Since July, as the "semiconductor peak" argument has resurfaced, semiconductor stocks such as Samsung Electronics and SK hynix have continued to plunge. What the market worries about most is how long hyperscalers' artificial intelligence (AI) capital expenditures (CAPEX), which support memory demand, can continue.

In particular, as major big tech companies have recently begun amassing investment firepower by tapping bond issuance beyond their own cash, some in the market warn that the AI investment cycle could end earlier than expected. Still, brokerages say the hyperscalers' financial structures remain solid, putting more weight on the possibility that the current aggressive investment stance will continue for at least two to four years.

SK hynix HBM plays in Times Square (Seoul=Yonhap News) On the 10th (local time), when SK hynix began Nasdaq ADR trading, a video of SK hynix's flagship product HBM plays on the anamorphic billboard in New York's Times Square./Courtesy of SK hynix.

◇ Structural expansion of AI investment… "CAPEX to increase through 2027"

As U.S. big tech corporations wage an all-out AI war, AI CAPEX is surging. According to KB Securities, CAPEX by the five hyperscalers (Alphabet, Microsoft, Amazon, Meta, Oracle) in 2026 is $725.1 billion (about 1,081 trillion won), a 4.9-fold increase from 2023 ($148.1 billion).

The CAPEX expansion stems from rising inference demand driven by agentic AI and higher expense due to memory bottlenecks. For example, Microsoft's graphics processing unit (GPU) expense accounts for 60%–70% of CAPEX, and memory expense makes up 13% of CAPEX. For now, the memory bottleneck is spreading to power infrastructure, liquid cooling systems and networking equipment, pushing up the cost per unit.

KB Securities expects the AI investment expansion to continue through 2027. Kim Se-hwan, an analyst at KB Securities, said, "Hyperscalers' CAPEX-to-revenue ratio will peak at about 38% in 2027 and then ease to around 34% in 2028," adding, "Because AI infrastructure is invested 6–24 months ahead of revenue generation, CAPEX pressure will gradually normalize as earlier investments start converting into revenue in earnest."

◇ Bond issuance has increased, but… "at least 2–4 years of funding capacity"

The issue is that hyperscalers' funding is shifting from equity-like sources such as cash to liability-based funding like corporate bond issuance. As debt-funded financing grows, interest burdens rise, and if the financial structure deteriorates, credit risk increases. According to Bloomberg and Samsung Securities, as of July this year, hyperscalers' corporate bond issuance stands at $194.1 billion. It has been climbing rapidly, following $20.1 billion in 2024 and $108.4 billion in 2025.

In particular, as Oracle's corporate bond issuance has risen quickly in recent months, its credit default swap (CDS) premium has jumped, prompting concern that hyperscalers' capital capacity may be depleting rapidly. Oracle's corporate bond issuance amounts to $25 billion as of July this year, and its "total debt/EBITDA ratio," used as a U.S. credit assessment factor, has increased quickly to 5.6 times.

Brokerages judge that credit concerns for big tech, excluding Oracle, are premature. Kwon Beom-seok, an analyst at Samsung Securities, said, "Oracle's total debt/EBITDA ratio is 5.6 times, higher than peers, but as of March 2026 the levels are Microsoft (0.6 times), Alphabet (0.6 times), Meta (0.8 times) and Amazon (1.3 times), so funding conditions for hyperscalers excluding Oracle are stable." The total debt/EBITDA ratio is a representative metric used by credit rating agencies to evaluate the financial soundness of corporations.

They also analyzed that it could take at least two years to reach the threshold for corporate bond issuance. Kwon said, "It is expected to take at least two years before hyperscalers' credit risk materializes," adding, "Given that hyperscalers' AI-related revenue is maintaining growth in the 30% range, CAPEX pressure could gradually ease alongside profitability improvement."

Specifically, the remaining capacity is estimated at 4.0 years for Microsoft, 4.9 years for Alphabet, 3.9 years for Meta, 3.1 years for Amazon and 1.9 years for Oracle. This estimate is based on each company's cash generation capacity (EBITDA), current borrowings and annual corporate bond issuance, and reflects the time it would take for liabilities to grow to a level that burdens credit ratings if bonds continue to be issued at the current pace.

However, the sharp rise in U.S. Treasury yields is a headwind. As U.S.-Iran tensions have reignited, the 10-year U.S. Treasury yield recently topped 4.6%, nearing its previous high. But after the U.S. consumer price index (CPI) slightly undershot market expectations the day before, easing rate-hike pressure, the probability of a July rate hold by the U.S. Federal Reserve rose to 83.4% that day on CME Group's FedWatch.

※ This article has been translated by AI. Share your feedback here.