AI is an industrial revolution, not a software cycle.

We view the current AI transition as comparable in scope to the steam age: a step-change in productive capacity, as decisive as the move from horse-drawn transport to rail. The investment question is therefore not simply who builds models, but which physical systems allow AI to scale.

Over the past three years, the talk of the industry has shifted from single GPU clusters to thousand-GPU clusters, then to hundred-thousand-GPU clusters. Every eighteen months, another zero is added to the deployment plans. Behind the scenes, there is a fierce scramble to secure every HBM allocation still available for the rest of the decade, every optical module that can possibly be procured, every advanced packaging slot that can be reserved. Asian manufacturing is gearing up to pour hundreds of billions of dollars into a long-unseen mobilization of industrial capacity. By the end of the decade, semiconductor production in the Asian corridor will have grown by tens of percent; from the silicon fabs of Hsinchu to the optical module factories of Wuhan, millions of components will hum.

The physical AI infrastructure race has begun. We are building systems that can think and reason at scale, but the machines that enable this reasoning depend on a tightly coupled stack of physical components: compute accelerators, high-bandwidth memory, optical interconnect, advanced packaging, specialized materials, power delivery and thermal management. Each layer of this stack has its own constraints, its own bottlenecks, its own supply-chain vulnerabilities.

Everyone is now talking about AI models, but few have the faintest glimmer of what is about to hit the physical infrastructure layer. Wall Street analysts still think 2025 might be close to the peak for capex. Mainstream pundits are stuck on the willful blindness of "it's just software." They see only hype and business-as-usual; at most they entertain another cloud-scale technological change.

Before long, the world will wake up to the physical reality. But right now, there are perhaps a few hundred people, most of them in Taipei, Seoul, Shanghai and Tokyo, that have situational awareness of the physical constraints. Through years of supply-chain research and factory-floor due diligence, we have found ourselves amongst them. These are very smart people — the smartest engineers and supply-chain specialists we have ever met — and they are the ones building this infrastructure. If they are seeing the future even close to correctly, we are in for a wild ride.

Let us tell you what we see.

---

The next phase of AI value creation is likely to be constrained by throughput, energy, yield, qualification cycles and supply elasticity. This is not a speculative claim — it is the observable reality of every factory floor, every procurement meeting, every capacity allocation decision being made today across the Asian semiconductor corridor.

Consider the optical interconnect layer. As AI clusters scale from thousands to hundreds of thousands of accelerators, the data movement between nodes becomes as important as the compute itself. A 100,000-GPU cluster requires millions of optical transceivers, each operating at 800G or 1.6T speeds. The supply of these modules is constrained not by design capability but by component availability: lasers, modulators, DSPs, and the specialized assembly capacity required to package them. Lead times for critical components have extended from twelve weeks to thirty-six weeks and beyond. The companies that control these supply chains are not merely participants in the AI boom — they are structural enablers without which the boom cannot continue.

Or consider high-bandwidth memory. HBM3E and the coming HBM4 represent perhaps the tightest supply constraint in the entire AI stack. SK Hynix, Samsung and Micron together control essentially all production, and expanding capacity requires not just capital but years of process development, yield improvement and customer qualification. A single HBM stack contains thousands of through-silicon vias, each requiring sub-micron precision. The yield curves for these products are closely guarded secrets, but industry conversations suggest that even the leading suppliers are operating at the edge of their current capabilities. When demand outstrips supply in such a concentrated market, pricing power follows — not through collusion, but through the mathematics of scarcity.

The same logic applies to advanced packaging materials, high-speed PCB substrates, power delivery components and thermal management solutions. In each case, the barrier to entry is not capital alone but process know-how, material purity, qualification cycles and customer relationships built over years. These are not markets that can be entered by simply writing a check. They are markets where incumbents with deep technical expertise and established supply-chain positions enjoy structural advantages that will persist for years.

Our thesis is therefore not simply "buy AI stocks." It is: identify the specific physical bottlenecks where demand growth is structurally constrained by supply inelasticity, and invest in the companies that control those bottlenecks, with appropriate risk management and valuation discipline. This is a research-intensive, conviction-driven approach that requires deep supply-chain knowledge, extensive industry relationships, and the patience to hold positions through the inevitable volatility of semiconductor cycles.

We are long/short, not long-only. For every bottleneck we identify, we also identify the risks: valuation extremes, customer concentration, technology substitution, policy disruption, and cyclical downturns. Our portfolio construction reflects this duality — structural conviction balanced against risk awareness.

Our research coverage spans four critical nodes across the Asian semiconductor corridor. Each node represents a different type of bottleneck, a different risk-return profile, and a different set of structural dynamics.

Node 01 — China Optical Bandwidth

Optical modules, silicon photonics, co-packaged optics and high-speed interconnects are central to cluster-scale bandwidth, latency and energy efficiency. As AI systems scale, data movement becomes as important as raw compute. The leading Chinese optical module suppliers have spent a decade building relationships with hyperscalers and developing the specialized manufacturing processes required for 800G and 1.6T products. The transition from pluggable to co-packaged optics will create winners and losers, and our research is focused on identifying which companies have the technical depth and customer relationships to navigate this transition successfully.

Node 02 — China AI Interconnect / PCB

High-speed PCB, HDI, server backplanes and advanced substrates form the physical backbone of AI server connectivity. As signal frequencies increase and trace densities rise, the material science of PCB manufacturing becomes increasingly critical. Low-Dk, low-Df laminates, HVLP copper foil, and advanced glass fiber compositions are not commodity inputs — they are differentiated products with significant technical barriers. The companies that control the supply of these materials, and the PCB manufacturers that can work with them at scale, sit at a critical node in the AI infrastructure stack.

Node 03 — Korea / Japan Memory Bandwidth

HBM capacity, TSV processes, packaging yield and allocation cycles shape accelerator utilization. The larger and more inference-heavy AI becomes, the more visible the memory wall becomes. SK Hynix and Samsung dominate HBM3E production, with Micron a distant third. But the supply chain extends far beyond the memory manufacturers themselves: TSV equipment, EMC/GMC materials, low-alpha fillers, and advanced packaging capacity are all critical inputs. Japanese suppliers of packaging materials and equipment are particularly important, as they control many of the specialized chemicals and processes required for HBM assembly.

Node 04 — Japan Packaging Materials

Advanced substrates, CCL, copper foil, ceramics, resins and process materials with qualification moats represent perhaps the most underappreciated layer of the AI infrastructure stack. These are not glamorous products, but they are essential. A single AI GPU package may contain dozens of different materials, each with specific thermal, electrical and mechanical requirements. The companies that supply these materials have spent years — in some cases decades — developing the formulations and processes required by leading semiconductor manufacturers. Customer qualification cycles are measured in years, not months, creating significant barriers to entry and sustained competitive advantages for incumbents.

Our research process is designed to generate conviction in a domain where information is fragmented, opaque and often deliberately obscured. We do not rely on sell-side research or consensus views. Instead, we build our own understanding from the ground up, starting with the physical reality of manufacturing and working our way up to investment conclusions.

1. Supply-Chain Mapping

We begin by mapping the complete supply chain for each infrastructure node we cover. This includes not just the end-product manufacturers but their suppliers, their suppliers' suppliers, and the specialized equipment and materials companies that sit further upstream. We track capacity, lead times, pricing trends and competitive dynamics at each layer. This mapping process is ongoing and iterative — supply chains evolve, new entrants emerge, and relationships shift.

2. Capacity and Qualification Tracking

For each critical component, we track installed capacity, announced expansions, and realistic timelines for new capacity to come online. We pay particular attention to qualification cycles — the process by which new suppliers or new products are approved by leading customers. These cycles are often the binding constraint on supply expansion, and they create structural advantages for incumbents that have already been qualified.

3. Customer / Order-Book Triangulation

We triangulate demand signals from multiple sources: hyperscaler capex guidance, foundry utilization rates, equipment booking trends, and direct conversations with industry participants. We are particularly interested in order-book visibility, backlog trends, and allocation decisions — the micro-signals that reveal where demand is concentrated and which suppliers are gaining or losing share.

4. Margin and Cycle Analysis

We analyze margin structures, cost curves and cyclical dynamics for each node. Semiconductor supply chains are notoriously cyclical, and understanding where we are in the cycle is essential for timing and position sizing. We look for inflection points: when lead times begin to extend, when pricing power shifts from buyers to sellers, when margins begin to expand.

5. Valuation and Crowding Analysis

Even the best fundamental thesis can be undermined by excessive valuation or crowded positioning. We maintain rigorous valuation discipline, comparing current prices to historical ranges, peer multiples, and our own bottom-up estimates of fair value. We also track positioning data to identify when consensus has become too one-sided.

6. Pair Construction and Risk Overlay

Our final step is to construct long/short pairs that express our views while managing risk. A typical pair might be long a structural bottleneck supplier and short a commoditized competitor, or long a beneficiary of AI capex and short a company exposed to the same end markets but with inferior positioning. We overlay macro hedges, sector hedges and FX hedges as appropriate.

Conviction must be balanced against risk. Our risk framework is designed to ensure that our structural views do not lead to excessive exposure, concentration or vulnerability to tail events.

Gross and Net Exposure Discipline

We maintain strict limits on gross exposure (the sum of long and short positions) and net exposure (the difference between long and short positions). These limits are adjusted based on market conditions, volatility regimes and our own confidence in the current opportunity set. In periods of high uncertainty, we reduce both gross and net exposure, preserving capital for more attractive entry points.

Concentration Limits

No single position may exceed a predefined percentage of the portfolio, and no single sector may exceed a larger but still bounded percentage. These limits are designed to prevent any single company or segment from dominating our returns — for better or worse. Diversification across nodes, geographies and supply-chain layers is a core principle.

Liquidity Filters

We maintain minimum liquidity requirements for all positions, measured by average daily trading volume relative to our intended position size. This ensures that we can exit positions quickly if our thesis changes or if market conditions deteriorate. Illiquid positions require additional conviction and smaller sizing.

FX and Regional Exposure Monitoring

Our portfolio is exposed to multiple currencies and geopolitical jurisdictions. We monitor FX exposure continuously and hedge when appropriate. We also track regional political developments — export controls, subsidy policies, trade tensions — that could impact our positions. The Asian semiconductor corridor is particularly sensitive to US-China technology competition, and we maintain scenario analyses for various policy outcomes.

Scenario Analysis

We conduct regular scenario analyses around the key risks to our thesis: export control expansion, hyperscaler capex pauses, technology substitution (e.g., optical to electrical, HBM to alternative architectures), earnings revisions, and semiconductor cyclical downturns. Each scenario is assigned a probability and an estimated portfolio impact. This allows us to stress-test our positions and adjust our risk posture proactively.

Stance Perception LP is available to qualified purchasers, accredited investors and institutional allocators seeking exposure to Asian AI infrastructure through a hedged equity framework.

To request a confidential strategy memorandum or to schedule a due diligence call, please contact us directly. All inquiries are subject to verification of investor status and are not binding until confirmed in writing.

Contact: info@stanceperception.com / info@everrun.ing

---