Songyee Yoon on Power of Compounding

Interview with Songyee Yoon, founder and managing partner of Principal Venture Partners, who says the firm focuses on investing in AI-native companies and uses specific criteria to distinguish genuinely native AI startups from those opportunistically chasing the AI boom. The discussion is qualitative guidance for venture allocation and deal sourcing in the tech/AI private markets; no new financing, valuations, or market-moving data were disclosed.

Analysis

The non-obvious leverage in the current AI cycle sits squarely in the peripheral layers: data pipelines, model-ops, and inference-cost engineering rather than raw model size. Over the next 12–24 months, companies that compress per-inference cost by 3–5x (via better quantization, tiling, or sparse execution) will convert incremental demand into sustainable gross margins; that mechanic favors specialized infra suppliers and software vendors that embed cost-reduction into their stack, not generalist SaaS rebrands.

Second-order supply-chain winners include advanced-node foundries and memory suppliers (capacity-constrained vendors capture outsized pricing power), plus synthetic-data and labeling platforms that control high-quality training inputs — these firms create moats that are far stickier than ephemeral model architectures. Conversely, consultancies and legacy enterprise apps that layer ML as a feature without rearchitecting data flows will face shrinking implementation margins and higher churn as customers move to integrated model-ops providers.

Key downside catalysts are commoditization (open-source models + cheaper GPUs/TPUs), tightened late-stage venture funding that raises mortality for capital-intensive model plays within 6–18 months, and regulatory constraints around sensitive training data that could increase compliance costs by 20–40% for certain verticals. A reversal can occur quickly if inference pricing falls faster than demand growth (months) or more slowly if enterprise adoption and M&A continue to concentrate spend into a handful of platform winners (1–3 years).

The practical portfolio implication: bias into durable, margin-capturing infra and data-moat plays while financing optionality in private, conviction-stage AI-native startups via secondaries; underweight high-multiple “AI-labeled” software names lacking clear per-inference economics or proprietary data streams.

Key Decisions for Investors

• Buy a tactical 6–12 month call spread on NVDA (size 1–2% notional): asymmetric upside to capture continued GPU pricing/volume tail while capping premium. Risk: export controls or sudden GPU supply relief; reward: ~30–60% upside if training demand normalizes and ASPs hold.
• Initiate a 12-month relative value position long TSM (TSMC) vs short INTC (size 1–1.5% net): TSMC captures advanced-node scarcity and foundry economics while Intel remains capital-intensive. Risk/Reward: asymmetric — ~20–40% upside potential vs concentrated execution/regulatory risks.
• Allocate 5–10% of private/venture allocation to secondaries in model-ops, synthetic data, or inference-optimization startups (hold 3–5 years): targets those with recurring revenue and measurable per-inference cost savings; aim for 20%+ IRR. Tail risk: funding winter and high burn rates.
• Establish a 6–9 month hedged short on high-multiple public 'AI-wannabe' names (example: targeted puts or delta-neutral short equity positions, size 0.5–1%): pair short against cloud/infra longs to exploit valuation divergence. Risk/Reward: limited short exposure caps downside if the cycle re-rates broadly; reward is protection against narrative-driven overpayments.