Which Businesses Will Survive AI? | Michael Dempsey

Summary

• Compound evaluates AI businesses for survival across technology cycles. Dempsey’s latest venture portfolio contains 16 biotech companies out of roughly 23, with only three AI investments. The firm waited about 16 months after its 2021 fund close before investing, but he admits it was not aggressive enough in 2023. “Real pain is losing money.”
• Rapid AI revenue growth can conceal a weakening competitive position. In Dempsey’s illustrative inference market growing 10x annually, a company growing 6x is losing ground despite impressive numbers. Providers expanding into one another’s stack could threaten $30 billion businesses within nine months; the question is whether they perform “above replacement level.”
• Dempsey holds Figma long, but the host—not Dempsey—sets out the bull case around AI-enhanced creative work and usage-based spending. Dempsey says Figma’s products took 18 months too long to ship, the stock is still not cheap, and it is not among his four largest positions. He describes his firm moving from Figma to Claude Design, then says Figma MCP has changed the design experience. Salesforce has more inertia than bears expect; he is a little more bearish on HubSpot.
• AI’s political reception could generate volatility even as the technology advances. Dempsey expects government to play “a more radical role than people think” because opposing AI will become politically popular. He sees inequality, accelerating change and fear of being left behind behind the resentment; he doubts that funding cancer-related companies alone will repair technology’s standing.
• Hyperfinancialization is an area Dempsey refuses to monetize as an investor. Despite earning money through sports betting and trading options since 18, he says, “I hate prediction markets,” and rejects investing in Robinhood or the Pump.fun token. His objection is that hypergambling is short-term-oriented and undermines creation and fulfillment. His proposals include raising the gambling age to roughly 25 or 26 and a simple wealth-transfer tax.
• Dempsey favors scientific businesses that turn discoveries into products with control over commercialization. Orbital Materials embeds an AI-designed, heat-absorbing material inside modular data centers optimized for GPU cooling. The attraction is “greater control over your fate,” reducing dependence on someone else to turn a technical breakthrough into revenue.
• Dempsey is bullish on biotechnology’s outcomes while questioning whether AI labs will capture them. He thinks 50% of today’s major fatal diseases will stop killing people in about 20 years, and he is long Eli Lilly. Faster safety evidence, regulatory change and distribution matter alongside discovery; biotechnology’s “tacit knowledge” remains a constraint on the labs.
• Liquid crypto looks more attractive to Dempsey than new crypto venture investment. Venture has been “pretty much dead for two years,” while fully distributed tokens beginning to accrue value can be assessed fundamentally and narratively. After the host calls his forecast of over 20 crypto IPOs too optimistic, Dempsey agrees on 10 within 24 months. He has also spent roughly 18 months trying to buy Chainalysis shares.

Deep dive

1. Long fund cycles make patience necessary and mistiming costly

• Compound’s research starts with scientific or engineering risk across machine learning, robotics, biotech and cryptography. Adjacent areas include materials science, energy and space, but usually only when they carry substantial scientific or engineering risk. Dempsey thinks the firm can recognize an unusual “n of one” opportunity better than rank five similar startups. When an unrelated category suddenly becomes fashionable, the usual response is to skip it. Investing means identifying the particular places where the firm has an edge, then developing those capabilities.

• After raising in 2021, Compound began investing in March 2023, about 16 months after the fund closed. Dempsey distinguishes irritating LP pressure from the actual pain of losing money, but admits the firm emerged too cautiously. Some interesting 2023 companies slipped past because downstream investors were not yet providing convincing financing signals. “That’s the worst part of venture capital—you rely on other people” to invest after you.

• The host’s challenge: remaining relevant may reward participation in a bubble more than restraint. Dempsey says this is partly true, noting that LPs can also be reluctant to admit a commitment was a mistake. He describes fund-office investors committing to invest across several funds and vintages, which can make changing course professionally difficult. Compound raises every four to five years, versus roughly 24-month cycles elsewhere and nine-month deployment during 2021. By its next fundraise, weak companies have had time to fail: “We must be able to maintain the portfolio we build.”

2. AI growth can obscure weak moats and approaching competition

• Dempsey thinks the labs’ displacement of other businesses is “more true than not,” although he admits underestimating some companies’ short-term revenue growth. Continued scaling of compute, data and performance should break intermediate layers; enduring product destinations might still accumulate value. The missing work is defining medium- and long-term moats. “The only free thing in building a company is thinking.”

• Inference providers such as Fireworks, Together and Modal illustrate the measurement problem. If the market grows, say, 10x year over year while a provider grows 6x, headline growth conceals deteriorating relative performance. Once market momentum settles, companies genuinely “above replacement level” should separate from those carried by subsidies and exceptional demand. An apparently successful asset can already be underperforming.

• The stack is converging: inference vendors buy GPUs while neoclouds develop software to optimize inference. Dempsey’s warning is conditional but sharp: an asset valued around $30 billion could face displacement within nine months if a competitor decides to acquire a specific part of the stack it owns. Margin erosion or restricted GPU access could be equally damaging. Understanding the entire value chain matters more than accepting each company’s account of its supposedly distinct niche.

• Secondary sales offer less protection than seed investors suggest. A successful early investment can become hundreds of millions of dollars of exposure, with meaningful liquidity concentrated in very few assets. Compound therefore assumes it may “hold onto these things forever.” Its public-market work helps it understand the biggest technology companies, including which neoclouds might be attractive to own, short or avoid.

3. Figma tests whether applications can retain value beyond model access

• At an assumed valuation around $2 trillion each, OpenAI and Anthropic might interest Dempsey over ten years; over 24 months, he prefers to ignore them. He is neither long nor short Salesforce, whose inertia he thinks investors underestimate. HubSpot draws a more bearish response. He rejects a blanket short of SaaS.

• His work on roughly 100 public SaaS companies tracks credible open-source clones through GitHub repositories and stars. Activity turned sharply upward about six months earlier, and Claude Opus’s ability to decompile software strengthens the concern. The qualification matters: enterprises will not simply purchase random open-source forks. Nevertheless, falling replication costs create structural pressure on many incumbents.

• The host lays out Figma’s bull case: that creativity and design need grounding in a product environment, that AI improves the experience, and that usage-based spending could reach 10x expectations. Dempsey confirms he has a long position, says Figma’s products took about 18 months too long to ship, and says the stock has fallen substantially since its IPO but remains expensive. It is not among his four largest positions.

• Dempsey describes his firm moving from 100% Figma to Claude Design, saying, “I would say Figma died.” The host argues that people criticizing Figma may be behind because they have not used Figma MCP. Dempsey then says that with MCP, they are designing things where his “slopometer is off.” He objects to founders’ immediate surrender to the proposition that anything overlapping a model lab must inevitably lose.

4. Technology’s social backlash can become a political market risk

• Dempsey links AI’s reception to tech’s insularity, founder-led communications and promises of abundance delivered alongside threats to livelihoods. He sees society as a pendulum rather than a “runaway freight train,” making today’s cultural direction reversible. The host’s Cleveland observation sharpens the point: anger toward AI and data centers was substantially stronger offline than a Twitter-centered view suggested.

• In the short term, Dempsey expects government to play “a more radical role than people think,” because opposing AI and technology will become popular with politicians. He does not know whether that matters over the long term, but expects volatility. His darker forecast is that more attempts at violence against major technology leaders are “very likely” over the next five to seven years.

• The host asks whether financing more worthwhile companies, such as cancer-treatment businesses, would restore trust. Dempsey doubts it, attributing the backlash to the wealth gap, accelerating change and fear of being left behind. He says people adapt well but process change poorly in real time. The host adds that quality of life may improve while relative outcomes still leave people unhappy; Dempsey’s point is that technology is often treated as first-class and people as collateral.

5. Financial nihilism conflicts with creation and civic responsibility

• Dempsey’s opposition to hypergambling comes with an acknowledged contradiction: quantitative sports betting helped support his college living expenses, and he has traded options since 18. “I hate prediction markets,” he nevertheless says. He understands why someone who benefited can look like he is pulling up the ladder, but thinks today’s scale and influence on young people are destructive.

• The host’s pushback is worth keeping: an investor also allocates capital rather than creating a product, so where is the line? Dempsey says creation is also about human fulfillment. Writing, photography and painting get what is “floating around in your head and your heart out into the world.” His concern is losing the idea of life as a long craft that one is constantly working on, not merely failing to found a company.

• His proposed interventions are gambling eligibility around age 25 or 26 and a simple, substantial wealth-transfer tax on large inheritances with no way around it. He also declines Robinhood and Pump.fun investments. Robinhood’s history of gamification and push notifications troubles him because those mechanisms attract users and trigger behavior.

• The host proposes building more interesting or fun companies rather than racing from zero to $100 million ARR in 18 months. Dempsey partly agrees, but cautions against a purely European romantic framing: he describes the Gilded Age as having a technocratic romanticism grounded in civic duty. Romanticism also needs management alignment; he contrasts Figma’s design-centered rhetoric with its CFO’s emphasis on engineering growth. He says some contemporary tech leaders lack that sense of civic duty.

6. Materials become businesses when the invention controls a product

• Compound’s latest venture portfolio contains roughly 23 companies, including 16 biotech businesses and three AI investments. Those three are Prime Intellect, which moved from distributed compute toward open-source post-training; the General Intelligence Company, an agent platform for solo or agent-native businesses; and Orbital Materials. The allocation reflects a search for durable company structures rather than maximum exposure to the current AI boom.

• Orbital’s foundation model helps create a novel material that absorbs heat at a precisely designed temperature. Embedded in a modular data center, it can dissipate heat efficiently at the temperature GPUs operate, including for a particular class of GPUs expected soon. The physical laboratory matters: Compound defended generating physical data when others wanted everything done in silico. The commercially interesting output is the cooling product, rather than an isolated scientific discovery.

• This structure offers “greater control over your fate.” Dempsey contrasts it with robotics companies dependent on systems integrators or customers to move from pilots into broad deployment. A technically successful company can still have its commercial outcome determined by another party. He wants scientific capability connected to a route to market the business can influence.

• Research relationships can take years to mature. Dempsey read Alex Kendall’s Cambridge PhD thesis two and a half years before Kendall started his autonomous-driving company, then revived their correspondence after hearing about the venture. Long fund cycles permit that patience. The failure mode remains explicit: Compound could become “an overpaid research firm,” producing excellent ideas without converting them into good investments.

7. Biotechnology’s upside need not belong to the AI labs

• Dempsey’s explicit forecast: “I think 50% of the major diseases that are killing us won’t kill us in about 20 years.” He thinks he and the host will live beyond 120. His latest investment concerns AI-based psychiatry; he sees clear benefit in a hypothetical 100-fold increase in American mental-health services. These are the outcomes driving his optimism, without certainty about which institutions deliver them.

• Bringing biotech inside AI labs makes narrative sense, especially around an IPO, because the potential market may exceed software. Operational ownership is another question. Dempsey calls it “very likely” that everyone currently working in biotech at Anthropic will be elsewhere in five years. Biotechnology contains substantial “tacit knowledge” that cannot simply be recursively improved in the manner assumed for other AI capabilities.

• He is long Eli Lilly, highlighting its move toward direct-to-consumer distribution. More biotech companies may create valuable assets, but must choose between relying on incumbents’ clinical-trial infrastructure, regulatory expertise and scale or building independently. His caution against straight-line extrapolation is Novo Nordisk, which he says trades below where it stood when it began investing in GLP-1s. He thinks cash-rich, capable companies could grow at 50% year over year to hundreds of billions in revenue.

8. Faster safety evidence could change biotech’s financing structure

• The host describes biotech’s unusual public-market structure: tiny teams list early, then a Phase 2b result can determine whether there is a valuable company or effectively a shell. Dempsey expects AI to establish whether medicines are safe for humans more quickly. That could reduce cost, support less regulation and accelerate commercialization. His thesis extends beyond finding candidate drugs to changing the apparatus that develops them.

• He cites Montana’s “Right to Try Act” as an early experiment in access to drugs deeper into clinical trials. He also favors human challenge experiments: infect volunteers, administer a candidate drug against a blinded control, and keep participants at a center for two weeks. Such trials mainly address respiratory viruses today; he thinks the approach can extend to other areas.

• Producing an answer is insufficient when an intervention affects the brain; Dempsey wants the full trace behind it. He also expects many people to experience negative effects from biological self-experimentation over the next ten years. His investing stance sits between “hyper-naive, very optimistic people” and “hyper-blasé, very intelligent and talented people.” Knowing when to ignore these perspectives is part of the claimed edge.

9. Crypto’s liquid assets offer clearer opportunities than its venture market

• Crypto remains “one of the most asymmetrically tilted upward areas” for Dempsey. Organizing capital and programming more efficient systems should retain value through social pendulum swings; rare-disease research is one possible use. He is a Bitcoin enthusiast. Disclosed holdings include Hyperliquid, HairDAO and Kinetic, plus roughly 13 other crypto assets; Helium was a historical position.

• Crypto venture has been “pretty much dead for two years,” he argues. Regulatory confidence is improving, but AI has absorbed the talent that might build elsewhere. Tokenization also damages the long-term incentives he cares about. Liquid markets look more interesting because fully distributed tokens are beginning to accrue value and can be understood fundamentally and narratively. Beyond new perpetual DEXs, he sees few innovative venture attempts.

• Cash-generating businesses choosing tokens are widening the investable universe; potential equity listings add another route. Dempsey initially suggests over 20 crypto IPOs within 24 months. After the host calls that too optimistic and proposes 10, Dempsey agrees. His examples of public-market crypto exposure include Coinbase, BitGo, Circle and Securitize. Fireblocks and Ledger are possible future candidates, with the qualification that many such businesses are not ready.

• Chainalysis is the private company he has tried to buy for roughly a year and a half. He says he is “constantly under pressure by some investor,” without clarifying what that means. The attraction starts with a “very, very large discount” in the last round. If crypto becomes more important and cybercrime keeps growing rapidly, he thinks Chainalysis is one of three reasonable companies with a right to win. He cites proof points in crypto security and the company’s work to win difficult clients.