Gao Yi’s Technology News Bet on Micron Tests the AI Memory Boom

Gao Yi Asset Management increased its disclosed Micron position by 283 percent during one quarter, despite an extraordinary rally across AI memory stocks. The August 12 filing turns a routine portfolio update into consequential technology news. It shows a major Chinese investment manager adding exposure after the memory cycle had already produced record earnings.

Perseverance Asset Management International, Gao Yi’s overseas investment entity, reported 34,417 Micron shares as of June 30, 2026. Those shares carried a combined filing value of about $39.7 million. Three months earlier, the manager disclosed only 8,975 shares worth approximately $3 million.

The filing also revealed a larger Sandisk position, connecting Gao Yi’s investment thesis to both major branches of semiconductor storage. Micron supplies DRAM, NAND, and high-bandwidth memory, or HBM. Sandisk focuses on NAND flash products used in data centers, devices, and removable storage.

That combination matters more than either trade alone. Gao Yi was not simply buying one company after a favorable earnings report. It was placing a broader bet that AI infrastructure has changed the economics of memory and storage.

The tension is timing. Micron and Sandisk had already reported sharp revenue and margin expansion before the June reporting date. Gao Yi therefore bought more exposure when the cycle looked exceptionally profitable, but also increasingly vulnerable to expectations, new capacity, and pricing reversals.

The Filing Confirms a Much Larger Micron Position

Gao Yi’s disclosed Micron share count rose from 8,975 to 34,417 between March 31 and June 30, a net increase of 25,442 shares.

The figures come from two Form 13F reports filed with the US Securities and Exchange Commission. Form 13F is a quarterly disclosure covering specified US-listed securities held by qualifying institutional investment managers.

Gao Yi’s latest 13F filing listed Micron twice because the manager reported holdings under different investment-discretion categories. One entry contained 32,589 shares under sole discretion. Another contained 1,828 shares under defined discretion.

Combining those entries produces the 34,417-share total. The SEC table assigned the position a value of $39,727,199 at the June 30 reporting date.

The corresponding prior-quarter table listed 5,176 shares under sole discretion and 3,799 under defined discretion. Together, those entries represented 8,975 shares valued at $3,032,114.

The quarter-to-quarter change is significant by several measures. Gao Yi increased the share count by about 283 percent. The reported value expanded more than twelvefold, reflecting both the additional shares and Micron’s market appreciation.

However, the filing does not identify Gao Yi’s purchase dates or execution levels. It only provides positions held at the end of each quarter. The manager might have accumulated shares gradually, purchased them after earnings, or adjusted the position several times.

The filing also does not disclose short positions, securities outside the 13F reporting universe, or most hedging arrangements. Readers should therefore treat the position as confirmed long exposure, not as a complete picture of Gao Yi’s Micron strategy.

The broader portfolio provides useful context. Gao Yi reported 28 entries with a combined value of approximately $979 million for the June quarter. Its largest technology exposure remained Taiwan Semiconductor Manufacturing Company, with about 498,500 American depositary shares across both discretion categories.

Micron represented roughly 4.1 percent of the reported portfolio by value. That made it meaningful, but not dominant. Gao Yi retained much larger positions in companies including TSMC, H World Group, Kanzhun, and PDD Holdings.

The manager also reduced several prominent technology holdings during the quarter. Its disclosed Nvidia position fell from 290,000 shares to 80,000. AMD and Lumentum disappeared from the June filing, while the reported TSMC share count increased.

These changes suggest rotation within technology rather than a simple increase in all AI-related exposure. Gao Yi shifted capital away from some compute names while expanding positions connected to semiconductor manufacturing and data storage.

That distinction creates the central conflict in this technology news story. The manager did not appear to chase every part of the AI trade equally. It concentrated new exposure where constrained supply and rising data intensity were producing unusually strong operating leverage.

Why AI Memory Became the Technology News Signal

Gao Yi’s purchases followed evidence that memory had moved from a replaceable component to a major constraint on AI system performance.

Traditional memory investing revolves around a familiar cycle. Producers expand capacity when prices and margins rise. Supply eventually catches demand, inventories increase, and pricing weakens. Capital spending then slows until the next shortage begins.

AI infrastructure has not removed that cycle. It has changed its most valuable products and the urgency of customer demand.

HBM stacks multiple DRAM layers close to an accelerator, increasing the rate at which processors can access data. That bandwidth is essential because an expensive AI accelerator delivers limited value when memory cannot feed it quickly enough.

Training large models requires immense volumes of fast memory. Inference also consumes more memory as models, context windows, and concurrent user workloads expand. The resulting demand extends beyond HBM into server DRAM, enterprise solid-state drives, and storage systems supporting data preparation.

Micron’s fiscal Q3 results made that change visible. The company reported revenue of $41.46 billion for the quarter ended May 28, 2026. Revenue had been $23.86 billion in the preceding quarter and $9.30 billion one year earlier.

Micron attributed the performance and its stronger outlook to the strategic value of memory in AI infrastructure. Management also said development of HBM4E using its 1-gamma DRAM technology was underway, with volume production expected during calendar 2027.

These are company statements, not independent guarantees about future demand. Still, the realized revenue provides stronger evidence than a product roadmap alone.

The increase also explains why Gao Yi’s timing differs from an ordinary recovery trade. The manager initiated its Micron position in the March quarter, then added substantially during a period of accelerating results. It was increasing exposure after the recovery became obvious.

That can be rational when earnings expectations are still adjusting faster than valuations. A cyclical company may look expensive relative to past profits while remaining less demanding against current cash generation. It can also be dangerous when investors extrapolate peak margins too far.

Gao Yi’s Sandisk purchase reinforces the first interpretation. The manager increased its reported Sandisk position from 5,479 shares in March to 16,000 shares in June. The position’s disclosed value rose from about $3.5 million to approximately $36.4 million.

Sandisk supplies NAND flash, which retains data without power. NAND serves different workloads from DRAM, but AI infrastructure needs both. Accelerators require fast working memory, while data centers require large storage pools for models, checkpoints, datasets, logs, and generated content.

The separate positions therefore cover two distinct bottlenecks. Micron offers exposure to DRAM and HBM alongside NAND. Sandisk provides a more focused route into flash storage and enterprise solid-state drives.

This allocation is more informative than a generic semiconductor bet. Gao Yi appears to be treating data movement and retention as investment themes alongside computation.

That shift matters to developers and enterprise buyers. Faster accelerators do not automatically produce faster AI applications. System performance depends on how quickly data moves through memory, storage, networking, and software.

It also matters to cloud customers. When a scarce component absorbs more of a server’s cost, infrastructure providers must decide whether to raise service rates, accept lower margins, or redesign deployments. Higher memory costs can reshape the economics of model training and inference even when accelerator availability improves.

Gao Yi Is Betting on Scarcity Over the Old Cycle

The primary contest is between durable AI-driven scarcity and the memory industry’s history of producing too much capacity after profitable periods.

Gao Yi’s increased position supports the scarcity case. Micron’s revenue acceleration shows that strong demand and constrained supply were affecting reported results, not merely forecasts.

Sandisk supplied another piece of evidence. Its flash results showed fiscal third-quarter revenue of $5.95 billion, up 97 percent sequentially. Data-center revenue increased 233 percent from the previous quarter.

Sandisk reported a GAAP gross margin of 78.4 percent, compared with 50.9 percent in the preceding quarter. It attributed the revenue increase to higher pricing and a shift toward higher-value customers.

Those figures are remarkable for an industry known for commodity-like pricing. They indicate that customers were paying for scarce capacity and that suppliers captured a substantial portion of the resulting value.

Sandisk also said it had signed customer agreements supported by firm financial commitments. Such arrangements can improve demand visibility and reduce some exposure to short-term spot pricing. They do not eliminate the risk of oversupply or customer renegotiation.

The durable-scarcity argument begins with production complexity. HBM consumes more manufacturing capacity than conventional DRAM because it requires advanced dies, stacking, packaging, and extensive testing. Yield losses at any stage can constrain final output.

Suppliers cannot instantly convert every conventional memory line into qualified HBM production. Products must meet the thermal, performance, and reliability requirements of accelerator vendors and cloud customers.

That qualification process gives established suppliers leverage. Micron competes directly with SK hynix and Samsung, while all three must coordinate production transitions with packaging capacity and customer roadmaps.

The same logic increasingly applies to enterprise NAND. Large AI systems create demand for high-capacity, high-throughput drives, but customers also require consistent performance, endurance, firmware support, and qualification.

Storage capacity alone is not enough. Products must deliver predictable behavior in dense server environments where a failed drive or stalled workload carries significant operational costs.

These constraints help explain why Gao Yi could add after a large rally. If supply remains structurally tight, current earnings may represent an early stage of higher profit pools rather than a brief peak.

However, the old cycle remains a formidable opponent. High margins create an incentive for every supplier to expand production. Customers also respond by optimizing workloads, negotiating contracts, and qualifying alternative components.

Memory companies have often described favorable supply discipline near the top of previous cycles. That discipline becomes harder to maintain when each producer sees an opportunity to gain market share.

Technology transitions can create another form of supply growth. Better yields and denser products increase effective output even without a proportional increase in factory space. A shortage can ease quickly once several bottlenecks improve together.

The market therefore faces two competing clocks. AI demand is growing through larger models, more inference traffic, and new data-center deployments. Supply is responding through capital spending, process improvements, and next-generation products.

Gao Yi’s position says the demand clock was still moving faster at the end of June. It does not establish how long that advantage will last.

What the 13F Filing Cannot Prove

A larger disclosed position shows conviction at one reporting date, but it does not prove that Gao Yi expects an uninterrupted memory upcycle.

The SEC’s 13F framework exists to disclose institutional holdings, not investment theses. It offers a delayed snapshot and excludes important elements of portfolio construction.

The June positions became public on August 12. By then, Gao Yi could have sold, added, or hedged the shares. Any analysis published in August therefore describes the manager’s position six weeks earlier.

The filing also combines assets managed with different discretion classifications. The 1,828 Micron shares reported under defined discretion may reflect a different mandate from the 32,589 shares under sole discretion.

Adding them is appropriate when measuring total reported exposure. It does not mean one portfolio manager made every trade for the same reason.

The headline description of Gao Yi “greatly increasing” a storage leader is factually supported by the share-count change. It becomes misleading if interpreted as a direct public endorsement from the firm.

Gao Yi did not issue a statement explaining the purchase. It did not publish a Micron target, forecast the memory cycle, or describe the position as a permanent holding. The thesis in this article is an inference from the disclosed allocation and the operating environment.

Position size adds another restraint. Micron accounted for around 4.1 percent of Gao Yi’s reported US portfolio value at quarter-end. Sandisk accounted for about 3.7 percent.

Together, the positions were material, but they remained smaller than Gao Yi’s disclosed TSMC exposure. The manager was expressing confidence without making memory the portfolio’s only technology outcome.

Its Nvidia reduction also has several possible explanations. Gao Yi might have viewed memory as offering better risk-adjusted exposure to AI spending. It might have taken gains, reduced concentration, or rebalanced around other mandates.

The filing alone cannot distinguish those motives. It also cannot show whether the manager used index futures, options, or other instruments to reduce broader semiconductor risk.

Company results require similar caution. Micron’s record revenue does not guarantee that future quarters will preserve the same growth rate. Comparisons become more demanding as prior periods incorporate higher prices and stronger shipments.

HBM demand also depends on a concentrated customer base. Large accelerator companies and cloud providers possess significant negotiating power, even when supply is tight. Their architecture choices can redistribute demand among memory types and suppliers.

Competition adds execution risk. SK hynix has held a strong position in HBM, while Samsung continues investing in product qualification and capacity. Micron must execute new technology ramps while controlling yields and costs.

Sandisk faces its own concentration and manufacturing risks, including reliance on its production relationship with Kioxia. NAND also serves consumer devices, where demand can behave differently from AI data-center spending.

Geopolitics remains another uncertainty. Micron operates across a globally distributed supply chain affected by export controls, trade policy, manufacturing incentives, and customer restrictions.

None of these risks invalidates Gao Yi’s allocation. They explain why the trade should be read as a calculated exposure to scarcity, not proof that memory has escaped cyclicality.

For readers following technology news, the useful conclusion is narrower. A major institutional investor increased direct exposure to two storage suppliers while reducing positions in some other AI-linked chip companies. That rotation suggests the market’s attention is moving deeper into the infrastructure stack.

Three Signals Will Decide Whether Gao Yi Was Early or Late

The next phase depends on HBM supply, enterprise NAND commitments, and whether memory margins remain elevated as new production reaches customers.

The first signal is Micron’s order visibility for HBM4 and HBM4E. Product qualification and committed volume matter more than broad statements about AI demand.

If Micron reports firm customer commitments extending into 2027, Gao Yi’s purchase will look aligned with a durable capacity shortage. Strong commitments would show that customers are securing supply before production becomes widely available.

The thesis weakens if qualification delays push volume production later, or if customers shift more orders toward SK hynix and Samsung. A roadmap without accepted products does not create the same economic value as contracted shipments.

Investors should also separate unit demand from pricing. HBM shipments can grow while profitability disappoints if competition intensifies or yields remain costly.

The second signal is Sandisk’s customer agreement structure. Multi-year commitments can make NAND revenue more predictable, especially when data-center buyers reserve capacity.

The critical questions concern duration, purchase obligations, pricing adjustments, and cancellation protections. Stronger contractual support would indicate that customers view flash availability as strategically important.

Weak commitments would leave Sandisk more exposed to familiar NAND pricing swings. Any decline in data-center revenue or gross margin after recent gains would challenge the idea of a lasting structural reset.

The third signal is industry capacity discipline. Micron, Samsung, SK hynix, Sandisk, and Kioxia all have incentives to expand when returns are high.

Capital spending announcements should be evaluated alongside production timing. A factory announced now may take years to contribute meaningful output, while better yields can affect supply much sooner.

Evidence that suppliers are prioritizing HBM and enterprise products without flooding conventional markets would strengthen Gao Yi’s position. Broad capacity growth across DRAM and NAND would increase the risk of another inventory correction.

These signals should be considered together. Strong AI demand cannot protect margins indefinitely if supply rises faster. Controlled supply cannot sustain profits if cloud customers slow infrastructure purchases.

Developers and enterprise technology teams should watch the same indicators for operational reasons. Memory availability influences server delivery schedules, cloud capacity, hardware configurations, and the cost of running data-intensive applications.

An engineering team may optimize model size or context length when high-capacity instances remain scarce. A cloud buyer may commit to capacity earlier when memory components have long lead times.

Knowledge workers will experience the effects indirectly. Better memory and storage can support faster retrieval, larger working contexts, and more responsive AI systems. Higher infrastructure costs can also limit access or slow deployment.

Gao Yi’s reported trade does not settle the memory debate. It makes the debate harder to dismiss as a temporary technology news narrative.

The manager nearly quadrupled its Micron shares and almost tripled its Sandisk shares during a quarter of extraordinary operating results. It did so while reducing exposure to several other technology names.

That is a specific capital-allocation judgment. Gao Yi was betting that the infrastructure required to move and retain AI data still offered more upside than the market’s most visible compute winners.

The next filings will show whether that judgment survived the summer. Watch the September 30 share counts, Micron’s HBM qualification progress, and Sandisk’s contracted data-center demand.

If all three strengthen, memory scarcity will look increasingly structural. If positions fall while supply expands, this technology news event will instead resemble a profitable but late-stage cycle trade.