The rise of capable, low-cost Chinese AI models like Kimi forces a US debate. Policymakers and incumbents like OpenAI hint at security risks and advocate for bans. Meanwhile, free-market proponents argue that restricting access would stifle innovation and inflate costs for US companies, creating a core tension between national security and economic competitiveness.
VCs may analyze an acquisition based on a 3x return over their last round. For a founder, the math is different. A life-changing financial outcome is only worth passing up if they genuinely believe they can build a company 10x larger. A potential 3x increase isn't enough to justify the immense personal risk and multi-year effort.
Unlike traditional software with zero marginal costs, foundation models have massive, ongoing training and inference costs. This prevents them from simply slashing prices to crush cheaper open-weight competitors. They must maintain high prices to recoup R&D and satisfy investors, creating a permanent opening for lower-cost alternatives.
Applications relying solely on generic, off-the-shelf foundation models will eventually hit a performance ceiling. Achieving superior, order-of-magnitude better results for specific workflows requires building a "micro model" through custom data labeling, fine-tuning, and creating a unique reasoning layer to create a defensible product.
A valuation disconnect exists in the AI venture market. Companies raising a Series A on $2-5M revenue can command $300-500M valuations. In contrast, growth-stage companies with ~$100M in revenue raise at $1-1.5B, a much lower multiple. This makes later stages appear more attractive on a risk-adjusted basis.
OpenRouter is wise to explore a sale now. The market has validated the need for a model routing layer, but this functionality is rapidly becoming a feature that larger platforms will build themselves. Selling now captures maximum value before the layer becomes fully commoditized and embedded elsewhere, representing a classic "sell at peak hype" moment.
The vast majority of spending and market capitalization in AI today is in the infrastructure layer—compute (NVIDIA), foundation models (OpenAI), and data services. The entire application layer's revenue combined is a rounding error in comparison, highlighting a massive, though likely temporary, imbalance in where value is currently being captured.
The valuations of hyperscalers, NVIDIA, and the broader tech market are fundamentally dependent on the continued exponential ARR growth of the two leading foundation models. A slowdown in their growth would trigger a systemic market dislocation, as vast capital commitments are predicated on this trajectory continuing. Their growth is the lynchpin.
Stripe, a high-growth private company, is attempting to acquire PayPal, a larger but slower-growing and undervalued public competitor. This move, executed with PE firm Advent, aims to dramatically increase market footprint by acquiring assets at a low multiple, despite the risk of diluting Stripe's own growth rate and adding immense operational complexity.
The increasing use of tranched financing, where investors pay different prices in the same deal, signals a market segment is overheating. It's a mechanism for lead investors to capture the excess returns that momentum investors were previously enjoying, signaling that the period of easy, outsized returns in that stage is likely ending.
A strange market inversion is occurring. Highly speculative, capital-intensive deep tech companies (e.g., nuclear) are going public via SPACs, while massively profitable, scaled companies like Stripe and Databricks remain private. This defies the traditional roles of public and private markets for managing risk and providing liquidity.