20VC: Venture Fund Size, AI Valuations and the Changing Software Market

Harry Stebbings introduces a discussion with SaaS investor Jason Lemkin and Scale general partner Rory O'Driscoll about venture financing, AI infrastructure and software markets. Their figures and market predictions are presented as views expressed in the conversation, not as verified financial data.

Stebbings describes a company with a reported $20 million in annual recurring revenue, rapid growth and profitability that he says has struggled to attract venture investors. He asks whether venture has become broken for companies that do not fit the categories currently attracting attention.

Lemkin describes a vaguely similar portfolio company that raised a small seed round and now plans not to raise again. Even with strong monthly growth and AI-related products, he says, few outside investors want to meet. In that situation, existing investors may have to support a company for longer than expected.

Stebbings sees capital concentrating in fewer names and a narrower definition of what investors find exciting. O'Driscoll agrees that some businesses with strong numbers face difficulty, but notes that highly speculative projects—from space-based data centers to quantum computing—are also receiving funding.

Their distinction is between projects with exceptionally large potential upside and software for specific markets, where investors may fear that AI could undermine the business. Stebbings describes a stronger-than-ever power-law mindset: some investors appear uninterested unless the potential outcome is enormous.

Lemkin says companies outside a small set of fashionable opportunities may have to meet a very large number of investors. He believes a strong company can still find a backer at a reasonable, though elevated, valuation; the difference is that strong metrics no longer produce an automatic term sheet from the most sought-after firms.

O'Driscoll asks whether the issue is a small group of elite funds focusing on extraordinary startups. Stebbings says his concern extends across the venture landscape. Lemkin then turns to fund size: in his view, a fund managing several billion dollars needs some investments that can become billion-dollar positions.

Lemkin argues that large funds cannot depend solely on small positions, because even a successful exit may not materially affect the fund. He describes a billion-dollar position as a necessary outcome for some large-fund investments, and stresses that ownership and dilution determine whether the position can reach that size.

O'Driscoll pushes back with historical data. He says very large outcomes have been rare and doubts that the whole industry can assume enough of them will appear to support many funds using the same strategy. Growth can expand the number of large exits, he says, but past scarcity is still relevant.

Stebbings asks whether the best seed investors can produce a higher fund multiple more easily than someone writing enormous growth checks. Lemkin says seed has become harder: entry valuations are high, ownership may be small, and later investors can reduce early backers' ability to maintain their stakes.

Stebbings offers an example of a relatively small seed check at a high entry price that, even after a large increase in company value, would not transform the fund. O'Driscoll accepts the arithmetic and says early-stage returns should generally compensate for the additional risk over long periods.

The panel nevertheless argues that recent late-stage returns may have inverted that relationship. The speakers cite the growth in the combined reported value of SpaceX, Anthropic and OpenAI as an example. Whether such gains continue at the same rate, they say, is a separate investment assumption.

The discussion moves to CoreWeave after a difficult week for its shares. Stebbings says he owns the stock. O'Driscoll describes AI cloud providers as structurally sensitive because they support expected growth with substantial capital expenditure and debt; their financial position depends on the timing of capacity and demand.

If a provider builds too slowly, it may disappoint customers. If it builds too quickly and demand slows, it may carry expensive capacity and financing commitments. O'Driscoll cites approximate figures from memory and emphasizes that the capital-intensive model does not have the same flexibility as an asset-light software company.

He says the crucial question is not simply whether demand is positive, but whether it grows enough relative to ongoing investment. Strong end-user demand could support both model providers and infrastructure suppliers; a slowdown makes it harder to justify capacity that is being financed in advance.

Lemkin agrees that markets may react sharply to changes in expectations for major AI providers. He argues that some infrastructure suppliers are more vulnerable than companies with distinctive silicon or software. O'Driscoll cautions that comparing a cloud provider directly with Nvidia mixes different types of business.

The investors disagree about the appeal of public-market companies outside the perceived leaders. O'Driscoll observes that less dominant companies can outperform over a given period, while Lemkin emphasizes the risk of holding businesses that lack a durable advantage. Both resist treating a simple ranking as a complete investment thesis.

The panel turns to reported Anthropic results and rumored IPO valuations. O'Driscoll says growth from an extremely high base is bound to slow; he sketches a hypothetical revenue path to illustrate how a large valuation might be rationalized. He explicitly presents that path as a scenario, not a forecast of results.

Lemkin says the reported growth is extraordinary but warns that model-provider revenue may be less predictable than conventional software subscriptions. Developers and applications can switch between providers, so market share, revenue growth and valuation could move sharply even while the overall AI market grows.

O'Driscoll adds that sustained shifts in market share could lead providers to compete more aggressively on price. Strong end-user demand might allow several participants to grow together; if demand slows, the need to win business from competitors could make the economics harder.

Stebbings raises Replit's decision to use an OpenAI model as a default. Lemkin, who says he uses Replit heavily, argues that the story is about routing rather than a permanent exclusive choice. In his account, Replit can choose models based on the quality and price required by a particular task.

Lemkin says application platforms have a strong incentive to balance model quality and cost because their customers are price-sensitive and variable inference bills matter. He treats their routing decisions as evidence about practical economics at a point in time, while acknowledging that the default could change again.

He suggests that if a model offered a sufficiently compelling cost and quality advantage, a platform with routing could direct more usage toward it. O'Driscoll agrees that platforms serving many customers provide a useful test of how much intelligence a task actually requires and what price is justified.

Stebbings asks about Anthropic's code-review product and the implications for independent companies. Lemkin says it is difficult to predict winners in a fast-moving market. Specialist products may have advantages, including the ability to use one provider's model to check another's output, but competition from a model provider is still a serious risk.

Lemkin argues that generation is only part of software development. A model may build most of a requested feature quickly, while engineers still spend substantial time testing and fixing it. He sees code review and the broader path to production as potentially more important than generating the first version.

O'Driscoll notes that the fastest-growing standalone AI applications often operate close to the foundation models themselves. That creates a tradeoff: they benefit from powerful new capabilities and rapid growth, but model providers can enter adjacent product categories. Avoiding that risk may mean pursuing slower-growing software markets.

Lemkin calls code review one of the most important categories in software, while distinguishing importance from the size of any particular venture outcome. His broader concern is how to turn code that works in development into software that is safe and ready to run in production.

O'Driscoll agrees that a dramatic increase in software production could change the whole development lifecycle. He sees related opportunities in testing, sandboxes and other tools required to deploy software successfully. The panel presents these as possibilities, not a guarantee that any individual company will win.

Stebbings cites a reported Modal financing at a $15.75 billion valuation and says its revenue run rate has surpassed $1 billion. Lemkin prefers the business to a provider he views as renting out hardware; O'Driscoll points to Modal's developer products and rapid reported acceleration.

O'Driscoll compares his estimate of aggregate inference-provider revenue with reported model-provider revenue, suggesting that open-weight inference has room to grow. The exact figures are the panel's estimates. The economic question is whether customers can get comparable results at a lower effective price.

The speakers also recognize competitive pressure. Infrastructure owners that buy compute and build data centers may decide to provide inference services themselves. Rapid growth can support ambitious valuations, but the durability of margins and differentiation remains uncertain.

Stebbings stages a hypothetical investment committee on Town, citing a reported financing at a $1 billion valuation. Lemkin says he would pass for now: the team and early revenue may be interesting, but he does not think there is enough evidence for him to have conviction about the outcome.

Lemkin characterizes the decision as an early-stage bet with a large headline valuation. He notes that investors who want exposure can choose to make such bets, but he finds it too early to predict whether the company will produce the scale of exit sought by a large fund.

O'Driscoll urges the group not to become numb to billion-dollar figures. A valuation can sound routine in a fast-moving market while still representing a substantial amount of capital. His point is that the scarcity of truly large outcomes remains relevant even when financing rounds are large.

Lemkin asks whether some private rounds have become disconnected from public-company comparisons. He cites Anduril and several software businesses as examples, contrasting high private revenue multiples with much lower reported multiples for mature public software companies.

O'Driscoll agrees that the gap is real, though rapid growth can make a high multiple plausible if a company continues to expand. He stresses how much of the future is already assumed in that price: a strong company can still turn out to have been purchased at the wrong valuation.

Stebbings notes that large investors may accept a lower multiple on invested capital if they can deploy a very large amount. Lemkin still worries that some AI versions of lower-multiple software categories could eventually trade more like their public peers once growth slows.

The conversation moves to Bending Spoons and a reported additional Berkshire investment. Stebbings asks why a holding company buying software assets with slower growth can receive such a premium. Lemkin compares its strategy to a long-term holding company that selects a few assets from many opportunities and tries to make them durable.

Lemkin argues that Bending Spoons has improved the economics of acquired products and may be buying higher-quality assets over time. O'Driscoll challenges the analogy with long-lived consumer brands: software products may change far more over decades. The panel leaves open whether the holding-company model can be as enduring in software.

The speakers discuss 37signals as an example of a profitable software business built without relying on the conventional venture path. They question the precise profit calculation mentioned in the discussion, but agree that a company can create substantial returns for its owners without seeking a venture-scale exit.

Lemkin compares this with a company that raises one early round and then grows for years. Such an outcome can be attractive to investors and founders, but he wonders whether taking no venture financing at all can be better for a business whose owners want to distribute profits over time.

Stebbings asks whether founders can simply buy out investors. The panel says buyouts are possible but can be expensive and difficult to negotiate. O'Driscoll explains that buying a large stake at a multiple of cash flow may consume years of available cash or create debt risk.

O'Driscoll's broader point is that valuable software companies do not all need to pursue the same enormous exit. Businesses that generate cash can be highly valuable to their founders even if they do not excite a venture fund seeking a very large outcome.

Stebbings asks whether more growth and venture firms will follow General Atlantic toward public markets. O'Driscoll says a firm generally needs substantial scale and fee income because public investors tend to place more predictable value on management fees than on uncertain future carried interest.

He offers a rough fee-income threshold and valuation framework as his own estimate, not a universal rule. On that basis, he expects only the largest firms to have a straightforward path to a public listing, although venture economics could lead investors to apply different assumptions.

O'Driscoll adds a cultural concern: public ownership can make an investment firm more risk-averse and change its incentives. A listing may provide liquidity for founders, but it can also turn the organization into a different kind of institution. Stebbings asks how much that distinction matters once a firm is already very large.

The panel notes that a public company faces continuing pressure to grow assets and launch products. A firm content to remain a small, stable partnership may not fit public-market expectations. The speakers distinguish that pressure from the purely financial question of whether an IPO is feasible.

Stebbings cites a reported 27.1% return for Harvard's endowment and its exposure to venture. O'Driscoll observes that a few very large winners can drive striking returns and that other institutions may also have benefited from concentrated positions. He is careful about which institution held which investment.

The panel speculates that distributions from successful private companies could be recycled into venture funds, increasing the capital available for new investments. Lemkin compares the reported performance with the exceptionally high university investment returns of the previous boom; the participants disagree in emphasis about how much scale changes the comparison.

In the final segment, Lemkin predicts that Tesla's new Roadster could precede a future combination of Musk's companies. He also predicts major changes in electric vehicles and autonomy. These are his forecasts, not announced transactions or established outcomes.

Stebbings describes what he calls an 'Elon premium': a belief that Musk's visibility and marketing can support unusually high valuations. The panel also discusses a proposed distribution of SpaceX shares to children's accounts, including its potential financial and reputational effects.

The speakers suggest that direct giving associated with a company share could make more people feel connected to a wealthy founder's success. They treat the possible distribution and its consequences as their interpretation of the proposal, then close the discussion.