The Venture Capital Cycle: AI Markets, Compute and the Cost of Moving Too Fast

Harry Stebbings brings together Jason Lemkin and Rory O’Driscoll to discuss the changing venture market, from partner compensation and the size of software markets to AI infrastructure and the difficulty of allocating capital through a boom.

The discussion begins with Everett Randle joining Benchmark and the economics of moving between venture firms. The panel considers how salary, ownership and carried interest can matter differently for a partner whose returns may take many years to materialize.

The investors debate whether venture compensation is transparent enough and how much a successful partner can earn. Some details are speculation rather than confirmed terms; their broader point is that a share of a fund’s profits can be more consequential than annual salary.

Lemkin distinguishes marks on paper from cash returned to investors. He discusses a 2017 fund whose estimated multiple he puts around five times while noting that distributions have not yet established that result. The long wait between an apparent win and a realized return changes how partners think about compensation.

Stebbings turns to Revolut and the challenge of assessing market size when a company reaches enormous scale. The panel discusses reported fundraising and valuation figures, using the company as an example of how the definition of a market can expand beyond an initial product.

Lemkin says he likes to see a company at $100 million in annual recurring revenue still holding less than 1% of the market. That is an investing heuristic, not a measurement of Revolut: a large remaining market offers room to compound rather than relying on taking every available customer.

O’Driscoll argues that founders can begin with a narrow product while entering a broad adjacent market. The relevant question is not whether the first feature alone supports a huge company, but whether customers and capabilities provide a credible route into neighboring products.

Stebbings emphasizes the founder’s ability to broaden the company as it scales. O’Driscoll adds that local conditions can confer an advantage: in discussing Spotify, he offers the possibility that the European licensing environment helped its early development, rather than treating that as an established explanation.

The panel questions how much investors can know about a future total addressable market. Existing categories are easier to count, but a product that changes behavior or replaces labor may create demand that historical software budgets fail to capture.

Lemkin worries that the rush into vertical AI could resemble the pandemic software boom. When a new technology creates an urgent reason for many potential customers to buy at once, the first cohort can produce exceptional growth without proving that the same pace will persist.

Stebbings pushes back on the analogy: pandemic demand had an unusually clear cause and expiration, while AI changes what software can do. The disagreement is about whether rapid adoption signals a durable new market or pulls years of demand into a short period.

O’Driscoll sees an important distinction between temporary pandemic conditions and a technological change. Even so, he warns against simply extrapolating the earliest adoption rate: implementation takes effort and customer organizations need time to absorb new workflows.

The question is not just whether customers want the software but whether it creates enough value to unlock a larger budget. An application that eliminates a painful task may reach buyers who would never have bought a conventional tool; that possibility must still be tested against actual adoption and retention.

Lemkin returns to the risk that an initially impressive vertical market turns out to have a limited number of buyers. Early revenue can make an investment look obvious while hiding the need to find new customers, expand the product or sell into adjacent workflows.

The panel contrasts applications with broad horizontal demand against products designed for one profession. The latter can be excellent businesses, but the venture outcome depends on the number of potential customers, the value of each account and the ability to grow beyond an initial use case.

Lemkin argues that if growth slows because the market is narrower than expected, acquisition can become a sensible outcome. His suggestion that a large share of B2B AI venture investments could lose money is a forecast, not a measured result.

The investors work through what a market-size calculation means in practice. Ten thousand customers paying $10,000 each imply $100 million in revenue; a much larger customer count at the same price implies a much larger opportunity. These are illustrative assumptions, not verified customer counts.

They discuss legal and other professional services, where AI might capture spending that has historically gone to people rather than software. Whether a product can do so depends on quality, trust and adoption, and the panel’s examples are predictions rather than guarantees about legal work.

Stebbings cites his investment in Solve Intelligence as an example of AI addressing patent work. He argues that a successful application can start with a specific job and then sell more broadly into a valuable workflow; the investment relationship is relevant to his perspective.

The discussion moves from application markets to the infrastructure underpinning them. The participants ask whether the economics of AI companies can support the rapidly increasing cost of computing, and who ultimately bears the risk of long-term commitments.

The panel considers arrangements involving OpenAI, Microsoft and Oracle. Their discussion of obligations, financing and risk allocation is an investor interpretation of reported deals, not an inspection of contracts or a verified calculation of any party’s exposure.

The investors distinguish a compelling commercial rationale from a guaranteed return. A provider may seek capacity to serve a fast-growing market while its partners expect a long-lived asset; both can still be vulnerable if demand or financing develops differently from their assumptions.

The conversation turns to reports about Poolside and a proposed data-center project associated with CoreWeave, including a two-gigawatt figure. The panel treats the scale and possible motives as topics for discussion; it does not establish that that capacity is operating or verify the contractual terms.

For a startup building models, securing sufficient compute can become a strategic constraint. O’Driscoll suggests that scarce capacity and large buyers’ prior commitments could push a company toward building infrastructure, but stresses that this is a possible explanation rather than confirmed inside knowledge.

The panel imagines how an investor would react if an ordinary startup pitched building its own data center. What began as a software investment can gradually become a far more capital-intensive business, raising the amount needed before a company reaches cash-flow break-even.

Stebbings discloses an early investment in the founder’s earlier business that rolled into Poolside and describes a roughly 50-fold paper gain. O’Driscoll draws a distinction between the rationale for seeking compute today and whether that decision will ultimately prove rational; the answer may take years.

O’Driscoll proposes that a shortage of available capacity might leave a model company with an unattractive choice: build, delay its ambitions or stop. He explicitly presents that account as an assumption, and says that a decision that makes sense under current constraints could still be regretted later.

Lemkin suggests the market has grown faster than founders originally anticipated, potentially making larger funding rounds possible. The panel agrees that expectations and risk appetite have risen, but the future availability of capital is uncertain.

O’Driscoll compares AI infrastructure with other boom-and-bust markets. During a shortage, buyers make commitments to secure supply; when new capacity arrives and demand weakens, prices can fall. He thinks a similar adjustment is possible, though its timing is hard to predict.

Stebbings questions whether an AI bust would involve capital leaving the sector for years. O’Driscoll replies that useful technology and a serious investment correction can coexist: a bubble can unwind if the pace of adoption falls short of what new infrastructure was built to serve.

As a hypothetical, a data center costing $2 billion to build might have to sell for $1 billion if demand disappoints. At that price, building another facility could be unattractive. O’Driscoll cites the historical bandwidth cycle to explain the mechanism, not to claim this outcome is certain.

Lemkin describes an early B2B AI investment he is making that could use continuous inference at a scale far beyond his previous investments. The desired workload is not economical at current prices, so the opportunity depends in part on compute becoming cheaper.

Lemkin points to Opus Clip, another investment, as an example of improving compute efficiency. The product initially offered a few clips before generating more on request, avoiding work customers might not need; he believes future applications could nevertheless use much more inference.

Asked whether this thesis points to Nvidia, the panel notes existing exposure through holdings such as retirement accounts and QQQ. The issue is how much additional concentration an investor wants, rather than a simple conclusion that an exciting technology makes any price attractive.

O’Driscoll believes powerful technologies often attract overinvestment and eventually a retrenchment. His practical question is how to participate in the upside while retaining the ability to survive a downturn and invest afterwards, not whether an investor can predict the exact turning point.

The panel calls this diversification across time. Stebbings says funds are again being raised on compressed cycles; investing too quickly can make a portfolio heavily exposed to one set of prices and assumptions even when it owns several different companies.

A simple example of funds returning five, three and one times capital does not automatically yield a three-times aggregate return: the answer depends on the relative fund sizes. The exchange illustrates why a sequence of funds and a weighted portfolio calculation are different things.

O’Driscoll argues that aggressive investors can look smartest just before a correction and suffer most afterwards. His aim is to take enough risk to benefit from the boom without taking so much that a downturn ends the ability to keep investing.

The panel asks whether limited partners will continue allocating to venture if returns lag public markets. O’Driscoll says that, over the preceding five years, venture returns had been much lower than public-market returns; that is his characterization, not a universal comparison for every fund or measurement period.

Lemkin adds the operational cost of venture investing: evaluating many managers, tracking funds and maintaining relationships requires time and staff. For an allocator with a small venture allocation, those costs make the hurdle for committing capital higher.

The discussion broadens to the market cycle and the consequences of rapid AI adoption. The panel considers how changing expectations can encourage investors to move faster, even when the underlying outcomes will take years to become clear.

They also discuss AI-generated content and the responsibilities of platforms. The concern is how a company should handle harmful or unauthorized material as tools become more accessible; the exchange describes possible risks and personal experiences, not adjudicated findings about particular incidents.

The speakers weigh the tension between enabling useful creative tools and preventing misuse. They recognize that moderation decisions and enforcement become harder when material can be generated and distributed rapidly, and that the consequences for individuals can be serious.

Stebbings begins an agree-or-disagree round with a hypothetical: will Replit reach $1 billion in annual recurring revenue by the end of the following year? Lemkin is optimistic, O’Driscoll is uncertain about market size, and Stebbings disagrees. None presents the target as an achieved result.

Stebbings argues that Replit’s addressable audience may be narrower than Lovable’s and warns that churn and maturing cohorts could slow growth. He has an investment in Lovable, which is relevant to his comparison.

Lemkin, a Replit user, says its agent improved markedly during the roughly 110 days he had used it. He believes easier software creation will enable new categories of applications and make evaluating early-stage products harder when a polished initial product is much easier to build.

O’Driscoll frames the market question as tools spending versus labor spending. If AI tools merely replace other tools, growth might eventually flatten; if they also replace spending on agencies and development work, the potential market could be much larger. That is a scenario, not a forecast of a specific winner.

Asked to choose between holding shares in Deel and Rippling, O’Driscoll initially hesitates because Scale has invested in Papaya, a competitor. He describes payroll as a large market: every employer must pay staff, but in the United States new vendors must often displace an incumbent.

O’Driscoll sees an international opportunity in helping employers manage workers across jurisdictions. He says Papaya focuses more on mid-market and larger customers while Deel started lower and expanded upward. The comparison reflects his judgment and his firm’s commercial interest.

Lemkin ultimately chooses Rippling while acknowledging uncertainty about which company will prove more defensible. He says Stebbings, who has invested in Deel, and O’Driscoll have information he does not; his preference is an opinion, not a finding about either company’s future performance.

Why invest in a company at $1 million in annual recurring revenue if even $1 billion can feel early? Lemkin says his edge is helping young businesses scale go-to-market, and that departing from his own area of competence has produced losses. A later-stage investment may have a different risk profile without matching his skills.

O’Driscoll agrees that early and late venture require different capabilities, even though both are grouped under one asset class. A small check into a young business and a huge investment into a mature company can have very different time horizons and sources of return. The panel closes by noting the long time it can take for venture outcomes to become clear.