Databricks Valuation, CoreWeave Debt and the Risks Behind AI Growth

Harry Stebbings, Jason Lemkin and Rory O’Driscoll discuss Databricks’ reported valuation, fintech performance, debt-funded AI infrastructure and the incentives of venture investors. Their numbers reflect the conversation at the time of recording, not current prices or independently verified company financials.

The panel starts with a reported $100 billion valuation for Databricks. Lemkin says the number no longer feels as astonishing in the AI market, while O’Driscoll compares it with Snowflake and cautions that he has not yet reviewed the latest Databricks figures in detail.

The speakers cite Databricks approaching a roughly $4 billion revenue run rate at about 50% growth and compare it with Snowflake’s $1 billion quarter and 26% growth. Quarterly revenue, annualized run rate and ARR are different measures; the exchange does not establish that they are directly comparable.

Stebbings asks whether roughly 25 times revenue could be reasonable given growth. O’Driscoll says such a valuation depends on growth persisting for years: a sharp slowdown could make the entry price painful. His examples of future growth rates are scenarios, not guidance.

Lemkin imagines a further wave of large IPOs creating more enthusiasm and returns for limited partners. O’Driscoll counters that the scarcity premium for fast-growing public companies might decline if more candidates list. Neither outcome is certain, and their references to future Canva and Databricks listings are hypothetical.

The speakers compare companies of different size and growth, using an illustrative Canva revenue scenario. O’Driscoll argues that the public market could absorb substantial new equity, but that a listing does not guarantee the same price surge seen in another company.

O’Driscoll says mid-sized venture funds also need meaningful outcomes beyond a handful of huge private companies. He mentions a profitable, growing portfolio company as an example, making clear that his fund has an interest in the category.

Stebbings asks what an early Databricks investor might earn if the company listed at $200 billion and the investor owned 15%. The implied $30 billion is illustrative: the participants are guessing at ownership, entry dates, follow-on investments and fund sizes, not reporting a verified holding or realized return.

O’Driscoll says a successful early stake held over many years can produce very large returns, but private investors lack the same easy exit as public shareholders if prices fall. A late entry at $100 billion carries meaningful valuation and liquidity risk.

He frames the bet in terms of growth persistence: if the business reaches a much larger revenue base over several years, an expensive multiple today may normalize. His future revenue and growth figures are hypothetical, and the outcome depends on continued demand.

A new SPAC prompts Stebbings to ask whether the market is at a peak. O’Driscoll says a peak can only be identified in hindsight. He criticizes the SPAC structure because sponsors can be rewarded for closing a transaction even if its quality or later performance is poor.

Lemkin questions the incentives behind sponsor economics, while O’Driscoll argues that public markets should channel savings to productive investments rather than be treated as a casino. Their analysis concerns incentives and potential conflicts; it is not a finding that a particular transaction is fraudulent.

The conversation turns to private equity and venture access in retirement accounts. O’Driscoll sees potential benefits from broader access to private-market returns, but warns that complex products, fees, valuations and adverse selection can harm less sophisticated investors.

Lemkin describes an example he says involved a discounted secondary portfolio being marked up soon after purchase. The details are not verified in the conversation. The broader question is whether disclosures explain valuation changes well enough for a retail investor to understand them.

O’Driscoll weighs disclosure and buyer responsibility against protective regulation. He says access to high-performing private companies and protection against opaque or unsuitable products are both relevant, and does not offer a simple universal answer.

Stebbings cites a reported $6 billion OpenAI employee secondary and asks how to assess liquidity at scale. O’Driscoll welcomes employees having an opportunity to sell, particularly when rivals are making large offers. A secondary sale transfers existing shares and is not necessarily $6 billion of new company financing.

He says a very large offer can change someone’s decision in ways that hypothetical conversations do not predict. Lemkin suggests more regular liquidity for vested shares could make long-lived private companies behave more like public companies, though such access depends on company and market arrangements.

The investors debate whether newly wealthy employees will keep working as hard. O’Driscoll thinks people who love the work may continue while others pursue different goals. Lemkin notes that easier liquidity can weaken the traditional retention effect of illiquid startup equity.

The panel discusses reports of a founder leaving a company after a large fundraising round. They do not verify the founder’s proceeds or any transaction terms. O’Driscoll says some venture investments fail and founders leave, while Lemkin worries that rushed deals can raise the likelihood of bad outcomes.

Stebbings cites Nubank results, $2.5 billion of net income and 123 million customers. The conversation does not clearly establish the net-income measurement period; it should not be read as $2.5 billion for a single quarter. The panel then compares Nubank, Revolut and Chime as businesses in different markets.

O’Driscoll argues each found weaknesses in existing banking services: Nubank expanded from deposits into lending in Latin America; Revolut used cross-border money movement as a starting point in Europe; and Chime built a US business around deposits and debit interchange. These are his broad descriptions of their strategies.

He sees Nubank’s lending and range of services as evidence of a broader banking franchise, while saying geographic conditions and incumbents differ. The company valuations and relative market positions cited in the discussion are time-specific and should not be treated as current prices.

Stebbings asks whether Nubank is undervalued. O’Driscoll sees possible expansion but warns that banks face credit, regulation and valuation risks; his hypothetical doubling and Stebbings’s higher valuation scenario are forecasts, not established future outcomes.

Lemkin admires early investors who identified the potential to win a significant share of financial services. O’Driscoll says a successful fintech can start with one valuable product and expand gradually into other services, citing international money transfer as an example.

Stebbings praises Revolut’s approach to testing new products. Lemkin says bringing experienced founders in to run product lines can help a company expand, citing other firms in which he has a board or portfolio interest. These are observations about management, not audited product economics.

The panel briefly considers On’s reported growth and margins. O’Driscoll admires the company but is uncertain that a single successful consumer brand provides a repeatable venture strategy; he sees broad digitization as a clearer investment trend in fintech.

The discussion shifts to CoreWeave’s reported $11.2 billion of debt. O’Driscoll cites roughly $1 billion in quarterly revenue, a loss and a $22 billion capital-expenditure plan as figures he recalls from a presentation. The discussion does not verify the reporting period or distinguish every accounting measure, so these are attributed approximations.

O’Driscoll describes CoreWeave as financing the infrastructure that customers want to use: borrowing, building capacity and signing long-term contracts. Debt is inherent to a capital-intensive business, in his view, but its risk depends on how liabilities and customer obligations align.

A long-term, enforceable take-or-pay customer contract matched to debt of similar duration could reduce exposure, he says. If the provider commits to long-term costs while customers can leave quickly, the mismatch can be dangerous. He explicitly says he lacks visibility into the exact contract terms.

Lemkin sees CoreWeave as a possible early indicator of stress in AI infrastructure because it has large financing needs and public-market exposure. He speculates that a financing or debt-service problem could appear there before it is visible in larger customers; he does not say such a problem has occurred.

O’Driscoll gives a hypothetical in which Microsoft continues paying on a matched contract despite using less capacity. That asset could still perform, but building the next facility without committed demand would become harder. Customer credit, enforceability, utilization and refinancing remain separate risks.

Lemkin says reported take-or-pay coverage sounds reassuring but asks what qualifications the contracts may contain. If a customer chose to pay rather than use capacity, he suggests that could signal slowing demand. The example is hypothetical, not a disclosed customer decision.

Stebbings recounts an unnamed executive describing strong demand for compute. O’Driscoll accepts that demand can be intense today while insisting investors should model both continued growth and a slowdown. He declines to make a confident one-year CoreWeave market-cap prediction.

O’Driscoll offers a personal estimate that there is at least a one-in-three chance demand slows enough over the next 12–24 months to compress some valuations. This is his subjective probability, not a statistical forecast or evidence that a downturn has begun.

Lemkin notes CoreWeave’s stock had fallen from an earlier peak despite strong AI enthusiasm. He and O’Driscoll also discuss their exposure to large technology companies through diversified funds and retirement accounts; the prices and index weights mentioned are historical and approximate.

The panel considers whether spending on chips and infrastructure can keep rising and how quickly the economics of customers and suppliers might change. O’Driscoll says investors must distinguish the current shortage from the possibility of future excess capacity.

Lemkin predicts people will use far more AI and compute over the following two years, raising the question of how much infrastructure will be required. His tenfold-use example is an extrapolation, not a measured commitment or assured demand.

The investors question how much value individual AI tools create and whether business customers can absorb and manage many separate applications. They consider how companies might prioritize products that solve a sufficiently large problem rather than merely demonstrate a good percentage return on a small task.

O’Driscoll gives a warehouse automation example: even a high return on a small number of workstations may not command management attention. He discloses a portfolio interest in Locus and says its focus on a larger warehouse activity illustrates his point. This is an investor’s account, not an independent outcome study.

Lemkin predicts businesses may eventually consolidate the AI agents they buy, just as they have consolidated software subscriptions. O’Driscoll thinks a strong vertical provider may be able to expand from one product into adjacent workflows, but faces competitors doing the same.

They discuss Abridge and Epic as an example of platform risk. O’Driscoll suggests a partner can become a competitor and says a startup may need a product users strongly prefer; the comments are strategic speculation, not evidence of a specific contractual decision or clinical performance.

Lemkin says investors may be overlooking risks they once debated more carefully because current growth is so attractive. O’Driscoll agrees that many strategies rely on continued AI adoption; if adoption slows, he expects the consequences to spread across valuations.

In a prediction-market round, O’Driscoll says a favorable payout can make a bet attractive even when he thinks an event is unlikely. He and Lemkin disagree over a hypothetical Claude 5 release; their comments are bets and opinions at the time, not release announcements.

The panel speculates about whether Mistral might receive an acquisition offer. They explicitly distinguish whether an offer would be attractive from whether one will actually arrive. The examples of potential prices and revenue are hypothetical, not verified negotiations.

Finally, they debate whether Deel or Rippling might list first. Lemkin says he does not have Deel’s financials and is unsure of its revenue mix; the profitability and timing discussion is based on public claims and speculation. Stebbings has discussed his investment in Deel elsewhere, though the relationship is not disclosed in this exchange.

O’Driscoll says either company’s decision would depend on readiness and access to capital, and that a more investment-heavy company might still seek public financing. Both speakers frame their timing assessments as guesses, not corporate plans.