Beyond Series · Part 1

BEAR AND FEAR
 

Deconstructing the Combined Bear Case on Korean Memory: Causality, Pricing Power, and the 2027-2028 LTA Watch Point

Editor's Note

Gorilla PE has been tracking the LLM and AI infrastructure transition from its early stages, both as an investor in U.S. private deep-tech assets and as a house focused on structural bottlenecks in the AI stack. In recent months, many investors have asked us a seemingly simple question: what do you think about Samsung Electronics and SK hynix?

This essay is not a buy-or-sell recommendation on any public security. It is our attempt to answer the larger structural question behind that inquiry: whether the current memory cycle is merely another cyclical peak, or whether AI infrastructure is shifting the causal structure of memory demand, customer lock-in, and supplier pricing power.

Abstract

Gorilla PE defines the Korean memory bear case that emerged in spring 2026 as a combined proposition made of two separate hypotheses.

The first hypothesis is that U.S. Big Tech AI capex can be cut abruptly and without warning. The second is that the cautious posture of Korea's two leading memory suppliers reflects trauma from prior hyperscaler order slowdowns. This essay separates the two hypotheses and tests each causal chain against primary-source data.
Our conclusion is fourfold.

First, Hypothesis A is not supported by the four historical examples most often cited by bears: Google in 2008, Meta and Amazon in 2022-2023, and Microsoft in 2025. In all four cases, the capex adjustment had clear leading causes: an exogenous macro shock, a failed internal thesis, a CEO regime change with pandemic overhang correction, or customer demand reallocation. In each case, Big Tech capex returned to record levels within one to two years.

Second, Hypothesis B contains a partial truth but misallocates causality. During the 2022-2023 memory downturn, hyperscaler order volatility did matter. But it was not the dominant driver. The larger shock came from the collapse in PC and mobile consumer demand, which accounted for an estimated 50-60% of the downturn, while hyperscaler spot-order volatility represented only a 15-25% layer.

Third, the real causal driver behind Korean suppliers' caution is not fear of adding capacity. It is the protection of pricing power. The caution is not about whether to expand production capacity. It is about whether to accept customer-funded dedicated capacity structures that could weaken bargaining power and reduce the suppliers to quasi-foundries for Big Tech customers.

Fourth, the real watch point is the 2027-2028 LTA renewal window. The key issue is not whether demand suddenly disappears in 2026. The key issue is how long-term agreements, take-or-pay clauses, HBM pricing spreads, and next-generation GPU schedules evolve as the AI infrastructure supply chain matures.
 

1. The Combined Proposition Behind the Market's Bear Case

In December 2025, reports emerged that Samsung Electronics and SK hynix had expressed caution about aggressive capacity expansion during global investment-bank meetings.[1] In March 2026, memory stocks fell roughly 10% in a single trading day.[2] These two events fused into a market narrative: the Korean memory cycle may have reached its peak.

The bear case is powerful because it combines two ideas that each contain an element of truth.

The first is that U.S. Big Tech AI capex can be cut suddenly. Bears usually cite four examples: Google in 2008, Meta and Amazon in 2022-2023, and Microsoft in 2025. The second is that Korea's leading memory suppliers are being conservative because they remember the pain of prior Big Tech order slowdowns, especially during the 2022-2023 memory downturn when all three major memory producers entered quarterly losses.

The combined proposition is strong precisely because it blends fact with inference. Each hypothesis sounds plausible. But the important question is not whether each contains a fact. The important question is whether the causal chain holds.

This essay therefore separates the two hypotheses and tests them one by one.

2. Testing Hypothesis A: Big Tech Capex Cuts Were Not Sudden, Uncaused Events

When the three major examples are examined chronologically, the claim that Big Tech capex can be cut without warning weakens substantially.

Google 2008-2009: an exogenous global financial shock. Google's quarterly capex fell from $842 million in Q1 2008 to $139 million in Q2 2009, a decline of roughly 83% over five quarters.[3] This was the sharpest reduction among the four cases. But the cause was clear: the Lehman Brothers bankruptcy in September 2008 and the global financial crisis that followed. Google's CFO at the time described the reduction as a capex breather. The company had built significant data-center capacity during the 2007 build-out cycle and could temporarily slow new investment while utilizing existing capacity. Recovery began in Q3 2009.[3]

In other words, this was not an unexplained capex cut. It was a capex adjustment following more than a year of macro stress and excess capacity absorption.

Meta 2022-2023: the Year of Efficiency. Meta lowered its 2023 capex guidance from $34-37 billion to $30-33 billion during its February 1, 2023 earnings call.[4] The company's 8-K stated that the reduction reflected lower data center construction spend as Meta shifted to a more cost-efficient data-center architecture capable of supporting both AI and non-AI workloads. The cause was not mysterious. Reality Labs had generated $13.7 billion of losses in 2022, while Apple's App Tracking Transparency policy weakened Meta's advertising business.

The capex cut was therefore the end point of a multi-quarter signal chain: a failed metaverse investment thesis, pressure on advertising revenue, and a management decision to re-architect infrastructure. It was not a sudden decision without prior warning.
 

Amazon 2022-2023: a CEO regime change and pandemic overhang correction. During the same period, Amazon also executed a capex cut. Total company capex fell from $63.6 billion in 2022 to $52.7 billion in 2023, a decline of roughly 17%.[19] But again, this cut was the end point of a multi-quarter accumulation of signals: Andy Jassy's appointment as CEO in July 2021 marked a transition from the Bezos Doctrine of minimizing profit and maximizing reinvestment to the Jassy Doctrine of financial optimization; the post-pandemic correction of fulfillment overcapacity (Amazon's headcount had doubled from 798,000 in 2019 to 1.6 million in 2022); and the largest layoffs in Amazon's history, eliminating roughly 27,000 corporate positions with a $1.8 billion severance charge.[20] AWS capex alone declined by only about 10%, and CFO Brian Olsavsky stated on the Q3 2022 earnings call that “we cut about one-third of our overall Amazon budget while still focusing our capital dollars really on the AWS business. ”Over the following two years, Amazon's total capex went from $52.7 billion to $131.8 billion, a roughly 2.5x expansion" — almost identical to Meta's recovery pattern (2023 $27.3B → 2025 $72.2B, roughly 2.6x). In both cases, the 2023 cut was the end point of pre-existing signals, and both companies returned to record-scale capex within one to two years.

Microsoft 2025: customer demand reallocation, not demand disappearance. On February 21, 2025, TD Cowen channel checks reported that Microsoft had walked away from a few hundred megawatts of U.S. data-center leases.[5] The report included lease cancellations with private data-center operators, Statements of Qualifications that did not convert into formal contracts, and a reallocation of international capex into the United States. A month later, on March 26, 2025, an additional roughly 2GW of capacity abandonment was reported across the U.S. and Europe.[6]

Analysts attributed the move to OpenAI demand shifting toward Oracle following the January 2025 announcement of the $500 billion Stargate Initiative, in which Oracle became a new infrastructure partner. Microsoft did not lose AI infrastructure demand as a category. A major customer reallocated demand from one infrastructure provider to another, leaving Microsoft with surplus capacity in certain locations. In the same period, Microsoft's fiscal-year capex guidance of $80 billion remained intact.[5]

The post-event evidence is even more important. By 2026, combined capex guidance for the four major U.S. Big Tech companies — Google, Microsoft, Meta, and Amazon — reached approximately $725 billion, up 77% from $410 billion in 2025.[7] Microsoft's standalone 2026 capex guidance reached $190 billion, well above analyst consensus of $152 billion. CFO Amy Hood indicated that approximately $25 billion of incremental spending was related to rising memory and component costs.[7] Alphabet raised capex guidance by another $5 billion to $185-190 billion, while Google Cloud backlog reached $460 billion, up 92% from $240 billion in Q4 2024.[7]

The causal pattern is clear. In all four cases, the capex cut had a leading cause: an exogenous macro shock in 2008, a failed business thesis in 2022 (Meta), a CEO regime change with pandemic overhang correction in 2022-2023 (Amazon), and customer demand reallocation in 2025 (Microsoft). In all four cases, capex returned to record levels within one to two years.

The phrase without warning is best understood as hindsight bias. It may feel true after the event, but it is not supported by the primary-source timeline.
 

3. Testing Hypothesis B: Decomposing the Memory Downturn by Demand Source

Hypothesis B is not entirely wrong. It is wrong in weight.

When the 2022-2023 downturn in Korean memory is decomposed by primary demand source, three sources of revenue contraction emerge.

PC and mobile consumer demand cliff, estimated at 50-60% of the downturn. During the pandemic, consumer electronics companies accumulated excess memory inventory. As that inventory normalized, additional orders stopped. By late 2022, memory revenue had fallen to roughly one-third of the prior year’s level, and Samsung, SK  ynix, and Micron all entered quarterly losses. The largest driver was not Big Tech AI infrastructure. It was a consumer-electronics demand cliff after pandemic-era overstocking.

Enterprise server demand slowdown, estimated at 20-25%. The Federal Reserve began its rate-hiking cycle in March 2022. Non-AI enterprise IT capex became more conservative, and traditional server DRAM demand weakened on a quarterly basis.

Hyperscaler spot-order volatility, estimated at 15-25%. This is where the bear case contains a partial truth. Samsung’s April 2023 earnings-call commentary stated that continued conservative purchasing from hyperscalers is expected to slow NAND recovery.[8] Micron’s September 2022 8-K filing also stated that recently, the market outlook for calendar 2023 has weakened and guided DRAM bit supply growth to negative territory.[9] Given that data centers accounted for roughly 20-30% of global DRAM consumption in 2022, hyperscaler order volatility clearly mattered.[10]

But the critical distinction is this: the hyperscaler source was a quarterly spot-order adjustment. It was not a collapse of multi-year LTAs.

So Hypothesis B is best described as a partial truth with a causality-weighting error. The trauma existed. But labeling the downturn as a Big Tech-driven event is misleading. The larger cause was the PC and mobile consumer-demand cliff.

4. 2022 vs. 2026: Four Structural Differences

For the 2022 trauma to apply directly to the 2026 AI capex cycle, the ordering mechanism would have to be similar. It is not.

The October 1, 2025 OpenAI Stargate LOI with Samsung and SK hynix is a particularly important signal.[12] The LOI targets up to 900,000 DRAM wafers per month. This is roughly 40% of global DRAM capacity and more than twice current HBM capacity. An LOI is not a final contract. But it is strong evidence that the market structure is moving from short-duration spot orders toward multi-year LTA-style customer lock-in.

The conclusion is straightforward. The 2022 cycle was a spot-order cycle hit by a macro and consumer-demand shock. The 2026 cycle is being built on multi-year customer lock-in, backlog visibility, and strategic-security motivations. The two cycles differ across order form, inventory position, demand motivation, and backlog visibility.
 

Applying the 2022 trauma directly to the 2026 AI infrastructure cycle is therefore a causal analogy error.

5. The Real Cause of Supplier Caution: Protecting Pricing Power

If the Korean suppliers are cautious, what are they cautious about?

The answer becomes clear when we compare two facts: SK hynix's cautious posture toward Big Tech-funded dedicated facilities, and the record-scale capex already underway at both Samsung and SK hynix.

Primary evidence 1: SK hynix's caution toward customer-funded dedicated capacity. In May 2026, multiple Big Tech customers reportedly offered to finance dedicated production facilities for SK hynix. SK hynix maintained a cautious position because accepting such funding could leave it beholden to specific customers and potentially force it to lower chip prices in exchange for longer-term, more stable orders.[13] Analysts added that this structure could reduce SK hynix to a foundry for its customers, leading to a loss of bargaining power.[13]

Primary evidence 2: both Korean suppliers are already expanding capex at historic scale. SK hynix's M15X fab investment is KRW 20 trillion, or approximately $14.7 billion — the largest single capex decision in memory-industry history.[14] Capex as a percentage of 2026 revenue is expected to be around 30%, far above historical averages. The first cleanroom operation was moved forward by four months to May 2026, and HBM4 production timing was accelerated to February 2026. Samsung is also planning a 50% increase in HBM capacity in 2026.[15] Micron has raised 2026 capex to $20 billion.

The intersection of these two facts reveals the real causal mechanism.

The Korean suppliers are not refusing to expand production capacity. They are already expanding at record levels. What they are cautious about is accepting financing and contract structures that could weaken their pricing power and increase customer dependency.

The caution is not about capex. It is about contract structure.

Therefore, the claim that Korean memory suppliers are cautious because they fear sudden Big Tech capex cuts is not the right causal reading. The more accurate interpretation is that they recognize the risk of ceding pricing power to Big Tech and are trying to defend bargaining leverage in the next LTA cycle.

6. How the Bear Case Could Evolve: Four Cards to Watch

The bear case is not static. If the two initial hypotheses are challenged, bears can bring new cards. The important question is whether those cards are supported by current primary-source evidence.

Card 1 — "What if AI infrastructure ROI fails to prove itself in 2027?"

The weakness of this card is the funding structure of Big Tech capex. In 2026, more than 95% of Big Tech capex is funded by operating cash flow, with limited dependence on incremental debt.[16] Google Cloud revenue reached $20 billion in Q4 2025, up 63% year over year, and GCP operating margins continued to expand.[7] Even if ROI proves uneven for some workloads, the more likely outcome is company-specific guidance adjustment, not a systemic cut to AI capex.
 

Card 2 — “What if a DeepSeek-style efficiency shock reduces capex demand?”

The evidence points in the opposite direction. After the January 2025 release of DeepSeek R1 — which triggered a 17% one-day decline in NVIDIA's stock and erased roughly $500 billion of market value — Big Tech capex guidance increased or remained intact.[17] Microsoft guided to $80 billion, Meta to $60-65 billion, Amazon to more than $100 billion, and Alphabet to $85 billion. The efficiency shock changed the motivation structure. Before DeepSeek, capex was primarily an ROI question. After DeepSeek, it became a national-competitiveness question. Capex driven by strategic-security logic is less sensitive to near-term ROI volatility.
 

Card 3 — "What if dependence on NVIDIA creates a signal-disruption risk?"
 

Memory demand is not limited to NVIDIA. Broadcom, AMD MI400, Google TPU v6, AWS Trainium2, Microsoft Cobalt, and Apple M-series all use HBM and high-capacity server DDR5. Even if NVIDIA's Vera Rubin schedule shifts, HBM3E demand for Blackwell has strengthened. Customer diversification reduces the risk that a single-customer signal disrupts the entire revenue base of the Korean suppliers.

Card 4 — "What if AI capex turns out to be like metaverse capex?"

This comparison fails at the category level. In 2022, Reality Labs had effectively no recurring revenue base, and the user-adoption thesis behind metaverse capex remained unproven. In 2026, AI capex rests on a very different revenue structure. Anthropic's ARR is estimated at $30 billion-plus, and OpenAI's ARR at $25 billion-plus.[21] Microsoft RPO reached record levels, and Google Cloud backlog reached $460 billion, up 92% year over year.[7]

Metaverse capex was capex without proven demand. AI capex is capex backed by binding contracts, backlog, and strategic competition. The two are not the same category.
 

These four evolved bear cards cannot be dismissed forever. But at the current point in the cycle, available primary-source evidence pre-validates much of the counterargument.

7. The Real Watch Point: 2027-2028 LTA Renewal

The more important output of this exercise is not the negation of the bear case. It is a better monitoring framework — five axes that will be most informative for tracking the 2027-2028 LTA renewal window.

The real watch point is the 2027-2028 LTA renewal window.

We identify five axes to monitor.

Axis 1 — quarterly signals on AI capacity utilization. The strongest negative signal would be a shift by Big Tech companies from separate AI revenue disclosure toward more opaque, integrated reporting. Sun Microsystems and Cisco followed similar patterns during the dot-com cycle.

Axis 2 — the HBM3E vs. DDR5 ASP spread. The ASP gap between HBM3E and DDR5 is expected to narrow from 2026.[18] A narrowing spread would compress HBM margins, potentially making supplier capex decisions more conservative. That, in turn, could constrain AI infrastructure supply and eventually slow AI capex itself.

Axis 3 — NVIDIA Vera Rubin and Vera production schedules. Changes in the production schedule for NVIDIA Vera Rubin are a critical watch point for HBM4 timing.[14] Any additional schedule shift in next-generation GPUs could directly affect SK hynix's quarterly HBM4 supply timeline.

Axis 4 — the take-or-pay ratio inside hyperscaler backlogs. Take-or-pay clauses force customers to pay penalties if they cancel committed orders. Microsoft, Google, and Amazon report record commercial RPO and backlog, but the percentage protected by take-or-pay enforcement is not publicly disclosed. This is a key item for primary-channel diligence.

Axis 5 — LTA renegotiation signals from the memory suppliers. The 2027-2028 period will likely concentrate the formal conversion of Stargate-related LOIs into LTAs and the renegotiation of multi-year hyperscaler contracts. Supplier IR language around contract-structure changes should be tracked closely.
These five axes form Gorilla PE's monitoring framework for the next two years.
 

Conclusion

The market question — have Korean memory suppliers already reached the peak? — is not the right question.

Hypothesis A, the idea of a sudden and uncaused Big Tech capex cut, is not supported by the four main historical examples. Google 2008, Meta and Amazon 2022-2023, and Microsoft 2025 all had identifiable leading causes and were followed by record-level capex recovery.

Hypothesis B, the idea that the memory suppliers' trauma was primarily caused by Big Tech, misallocates causality. Hyperscaler order volatility mattered, but it was a 15-25% layer, not the dominant driver. The larger shock was the PC and mobile consumer-demand cliff.

The true cause of supplier caution is not capacity fear. It is contract-structure negotiation and pricing-power protection. SK hynix's M15X investment and Samsung's planned HBM capacity expansion show that capex is already expanding at historic scale. The caution lies in refusing customer-funded structures that could compromise bargaining power.

The 2026 cycle is structurally different from the 2022 cycle. It differs in order type, inventory position, demand motivation, and backlog visibility. The bear case's initial combined proposition breaks down once the two hypotheses are separated.

The market sees danger where the danger does not yet reside. The real risk may arrive through a different mechanism in the 2027-2028 LTA renewal window. That is why the right question is not whether to fear Big Tech capex in the abstract. The right question is what happens to pricing power, contract structure, take-or-pay enforcement, and next-generation HBM supply when the current AI infrastructure build-out enters its first major renewal cycle.

The combined proposition works only when bear and fear remain fused. Once separated, both become partial truths — and partial truths are not enough to define a cycle.

Sources

[1] Korea Economic Daily, December 2025 — reports on Samsung and SK hynix expressing caution during global IB meetings.
[2] Reuters and Finance Spot News, March 2026 — memory stocks falling roughly 10% in a single trading day.
[3] Data Center Knowledge — Google quarterly capex series from Q1 2008 to Q3 2009, based on SEC 10-Q tracking.
[4] Meta Platforms earnings call, February 1, 2023, and 8-K Newsroom Post, March 14, 2023 — capex guidance reduced from $34-37 billion to $30-33 billion.
[5] TD Cowen channel-check report, February 21, 2025 — Michael Elias, Cooper Belanger, and Gregory Williams; cited by Bloomberg and Yahoo Finance.
[6] TD Cowen channel-check report, March 26, 2025 — additional 2GW of U.S. and European capacity abandonment; cited by Bloomberg and Datacenter Dynamics.
[7] Financial Times, Tom's Hardware, and Wall Street Journal Q1 2026 earnings aggregation — Big Tech 2026 capex of $725 billion, Microsoft capex of $190 billion, Alphabet capex of $185-190 billion, and Google Cloud backlog of $460 billion.
[8] Samsung Electronics earnings call, April 28, 2023 — cited by The Register.
[9] Micron Technology September 2022 8-K — SEC filing.
[10] IDC — 2022 global DRAM consumption share for data centers.
[11] Reuters and TrendForce — global DRAM inventory declining from 31 weeks in Q1 2023 to 8 weeks in Q4 2025.
[12] OpenAI official announcement, October 1, 2025 — Stargate LOI with Samsung and SK hynix; cited by TechCrunch, Reuters, and Astute Group.
[13] TradingKey, May 2026 — reports on SK hynix's cautious posture toward Big Tech dedicated-facility funding proposals.
[14] SK hynix official announcements, April 24, 2024 and 2025; Financial Content, February 2026 — M15X KRW 20 trillion investment, accelerated cleanroom schedule, and NVIDIA Vera Rubin production schedule watch point.
[15] The Elec, December 30, 2025 — Samsung's plan to increase HBM capacity by 50% in 2026.
[16] JP Morgan, November 2025 — Big Tech capex funding mix, with more than 95% funded by operating cash flow.
[17] Gorilla PE Insights, January 2025 — analysis of Big Tech capex changes after DeepSeek R1.
[18] TrendForce 2026 outlook — expected narrowing of the HBM3E versus DDR5 ASP spread.
[19] Amazon 10-K FY2022 and FY2023 (SEC EDGAR) — total company capex of $63.6 billion in 2022 falling to $52.7 billion in 2023 (-17%). AWS segment capex estimated at $27.5 billion to $24.8 billion (-10%), per Platformonomics "Follow the CAPEX: Cloud Table Stakes 2023 Retrospective" (February 2024).
[20] Amazon CFO Brian Olsavsky, Q3 2022 earnings call (October 27, 2022) — "we cut about one-third of our overall Amazon budget while still focusing our capital dollars really on the AWS business." Andy Jassy CEO appointment: July 2021. 27,000 layoffs (late 2022 through 2023) — CNBC, April 11, 2024; Amazon Annual Shareholder Letter.
[21] OpenAI ARR $25 billion-plus: The Information (March 6, 2026) "OpenAI Tops $25 Billion in Annualized Revenue as Anthropic Narrows Gap"; Reuters (March 2026) reported the same. OpenAI CFO Sarah Friar disclosed $20 billion ARR for 2025 (vs. $6 billion in 2024 and $2 billion in 2023, a 10x expansion) on January 20, 2026. Anthropic $30 billion-plus ARR: company announcement, April 7, 2026.

Disclaimer

This essay is provided for informational and discussion purposes only. It does not constitute investment advice, an offer to sell, or a solicitation to buy any security. The analysis reflects Gorilla PE's market view based on public sources, internal research, and industry diligence as of the date of writing. Forward-looking statements are inherently uncertain and should not be relied upon as guarantees of future outcomes.