Investors should reduce exposure to broad AI infrastructure and high-end hardware makers like NVIDIA (NVDA) over the next 2 to 3 years due to elevated valuations, excessive private debt, and emerging foreign chip competition.

Reallocate capital toward energy and power infrastructure, specifically targeting low-cost utility generation and electrical grid equipment that solve critical data center electricity bottlenecks.

Hold Taiwan Semiconductor Manufacturing Company (TSM) as an essential semiconductor manufacturing play, using its CHIPS Act-funded US factory expansion over the next 2 to 3 years as a hedge against geopolitical risk.

Capitalize on open-source software growth by investing in Alibaba Group Holding Limited (BABA), which is rapidly gaining global developer market share through its Qwen model ecosystem without incurring heavy hosting costs.

Prioritize investments in fixed industrial automation and domain-specific enterprise software applications over generic AI models and complex bipedal humanoid robotics, which face severe industry consolidation over the next 1 to 2 years.

AI Infrastructure & Private Credit (Macro Sector)

• The AI sector represents approximately 45% of total US stock market value, compared to roughly 30% at the peak of the dot-com bubble.
• The Buffett Indicator (market cap to GDP ratio) sits at 240%, well above the 120% recorded during the dot-com era, signaling high vulnerability to an economic correction.
• Major hyperscalers carry an estimated $1.6 trillion to $1.7 trillion in off-the-books debt and private credit to finance data center construction.
• The rapid rise of open-source and edge-deployed models threatens the premium subscription and cloud margins that data center builders rely on to service their massive capital expenditures (CapEx).

Takeaways

• Exercise caution regarding broad AI equity exposure, as overleveraged infrastructure spending and lofty valuations heighten the risk of a market correction over the next 2 years.
• Watch for enterprise AI value to accrue in real-world software applications and efficiency gains rather than pure cloud infrastructure and raw compute hosting.

NVIDIA Corporation (NVDA)

• NVIDIA has participated in structured compute deals, including an estimated $500 billion partnership with private equity and asset managers like BlackRock (BLK), Blackstone (BX), and Carlyle (CG) to securitize chips and data center capacity.
• US export controls limiting high-end GPU sales (such as the H-200 and Blackwell lines) to China have unintentionally accelerated the domestic Chinese chip industry.
• Chinese competitors such as Moore Threads, Cambricon, and Biren were forced to develop indigenous chips due to US sanctions and are expected to become competitive exporters within 2 to 3 years.
• Despite restrictions, international data centers are frequently used by offshore labs to run training workloads, mitigating the long-term barrier created by hardware embargoes.

Takeaways

• Long-term revenue growth from high-end GPUs may face headwinds as Chinese alternatives mature and hyperscaler capital expenditure cycles decelerate.
• Keep track of private credit securitization structures backing GPU deployments, as any debt distress in data center operators could dampen forward hardware demand.

Taiwan Semiconductor Manufacturing Company (TSM)

• Geopolitical tension around Taiwan remains central to the semiconductor supply chain, but physical capture of operational fabrication plants (fabs) is unlikely due to their reliance on global maintenance, specialized supplies, and overseas engineers.
• The US CHIPS Act is deploying billions of dollars to replicate advanced node manufacturing domestically (such as facilities in Arizona) over the next 2 to 3 years to mitigate single-point-of-failure risks.

Takeaways

• TSMC remains the critical bottleneck for advanced AI chips worldwide, but geographical diversification into the US provides a medium-term hedge against cross-strait geopolitical risks.
• Monitor execution timelines for domestic US fab facilities, as these will dictate when Western reliance on Taiwanese manufacturing begins to decrease.

Alibaba Group Holding Limited (BABA)

• Alibaba’s open-weight AI model family, Qwen, has achieved over 1 billion downloads with more than 180,000 derivative models created by the developer ecosystem.
• Chinese open-weight models have expanded from 2% to 61% of OpenRouter traffic over a two-year period, closing the performance gap with leading closed-source models to just 2 to 3 months.
• The open-source distribution strategy allows Chinese hyperscalers to expand global software market share without bearing the full CapEx burden of running global inference servers.

Takeaways

• Alibaba is cementing a dominant position in the global open-source AI ecosystem, which builds long-term developer mindshare and software distribution reach.
• Open-weight software strength offers alternative monetization routes through enterprise tooling and cloud integration rather than direct API token markups.

Frontier AI Labs & Closed-Model Ecosystem

• The business model of spending $1 billion+ to train closed frontier models is under economic pressure, as competitors can distill and match frontier reasoning capabilities for $2 million to $10 million.
• Leading private labs face heavy revenue concentration, with a small number of enterprise buyers driving the majority of API volume (e.g., Meta spending $100 million to $200 million per month on Anthropic tokens).
• Enterprise adoption surveys show that 80% to 90% of successful enterprise AI deployments rely on small, specialized models (millions to low billions of parameters) rather than multi-trillion-parameter frontier systems.

Takeaways

• Closed-source frontier AI providers risk rapid commoditization from open-weight models, challenging premium API pricing models.
• Investors should favor companies building proprietary enterprise workflows, domain-specific smaller models, and practical application layers rather than generic foundational model wrappers.

Humanoid Robotics & Industrial Automation

• Over 150 to 200 humanoid robotics companies are actively developing systems, but aggregate global commercial deliveries remain low (in the tens of thousands).
• High hardware costs, balance instability, and maintenance challenges of bipedal designs are prompting manufacturers (such as Unitree) to prioritize upper-torso and wheeled configurations for commercial environments.
• High levels of factory automation (such as "dark factories" in China) rely heavily on single- or dual-arm fixed industrial automation rather than full humanoid forms.

Takeaways

• Expect severe consolidation in the humanoid robotics space over the next 1 to 2 years, with the market likely narrowing down to single-digit winners.
• Near-term capital and commercial traction will favor specialized industrial automation and wheeled torso robots over complex bipedal walking humanoids.

Energy & Power Infrastructure

• Power availability and electrical transmission are emerging as primary structural constraints for large-scale AI compute clusters.
• Ultra-high-voltage power transmission and massive utility-scale solar and wind buildouts in western regions have enabled energy costs of $0.02 to $0.03 per kilowatt-hour, roughly 10 to 15 times cheaper than many US regions.
• Deploying compute capacity near generation sites reduces power transmission loss and lowers total operating expenditure for data center operators.

Takeaways

• Low-cost power generation, utility infrastructure, and advanced electrical grid equipment represent critical value capture points in the AI technology stack.
• Look for investment opportunities in power infrastructure, energy transmission, and data center efficiency solutions as grid constraints become more binding.