A Direct Rebuttal to the Semiconductor Bubble Thesis: Hyperscalers' Shift to Positive Free Cash Flow Holds the Key to a Second-Half Rebound

A framework was proposed for viewing the second-half market direction as simplified into the relative magnitude of growth (growth momentum) versus risk. When growth outweighs risk, the market rises; when risk overwhelms growth, it falls. Risk factors ultimately originate from interest rates, and the assessment was that, unless current rate policy turns more hawkish, risk has effectively already reached its peak level heading into year-end.

On the growth side, it was pointed out that, unlike the cyclical cycles of consumption, production, and investment in the past, a specific factor — AI capex — is now driving growth. This is supported by the fact that the US Conference Board Leading Economic Index has remained in negative territory continuously since 2022. Restoring confidence in whether this growth is sustainable, and confirming that Fed and Bank of Korea monetary policy is not more hawkish than expected, were presented as the key variables for the second half, with big tech earnings releases in late July-early August, the FOMC, and the Bank of Korea's August Monetary Policy Board meeting cited as key checkpoints.

If risk containment and earnings improvement are confirmed simultaneously, the possibility was raised of a rotational trade in which attention partly diversifies into dividend and low-volatility names such as financials, rather than a repeat of the semiconductor concentration seen in the previous rally. Still, from the view that growth will ultimately win out over risk, the opinion was emphasized that attention to semiconductors and hardware should be maintained. The outlook was also offered that August as a whole would see a gentle rise accompanied by a recovery of prior losses.

Three reasons were offered for why investors distrust semiconductor earnings even though the figures are clearly confirmed. First, the burden stemming from the fact that both earnings and share prices have risen at an unprecedented pace. Second, a sense of déjà vu from the past IT bubble. Third, the fact that hyperscalers are continuing data center investment by relying on external borrowing rather than their own cash flow, which is feeding concerns about a longer-term pullback in investment.

A counterargument was also raised that equating the current situation with the early-2000s IT bubble is difficult. At the time, Cisco surpassing Microsoft's market capitalization is cited as the symbolic case of the bubble, but Cisco's earnings then were only about 20% of Microsoft's. By contrast, the recent case of SK Hynix surpassing Samsung Electronics' market capitalization is hard to classify as a bubble, since the earnings valuation itself (a forward P/E in the low 4-5x range) is low.

The 2026 consensus operating profit for the semiconductor sector is about KRW 970 trillion, but given this year's export growth trend and exchange-rate conditions, the actual figure could approach KRW 1,000 trillion, with an optimistic outlook that it could grow to KRW 1,300 trillion next year. This is said to represent a high growth rate even relative to the average since the 2010s. The fact that US companies' 12-month forward capex projections are rising more steeply than trailing figures was also cited as evidence suggesting a continuation of planned investment rather than a pullback.

Ultimately, the market is said to be home to both an optimistic view that 'semiconductors are earning well' and a pessimistic view that '2027-28 earnings are already priced in,' and it was diagnosed as normal for these two views to keep clashing. Still, while it remains necessary to keep monitoring for the point at which investment momentum eventually turns, the conclusion offered was that at the current juncture, prices have come down sufficiently relative to earnings, and there is room for earnings improvement to continue at least through the rest of this year and into early next year.

The diagnosis was that the US 10-year Treasury yield is holding at around 4.5-4.6% and Korea's market interest rates are also moving at elevated levels. In addition to uncertainty over whether the Fed will cut rates further, the war concerns that have persisted since February-March were cited as a factor keeping rates pinned to the upside. As oil prices have risen from the $60 range to the $100 range, their impact on inflation has grown, and despite recent signs of stabilization, uncertainty remains, keeping rates from easily coming down.

In Korea's case, even though the policy rate has already been raised, market rates are pricing in a much larger spread than during the 2017-18 hiking cycle, and the bond market has already priced in the possibility of roughly three additional hikes, the diagnosis noted. The year-end and next-June policy rate outlook implied in US futures markets also points to the 4% range, higher than current levels, but the assessment was that since the probability of a further rise has recently stalled, this has already been largely reflected in market rates. The outlook presented for the second half was that rather than rising further from peak levels, rates would likely decline gradually as time passes.

Hyundai Motor Securities Center Head Roh Geun-chang appeared to share his diagnosis of the recent sharp semiconductor selloff. He noted that while a similar correction was experienced in 2017-18, this is the first time a sharp drop has occurred alongside such strong earnings, and assessed last month as a phase that will long be remembered by capital market participants. He mentioned that Samsung Electronics fell to a level close to the KRW 183,000 range cited as the next-year-earnings-based target price.

He diagnosed that while the semiconductor industry tends to be analyzed from a supplier-centric perspective, the recent decline was triggered by concerns on the demand side — namely, hyperscalers. He explained that the background to Amazon and Google's free cash flow turning negative lies in cheap long-term GPU contracts signed in 2024, and that as these contracts come up for renewal starting in 2027, prices could be reset.

In particular, he presented figures showing that Amazon's in-house chip (Trainium, Graviton) revenue is expected to reach $25 billion this year, about 13% of cloud revenue, while Google's TPU revenue is $20 billion, about 18% of cloud revenue. He stated that by supplying their own chips to OpenAI, Anthropic, and others instead of Nvidia's, hyperscalers are offsetting contract price increases, and that if this trend continues, hyperscalers' free cash flow could turn positive around 2027-28. He also cited as supporting evidence that Anthropic has already turned profitable with an operating margin in the 5% range, and that OpenAI's active users have surpassed 1 billion.

He pointed to the recent widening of CDS (credit default swap) spreads for major companies such as CoreWeave, Oracle, and Nvidia, and the fact that a specific leveraged product in Korea amplified the semiconductor valuation debate, as the backdrop to this selloff. However, he forecast that since hyperscalers' data center investment is already booked through 2028, the capex cycle is more likely to continue gradually than to turn sharply downward.

Center Head Roh addressed concerns over China's CXMT (ChangXin Memory Technologies), citing changes in ASML's revenue mix by region as evidence. He noted that China's share, which was 41% in 2024, fell to 33% last year and to 14% this year, while Korea and Taiwan rank first and second at 43% and 30% respectively, and that this structure is likely to persist for the next five years. He explained that CXMT's current monthly wafer capacity of 270,000 units is significantly below Samsung Electronics (710,000), SK Hynix (610,000), and Micron (400,000).

He diagnosed that due to US export restrictions, procurement of key process equipment — Lam Research's etch equipment, Applied Materials' deposition equipment, and ASML's DUV equipment — is constrained, making it difficult to actually realize capacity expansion plans even if they exist. In particular, he pointed out that active investment by existing players — TSMC ($65 billion in annual capex, rising to $180 billion next year) and Korea's three memory makers (a combined roughly $100 billion) — means equipment supply itself is in short supply.

The case of Apple reportedly reviewing CXMT NAND, which drew pushback from the US House and Senate and required Apple to submit a plan by August 21, was also introduced. The outlook is that Apple would find it difficult to easily adopt this given the conflict with its investment in Micron's US plants. It was also noted that CXMT's HBM-related logic dies are foundried by SMIC and used in Huawei's Ascend accelerators, meaning there is a risk of being placed on the US Department of Defense blacklist if used in autonomous driving or military equipment, with the possibility of expanded sanctions. Still, since data-center-bound memory demand accounts for over 60% of the total and more than 60% of data centers are set to be built in the US by 2030, the conclusion was that while Chinese memory could be a threat in the consumer electronics market, its impact on current share prices is limited.

Center Head Roh explained changes in the memory industry's long-term agreement (LTA) contract structure. He noted that contracts are typically for five years, with roughly half of the contracted volume actually fulfilled, and that a penalty structure has been strengthened through the introduction of upfront payment clauses that did not previously exist. He explained that server DRAM and SSDs are supplied directly to hyperscalers under three-to-five-year contracts, whereas HBM supply is overwhelmingly channeled through Nvidia, operating in the form of long-term contracts with Nvidia. He diagnosed that because hyperscalers approach this from a total cost of ownership (TCO) reduction perspective spanning compute racks, switching racks, and storage racks rather than scrutinizing HBM cost directly, a more flexible structure than in the past has formed.

On valuation, he cited next-year book value per share of about KRW 1 million for SK Hynix and about KRW 180,000 for Samsung Electronics, implying levels of roughly KRW 1.5 million (1.5x P/B) and roughly KRW 280,000 (1.5x P/B) respectively, while taking a cautious stance, saying it is hard to be confident that domestic semiconductor stocks are in a cycle that warrants such a premium compared to TSMC's 10x P/B. He also offered the interpretation that last Friday's sharp rally was largely short-covering in nature.

On shareholder return policy, he expects SK Hynix to unveil concrete plans once the ADR listing issue is resolved, and forecast that since Samsung Electronics' three-year shareholder return policy ends this year, a new policy is likely to be announced alongside the Q4 earnings release in January next year. He also offered the view that, unlike the chicken game of the past, there is currently no clear target company to push out of competition, so a price-taking strategy through competition on new-product specs — low power, high performance — rather than price competition, is likely to remain effective. Finally, he offered the forecast that AI-driven cloud and LLM revenue will surpass capex around Q4 2027, advising that attention should focus on earnings and direction rather than current noise.

The structure of the subprime mortgage crisis, the subject matter of the film The Big Short, was explained. Originally, banks were supposed to lend only to prime borrowers with high creditworthiness and reliable repayment ability to avoid defaults, but once lending to the highest-credit borrower pool became saturated, banks expanded lending to the next tier down — subprime borrowers. In this process, a method emerged of stacking prime and subprime bonds in layers and then slicing them vertically to bundle them into a single product; the core structure presented was that this allowed the entire product to be rated highly on the basis of the top-tier prime bonds' creditworthiness, effectively letting subprime debt circulate as if it were a prime product. The diagnosis given was that once defaults occurred in the upper-tier bonds, the lower-tier bonds were dragged down together without question, and this chain-reaction structure ultimately magnified the crisis.

It was pointed out that the entertainment value the film offers ultimately lies in the thrill of 'I made money that others couldn't.' The narrative of a small number of figures, including Michael Burry, who predicted the crisis and bet on it to make money, leaves a strong impression on viewers, but the message the film actually intends to convey is said to differ from that. The interpretation offered was that, through the director's closing lines, the film sought to raise a question about the very fact that a few people made money while the collapse of the financial system cost others their jobs and homes.

However, a concern was raised that many viewers consume the film not for this ethical message but in the direction of 'let me make money by betting against the market like Michael Burry did.' It was pointed out that investment habits such as betting on inverse products when stock prices rise may have spread following the film's box-office success. It was also noted that the handful of successful cases shown in the film are merely survivorship bias, and that failed short sellers never appear in the film in the first place.