Ex-OpenAI Researcher Warns AI Labs Face Overvaluation Risk Amid Rising Infrastructure Costs

Andrew Ho, a researcher who recently exited OpenAI after an eight-month tenure, has urged tech workers to cash out their equity in private artificial intelligence firms, warning that current private valuations may be unsustainable.

Speaking shortly after announcing his departure to launch an AI data startup, Ho advised former colleagues eligible for secondary tender offers to take liquidity while market conditions permit. He noted that while a doubling of current valuations following a future initial public offering appears unlikely, a 50 percent valuation reduction remains a plausible risk.

Private tender offers have increasingly served as the primary mechanism for employees at late-stage AI startups to monetize their equity prior to a public listing. However, non-eligible staff or recent departures often remain bound by strict post-employment exercise windows and lockup periods. Ho acknowledged holding approximately $700,000 in equity that he cannot legally sell until after an eventual IPO, leaving his personal balance sheet exposed to shifting market sentiment.

The timing of Ho’s comments coincided with heightened market volatility across the broader technology sector, where heavy capital expenditure on AI infrastructure has sparked investor anxiety. Recent market shifts pushed the Nasdaq 100 into correction territory, driven in part by multi-billion-dollar infrastructure spending plans at major tech giants. Shares of Meta Platforms dropped 10 percent in a single trading session following earnings disclosures, while Alphabet recorded an 8 percent decline the previous week over similar concerns regarding return on investment.

At the core of Ho’s financial critique is what market analysts describe as a “Red Queen’s race”—an economic dynamic where frontier labs are forced to spend exponentially more capital on computation and infrastructure simply to maintain their competitive positions. While training costs escalate with each successive model generation, low-cost competitors are increasingly able to replicate advanced capabilities at a fraction of the cost using distillation, a process where smaller models learn from the outputs of larger, more expensive systems. Chinese AI startup Moonshot, creator of the Kimi model, represents one such player utilizing cost-effective approaches to narrow the gap with top labs.

Ho voiced skepticism regarding Recursive Self-Improvement, or RSI—a theoretical concept favored by industry optimists where AI models perform their own research, rapidly accelerating software capabilities and justifying massive compute expenditures. According to Ho, AI development remains constrained by a lack of intuitive research judgment rather than raw reasoning power, noting that models struggle to formulate meaningful experiments or evaluate subtle empirical findings.

While machine learning progress has accelerated rapidly in domain-specific areas with easily verifiable outcomes—such as mathematical proofs and code compilation—progress on messier, unstandardized intellectual tasks has lagged. This performance gap informed Ho’s decision to launch a new venture focused on producing specialized, high-end reinforcement learning datasets designed to improve model capabilities in complex scientific analysis and long-horizon reasoning.

From a market perspective, the primary financial beneficiaries of the current compute buildout remain hardware vendors such as Nvidia and Micron, which supply the high-bandwidth memory and graphics processing units fueling server infrastructure. To escape the financial pressures of escalating training costs, frontier labs will likely need to develop proprietary silicon to bypass chipmaker pricing power or expand higher up the software stack into user-facing applications to capture greater economic value.