Nvidia’s OpenAI Commitment, H-1B Policy and the Venture Investor’s Risk

In this September 2025 roundtable, Harry Stebbings, Jason Lemkin and Rory O’Driscoll debate AI infrastructure financing, venture returns, Navan’s IPO filing, H-1B policy and the meaning of founder-friendly investing. Their financial figures and policy interpretations reflect the discussion at that time, not independently verified current facts.

Stebbings opens with Nvidia’s proposed investment of up to $100 billion in OpenAI and asks whether capital can circulate from an investor to a customer and back to a supplier. The headline describes a potential, staged commitment, not evidence that the full $100 billion had already been invested or spent.

O’Driscoll says financing allows OpenAI to pursue an unusually large bet on computing and model scaling. If the underlying business produces the expected returns, the financing can look sensible; if it does not, the commitments create risk. He is skeptical that every additional dollar of infrastructure spending will earn a return.

Stebbings questions whether scaling will continue. Lemkin points to Sam Altman’s public ambitions for much greater compute and argues that leaders intend to keep spending. O’Driscoll stresses that willingness to finance the experiment is not proof the projections or the economics are correct.

The panel says investors tend to back a successful company for as long as the story and financing remain compelling. O’Driscoll cannot say when that cycle will stop. The scale of the proposed capital gives OpenAI an opportunity to test its strategy, but does not eliminate the possibility of a painful reversal.

Stebbings asks whether Anthropic is disadvantaged by the proposed Nvidia relationship. O’Driscoll says the answer depends on whether capital or access to chips is genuinely constrained; he does not know whether OpenAI received preferential access. The guests’ claims about competitors’ fundraising capacity are their contemporary assessments.

Lemkin discusses competition among AI providers and the possibility of firms developing alternatives to Nvidia chips. He also speculates about concentration in consumer AI. The panel’s use of the word monopoly is conversational opinion about market position, not a legal finding or a verified market-share calculation.

O’Driscoll wonders whether OpenAI’s own chip ambitions affect its relationship with Nvidia. He explicitly says he does not know what, if any, agreement exists on the subject. The panel then considers how supplier and customer concentration can create bargaining power and risk on both sides.

The guests cite an estimate that a small number of large customers account for much of Nvidia’s revenue and compare that concentration with companies serving broader customer bases. These figures are presented as the speakers’ recollections; the relevant customer categories and calculation are not established in the episode.

Stebbings says a small number of customers also make up a large share of revenue for AI data-labeling businesses. O’Driscoll distinguishes the application boom from the capital-expenditure boom, arguing that large buyers are racing to build capacity. Their customer and spending estimates are not independently verified financial disclosures.

O’Driscoll compares the atmosphere to the late 1990s: transformative long-term trends can coexist with financing that gets ahead of near-term returns. He recalls vendor financing and the subsequent downturn while emphasizing that today’s amounts and structures are different. The analogy does not establish that a crash is imminent.

Lemkin observes that much more capital appears available than in the earlier internet era, while the guests question how long interlocking commitments can continue. A public announcement, a financing commitment, an actual payment and realized revenue are different events; the panel does not establish every cash flow in these arrangements.

The speakers cite Nvidia free-cash-flow figures and discuss buybacks. O’Driscoll questions spending heavily on repurchases if the stock is expensive or future conditions change. Lemkin notes that companies may also buy shares to offset employee equity dilution; their remarks are investment views, not a valuation conclusion.

O’Driscoll argues that buybacks should be evaluated against price and alternative uses of cash, rather than mechanically matching dilution. The guests discuss how stock compensation affects the economics shareholders experience. Their numbers and interpretations are examples, not a complete analysis of Nvidia’s accounts.

Stebbings cites a claim that 75% of venture dollars in 2025 went to 19 companies and asks whether the industry has fundamentally changed. O’Driscoll treats ultra-late-stage private financing as a distinct layer on top of earlier-stage venture. The cited concentration statistic is not independently checked or fully defined in the conversation.

Stebbings says his team has trouble predicting what the next financing round will reward. O’Driscoll likes Lemkin’s practical rule: consider what a company can achieve in 18 to 24 months and whether that progress can support a meaningful valuation step-up, while retaining a longer-term view of potential value.

Lemkin says the hottest companies are often easy to recognize, but the next tier is much harder to predict. Investors can disagree about churn, margins or the durability of a growth spurt. Strong figures alone do not guarantee that every fund will interpret a company’s risk the same way.

The guests debate whether the old software growth heuristic of triple, triple, double, double has become insufficient. Lemkin says companies growing quickly at meaningful scale can still raise money, although meetings and financing may take more work. O’Driscoll agrees that a few faster outliers do not make the older trajectory irrelevant.

Lemkin points to Netskope and DX as examples when asking whether strong but less spectacular exits matter to today’s very large funds. O’Driscoll says they can represent excellent returns: fund size, ownership and the amount invested determine how much an exit contributes to a particular portfolio.

O’Driscoll illustrates how a successful mid-sized IPO and a famous AI investment might generate similar multiples on invested capital, even when their absolute dollar gains differ. His seven- or eightfold examples are rough calculations from memory, not audited fund performance or guaranteed realized proceeds.

The panel asks whether an investor holding a large OpenAI position should sell some shares or let the bet ride. Lemkin describes the emotional appeal of retaining an apparent winner; O’Driscoll says expected upside must be considered alongside concentration, liquidity and the investor’s own risk tolerance.

O’Driscoll draws on a portfolio-management book to argue that wealth depends on sizing bets as well as picking assets. A position that has grown to dominate a portfolio can require a different decision than the original investment. The right answer depends on how much loss an individual or institution can bear.

Stebbings asks whether financially secure investors find it easier to keep taking risk than emerging managers who need distributions. O’Driscoll says prior success may increase confidence and access, though excessive confidence can also lead to mistakes. The discussion distinguishes personal risk tolerance from the obligations of a fund.

Stebbings introduces Navan’s IPO filing and cites approximately $613 million of revenue, 32% year-on-year growth and 110% net dollar retention. The panel discusses a prominent investor’s concentration across vehicles but acknowledges it does not know the exact allocation within individual funds.

Lemkin explains that he is comfortable concentrating more capital in a breakout investment. O’Driscoll says concentration can increase potential returns while substantially increasing risk; an individual partner’s exposure is not the same as a fund’s aggregate exposure. Both describe their own approaches rather than a universal allocation rule.

O’Driscoll suggests quantifying how much confidence an investor needs in an asset’s excess return before concentrating a large share of wealth. The exercise forces assumptions into the open, even though those assumptions can be wrong. Confidence alone does not ensure a concentrated bet will succeed.

Lemkin asks whether Navan’s IPO timing reflects a desire to reach public markets before Brex or Ramp. O’Driscoll notes that their businesses overlap but differ: Navan is strongly associated with travel, while card and payments are central for the others. He considers the timing argument plausible but does not claim to know the company’s motives.

The guests weigh going public before profitability against waiting for stronger financial results. O’Driscoll says market conditions, access to capital, liquidity and competitive positioning can make an earlier listing attractive. These are hypothetical tradeoffs based on the filing and contemporary conditions, not a prediction of the IPO outcome.

Stebbings asks what happens after an IPO when a venture investor owns a large stake. O’Driscoll says reported holdings attributed to a named partner may include shares economically belonging to the wider fund and its limited partners, so a filing should not be read as that individual’s personal wealth.

He describes the typical lockup, sometimes around six months, and the possibility of an early release or a registered secondary offering under particular circumstances. Actual restrictions vary by offering and contract. An IPO headline does not mean all investors can sell their entire position immediately.

After restrictions lapse, an investor may sell over time or distribute shares to limited partners. O’Driscoll notes that board membership, reporting obligations and trading windows can lengthen the process. His 18- to 24-month estimate is his experience, not a standard timetable.

The panel debates whether a venture manager should hold public shares or distribute them to LPs who may have automatic selling policies. They describe information and incentive tradeoffs; their conversation is not a guide to trading while in possession of material nonpublic information, which raises legal obligations.

O’Driscoll discusses the tradeoff between remaining on a public board and returning attention to private companies. Board access can provide context but also restrict trading and require significant work. His examples of long-held investments are anecdotal and do not establish a strategy appropriate for every portfolio.

Stebbings raises the announced $100,000 H-1B fee policy as understood in September 2025 and asks about startup hiring. The panel discusses possible effects and uncertainty, not a definitive explanation of who is covered or current immigration law; details and implementation can change.

O’Driscoll argues that skilled immigration has benefited technology and that higher costs could make hiring harder at the margin. He says the size of the impact is harder to determine. Application counts, visa totals and economic estimates mentioned by the speakers are not independently established in the episode.

He suggests some supporters of higher fees see them as a way to address concerns about substitution for domestic workers, but criticizes a blunt dollar threshold as a policy instrument. His comments are a policy argument rather than a finding about individual visa applications or employer conduct.

Lemkin says employees on H-1B visas were important to his first startup. He expects larger companies to be better able to absorb higher costs and mentions other visa routes as possibilities, while acknowledging their burdens. His prediction about the overall impact is conditional and is not individualized immigration advice.

O’Driscoll favors a more carefully designed skills-based approach and argues that a single fee is a crude proxy for an applicant’s contribution. The guests’ preferences and descriptions of alternatives reflect the 2025 conversation; eligibility and fees require current authoritative guidance.

Stebbings raises a reported $500 million annual recurring revenue figure for Notion and points to accelerating growth. O’Driscoll treats the business as a counterexample to the idea that only the very fastest-growing companies matter. The revenue and growth figures are cited in conversation, not verified here.

Lemkin discloses that he received Notion stock when it acquired some of his companies, so the prior valuation matters to him personally. He and O’Driscoll discuss how AI features might help an established software company reaccelerate, without asserting that a particular growth rate or public valuation has been achieved.

Asked to estimate a hypothetical IPO valuation, O’Driscoll applies rough public-market multiples and explicitly says he does not know the precise growth rate. His calculation is a scenario, not an appraisal or an announced IPO price. The panel also discusses the complications of prior preferred financing.

Lemkin says he also holds Airtable shares acquired through company sales and asks about their prospects. O’Driscoll says mature software businesses will ultimately be judged on fundamentals such as revenue, growth and cash flow. His illustrative multiples are not actual offers or confirmed company results.

Lemkin argues that investors should stop anchoring on exuberant 2021 private valuations. O’Driscoll jokes that the market may be making similar mistakes again. They agree that earlier marks can complicate decisions even if a business remains fundamentally valuable.

The discussion turns to compressed due diligence in hot AI deals. Stebbings describes a portfolio company receiving and then losing a term sheet, and worries that investors use term sheets to secure exclusivity before doing the work. This is his account of an unnamed transaction, not a conclusion about a named firm.

Lemkin says an extremely short decision window with limited information can make an investor more sympathetic to withdrawing if later diligence reveals important discrepancies. He also acknowledges that issuing a term sheet simply to hold a place is poor practice. O’Driscoll says his firm can decline to engage on a timeline that does not allow adequate work.

Asked what founder-friendly means, Lemkin says it should show up in difficult moments: supporting a company when others leave, helping hire and providing honest assistance rather than automatic praise. His examples express his view of an investor’s responsibilities, not a claim that every financing relationship works this way.

O’Driscoll prefers the phrase founder-honest: tell founders what you actually think and judge an investor by conduct when conditions are difficult. Stebbings recalls Lemkin helping during the Silicon Valley Bank crisis, and the guests say challenging situations reveal more than favorable market conditions.

In a quick-fire question, the panel speculates about when a TikTok transaction might be reached. Their different predictions are dated views from September 2025, not reports that an agreement was signed. They also refer back to uncertainty around the H-1B announcement.

The guests disagree on whether Meta’s new smart glasses will succeed. Lemkin, who says he owns several earlier pairs, doubts that a heavier display addresses a widespread need; Stebbings sees potential in integrating computing into daily life. Both are expressing product opinions and predictions.

Finally, they discuss Atlassian’s acquisitions, including DX. O’Driscoll thinks acquisitions can help an incumbent add AI capabilities and serve existing customers without necessarily making it the dominant AI coding company. Lemkin says the purchases are a start but doubts that they alone transform Atlassian.

The guests close with the observation that adapting an established software company to a new technology wave is difficult. Their acquisition, product and market assessments are opinions from the recording, not verified forecasts of future company performance.