Samsung and TSMC Deliver Outstanding Earnings but Fail to Ignite AI Frenzy! US Treasury Yields Reach 24-Year High, AI Bull Market Faces a "Computing Power Cash Return Test"

According to Zhitong Finance APP, while Samsung and TSMC posted strong results, the flow of capital is revealing a pricing shift amidst the AI investment frenzy: investors are beginning to demand that data center operators actively prove that their ongoing AI compute investments can generate enough cash returns to outweigh rising capital costs. On Thursday, Samsung announced its preliminary Q3 operating profit of 107.4 trillion KRW, an increase of 782.5% year-on-year, which is 8.825 times the same period last year, while revenue rose 126.59%. TSMC's Q3 revenue grew about 51%, both confirming that AI infrastructure procurement remains robust. However, against the backdrop of 10-year and longer-term US Treasury yields climbing to the highest in 24 years, persistently high energy costs, and the continued expansion of financing needs for AI tech firms such as cloud computing companies, record-breaking performances are still not enough to dispel market doubts about the sustainability of the AI compute boom.

On October 8th, Samsung closed down 2.42% at 262,000 KRW, SK Hynix fell 2.44%, and the KOSPI dropped 2.62% to 6,625.93 points, marking the second consecutive trading day with a decline of more than 2%. That day, foreign institutions and domestic organizations had net sales of about 2.42 trillion and 2.09 trillion KRW respectively, while individual investors had net purchases of about 3.74 trillion KRW. This shows the market is experiencing a divergence where institutions are reducing holdings and retail investors are absorbing stocks; options expiration, semiconductor ETF rebalancing, and position trimming ahead of Korean holidays also amplified the pressure from interest rates and oil prices.

The core issue behind Samsung's weak share price lies in the gap between record profits and the sustained profitability demanded by investors. Bloomberg's analyst survey shows profits were slightly below Bloomberg's expected 108.7 trillion KRW; LSEG expectations cited by Reuters were about 106.1 trillion KRW, and the actual profit was slightly above this level. Therefore, the explanation closer to market performance is: capital is re-evaluating how long memory price increases can last, how much HBM share gain can contribute incrementally, and whether high returns can still be maintained if memory chip expansion continues amid weak monetization from cloud AI applications. At the same time, the appreciation of the KRW has also significantly reduced profits when USD income is converted to KRW.

The demand within the AI compute industry still has clear technological support. A single agent request may trigger multi-step reasoning, retrieval, and tool invocation, increasing the need for computation, memory capacity, and data transfer; AMD announced that each MI455X GPU is equipped with 432GB HBM4, and a Helios rack with 72 GPUs has about 31TB HBM4, demonstrating the tangible demand for high-bandwidth memory in compute expansion. Counterpoint raised its Q3 DRAM price hike forecast to 10%-20% quarter-on-quarter, and Citibank foresees 12-layer HBM4 prices rising by 100%-150% next year, indicating memory manufacturers still have strong profit elasticity.

Undoubtedly, leading chip designers and manufacturers, including Samsung, SK Hynix, Nvidia, AMD, TSMC, and Intel, are cashing in on profits driven by the short supply of core AI compute hardware, but whether these profits can be sustained ultimately hinges on whether downstream cloud customers and broad data center operators can turn expensive compute into sustainable AI monetization data.

Samsung and TSMC fail to ignite investor enthusiasm; record results seem to lose sensational effect

Samsung Electronics' quarterly operating profit was nearly nine times last year’s figure, and TSMC's Q3 revenue surged 51%, reflecting the hefty profit returns from ongoing AI compute infrastructure spending.

However, investors responded tepidly to these back-to-back record results, mainly due to growing market concerns about whether the investment boom, increasingly dependent on debt, can last. On Thursday, Samsung's share price fell 2.4% in Seoul. In Tokyo, supplier stocks including Tokyo Electron, Advantest, and Ibiden also declined.

Investors are questioning how long AI spending can last, especially amid rising global borrowing costs and pressure on companies’ hefty capital expenditure plans.

“Can we see the kind of extremely strong and sustained returns currently promised by data center operators, including cloud giants?” said Jun Bei Liu, co-founder and chief investment portfolio manager at Ten Cap Investment Management.

Samsung, the world's largest memory chip producer, announced record preliminary operating profit of 107.4 trillion KRW (about $80.1 billion) for the quarter ended September, with revenue more than doubling. However, profit still fell slightly short of the analyst average forecast.

Josh Gilbert, chief analyst for Asia-Pacific at eToro, said: “Samsung delivered record profits but still missed expectations, which shows how demanding expectations have become for AI investment trades.”

Tech companies are racing to secure the memory supply required for running AI services. High Bandwidth Memory (HBM) components—key chips paired with Nvidia and other companies’ AI accelerators—are attracting massive investments, corporate capital expenditure, and a rapid shift and expansion of memory capacity. Flash memory linked with NAND and low-power chips are also in strong demand. Outstanding results and high expectations highlight the significant economic benefits that the AI boom is bringing to companies like Samsung, SK Hynix, and Micron.

As shown in the image above, the AI investment frenzy has driven Samsung's profits to new highs. Note: The Q3 2026 data is preliminary operating data announced by Samsung management.

Counterpoint Research raised its Q3 DRAM price hike forecast from 5%–10% previously to 10%–20%, due to customers placing orders ahead of time.

Neil Shah, vice president of research at Counterpoint, said: “Samsung's current valuation is quite low. Next year, they still have a lot of pent-up potential in HBM, mainly because related benefits haven't fully materialized yet.”

Citibank analysts Peter Lee and Jayden Oh expressed similar optimism, expecting 12-layer HBM4 prices to surge 100%–150% next year, significantly boosting Samsung's average selling price and profit margin. However, the firm has cut Samsung’s full-year profit forecast due to the stronger KRW.

This boom has transformed Samsung’s semiconductor business. Just a few years ago, the segment was losing money due to a severe demand slump. Now, AI infrastructure construction is consuming large volumes of advanced DRAM, and Samsung is closing the gap with SK Hynix in the HBM sector. The impressive performance of this cycle is also reflected in Korea’s trade data: in September, semiconductor exports were over three times higher than the same period last year.

Wall Street analysts expect, on average, that Samsung’s chip division will deliver operating profit of 110 trillion KRW, while the electronics division will incur a loss. The Suwon-headquartered company will announce its full financial report on October 29.

Nevertheless, market caution remains. Even after announcing record results in July, Samsung’s share price is still about 25% lower than its June peak. In an industry known for cycles of boom and bust, investors remain somewhat skeptical about profit surges. Samsung and SK Hynix are currently expanding capacity and increasingly supporting AI sector enterprises to drive demand for their products.

As shown in the image above, after a recent rebound, Samsung’s share price remains more than 20% below its all-time high.

A highlight of the current market is that AMD CEO Lisa Su, during her visit to Seoul this week, emphasized AMD’s long-term strategic partnerships with Samsung and SK Hynix. Su confirmed that AMD has started volume delivery of the next-generation Instinct MI455X and Helios systems using HBM4.

Su told reporters: “Memory is absolutely critical for the core computational processes of overall AI data centers; supply has remained very tight.” She added in the interview that AMD’s HBM cooperation with Samsung and SK Hynix will last for several product generations. “We are encouraging partners to build as quickly as possible.”

From energy shocks to the market's urgent pursuit of cash returns on compute investment

The US-Iran situation continues to affect asset prices through shipping, fuel, and financing costs. Iran recently reiterated that it will not resume normal navigation in the Strait of Hormuz until certain conditions are met; from September 28 to October 5, there were at least 12 attacks, attempted attacks, or harassment incidents against oil and gas tankers in the area—the highest weekly number since the war broke out. Persistent shipping risk makes it difficult for the market to believe that energy supplies can reliably recover.

As of 18:41 (Beijing time) on October 8, Brent crude futures were at $105.20 per barrel, up 4.99% on the day, and WTI at $92.75, up 5.06%. Compared with the last trading day before the outbreak of war on February 27—when Brent and WTI settled at $72.48 and $67.02—these are increases of about 45.14% and 38.39%, respectively. Hurricane-related production halts in the US Gulf of Mexico also contributed to the day's gains, putting energy markets under the dual pressure of geopolitical and weather disruptions.

Energy inflation pressure is even more pronounced in refined oil. The International Energy Agency points out that close to 3 million barrels/day of Middle Eastern refining capacity is offline due to attacks and export delays; US diesel prices have almost doubled compared to pre-war levels, and Europe and Asia are close to the same increase. Rising diesel prices directly push up costs for transportation, engineering, and industrial production, while disruptions in natural gas supply add to power cost pressures in some regions. The IEA has decided to speed up the previous commitment to release reserves, prioritizing diesel, but the release mainly fills short-term gaps—recovery of shipping and refining capacity remains the fundamental determinant of whether supply can sustainably improve.

Long-term bond pricing contains the market's average expectation of short-term rates for years to come, as well as the term premium investors demand for holding long bonds. Lower odds of a rate hike at the next meeting do not mean conditions for long-term borrowing will simultaneously relax. Energy shocks make it harder for inflation to ease, fiscal deficits and expanding bond supply put pressure on financing, and AI companies issuing debt are competing with governments for long-term capital. The UK also faces a credibility test for its pre-budget fiscal policy, while Japan is dealing with both fiscal expansion concerns and expectations for policy normalization by its central bank.

In the global sovereign debt market, the US 10-year Treasury yield—known as the "anchor of global asset pricing"—jumped to ~5.36% intraday on October 7, its highest since 2002; the 30-year yield rose to ~5.73%, remaining at its highest levels in about 24 years. The UK 30-year bond yield also hit a record high since January 1998, hovering near 6.05%.

This climb in US Treasury yields has a direct characteristic for AI valuations: real rates and term premiums are rising, pushing up bond yields of AI tech companies—who, like the Treasury, compete for long-duration pools of capital—and continually driving up their financing costs. According to Reuters analysts, citing the New York Fed model, the 10-year US Treasury term premium is at its highest since 2014, and recent nominal yield increases have mostly come with rising real yields, not higher long-term inflation expectations; Treasury data shows the real yield on 10-year TIPS reached 2.92% on October 7. All of this means that capital providers' required returns in the credit market are rising, and companies’ future profits must meet higher valuation bars.

Persistently rising long-term yields are becoming a powerful catalyst for pricking parts of the AI asset bubble, with a clear sequence of transmission. First, the discount rate rises, dragging down the present value of future cash flow and putting immediate pressure on tech stock valuations. Next, new financing costs, determined by bond benchmarks and credit spreads, increase, so data centers need higher utilization, more stable client payments, and faster cash recovery to cover construction, chip procurement, operations, and debt costs. If these conditions are not met, project delays and cutbacks in capex gradually cascade to GPU, HBM, advanced packaging and equipment orders. The profit Samsung realizes today comes from past procurement; the stock market tends to pre-price how quickly the next round of purchasing might continue.

The economically meaningful “AI kill line” is when new compute projects’ expected returns stay below their capital cost, which further erodes financing and expansion ability. This also explains why Panmure Liberum's senior strategist Joachim Klement links high financing costs to AI investment sustainability and pessimistically suggests an AI bubble burst scenario in 2027–2028. For hot AI tech stocks, the next stage of differentiation will focus even more on: who can turn token growth into revenue growth, who can raise per-unit compute profits via software/hardware efficiency, and who can back expansion with continuing cash returns. Record profits from AI compute leaders are still important, but the sustainability of capital returns is increasingly dictating how much the market is willing to pay for those profits.