"This cycle is different from previous ones."

This has become a dominant market sentiment regarding the global computing power supply chain amid the current AI wave.

That’s partly true, and partly not. The long-term trajectory of AI as a transformative productivity driver is certain, but this does not alter the fundamental laws of cycles—what has changed is that the systemic parameters and pressure (or amplitude) of this super-cycle may be far greater than in the past.

South Korea and SK Hynix, positioned at the core of this AI surge, likely share this view.

On June 23, South Korea’s stock market underwent severe correction, with indices dropping nearly 10% in a single day, triggering circuit breakers again. Shares of Samsung Electronics and SK Hynix both fell over 12%, marking their largest single-day losses in 17 years. Subsequently, the pullback in Korea rippled across multiple global AI-themed markets and companies.

The immediate catalyst for this adjustment was a report from Korean media stating that SK Hynix decided to slow down its transition to HBM4 production lines, redirecting more resources toward standard DRAM markets. The rationale cited was the downward revision of expected output for NVIDIA’s next-generation Rubin platform, which uses HBM4, leading the company to conclude that accelerating HBM capacity conversion was no longer necessary.

Meanwhile, the head of South Korea’s Financial Services Commission publicly expressed regret over approving a leveraged single-stock ETF linked to Samsung Electronics and SK Hynix in late May, as the product’s scale surged from around $3 billion to over $9 billion within just one month, with 92% of holders being retail investors—an amplified volatility effect now clearly evident.

This starkly contrasts with earlier optimistic narratives surrounding the AI super-cycle.

For over a year, driven by AI capital expenditure, HBM-related themes have been intensifying, fueling explosive growth in leveraged financial products in the Korean market, rapidly inflating stock prices and valuations. The wealth effect displayed has even spilled over into Korean society and global markets. Yet SK Hynix, the world’s leading supplier in HBM market share, is now reconfiguring its capacity—directly prompting market reassessment of demand sustainability. The regulatory warning from Seoul appears more like a preemptive signal of future risks.

So why, amid widespread belief in an ongoing AI super-cycle, would Hynix proactively adjust its capacity allocation between HBM and DRAM?

In my view, this move is not a rejection of AI’s long-term trend. Nevertheless, it serves as a reminder to remain vigilant about the extraordinary risks embedded within super-cycles. Facing escalated cycle parameters—higher capital expenditure, faster technological iteration, and massive depreciation costs—rebalancing capacity and product mix may represent a more rational commercial strategy.

This partially validates our prior analysis on SK Hynix’s structural contradictions and Longxin Technology’s catch-up trajectory: the fundamental cyclical patterns in memory markets haven’t changed; what has changed is the extreme amplitude.

Currently, HBM accounts for roughly 40% of SK Hynix’s business, yet contributes the vast majority of its profits.

However, such high profitability does not equate to superior free cash flow conversion. HBM’s highly customized nature requires strict adaptation to specific chip design specifications from NVIDIA, AMD, and others—far less universal than standard JEDEC-compliant DRAM. Moreover, HBM’s generational upgrade cycle is only 1–2 years, whereas traditional DRAM typically has a lifecycle exceeding five years. This means HBM manufacturers cannot leverage long-term economies of scale to amortize fixed assets as conventional DRAM producers do; instead, they must rely on sustained high ASPs and deep customer lock-in to offset enormous equipment depreciation costs.

To maintain technological leadership and market position, SK Hynix must reinvest at unprecedented speed while simultaneously rebuilding an entire supply chain tailored to HBM. Equipment suppliers along this chain must also follow the same high-intensity rules. Consequently, the operational rhythm, intensity, and management complexity of this supply chain are immense—only high margins can sustain such high investment; otherwise, the cycle cannot be maintained.

Currently, SK Hynix’s HBM supply chain is not operating as smoothly as anticipated. Recently, several tier-one semiconductor equipment suppliers have submitted price increase requests of 3%–4%, citing tight delivery capacity for key equipment. Notably, SK Hynix unusually granted these requests.

These supplier price hikes are unlikely due to opportunistic “price gouging,” but rather reflect rising business risk. Equipment vendors often finance production upfront and recognize revenue later. The higher capital intensity and operational stress inherent in HBM supply chains place greater forward risk on upstream suppliers. Thus, SK Hynix’s approval of price increases is essentially a measure to stabilize the supply chain.

Yet, when dealing with NVIDIA, SK Hynix may lack negotiating power—primarily because NVIDIA is nearly Hynix’s sole customer, while NVIDIA actively supports Samsung and Micron. As a result, SK Hynix will likely face escalating cost pressures going forward, including upward pressure from suppliers and its own need for continuous fixed asset reinvestment.

If a cyclical inflection point arrives, SK Hynix will struggle to pass on pressure, leaving it highly vulnerable.

In fact, NVIDIA could pass upstream supply chain costs downstream to AI model developers—but the issue is that, aside from Anthropic achieving quarterly profitability, all major model developers are currently operating at a loss. Furthermore, all are raising prices. The underlying pressure stems from the risk that if end consumers refuse to pay, the commercialization logic of AI collapses, rendering the current large-model narrative unsustainable. This, in turn, would trigger a sharp contraction in demand for AI hardware infrastructure, completely bursting the temporary “bubble.”

Therefore, NVIDIA has no incentive to transfer infrastructure-level pressures to the application layer.

Hence, SK Hynix must adopt a more cautious approach to recalibrate the trade-off between today’s high HBM profits and tomorrow’s high costs. Adjusting its business structure—shifting toward general-purpose DRAM—is thus a rational strategy to mitigate HBM’s high-risk exposure and capture higher marginal returns from the current shortage in standard DRAM.

Further, any adjustments SK Hynix makes in HBM will inevitably influence Samsung and Micron’s stance on HBM. If the reason is indeed NVIDIA’s downward revision of Rubin platform output expectations, Samsung and Micron may also reconsider prioritizing DRAM, potentially reshaping the future supply-demand balance in the memory industry.

Since HBM capacity diverts from DRAM supply, Longxin Technology has capitalized on this DRAM scarcity through massive expansion, capturing the largest windfall of this cycle. Its profit structure has reached parity with Samsung, SK Hynix, and Micron, while its market share has risen from 4% in H1 2025 to approximately 10% today—though not due to technological moat premiums.

Among Longxin’s three factories, the Xin Qiao and Ji Dian plants are massive in asset size but have long retained at least half of their capacity dedicated to R&D and process iterations. This creates an inherent tension between commercialization and R&D during the catch-up phase. In the current upcycle, Longxin has proactively increased commercial capacity to capture market share and generate cash flow, while slowing overall fixed asset investment to avoid reckless expansion before technology and yield gaps are closed.

Nevertheless, as a latecomer, Longxin must still reserve significant capacity for R&D catch-up. Its implicit cyclical pressure may be even more pronounced. Once the three giants reallocate resources back to DRAM, the price decline channel will accelerate, and Longxin’s profit erosion pressure will exceed that faced by the established players with structural technological advantages.

Therefore, Longxin must treat this super-cycle with greater caution. If it can preserve its current market share through the downturn, its commercial viability and strategic path will be validated. Consequently, its strategic choices must continuously involve difficult trade-offs between commercialization and R&D—and build buffer capacity during the upcycle to prepare for the inevitable downcycle.

This mirrors SK Hynix’s rebalancing between HBM and DRAM: both responses to evolving cycle parameters. Even within the narrative of a super-cycle, leaders and followers alike must confront the heightened risk exposure brought by upgraded cycle dynamics. When the industry returns to normalcy, what will ultimately determine long-term positioning are technological iteration capability, capacity allocation efficiency, and the ability to defend market share and cash flow during downturns.

In essence, the rapid pace driven by AI has had profound impacts—dramatically compressing the semiconductor industry’s original depreciation models and economies of scale, placing companies across the supply chain in a state of hyper-profit euphoria and chronic pressure.

This easily fosters a sense of “building a new world/order.” But the most profitable player in this round—SK Hynix—has demonstrated through action that it’s time to slow down and regain clarity.

Source: LatePost