Alpha Call | The Truth About AI Storage Demand Behind KV Cache Compression Panic Summary Report

China's daily token usage has exceeded 140 trillion, growing over a thousand times in two years. However, a research paper on KV Cache compression triggered a sharp decline in the storage sector, followed by SK Hynix$CSOP SK Hynix Daily (2x) Leveraged Product (07709.HK)$ securing a nearly 6.9 billion euro EUV equipment order.$ASML Holding (ASML.US)$ Between emotional panic and industrial reality, which one should we believe?

On April 1, Tim Yang, Senior Industry Analyst at Futu Securities, provided a systematic analysis during the Alpha Call live broadcast regarding the core logic of the storage sector, the essence of the recent pullback, and investment directions following the OFC Global Optical Communication Conference.

I. AI Agents Push Token Consumption to New Heights, Transforming Storage Demand from Cyclical to Essential

First, let’s revisit the origin of demand. According to OpenRouter data, last week, the total global AI large model token usage reached 22.7 trillion tokens,a week-on-week increase of 11%.The Chinese market contributed 9.86 trillion tokens,Up 34% from the previous period, surpassing the US for four consecutive weeks.

The core catalyst driving this surge is AI agents.Unlike traditional chatbots, an AI agent requires multiple cycles of perception, planning, execution, and reflection in a single deep research task, consuming millions of tokens per session — hundreds of times more than traditional conversations.Industry data shows that token consumption on leading agent platforms doubles every two weeks, with growth exceeding tenfold within half a year.

This has directly changed the nature of demand in the storage industry. In large model inference, KV Cache is used to store key-value pairs of historical conversations to avoid redundant calculations, with its capacity requirements growing linearly with the length of the token sequence and concurrency. Currently, the context window of leading large models has generally broken through the million-token mark, and token consumption for video generation tasks is dozens of times higher than for text.The storage demand for a single AI server has reached 8-10 times that of a traditional server, pushing HBM, high-bandwidth DDR5, and enterprise-grade SSDs into full-scale production increases.

Tim believes this signifies that storage demand is transitioning from past cyclical fluctuations to long-term inelastic growth.

Two, analysis of two major causes for the pullback: technical disruptions and price volatility do not undermine the core logic.

The storage sector has recently experienced a significant pullback. In addition to the impact of geopolitical conflicts on the broader market, the main factors weighing on the market were twofold, which Tim dissected individually.

First, Google.$Alphabet-C (GOOG.US)$Turbo Quant paper triggered panic over KV Cache compression.

Tim pointed out that the core implication of this paper is precisely that KV Cache compression within the 3-4 bit range is increasingly approaching the practical engineering limit. The mainstream compression rate currently implemented in practice is only about 2x, while the upper limit of cutting-edge solutions can reach 4-6x, but the industry average is far from that. The best results shown in the paper do not equate to the actual implementation level across the entire industry.

He further noted that this paper illustrates two key points:KV Cache CompressionThis path still has some way to go, but the remaining potential is not as substantial as in the early stages—further gains will be increasingly limited and achieved at higher complexity, with significantly rising adaptation difficulties, validation costs, and scenario-related loss risks.This area resembles a mature optimization aspect rather than a track maintaining an ultra-high growth trajectory over the long term.

More crucially, the validation of historical experience. Tim used DeepSeek's MLA technology as an example—Even with a 90% level of compression achieved, the total memory pressure in the industry has not been alleviated because the saved resources are quickly 'consumed' by longer contexts, higher concurrency, and more complex tasks.The real impact of MLA is not making memory abundant but pushing down the cost curve of AI, thereby extending capability boundaries. Therefore, high-bandwidth HBM will not lose demand due to KV compression; instead, it will become more valuable due to higher token economic efficiency.

Second, concerns about a cyclical peak triggered by a widespread price drop in consumer-grade memory.

In the US, multiple retailers have seen DDR memory kits drop by up to 100 USD. Domestically, DDR4 16GB memory sticks fell from a wholesale price of 500 RMB to below 300 RMB within three days, while DDR5 32GB kits retreated from 1,000 RMB to around 700 RMB.

Tim believes this issue needs to be analyzed on different levels. The price cuts are concentrated in the consumer-grade spot market, where excessive channel stockpiling combined with overly rapid price increases exhausted consumer willingness, leading to significant selling pressure on DDR4. However, in the OEM market, domestic smartphone and PC manufacturers still accepted substantial price hikes for DDR5 in the first and second quarters—as DDR5 is a core component in AI and high-end device ecosystems, supply must be secured even at a premium. In enterprise-level server storage, price fluctuations were minimal, with 64GB DDR5 maintaining above 13,000 RMB.

The industry currently shows clear stratification in the short term: the DDR4 market is accelerating its clearance, while high-end categories such as DDR5 and HBM remain in structural tight balance.

Tim emphasized a cognitive shift: the downstream market is no longer dominated by consumer electronics but by AI. It is no longer appropriate to judge the cycle’s peak based on falling prices and shipments from consumer-grade suppliers as before. Memory manufacturers have learned painful lessons from the past, and their capital expenditures are far more cautious than before, moving closer to Taiwan Semiconductor's business model—only expanding capacity after ensuring prepayments and long-term demand visibility from key customers. Downstream AI giants have also shifted from adjusting orders quarterly to long-term infrastructure-oriented procurement models.

Third, Hynix’s 6.9 billion euro EUV order: the industry is answering market concerns with concrete actions.

At a time when market sentiment was most shaken, the industry sent a completely different signal.

SK Hynix recently placed an order for EUV lithography machines worth approximately 6.9 billion euros. Estimating the cost of a single Low-NA EUV unit at around 230 to 250 million euros, this corresponds to about 25 to 30 units. The order was finalized in March this year and is expected to be delivered by the end of 2027. Both the scale and timing significantly exceed market expectations.

Tim believes there are two key signals behind this order.

First,This is not a routine purchase but rather a pre-order with strategic implications to lock in production capacity. ASML’s current annual production capacity is about 90 units, and based on expansion plans from logic chip and memory clients, global EUV demand by 2027 may approach the supply limit. SK Hynix's bulk order likely aims to secure critical equipment within the next two years.

Secondly,The timing of the order points to process integration and capacity ramp-up starting in the second half of 2027. In other words, this investment is essentially preparation for the next round of more advanced DRAM upgrades. Due to the difficulty of cleanroom expansion, which can even surpass the challenge of obtaining EUV equipment itself, capital expenditure increases over the next two years will focus on transitioning to higher EUV density and more advanced processes within limited cleanroom space.

Tim pointed out that this round of capital expenditure growth does not reflect traditional extensive capacity expansion but rather tilts toward advanced process upgrades. Memory giants remain highly cautious about increasing capital spending, indicating strong certainty on the demand side. Since January, manufacturers have gradually signed multi-year LTA agreements with hyperscale companies, providing demand endorsement for capacity expansion. Increased capital spending does not necessarily equal runaway supply; instead, it reflects accelerated adoption of advanced nodes and structural upgrades.

Under this logic, despite adjustments, the memory sector remains robust. Companies like SK Hynix, which focus on high-end products, deserve attention. Meanwhile, upstream semiconductor equipment suppliers such as ASML, Lam Research, and Applied Materials will benefit first from upward revisions in capital expenditures.

Fourth, Core Directions Post-OFC Conference: DCI and OCS Enter Order-Driven Phase

In the second part of the live stream, Tim turned to investment opportunities following the OFC Global Optical Communications Conference.

He noted that DCI (Data Center Interconnect) and OCS (Optical Circuit Switching) have become the two most consensus-driven directions in the industry. In Q2 2026, North American CSP cloud service providers will release DCI bidding demands totaling approximately 10 billion USD — the industry's first large-scale tender, marking the shift from sentiment-based investments to order and performance-driven phases. Optical modules and systems account for about 40%, while switches make up roughly 25%.

Regarding relevant targets, Ciena$Ciena (CIEN.US)$ Valuation is now relatively full; Lumentum ($Lumentum (LITE.US)$ ) and Coherent ($Coherent (COHR.US)$ ) remain the main beneficiaries in the OCS space. On the other hand, Nokia (NOK) shows significant expectation gaps, making it the key stock Tim recommended this session.

Five, Core Stock Analysis: Nokia$Nokia Oyj (NOK.US)$ —— An Undervalued New Force in Optical Networking

Business Overview.After selling its mobile phone business and undergoing acquisitions and restructuring, Nokia now focuses on network infrastructure, optical networking, IP routing, and enterprise solutions. By 2025, the revenue structure will consist of approximately 40% from network infrastructure (including optical networking) and 40% from mobile infrastructure, with patent licensing contributing around 7.5% but accounting for nearly half of operating profits, providing stable cash flow.

Optical Networking: The Core Growth Engine for the Future.After acquiring Infinera in 2025, Nokia became the only Western company with a vertically integrated indium phosphide wafer fab, gaining control of key materials and manufacturing capabilities for optical communications. The 800G and above coherent optical modules are expected to be sampled by summer 2027, with commercial availability in the second half of the year. At the OFC conference, management projected the total addressable market (TAM) for optical networks to grow from $13 billion in 2025 to $24 billion by 2030, representing a compound annual growth rate (CAGR) of approximately 13%, exceeding previous mid-term guidance.

Order validation has been completed.Nokia has secured a switch order worth about $1 billion from Google, establishing itself as a competitor on par with Arista and Cisco. In terms of optical transmission systems, its Multi-Haul multi-track parallel expansion solution boasts industry-leading density specifications, targeting a market space of approximately $7 billion annually.

Telecom business bottomed out and is recovering.Investment in 5G core networks by communication service providers surged by about 80% year-over-year in the fourth quarter of last year. AT&T indicated that its total capital expenditure over the next five years will be approximately $250 billion, averaging $500 billion annually, significantly higher than the $200 billion spent each year over the past two years. Investment in 5G core networks is expected to maintain double-digit growth through 2026. Additionally, global 6G communication standards are anticipated to be finalized by the end of 2026, with initial deployments in Asia starting in the second half of 2027, followed by Europe and the US in 2028, opening up medium- to long-term opportunities for telecom equipment vendors.

AI-RAN collaboration holds significant potential.Last year, NVIDIA acquired shares in Nokia, and their collaboration involves deploying GPUs at telecom base stations, which are naturally suited for edge computing applications. However, Tim noted that this segment carries less certainty compared to optical networking and should not be assigned overly optimistic expectations.

Valuation and recommendation.Nokia's forward-looking price-to-earnings ratio for 2027 is currently around 18.8 times, slightly above the one-year average plus one standard deviation, with some catalysts already priced in. Tim believes valuation could recover to above 20 times, offering limited short-term upside but making it suitable as a medium- to long-term investment target.

Six, selected Q&A

Regarding NVIDIA's Rubin Ultra architecture adjustment.Tim stated that transitioning from a four-die package to a standard dual-chip version and an enhanced 2×2 solution is a reasonable choice to reduce packaging complexity and improve mass production feasibility. The total computing power and specifications remain unchanged, posing no substantial impact on the demand for HBM4/HBM4E. The market's overreaction mainly stems from the current extremely low risk appetite; if geopolitical tensions ease, memory could become one of the primary rebound sectors.

Regarding Nokia and InnoLight$Zhongji Innolight (300308.SZ)$ and other optical module manufacturers' competition.Tim clearly pointed out that these two companies operate in different markets and do not pose direct competition. InnoLight and Eoptolink focus on medium- and short-distance transmission within data centers; Nokia specializes in long-distance coherent transmission over 100 kilometers and sells an integrated solution including optical modules, switches, and optical transmission systems, which pure module manufacturers cannot achieve. The optical networking business has ample room for growth, marking the beginning of Nokia’s long-term earnings expansion.

Key conclusions:The exponential growth of tokens driven by AI underpins the long-term prosperity of both storage and optical communication sectors. KV Cache compression and price reductions in consumer-grade memory reflect normal technological optimization and channel adjustments, without altering the fundamental shift from cyclical fluctuations to sustained growth. The industry has validated demand certainty through large-scale orders backed by real investments. It is recommended to focus on core assets with technological barriers and expectation gaps, positioning for medium- to long-term investment opportunities in the AI computing power supply chain during pullbacks.

509K Views