AI Burn Multiples, Infrastructure Financing and the Venture Market

Harry Stebbings, SaaS investor Jason Lemkin and Scale Venture Partners general partner Rory O’Driscoll examine how AI changes venture investing. They discuss burn multiples, financing, the energy needed for AI infrastructure, consolidation and the responsibilities of founders speaking publicly.

Lemkin points to an Iconiq software report that, he says, found AI-native companies below $100 million ARR had free-cash-flow margins of minus 126%, versus minus 56% for non-AI companies. He says their faster growth nevertheless produced lower burn multiples. These are report figures as discussed, not universal statistics for every company.

The burn multiple compares cash burned with net new annual recurring revenue: spending $2 for each $1 of added ARR produces a multiple of two. Lower can indicate that a company adds revenue more efficiently, but the number does not establish profitability or show how much cash remains.

Lemkin says AI introduces different costs and growth rates, including model token costs, so comparing the ratio across businesses can mislead. His example of a company burning $6 million a month while adding $300 million in ARR is hypothetical, not a verified forecast of Lovable’s results.

O’Driscoll still finds the burn multiple useful, but only with its assumptions visible. ARR must represent real recurring revenue; fast growth can make churn appear smaller than it will later; and gross margins can shift. For capital-intensive model companies, the ratio also does not capture capital expenditure.

To test whether recurring revenue is real, O’Driscoll suggests comparing reported ARR growth with changes in recognized GAAP revenue. The historical SaaS businesses for which simple comparisons worked often had similar margins, sales models, churn and capital needs; those conditions cannot be assumed today.

The valuation logic also depends on customer retention. A hypothetical $2 million of burn that adds $1 million ARR may look attractive at a ten-times revenue multiple, but if that revenue disappears a year later, the assumed value creation disappears with it.

Growth alone is a simpler scorecard than growth plus profitability, O’Driscoll says. Spending more on sales or free users can buy additional revenue, but not necessarily enough to justify the marginal cost. At some point a company must demonstrate sustainable economics.

Even a good burn multiple does not put cash in the bank. The theoretical argument that efficient growth should attract funding is different from having enough runway to make payroll. O’Driscoll wants to know the cash balance, particularly if a company has less than six months left.

Stebbings says some portfolio companies have strong growth and burn metrics but are struggling to raise, describing a divide between companies investors pursue and those they overlook. He asks what founders should do when numbers that once seemed sufficient no longer secure funding.

O’Driscoll argues that venture investors price some deals on future possibilities and others on current multiples. A solid $15 million revenue company can have little appeal to a particular VC if that investor sees no path to a very large exit; that is a statement about venture incentives, not the company having no economic value.

Lemkin says even companies exceeding the familiar triple-triple-double-double growth pattern can struggle if they are not perceived as breakout AI winners. He worries that some boards still offer fundraising advice based on earlier markets and underestimate the difficulty of raising today.

His example is a hypothetical company at $15 million ARR, growing 100%, with a good burn ratio but a limited market. If a credible investor offers a reasonable round, he would take it rather than assume time and growth guarantee a better price. O’Driscoll agrees that cash and capital efficiency deserve priority.

Lemkin argues that the distinction between AI and non-AI software is becoming less clear as established vendors add AI features. He cites a report’s claim that 94% of public software companies describe themselves as AI companies; that characterization should not be confused with a measure of their AI revenue.

Stebbings says investors seem reluctant to fund companies competing with a well-funded, prominent market leader. He has portfolio companies he describes as second or third in their categories, so he has a direct investment interest in the question.

O’Driscoll agrees that a top investor’s backing is a factor but not an automatic reason to reject a competitor. Brand can matter more when the prospective customers themselves are venture-backed and know the investors; a buyer in another industry may care far less.

Lemkin argues that nervous enterprise buyers facing a new AI category may use a familiar leader as a shortcut. He expects the situation to settle over time, but says responsiveness and customer trust can still let a rival win. He cites an anecdote about a customer choosing Bolt over Lovable, not a systematic market study.

The bigger problem, O’Driscoll adds, can be the second financing round attracted by a famous investor’s first check. His illustrative company with $20 million may raise another $60 million because others follow the lead. More capital changes the competitive landscape, although money alone does not ensure a winner.

Stebbings asks whether extreme AI valuations are a peak or a rational response to a much larger market, including spending previously allocated to labor. O’Driscoll frames two scenarios: transformative productivity and enormous revenue, or valuable technology accompanied by a painful valuation adjustment. Neither outcome is established.

Lemkin believes B2B AI is still early and describes his own use of AI agents and coding tools. At the same time, he worries that investors may be underestimating loss rates and overpaying, just as he believes they did for many software unicorns in 2020–2021.

O’Driscoll separates two ways a fund can fail. Investors can choose companies that never find product-market fit, or pick good companies but pay such high entry prices that even successful outcomes cannot compensate for losses elsewhere. Both selection and price matter.

Stebbings worries that exceptional transactions involving prominent AI leaders can become precedents for increasingly high entry prices. O’Driscoll cautions against treating a market regime as permanently different; the panel’s valuation examples and predictions are opinions about a particular moment.

The panel discusses Figma trading below its post-IPO peak and cites a roughly 26-times revenue multiple at the time of recording. Lemkin says the company can be excellent while its share price still falls after a very high opening; these figures are historical observations, not current market prices.

O’Driscoll describes a valuation ladder anchored to public SaaS peers. He recalls a historical benchmark of roughly 30% growth at six or seven times next-twelve-month revenue, and worries that if today’s higher benchmark falls, private valuations built on it could fall too. Both the benchmarks and resulting risk are his analysis.

After discussion of Klarna and StubHub, the panel considers IPO pricing. O’Driscoll expects buyers who have suffered losses to ask for more room between an IPO price and comparable public companies. He thinks the market can remain open, although pricing may become more cautious.

They discuss the proposed Electronic Arts take-private and its reported $55 billion size. O’Driscoll says leverage must be viewed in the context of a hit-driven games business, while Lemkin questions whether an AI strategy alone will revive growth. Their comments are judgments on a reported transaction, not verified deal underwriting.

Stebbings turns to the scale of OpenAI’s possible energy and data-center requirements, citing a potential $1 trillion infrastructure figure and comparisons with national electricity capacity. The panel treats these as scenarios and discussed plans, not established spending, available power or a confirmed funding commitment.

O’Driscoll says compute demand follows the company’s ambition, and financing may continue as long as revenue and capital support it. He suspects the scale could become difficult to fund within two or three years, but explicitly distinguishes a technology that may work from the economics of funding its expansion.

Lemkin imagines data centers with power needs comparable to major cities and few on-site workers. He believes Altman may find solutions, including future energy advances, but the images of AI cities and fusion are speculative predictions, not descriptions of operating facilities.

O’Driscoll predicts that practical constraints on power, construction, sales and enterprise adoption will force some infrastructure forecasts down over five years. His concern is the speed implicit in investment plans, not a claim that the technology has no value.

Stebbings asks where a trillion dollars would come from. O’Driscoll agrees that the magnitude is exceptional; Lemkin suggests the full target might not be necessary, and that progress could continue more slowly with less capital. The speakers do not establish an actual funding source or final capital requirement.

O’Driscoll admires Altman’s ability to pursue a large vision and predicts OpenAI will win the leading AI prize. He warns separately that a visionary target should not automatically be treated as a purchase order or used as certain revenue when lenders and suppliers plan their own commitments.

He offers a scenario in which OpenAI remains highly successful but grows more slowly and needs less capital expenditure than the most aggressive plans imply. In that case infrastructure investors could still face consequences. The growth and revenue figures he mentions are illustrative, not guidance.

Stebbings discloses that Meta is his largest public-market position, while saying his confidence in the company’s AI strategy has declined. The panel discusses its hiring and spending as observers; their judgments about management and future results are opinions, colored in Stebbings’s case by his holding.

O’Driscoll says a leader’s past success can justify another ambitious bet without proving that the new bet is correct. He speculates that a large AI investment could disappoint while earlier successes such as Instagram and WhatsApp still dominate the overall record.

The panel considers the competition for users’ attention: time spent with an AI assistant can come at the expense of social platforms. They see this as one reason an incumbent might spend heavily to develop new products, even when its business case is uncertain.

On buying inside ChatGPT, O’Driscoll expects platforms to explore commerce and advertising as potential ways to monetize free users. Lemkin treats the feature as an experiment and says press attention does not establish that it will become a material revenue stream or transform shopping.

At very large companies, a new product may need billions in revenue before it materially affects the business. The panel contrasts that challenge with the much larger hypothetical infrastructure cost and emphasizes that a trillion dollars contains a thousand billions.

The conversation turns to reports that Fivetran was in talks to buy dbt Labs. The speakers cite reported ARR figures of roughly $400 million and $100 million respectively and speculate that their combined revenue could exceed $500 million. The discussion does not establish that the transaction had closed.

O’Driscoll says adjacent products can be more valuable together if their customer bases and offerings fit. He sees consolidation as one possible path for venture-backed companies to reach the scale needed for an IPO, while acknowledging integration and deal-making are difficult.

Lemkin says common investors may make private-company combinations easier to negotiate. Using hypothetical ownership stakes, he explains why a holder of 20% in one company may resist ending up with a smaller share of the merged business, even if the combination improves the prospects of an exit.

O’Driscoll agrees the reduced ownership can feel costly, but argues that a smaller stake in a company able to go public may be preferable to a larger stake in one that cannot. The percentages in their exchange are examples, not disclosed terms of the proposed transaction.

The panel considers the incentives and complications in combining portfolios, including dilution and investor preferences. A deal that looks persuasive on a spreadsheet still requires agreement among founders, boards and shareholders with different stakes.

They discuss combining complementary software products and the limitations of simply aggregating slow-growing businesses. Lemkin argues that a coherent product relationship matters and speculates about the prospects of individual companies; those comments are his assessment, not verified operating results.

Stebbings asks whether private equity firms should reconsider buying established SaaS companies as AI increases disruption risk. O’Driscoll thinks investors should explicitly underwrite that risk, but warns that moving into an entirely different investment strategy merely because it looks attractive can itself be a mistake.

Lemkin says many enterprise software products changed slowly for years, which made their revenues seem dependable. He discloses that he led Pipedrive’s seed round and uses it as an example: the pace of product change he recalls may no longer be sufficient in an AI market.

O’Driscoll agrees that technological change poses risks for both established private-equity holdings and startups. In his view, product-market fit that once persisted for years may become less stable as models and customer expectations change. This is an investment judgment about future risk.

Lemkin points to RevenueCat, a company in his portfolio, while discussing API-based businesses. He argues that AI agents may reduce the number of human software seats a company needs, though seat-based pricing is not necessarily dead. If vendors provide valuable agent services themselves, customer spending could change in the other direction.

O’Driscoll says seats need not fall to zero to become much less predictable. The panel also reacts to a corporate statement about layoffs and reskilling, criticizing the euphemistic language used to describe a painful employment decision.

Stebbings asks whether a founder’s obligations to employees and shareholders should outweigh personal political expression. O’Driscoll initially resists the idea that a CEO must surrender personal opinions, while suggesting companies should think carefully before taking positions outside their mission.

O’Driscoll acknowledges the tension: some institutional roles have traditionally required political restraint, even if he values personal expression. He does not resolve the trade-off into a universal rule, and distinguishes a company’s official stance from a leader’s personal view.

Lemkin describes occasionally contacting executives privately when a post may have landed differently than intended. He says most people he approached understood the potential business costs but felt strongly enough to accept them; his anecdotal experience is not a general survey of executives.

O’Driscoll says a board should be cautious about blaming someone for a clearly personal opinion, while recognizing a hypothetical extreme in which customers or employees leave. Lemkin replies that some leaders may knowingly accept such consequences.

Stebbings suggests the attention cycle is fickle and controversies can fade quickly. O’Driscoll agrees that the public often moves on rapidly, and the discussion closes without a single prescribed answer for every founder.