Woofun AI reports that investor Leopold has grown a fund from $225 million to $13 billion in just 12 months by executing a strategy that completely excludes NVIDIA shares while shorting the entire chip sector with $8.46 billion in put options. This approach, mirrored by Intel CEO Chen Liwu, targets nine specific physical bottlenecks ranging from EDA software to power generation, achieving an annualized return of 225 times for Leopold and a 32-fold stock increase during Chen Liwu's twelve-year tenure at Cadence.

The investment thesis relies on a granular dissection of the AI hardware supply chain, starting with the pre-manufacturing phase where Electronic Design Automation (EDA) tools face a critical efficiency gap. With verification accounting for 60%-70% of the chip development cycle, a single tape-out failure can cost tens of millions of dollars, a risk amplified as AI accelerators integrate hundreds of billions of transistors. The EDA market is projected to reach $14.5 billion in 2025 and $18 billion in 2026, dominated by Synopsys, Cadence, and Siemens who collectively hold over 65% market share. Chen Liwu, having led Cadence for twelve years, identifies this segment as a 'gold mine,' noting that while design complexity explodes, tool efficiency lags, with Cadence improving convergence speed fivefold and Siemens achieving tenfold acceleration in specific tasks.

Material science presents the next layer of constraint as traditional silicon approaches physical limits in power consumption and heat dissipation. Five new materials are now critical: GaN for high-frequency power, SiC for high-voltage currents, InP for optical communication, synthetic diamond for thermal conductivity, and glass substrates for advanced packaging. The demand gap for AI optical interconnects relying on InP materials currently stands at 40%-60%, while Wolfspeed and Infineon are committing over $15 billion to expand SiC capacity between 2025 and 2027. Simultaneously, helium supply has emerged as a volatile variable; early 2026 disruptions in Qatar's Ras Laffan impacted 27%-30% of global supply, spiking spot prices by 40%-100%. The South Korean semiconductor industry, which relies on Qatari helium for 64.7% of its needs, faces direct production risks for Samsung and SK Hynix HBM lines, as helium consumption per unit in the 2nm process is expected to rise 20% compared to 3nm.

On the circuit board itself, the bottleneck shifts to memory and packaging, where High Bandwidth Memory (HBM) supply remains critically tight. The global HBM market is forecast to grow from $9.2 billion in 2026 to nearly $70 billion by 2035, representing a compound annual growth rate exceeding 25%. SK Hynix, Samsung, and Micron dominate this space, with SK Hynix serving as a core supplier for NVIDIA due to its leading capacity. Even when GPUs and HBM are produced, advanced packaging remains the gatekeeper; TSMC's CoWoS capacity is described as 'extremely tight, sold out for 2026.' Capacity targets have risen from 35,000-40,000 wafers monthly at the end of 2024 to 120,000-140,000 by 2026, yet global demand is expected to approach 1 million wafers, with NVIDIA alone locking in 60% through long-term contracts. Intel is countering with EMIB and glass substrate solutions, while ASE and Amkor expand their own packaging capabilities.

Between boards, the interconnect bottleneck forces a transition from copper to photonics as copper cables approach physical bandwidth limits. Large model training requires thousands of GPUs to synchronize, and electrical signal attenuation over long distances drags down cluster utilization. Photonic technologies like silicon photonics and CPO are projected to reduce interconnect power consumption by 30%-50%, though manufacturing maturity remains low. The optical interconnect market is expected to expand from $15 billion in 2025 to $43 billion by 2034. Jensen Huang has aggressively backed this shift, with NVIDIA investing over $6.5 billion in photonics since 2026, including approximately $2 billion each in Lumentum and Coherent, and $500 million in Ayar Labs.

Power conversion and thermal management create further friction around the board. AI servers must step down 48V to less than 1V, a process where traditional silicon devices fail, necessitating GaN and SiC solutions. onsemi estimates power semiconductor value in a 1MW AI rack will double from $50,000 to $100,000, with the GaN/SiC market growing from $2 billion in 2025-2026 to over $8 billion by 2030. Thermal constraints are equally severe; NVIDIA's GB200 NVL72 cabinets consume over 120kW, rendering air cooling obsolete. The global liquid cooling market is projected to surge from $5 billion in 2025 to $27.1 billion by 2035, with adoption in new AI data centers rising from 35% in 2025 to 55% by the end of 2026. Chen Liwu is specifically investing in synthetic diamond technology to address localized heat concentration in high-power chips.

The most enduring bottleneck lies outside the board: electricity. Amazon, Microsoft, Google, and Meta are expected to spend $700 billion on capital expenditure in 2026, a significant portion directed at energy support, yet grid expansion cannot match this pace. Leopold argues the real constraint is not algorithms but electricity, prompting a scramble for power contracts and transformers. Williams is investing $5.1 billion in modular natural gas facilities, GE Vernova holds a 100GW backlog of gas turbine orders, and NVIDIA is backing TerraPower for small modular nuclear reactors. Unlike semiconductor fabrication, power infrastructure involves land, grid access, and approvals, creating construction cycles measured in years rather than months.

Woofun AI data shows that the supply-demand imbalance will persist through specific release nodes, with the first wave of new capacity arriving in the second half of 2027 from SK Hynix's M15X factory and Micron's Singapore and Taiwan facilities. A second wave is scheduled for 2028, featuring Samsung's Pyeongtaek P5, SK Hynix's Indiana, and Micron's Hiroshima factories, leading Chen Liwu to conclude there will be 'no relief before 2028.' However, new capacity does not guarantee a resolution; each GPU generation doubles HBM demand, and hyperscalers have already locked in new capacity via long-term contracts. Historical precedent from 2017-2018 saw DRAM prices collapse 18 months after capacity investment, but the current cycle is three times larger in revenue base with DRAM prices expected to rise 275%-300% from 2025 to 2027. With HBM profit margins reaching 60%-70% and memory manufacturers surpassing $1 trillion in market cap, the potential price reversal window is projected for late 2028 to mid-2029. Leopold's massive short position suggests he anticipates a margin compression event once the infrastructure cycle peaks, betting that physical scarcity in power and materials will outlast the semiconductor boom. This divergence marks a strategic shift where infrastructure scarcity is viewed as a more durable asset class than the chips themselves.