Brex, AI Inference Costs and the IPO Market: A 20VC Roundtable

Harry Stebbings opens with the announced acquisition of Brex by Capital One, citing a $5.15 billion price split between cash and shares. Rory O’Driscoll calls building a company to that scale an achievement, while distinguishing the absolute transaction value from the outcome for investors who entered at different prices. The panel does not review the signed agreement or closing conditions.

O’Driscoll asks how to evaluate a sale below Brex’s earlier reported $12 billion valuation and whether it changes the competitive picture for Ramp. Jason Lemkin describes a conversation with a smaller investor who regarded the result as good in the circumstances. That individual account does not establish returns for all shareholders or employees.

Lemkin says a lower sale price than a previous private valuation can feel disappointing despite an impressive absolute outcome. He distinguishes late investors from earlier stakeholders but does not have a complete capitalization table. Their respective proceeds cannot be inferred from the headline sale price alone.

Lemkin calls very high financing prices a form of ambitious fundraising: capital and momentum can help a company compete, but the implied expectations can make an eventual lower-priced sale feel disappointing. He says early stakeholders may still benefit. Individual proceeds depend on actual securities, preferences, debt and deal terms, none of which are analyzed in full here.

O’Driscoll argues that when a company needs capital, accepting a high market price may reduce dilution and fund competition, even though later market conditions may change. He calls the reported sale an achievement while recognizing that later investors may see a different result. His claim about what the company needed is an inference, not verified cash-flow data.

Lemkin recalls a founder’s rule of avoiding a valuation more than roughly two years ahead of an achievable business. O’Driscoll says extrapolating very high growth can make a valuation appear reasonable until growth or financing conditions change. Their example is a risk framework, not evidence of Brex’s internal forecast.

O’Driscoll describes how a valuation can be based on an assumption that exceptional growth continues for several years. When growth slows, investors may change their expectations and companies may need to pursue profitability sooner. This is a general illustration of financing risk rather than a verified account of Brex’s decisions.

O’Driscoll discusses the possibility that a late-stage portfolio can absorb some low-return investments if other positions perform well. He uses hypothetical multiples and assumptions about liquidation outcomes. Preferred-share terms and proceeds can vary, so a one-times outcome is not guaranteed.

Asked whether the Brex deal helps or hurts Ramp, O’Driscoll sees operational validation for a competitor but says an acquisition multiple might cause investors to revisit private marks. He compares rough reported revenue and valuation figures and notes the businesses have different growth. His valuation arithmetic is an illustration, not a reliable price for Ramp.

O’Driscoll says private funding rounds can be thin transactions that do not necessarily determine what a buyer would pay for an entire company. He speculates about a future public valuation and growth rate for Ramp while acknowledging different outcomes are possible. The figures are not verified and should not be treated as a forecast of the share price.

The group emphasizes that a reported purchase multiple is one data point, not a definitive mark for a faster-growing competitor. A private round and acquisition can differ in rights, growth expectations and the portion of a business purchased. O’Driscoll’s hypothetical calculations do not establish Ramp’s fair value.

Lemkin considers a founder’s choice between accepting a sale and continuing through several uncertain years toward an IPO. He says motivation to build a larger independent business matters as well as the potential economics. His discussion is hypothetical and does not establish the Brex founders’ private reasons.

O’Driscoll believes Capital One could benefit from combining Brex with its payments infrastructure and its Discover relationship. He suggests directing more transactions through owned rails could affect economics, but the interview does not establish the integration plan, fees or realized synergies. He previously worked with BILL, a company mentioned in the comparison.

The panel weighs three possible implications for Ramp: competitive validation, a transaction multiple that may challenge some private assumptions, and a larger competitor with payments infrastructure. These are investor interpretations based on incomplete deal data, not evidence of Ramp’s future performance or Capital One’s exact cost advantage.

Stebbings turns to a reported TikTok US transaction and says US investors would own a majority while the algorithm remains associated with Chinese owners. O’Driscoll frames it as a geopolitical as well as economic arrangement. The panel does not examine definitive ownership, licensing or control documents, so the precise structure remains unverified here.

O’Driscoll estimates US revenue and a purchase multiple and says licensing or other payments could affect the economics. He acknowledges that he does not know the complete terms. His favorable assessment is an investment opinion, not proof that the deal is inexpensive or will generate a specific return.

Lemkin compares the unusual structure with an earlier Skype transaction and wonders why some prominent investors did not participate. O’Driscoll says he does not know and recalls that the earlier deal had licensing issues. Their discussion of absent investors’ motives and historical returns is speculation or recollection, not transaction evidence.

The panel recalls an earlier Skype deal involving eBay and subsequent investors as an example of a complicated transaction. Their account is from memory, including its timing and return. They do not know why specific firms did or did not participate in the TikTok arrangement.

Stebbings cites a report that Anthropic inference costs were 23% higher than expected and asks about economies of scale. Lemkin warns companies against cutting model usage so aggressively that their product becomes less competitive. The percentage is a reported comparison with an expectation, not a verified current cost structure.

O’Driscoll cites approximate historical and current gross margins and argues the reported direction suggests improving economics despite a miss against expectations. He says ultimate margins and profitability remain uncertain. The specific negative and positive margin figures are his unverified account and should not be treated as audited financial statements.

Lemkin predicts more continuous inference and higher token consumption even if the cost per token falls. He describes building a personal chat product that retains conversation context, and expects such features to use more tokens. That is his experience and forecast, not proof that every application’s total bill will rise.

Lemkin says a company’s inference budget may increase even as models become cheaper per token if users ask for more work or agents run more often. His own experiment motivates the concern, but its performance and cost are not independently measured. The future rate of adoption remains uncertain.

O’Driscoll says an abundant demand for a costly digital service creates a challenge in pricing and rationing usage. He distinguishes model-provider economics from application-company economics and gives illustrative ranges for inference as a share of revenue. Those ranges are not a benchmark for any named company.

Lemkin presents a hypothetical established software business that reached breakeven, built an AI agent and then faces a large inference bill while competing with well-funded startups. His interaction volumes, prices and funding needs are scenario inputs, not reported company results. He worries a weak product cannot simply cut model spending and remain competitive.

O’Driscoll agrees that companies must decide how aggressively to spend and says his portfolio faces similar questions. He suggests charging for differentiated customer value and using efficient models or data where appropriate, while acknowledging that a subsidized competitor can make the choice harder. His framework does not guarantee a positive margin or a successful exit.

O’Driscoll acknowledges the dilemma for an established company that has achieved profitability but must now invest in a new AI product. He says the ability to fund inference depends on customer value and competitive conditions. His examples from boardrooms are generalized and not disclosures of named companies’ budgets.

Lemkin argues that a strong agent may justify a price tied to the value it delivers, but says reaching that quality is difficult. O’Driscoll emphasizes that, eventually, a sustainable business needs revenue to exceed its costs. Their examples of labor savings and pricing are hypothetical, not proof that a particular agent replaces a specified number of workers.

Lemkin suggests some fast-growing AI firms can treat inference as an acquisition expense because they spend less on traditional sales and marketing. That is his interpretation of selected companies, not audited expense data or a rule for all AI businesses. He believes incumbents may face both inference expense and established sales costs.

O’Driscoll cites comments he recalls from TSMC about compute demand and capital expenditure as a signal worth watching. He distinguishes those plans from the interests of model providers and notes that fabrication investments have long lead times. The spending figures and his interpretation of demand are approximate, not independently checked in the conversation.

The guests consider why infrastructure companies might continue investing despite uncertainty about long-term demand. O’Driscoll treats a foundry’s capital plans as one indicator, not proof that every investment in the chain will earn a return. Project schedules, costs and customer commitments are not examined in the interview.

Stebbings asks whether strong infrastructure demand removes the possibility of an AI bubble. O’Driscoll says near-term spending plans can be credible even if overinvestment or a downturn occurs later. He uses approximate industry capital-expenditure and application-revenue numbers; comparing them directly does not establish profit, losses or the timing of a cycle.

The panel distinguishes falling cost per token from rising total demand and spending. O’Driscoll says aggregate investment may eventually face financial or resource constraints, while Lemkin believes there are reasons to plan for continued demand. Both are discussing uncertain scenarios, not a recommendation to buy or sell semiconductor securities.

O’Driscoll says a downturn may eventually create an opportunity for investors with a particular view but explicitly says he is not taking that position now. He advises businesses to consider alternative scenarios and funding needs if conditions change. His comments are personal market opinions and do not establish when a cycle will turn.

Stebbings raises a reported OpenEvidence financing at a $12 billion valuation and describes a possible pharmaceutical-advertising opportunity. The cited step-up, market size and future $4–5 billion revenue possibility are reporting or hypotheses, not independently verified company financials. The panel has not established the terms or the company’s future sales.

O’Driscoll describes OpenEvidence as a tool for medical professionals and discusses relationships with medical publishers and advertising. These are his product and commercial claims. The interview provides no independent clinical accuracy, safety, privacy-compliance or patient-outcome evaluation, and the discussion is not medical advice.

He questions whether all pharmaceutical advertising budgets are addressable: some spending targets consumers directly, while other spending reaches physicians through representatives or other channels. He suggests growth could require shifting those budgets online or providing additional services. The dollar estimates and possible shifts are his assumptions, not independently verified market measurements.

In discussing medical advertising, O’Driscoll distinguishes consumer television promotion from marketing directed at physicians. He believes a platform may need to capture a different budget or add services to expand substantially. His estimates are assumptions, and the conversation does not establish clinical benefit or recommend treatments.

Lemkin says price is central even when a product and founder look attractive. He predicts another financing at a higher valuation could happen, but asks which round might eventually be too expensive if growth slows. This is a speculative comparison with Brex; it does not establish OpenEvidence’s future valuation or investors’ returns.

O’Driscoll says a company can maintain an apparent revenue multiple as revenue and valuation grow together, yet a late investor still risks a disappointing outcome if assumptions change. Both guests describe how easy it can be to rationalize a popular high-growth deal. Their examples are hypothetical, not evidence that a named investor overpaid.

The investors compare a popular financing with earlier transactions that later looked expensive after growth slowed. They acknowledge that a company can be strong even when a particular entry price yields a weak return. The future value of OpenEvidence and the outcome for any investor remain unknown.

Stebbings cites a venture firm’s report claiming that companies it backs generate two-thirds of private AI revenue and that it invested $8 billion in 2025. O’Driscoll says a few very large companies may dominate the calculation. The panel does not inspect the report’s definitions, inclusion rules or revenue sources, so the share is not independently verified.

Lemkin asks whether venture has become a broader asset class and cites a proposal for substantially more capital. O’Driscoll distinguishes early-stage venture from later-stage private growth and cautions that excess capital can erode returns. Both have professional interests as venture investors; their statements about fund performance are opinions and rough examples.

O’Driscoll says encouraging more people to try startups can be valuable, while cautioning that investing more money does not automatically improve a fund’s results. He recalls wishing his firm had been less active during a highly priced period. His retrospective examples do not establish the performance of every investment vintage.

He describes a multistage firm’s strategy of using early investments and relationships to support later investments in the same companies. Lemkin highlights the importance of patience as well as activity. Their estimated funding shares and conclusions about repeatability are interpretations, not independently audited fund data.

The guests consider whether a prominent venture firm could retain its position after a leadership transition. O’Driscoll argues institutions can plan for succession, while Lemkin questions how much influence particular leaders have. They do not report an actual transition or know the firm’s succession plan.

O’Driscoll contrasts a small partnership whose assets depend on a few decision-makers with a larger institution seeking to operate beyond its founders. He draws an analogy with investment banking. That is a theory about organizational design, not proof that either model will outperform or that a specific firm will remain successful.

The speakers disagree over how dependent an investment institution is on individual partners. O’Driscoll thinks deliberate succession and an institutional platform can help, while Lemkin stresses the difficulty of replacing prominent leaders. Neither has access to an announced succession decision in the discussion.

The discussion turns to EquipmentShare’s reported IPO and first-day trading. O’Driscoll describes it as a construction-equipment rental business using technology, and Lemkin cites approximate growth, revenue and profitability. Their characterization of the offering and trading is time-specific and the figures are not verified against filings here.

They contrast EquipmentShare with Wealthfront, whose shares they say had fallen after its IPO. Lemkin worries that a small public-market valuation can reduce liquidity and analyst attention. O’Driscoll says he retains affection for the company and may examine it as a possible investment; his interest and valuation views are personal, not a recommendation.

Stebbings asks whether a weak IPO makes hiring difficult. O’Driscoll thinks people interested in finance may still find the company compelling, while Lemkin believes strong leaders can recruit through a difficult period. These are opinions about recruiting, not evidence about any company’s actual employee quality or intentions.

Lemkin suggests a determined founder may find a small number of leaders who can help a company through a difficult period. O’Driscoll challenges the idea that employees at companies with less celebrated outcomes are necessarily mediocre. Their exchange reflects personal views on talent rather than a measurement of a workforce.

O’Driscoll pushes back on the idea that all exceptional people work only at exceptional companies. He says company outcomes and individual ability are different distributions, though he expects strong leadership to matter. Lemkin agrees that capable leaders can help a business regain momentum; neither has surveyed the named company’s staff.

The panel discusses a planned Ethos listing at a reported valuation below an earlier private price. Stebbings questions whether a smaller IPO is desirable; O’Driscoll asks what alternatives are available and says public markets can provide a path to compound over time. The listing and subsequent trading are not confirmed in this contemporaneous discussion.

Lemkin predicts the IPO may trade down and interprets a lower valuation as a possible acceptance by earlier investors that market prices have changed. O’Driscoll cautions against dismissing a $1–2 billion business and argues price can attract buyers. Both are predictions and interpretations, not evidence of later trading or realized proceeds.

O’Driscoll argues that a public listing at a lower valuation than an old private round can still be a meaningful company achievement and a starting point for future growth. Lemkin asks whether more companies will take that path. Their discussion does not predict a specific listing’s price or future performance.

The guests debate whether more companies financed at high valuations will pursue listings or acquisitions at lower prices. O’Driscoll says a lower price can still clear a market and give a company a chance to grow. His estimates of the number of private unicorns and their growth profiles are rough and unsourced in the interview.

Stebbings cites a reported $5.6 billion Salesforce Army contract over ten years. The panel treats it as evidence that established systems can remain useful alongside AI, but does not inspect the contract’s ceiling, obligations or payment schedule. A headline contract value is not automatically revenue recognized or cash received.

O’Driscoll says replacing an existing system of record is different from adding AI capabilities around it. Stebbings is optimistic about Salesforce’s distribution, while O’Driscoll says renewed growth still requires execution. Their views are strategic opinions, not proof of future company growth.

Lemkin argues that established software companies face several pressures at once: seat counts, pricing, customer budgets and the need to build AI products. He cites examples and approximate price changes from his own observations. These are not a representative dataset, and he does not prove that all workforces or software budgets are shrinking.

He says changing pricing units alone cannot force customers to spend more if the product does not deliver additional value. O’Driscoll summarizes a possible middle outcome in which established software remains useful but grows more slowly. The conversation ends with differing views on the pace and scale of that transition.