins that determine whether a chip is commercially viable or merely technically possible.

TrendForce estimates CXMT currently holds roughly 9.5 percent of global DRAM market share, built largely on commodity products. A jump to cutting-edge architectural competitiveness would represent a qualitative shift, not merely a quantitative one.

Samsung and SK hynix have both presented 4F² research at conferences like ISSCC and IEDM. Neither has committed to mass production before 2028. The Korean firms possess EUV tools, decades of yield-engineering expertise, and entrenched relationships with hyperscaler customers. The question is whether CXMT's DUV-only approach can replicate those advantages through sheer architectural cleverness — or whether the announcement is, as several analysts have labeled it, a "paper tiger" designed to demonstrate capability under sanctions pressure without delivering commercial volumes.

The market sold first. Verification can wait.

TAIZ, Yemen — Houthi forces have effectively encircled Taiz, Yemen's third-largest city, cutting supply routes from the north, east, and west and trapping more than a million residents inside a tightening siege that aid agencies warn could produce a humanitarian catastrophe within weeks.

The offensive, launched in early September and escalating sharply through October, has seized key towns on the city's outskirts. Fierce fighting has concentrated around al-Dhabab to the southwest, al-Shaqab and al-Aqrud to the southeast, and the districts of Hayfan and al-Misrakh. Government forces backed by Saudi airstrikes have mounted counteroperations using artillery, drones, tanks, and missile fire, claiming heavy casualties on both sides.

The military claimed this week that more than 200 Houthi fighters were killed or injured in the southwestern Taiz province. The figures could not be independently verified.

Inside the city, the consequences of the encirclement are immediate and physical. Essential supplies — wheat, flour, fuel, cooking gas, and medical treatments — have plummeted. Prices have spiked. Civilians report burning wood and cardboard for cooking. Tens of thousands have been displaced from outer districts; others are attempting to flee the city itself, though exit routes are increasingly contested.

On October 14, Yemen's Foreign Minister Afrah Abdulaziz al-Zouba formally appealed to the UN Security Council and international humanitarian agencies, demanding pressure on the Houthis to lift the siege and open safe corridors for the more than four million people in the broader province.

The timing is not accidental. The Houthi offensive on Taiz coincides with the group's intensified aerial campaign against Saudi Arabia, including the devastating strikes on King Khalid International Airport in Riyadh on October 10–11 that killed and injured hundreds. Iraq's new Prime Minister Ali al-Zaidi visited Riyadh this week and pledged that Baghdad would not allow Iraqi territory to be used as a launchpad for attacks on Saudi soil — a statement notable for the fact that it needed to be made at all.

Trump said this week that the war involving Iran will end "very soon" and that oil prices will "drop like a rock" when it does. In Taiz, where the flour is gone and the roads are closed, the timeline for "very soon" is measured differently.

ROME — Pope Leo XIV was walking the corridors of Gemelli Hospital on Wednesday, two days after surgeons removed the lower lobe of his right lung in a 90-minute operation to excise a malignant tumor.

Post-operative pathology confirmed a diagnosis of lung adenocarcinoma, the most common form of lung cancer. The Vatican's medical bulletins, issued under the names of Dr. Luigi Carbone and thoracic surgeon Dr. Stefano Margaritora, report that the Pope's recovery is proceeding within expected parameters. Blood tests and chest X-rays are normal. He has resumed a regular diet and begun respiratory physiotherapy.

Doctors expect the 71-year-old to remain hospitalized for approximately one week from his October 13 admission. The Vatican has cancelled general audiences through October 20. Further tests on the removed tissue will determine whether additional treatment — chemotherapy or radiation — is necessary. No announcement on staging or spread has been made.

The Vatican's handling of the diagnosis has been remarkably direct by institutional standards. Where previous papacies treated the pontiff's body as nearly sacrosanct — concealing John Paul II's Parkinson's for years, obscuring Benedict XVI's declining health — Leo's office disclosed the cancer in clinical terms, named the tumor type, described the surgical approach, and published daily medical bulletins with the treating physicians' names attached.

The transparency has drawn praise from Catholic commentators and medical ethicists alike. It has also surfaced an uncomfortable truth about institutional authority: the reason the Vatican can be honest about lung cancer is that honesty, in this instance, costs nothing politically. The harder transparency tests — on finances, on abuse investigations, on doctrinal disputes — remain precisely as opaque as they were before the diagnosis.

For now, the Pope walks. The pathology will speak next.

Donald Trump told the nation this week that the war with Iran will end "very soon." Oil prices, he added, will "drop like a rock." There would be no new military strikes before the November 3 midterm elections.

Set aside, for a moment, whether any of this is true. Focus instead on the grammar. "Very soon" is the president's most reliable temporal construction because it occupies no fixed position on a calendar. It is a promissory note drawn on an account that never matures. It sounds like a deadline but functions as an indefinite deferral, the political equivalent of a check postdated to an address that doesn't exist.

We have heard this phrase before. The pandemic would end very soon. The wall would be built very soon. Infrastructure week was very soon. The common thread is not dishonesty — though that is often present — but a more fundamental category error: the substitution of narrative momentum for temporal specificity. "Very soon" feels like progress because it implies proximity. It is, in practice, a device for purchasing the present with a claim on the future.

In the case of Iran, the stakes of the ambiguity are unusually high. Brent crude remains above $100. The Strait of Hormuz is contested. Houthi missiles have struck Riyadh's main airport. In Taiz, a million Yemenis are besieged. Saudi refining infrastructure remains damaged. The physical conditions for a resolution — security in the Gulf shipping lanes, a ceasefire in Yemen, some framework for Iranian nuclear compliance — are not "very soon" conditions. They are conditions that require negotiation, verification, and enforcement, each of which operates on timelines measured in months or years.

But the midterms are in nineteen days. And nineteen days is, in the president's arithmetic, precisely the length of "very soon" — long enough to promise, short enough that no one can prove the promise broken before the ballots are counted.

The pledge not to strike Iran before November 3 is revealing. It implies that the decision to launch military action is governed not by intelligence assessments or strategic necessity but by the electoral calendar. If a strike would be justified on November 4, it is justified on October 15. The intervening three weeks are not a window of diplomacy. They are a window of political convenience.

Presidents have always timed war around elections. Johnson escalated after 1964. Nixon bombed before 1972. The innovation here is not the timing but the transparency: Trump is telling you, in plain language, that the war pauses for the vote. He is not even pretending otherwise. The candor is its own kind of honesty — the honesty of a man who has concluded that the electorate either doesn't mind or won't remember.

"Very soon" is the sound of a bet that memory is shorter than a news cycle. So far, it has paid off every time.

In 2006, the financial system discovered that you could package risky mortgages into securities, assign them investment-grade ratings, and sell them to institutions that didn't look too carefully at the collateral. The collateral was houses. The assumption was that housing prices always go up.

In 2026, the AI infrastructure sector has performed a structurally similar operation. The collateral is GPUs. The assumption is that AI demand always goes up.

CoreWeave's unsecured bonds now yield above 15 percent. The company carries $35 billion in debt and $52 billion in total obligations. Its quarterly interest bill runs $640 million. It is cash-flow negative. Its pitch to investors is a $104 billion revenue backlog — a number that sounds enormous until you realize that backlog is a contractual promise, not money in hand, and that the customers responsible for fulfilling those contracts are themselves making bets on AI revenue that may or may not materialize.

This is the structure of a leveraged intermediary. CoreWeave borrows money to buy NVIDIA GPUs. It rents those GPUs to companies building AI applications. Those companies charge users for AI services. Each link in the chain assumes the next link will hold. If end-user AI revenue disappoints — if the $3.6 trillion that Gartner projects through 2030 turns out to be the ceiling rather than the floor — the entire chain reverses. The customer can't pay CoreWeave. CoreWeave can't service its debt. The bondholders take the loss.

The resemblance to subprime is not exact, but the family likeness is unmistakable. In both cases, the underlying bet is a directional wager on a single asset class disguised as infrastructure investment. In both cases, the leverage amplifies returns on the way up and destruction on the way down. In both cases, the intermediaries insisted that the math worked because demand was structural and permanent. In both cases, the rating agencies and research analysts nodded along until the market started pricing the risk.

The market is pricing the risk now. Oracle fell 9 percent. Credit default swaps on AI issuers are at record highs. Gartner says 80 percent of neoclouds won't survive independently to 2030.

There is a crucial difference from 2008: the banks are not yet loaded with neocloud debt in the way they were loaded with mortgage-backed securities. The systemic risk is smaller. The losses will be concentrated among private credit funds, venture investors, and high-yield bond holders, not depositors. This is a feature, not an accident — the financial system learned something from the mortgage crisis, even if the lesson was mostly about where to put the risk rather than how to measure it.

But the human pattern is identical. When an industry is growing fast enough, the capital markets will finance it past the point of rational return, because the fees are too good to refuse and the models say the music doesn't stop. The music always stops. The question is only what it takes to make the band quit playing.

Fifteen percent yields on CoreWeave bonds is the sound of the drummer looking at the door.

WASHINGTON — The Bureau of Labor Statistics reported Tuesday that the Consumer Price Index rose 3.7 percent year over year in September, above the 3.6 percent consensus and up from 3.4 percent in August. Core CPI, excluding food and energy, rose 0.3 percent month over month, pushing the annual core rate to 2.6 percent — a tenth above expectations.

The headline number was driven almost entirely by energy. The overall energy index jumped 5.2 percent for the month. Gasoline prices surged 9.1 percent, accounting for more than half of the total monthly CPI increase. Year over year, gasoline is up 34.2 percent, a figure that traces directly to the disruptions in Gulf shipping lanes, Houthi attacks on tanker routes, and the broader Iran conflict that has kept Brent crude anchored above $100 a barrel.

Shelter inflation continued its slow grind, rising 0.4 percent for the month and 3.2 percent year over year. Recreation, used cars, airline fares, medical care, and apparel all contributed smaller increases. Food prices remained relatively steady.

Treasury yields spiked immediately on the release. The 10-year jumped sharply as fixed-income markets priced out the remaining probability of near-term rate cuts. For the technology sector — particularly the capital-intensive AI infrastructure buildout — the implications are direct: every percentage point of additional borrowing cost translates into billions of dollars of incremental debt service across an industry that has financed its expansion overwhelmingly through leverage.

The September print arrives at a politically loaded moment. Trump has pledged that oil prices will plummet once the Iran conflict ends, but the conflict has not ended, and the CPI reflects the physical reality of a war economy: damaged refineries, contested shipping lanes, and energy prices that respond to munitions, not rhetoric.

The Federal Reserve meets next in early November. The September data makes a rate hike more likely than a cut — an outcome that, six months ago, almost no forecaster considered plausible.

In late September, during a basketball game at the Asian Games in Japan, Canon placed roughly 100 cameras around the court, processed the footage into three-dimensional spatial data at its Kawasaki office, compressed it in real time, and streamed it through a cloud distribution platform to approximately 750 individual consumer devices simultaneously.

Each viewer could freely choose their own perspective — overhead, courtside, behind the basket, tracking a single player — while the game was still being played. This was not a replay feature. It was live.

Canon calls it volumetric video, and the company announced the demonstration as a world first: the initial successful test of streaming live volumetric capture to a large number of independent devices. A separate but related demonstration, conducted with NTT's IOWN All-Photonics Network, let remote spectators view a live game through mixed-reality headsets at a satellite viewing site.

The technology is real, and it is spectacular in the way that early demonstrations of genuinely new media always are. The first television broadcast was spectacular. The first live satellite feed was spectacular. Spectacle, however, is a necessary but insufficient condition for changing how people actually watch things.

The barriers are substantial and mostly un-glamorous. Volumetric capture requires roughly 100 cameras per venue — a hardware deployment that is economically rational for a World Cup final but absurd for a Tuesday-night league match. The data volumes are enormous; the compression and real-time processing requirements are at the edge of what current compute can handle even in a demonstration. The consumer devices that received the stream were presumably pre-configured; deploying this to a mass audience on heterogeneous hardware is an entirely different engineering problem.

And then there is the viewing question. Television succeeded not because it gave viewers control but because it took it away. A great sports broadcast is a narrative, shaped by a director who knows when to cut to the wide shot, when to zoom on the striker's face, when to replay the foul in slow motion. Giving every viewer their own camera is not obviously better; it may simply be more overwhelming. The history of interactive television is littered with technologies that were technically impressive and experientially exhausting.

But Canon has done something that matters: it has proved the plumbing works. The rest is taste, economics, and time.

BANGKOK — Nepal and the World Bank signed a mandate agreement this week to develop what would be one of the first earthquake catastrophe bonds issued by a South Asian sovereign: a three-year parametric instrument targeting $80 million to $190 million in financial protection against major seismic events.

The structure is elegant and unsentimental. Investors buy the bonds and collect a premium — essentially an insurance fee — for the duration. If an earthquake exceeding pre-defined physical parameters (magnitude, ground-shaking intensity) strikes Nepal during the bond's life, the principal is automatically redirected to the Nepalese government. If no qualifying earthquake occurs, investors keep their principal plus the premium. No damage assessment is required. No aid bureaucracy intervenes. The payout is triggered by physics, not paperwork.

The World Bank Treasury is acting as intermediary, structuring the risk analysis and connecting Kathmandu with international capital-market investors. Nepal's Council of Ministers approved the exploration earlier this year.

For Nepal, the logic is straightforward. The country sits on the boundary between the Indian and Eurasian tectonic plates. The 2015 earthquake killed nearly 9,000 people, displaced 3.5 million, and caused damages estimated at $7 billion — roughly a third of GDP. Reconstruction relied overwhelmingly on donor pledges that arrived slowly and often with conditions. A cat bond converts that uncertain, politically mediated aid flow into a pre-arranged financial mechanism with a contractual trigger.

For investors, the attraction is diversification: earthquake risk is uncorrelated with equity markets, interest rates, or corporate earnings. A Nepal cat bond would trade on geology, not macroeconomics. The premium — likely several percentage points above risk-free rates — compensates for the small but nonzero chance that the Himalayas collect.

The instrument is not without critics. Parametric triggers can miss the mark: an earthquake that devastates Kathmandu but falls slightly below the magnitude threshold would leave Nepal with casualties and no payout. Conversely, a deep, high-magnitude quake in an unpopulated area could trigger a payment where little damage occurred. The precision of the physics is both the product's strength and its limitation.

But the broader significance is harder to argue with. Nepal — a nation with a GDP smaller than some American counties — is preparing to access the same capital markets that price CoreWeave bonds and Treasury futures. If it works, it will be the most direct proof yet that catastrophe bonds can serve not just Caribbean islands and Japanese insurers but the poorest, most seismically exposed nations on earth.