TSMC's revenue jumped 51% on AI demand, Samsung forecast a record profit, GlobalFoundries won a $2B TSMC packaging deal, and Huawei claimed its AI chips outsell Nvidia's in China.
Published 8 October 2026 · 18:00 GMT
Thursday was the busiest day of the year for the silicon that powers artificial intelligence. TSMC, the Taiwanese foundry that makes most of the world's advanced chips, said third-quarter revenue rose to NT$1.49 trillion, up roughly 51 percent from a year earlier and ahead of the NT$1.46 trillion analysts expected. Samsung forecast a record quarterly operating profit of about $80 billion, up nearly ninefold on AI-driven chip demand. GlobalFoundries announced a $2 billion, five-year manufacturing agreement with TSMC to build advanced-packaging components in New York. And in Shenzhen, Huawei's rotating chairman claimed the company's Ascend AI chips have now overtaken Nvidia's sales inside China. Four announcements, three continents, one message: the AI buildout is still accelerating, and the shortages are getting worse before they get better.
Start with the number that matters most. TSMC, the Taiwanese foundry without which the AI revolution does not happen, reported third-quarter revenue of NT$1.49 trillion, up roughly 51 percent from a year earlier and comfortably ahead of the NT$1.46 trillion analysts had expected. September alone brought in NT$511.86 billion, up 54.6 percent year over year, keeping the company within a hair of August's all-time monthly high of NT$514.8 billion. This is not a company riding a wave. This is a company that is the wave: TSMC controls roughly 72.5 percent of the global foundry market, its 5-nanometer, 4-nanometer and 3-nanometer lines are fully booked, advanced capacity is heavily booked into 2028, and early 2-nanometer wafers are priced around $30,000 apiece.
The customer list explains the numbers. Nvidia and Apple remain TSMC's two biggest customers, with neither showing any sign of pulling back orders. Nvidia is reported to have taken about 60 percent of TSMC's CoWoS advanced-packaging capacity and to have overtaken Apple as the single largest customer, though that claim is unverified elsewhere. Apple, meanwhile, is reported to have secured roughly half of early 2-nanometer output for chips headed into the iPhone 18 generation. When the two most valuable companies in the world are fighting over your factory slots two years in advance, you do not have a demand problem. You have everyone else's supply problem.
One in five AI packages the world wants next year cannot be built for lack of packaging capacity. That is the whole story in a single shortage.
But the market's reaction was a lesson in expectations. Samsung's preliminary numbers missed the lofty consensus, KRW 108.7 trillion in profit and 199 trillion in revenue, and the KOSPI fell 2.04 percent on Thursday. In the AI hardware trade, even a record is a disappointment if it is not a bigger record. The deeper signal is the pricing power: when memory makers can push through double-digit quarterly price increases and still sell every wafer, the shortage is structural, not cyclical. The buyers cannot wait, because their own data-center buildouts cannot wait.
The Malta facility is expected to become the first U.S.-based source of silicon interposers for advanced packaging, which is why GFS shares rose nearly 6 percent in morning trade even as the broader market sold off. Counterpoint Research estimated last month that the supply-demand gap for TSMC's CoWoS packaging capacity could remain around 20 percent in 2026 before narrowing to roughly 10 percent in 2027. In plain language: one in five AI packages the world wants next year cannot be built for lack of packaging capacity. Moving some of that capacity onto American soil is industrial policy made silicon, a hedge against the concentration of the entire AI supply chain on one island in the Taiwan Strait.
Huawei also unveiled the Atlas 950 SuperPoD on its new Peerium architecture and is pushing LogicFolding, a DUV-based workaround targeting 1.4-nanometer-equivalent transistor density by 2031. The technical details matter less than the strategy they reveal: barred from the extreme-ultraviolet lithography machines that make the world's best chips, China's champion is trying to innovate around the blockade with clever engineering on older tools. Whether LogicFolding can deliver 1.4-nanometer-equivalent density by 2031 is an open question. That Huawei feels compelled to try answers a different question: how seriously Beijing takes technological self-sufficiency.
There is a second ledger beneath the first: who pays for all of this. TSMC's answer comes October 15, when the market will parse its capital-expenditure guidance for 2027 the way it once parsed oil inventories. Every leading-edge fab costs $20 billion or more, and the company is being asked to build them while its customers, Nvidia, Apple, the hyperscalers, compete to reserve capacity years in advance. Samsung's ninefold profit jump shows the memory side of the same bet: when DRAM prices rise 10 to 15 percent in a quarter, the windfall funds the next generation of fabs, which is exactly how a structural shortage finances its own eventual cure, slowly. The risk is timing. If AI demand pauses before the 2028 capacity arrives, today's pricing power becomes tomorrow's glut, the classic semiconductor cycle wearing a new costume. Nobody in Thursday's announcements behaved as if a pause were coming. Nvidia is reported to have taken 60 percent of TSMC's CoWoS capacity. Apple is reported to hold half the early 2-nanometer output. Huawei cannot make Ascend chips fast enough for its own home market. The ledger says the only safe bet in semiconductors right now is that the customer with the deepest pockets eats first.
Put the four stories together and the picture is stark. Demand for AI compute is outrunning every link in the chain: the leading-edge wafers, the memory, the packaging, the lithography. South Korea's memory makers cannot build fast enough. Taiwan's foundries are booked into 2028. America's onshoring will not produce its first interposers until 2028. And China, cut off from the top of the stack, is building a parallel stack of its own. TSMC will give fuller third-quarter detail and its fourth-quarter outlook at an investor call on October 15. Expect the only number anyone cares about to be capital expenditure: how fast the industry's most important company thinks it can pour concrete.
From the Western reading, Thursday was vindication. The export-control regime was supposed to slow China's AI advance while the West and its allies built an unassailable lead, and the numbers say the lead is real: TSMC booked into 2028, Samsung printing money on memory, American soil about to host interposer production. The GFS deal is the proof of concept for onshoring, moving a chokepoint technology, advanced packaging, out of exclusive dependence on Taiwan. Washington reads this as deterrence by supply chain.
The Western worry is complacency. Huawei's claim, unverified or not, shows the blockade is leaking at the level that matters: Chinese data centers are learning to live without Nvidia's best chips, and every quarter they do, the export controls buy less leverage. The West's lead is measured in years, not decades, and the 2028 dates on every new fab are a reminder that concrete dries slowly while software iterates fast.
From the Eastern reading, the day belonged to Shenzhen. Even discounted as corporate theater, Eric Xu's claim marks the moment China's AI industry stopped apologizing for the blockade and started marketing around it. The Atlas 950 SuperPoD and LogicFolding are the point: denied the best tools, Chinese engineering is routing around the wall, and the 2031 target for 1.4nm-equivalent density is a declaration that the wall has an expiration date. Beijing reads American onshoring the same way Washington reads Huawei: as confirmation the decoupling is real and permanent.
Moscow and Beijing also note the geography of power. Every critical number in Thursday's ledger, TSMC's bookings, Samsung's memory, the CoWoS gap, flows through East Asia. The United States can sign $2 billion deals in New York, but the silicon still gets etched in Taiwan and Korea. In the Eastern view, the West is buying insurance against a concentration it cannot unwind.
From the Global South, the AI hardware ledger reads as a bill it cannot pay. The entire world is bidding for the same wafers, the same memory, the same packaging slots, and the price of admission to the AI economy keeps rising: $30,000 for an early 2-nanometer wafer, double-digit quarterly memory price hikes, capacity booked into 2028. Countries from Nairobi to Jakarta watch the buildout the way they once watched the oil shocks, as spectators to a market that will decide their digital futures without consulting them.
But the South also sees the opening. Huawei's pitch, an AI stack that does not depend on American silicon, is aimed squarely at the Global South's data-center ambitions. If Ascend chips are good enough and cheap enough, the export-control wall becomes a price umbrella under which a second, cheaper AI ecosystem grows. The South's interest is not in who wins the ledger. It is in whether the ledger's prices ever come down far enough to let everyone play.
An interposer is a thin slice of silicon that sits between a logic chip and its memory stacks inside an AI accelerator package, wiring them together at very high speed. Advanced AI chips cannot function without one. That is why a $2 billion deal to make them in New York is strategic: interposers are a chokepoint, and until now they were made almost entirely in Taiwan.
Because markets trade on expectations, not headlines. Samsung guided to KRW 107.4 trillion in operating profit against a consensus of 108.7 trillion, so the record was already priced in and then some. The KOSPI's 2.04% drop reflects disappointment relative to hopes, not doubt about the memory supercycle, which TrendForce's double-digit price forecasts confirm is very real.
Unproven. Eric Xu's claim was made on Huawei's own stage and is unverified elsewhere, and he admitted Ascend trails Nvidia on raw chip power and that supply cannot meet demand. What is verifiable is the direction: U.S. export controls bar Nvidia's best chips from China, Chinese data centers are adapting at scale, and Huawei unveiled new systems like the Atlas 950 SuperPoD to accelerate that shift.
Not soon. TSMC's advanced capacity is heavily booked into 2028, memory makers' 2027 DRAM and HBM output is already sold, Counterpoint sees a 20% CoWoS packaging gap in 2026, and SK's chairman talks of tightness through 2030. New fabs and the Malta interposer plant arrive in 2028. Until then, expect rising prices and allocation battles for every link in the chain.
