Which Businesses Will Survive AI? | Michael Dempsey

Michael Dempsey’s central test for an AI business is whether it can endure successive innovation cycles, not simply grow quickly in the current market. Compound seeks technically difficult companies with distinctive, potentially large opportunities, and is better at identifying an “N of one” business than picking among many similar startups in crowded categories. Dempsey says AI’s durable winners remain uncertain: multiple businesses may succeed, or leading model labs may capture much of the value. He leans toward lab-driven disruption, as larger models, more compute, and better data threaten companies that depend on product-layer advantages. Some canonical GPT wrappers, he argues, do not deserve to exist.

For inference providers, rapid growth is not proof of durability because demand is heavily subsidized. Dempsey’s test is whether a company remains “good above replacement” once the market settles; a fast-growing provider that performs worse than alternatives could then look far less attractive. He expects substantial demand for intelligence but says it is unclear which businesses can sustain their revenue trajectories. He also expects infrastructure companies to expand across the value chain: inference providers are buying GPUs, and some neoclouds are building software to optimize inference. Consolidation and broader control of the stack may matter, but even a roughly $30 billion asset could be outcompeted within nine months, suffer major margin erosion, or be constrained by GPU supply.

Dempsey is generally short toward SaaS, though he expects a small subset to remain attractive. He cites a sharp rise, about six months earlier, in GitHub projects cloning public SaaS products, and says improved software decompilation by Opus could accelerate open-source alternatives. Enterprise reluctance to adopt untested forks may slow the threat, but he sees the trend as structural risk to some SaaS businesses. He is neither long nor short Salesforce and thinks its inertia is stronger than many assume; he is somewhat more negative on HubSpot. He is long Figma, though it is not among his four largest positions. That view depends on design retaining a canvas, AI improving the workflow, and AI-native software usage and spending growing under usage-based pricing. Risks include agents displacing design, AI-generated “slop,” valuation, executive turnover, and competition from Anthropic’s Claude Design; he does not believe AI models winning means design is over.

Compound’s current venture fund has about 23 companies, including 16 in biotech and three in AI: Prime Intellect, General Intelligence Company, and Orbital Materials. Prime Intellect began with distributed compute and also works on post-training open-source models; General Intelligence is building an agent platform for single-person or agent-native companies. Orbital uses foundation models to develop novel materials and has a physical lab that can generate data for further model development. Dempsey stresses that a deep-tech company also needs a route to market. He finds Orbital’s case more compelling because its AI-developed material, engineered to extract heat at GPU operating temperatures, is combined with a modular data-center product. Compound’s earlier AI investments include self-driving company Wave and Runway, both of which became multibillion-dollar companies.

Dempsey sees potential for AI to make drug development cheaper by helping researchers determine more quickly and safely whether drugs are suitable for human testing. He expects regulatory reform to expand the number of possible assets for addressing disease, while acknowledging that biology relies on tacit knowledge that may resist the recursive self-improvement some expect. He points to Montana’s Right to Try Act and human challenge trials as possible ways to produce human evidence more quickly, while describing challenge trials as deliberate exposure to illness followed by comparison of treated participants with a blinded group. His latest investment is in AI-oriented psychiatry, and he argues that providing 100 times as much mental-health care in America would be beneficial.

Dempsey says research is the first-class citizen at Compound and informs its investments; historically, about 70% of the firm’s investments have come from outbound founder outreach. Because seed investors may need to hold successful companies for a long time, he focuses on whether they can survive successive innovation cycles rather than selling secondary shares during a frenzy. Compound waited roughly 16 months after closing a fund in 2021 before resuming investment around March 2023; Dempsey says it was not aggressive enough and missed some interesting companies, but describes the cost mainly as annoyance and rejects investing just because others are doing so. He values founders who understand a difficult business’s risks and history rather than presenting success as inevitable.