OpenAI’s AMD Deal and the Venture Math Behind AI’s Biggest Rounds

In this October 2025 roundtable, Harry Stebbings, Jason Lemkin and Rory O’Driscoll discuss OpenAI’s AMD partnership, AI product announcements and unusually large venture rounds. They also examine liquidity, possible IPOs and competition between startups. Their figures, reports and forecasts reflect the information available in the episode and are not independently verified here.

Stebbings introduces OpenAI’s announced chip supply partnership with AMD, describing plans for up to six gigawatts of AMD Instinct chips and warrants potentially representing up to 10% of AMD. O’Driscoll sees the deal as an example of bargaining power between chip suppliers and a major AI customer.

Stebbings presses on an important distinction: a warrant is a right to acquire shares, not immediate ownership. O’Driscoll says the potential warrants depend on conditions including chip purchases and stock-price milestones. The panel’s discussion of value is conditional; it does not establish that OpenAI already owns 10% of AMD.

O’Driscoll imagines how the parties might have negotiated: OpenAI could argue that its demand helps AMD’s market position and seek a share of the upside. This is his reconstruction, not an account of the actual negotiations. He notes that delivery and purchases still have to occur for the relevant conditions to matter.

Lemkin offers another possible interpretation: OpenAI wants an alternative to Nvidia without giving AMD a free commercial endorsement. He compares the logic with other platform-vendor arrangements. The guests do not claim to know the parties’ complete economic terms or intentions.

O’Driscoll compares the emerging AI supply chain with the historical relationship among Microsoft, Intel, AMD and IBM. In his analogy, OpenAI controls a consumer relationship, Nvidia supplies critical chips and AMD offers another source. The comparison is a framework for discussion, not proof that the industries will evolve alike.

He argues that Microsoft’s early OpenAI partnership may have helped create a powerful future competitor, while acknowledging a major difference from his IBM analogy: Microsoft has an economic interest in OpenAI. Ownership can produce financial gains even if competitive dynamics become more complicated.

Asked how Nvidia might view the AMD arrangement, Lemkin suggests Nvidia can tolerate some competition while defending its substantial position. His assessment of Nvidia’s margins, market share and the companies’ negotiation strategy is opinion based on the information then available, not a disclosed plan.

O’Driscoll says OpenAI’s user demand gives it leverage even while it incurs large costs: chip suppliers want customers able to use substantial capacity. The revenue trajectory and future purchases he mentions are expectations, not guarantees that OpenAI will fund or deploy every proposed commitment.

Lemkin contrasts Nvidia’s apparent pricing power with the ordinary economics of component suppliers. O’Driscoll says scarce competition and architectural lock-in can change that relationship. Both are describing market dynamics from their perspective, rather than establishing precise chip costs or future margins.

O’Driscoll notes that semiconductor startups have often been difficult venture investments even while some public chip companies have prospered. He contrasts concentrated supplier markets with more cyclical categories such as memory, cautioning that competition can reshape economics over time.

The conversation turns to OpenAI’s developer announcements, including using applications within ChatGPT and tools for building agents. Stebbings asks whether integrations could reshape software distribution and threaten existing application companies.

Lemkin sees promise but says the demonstrations did not yet show him a compelling new cross-application use case. He compares the idea with Slack integrations: an integration may exist without becoming the way most people perform complex work. His reaction is an early product judgment, not a measurement of adoption.

O’Driscoll distinguishes invoking an application inside ChatGPT from embedding AI functionality in another application. Searching for a home through a conversational interface might be useful, he says, while a dedicated interface could still be preferable when a user wants to compare many detailed options.

The panel expects some tasks to move into conversational interfaces but is unsure how many users will abandon standalone apps. Stebbings asks whether agent-building tools threaten specialist vendors. O’Driscoll says an eight-minute demo cannot establish that: enterprise agents may also require orchestration, management and integration.

Stebbings asks whether a reported billion-dollar fundraise at a $5 billion pre-money valuation for a startup associated with former Databricks AI leader Naveen Rao breaks venture math. He notes that dilution means a very large eventual company value could be needed for a tenfold investment return.

O’Driscoll explains why investors may pay up for a proven technical founder in an infrastructure market: the pool of people with relevant experience is small, and a track record can reduce some execution uncertainty. But even if the founder picks and solves the right problem, the market must be large enough to justify the entry price.

Lemkin argues that founders who have already seen a company reach a very high valuation may find it easier to imagine a similarly large outcome. He contrasts that mindset with earlier, cheaper seed investments. His examples illustrate the change in expectations, not a guarantee that every new company can reproduce a prior success.

The guests debate the claim that an AI wave could produce future Amazon-scale companies. O’Driscoll replies that a great company does not automatically create a great venture return: entry price, dilution and time to exit matter. A higher starting price requires more things to go right.

O’Driscoll calls venture unusually forgiving of some pricing mistakes because exceptional growth can outweigh a high entry price. He immediately adds that this forgiveness has limits: if only a minority of companies succeed and the price caps the upside of each success, the whole portfolio can struggle.

He cautions against relying only on comparable valuations. A company can look cheap next to an even more expensive peer while both are overpriced relative to what they will be worth years later. Investors need an independent view of durable revenue, future multiples and exit value.

Lemkin asks how very large funds can find enough exceptional companies to deploy their capital quickly. The panel sketches hypothetical fund sizes and check counts, then notes that fees, reserves and available opportunities affect the calculation. Their estimates about specific funds and prospective payouts are not verified results.

Stebbings raises reports that university endowments are selling venture fund interests. O’Driscoll suggests liquidity needs and pressure on universities may contribute to a reassessment of illiquid allocations. He regards such sales as a possible adjustment rather than evidence that every endowment will keep selling indefinitely.

Lemkin recounts an LP secondary sale from his own experience and criticizes the friction and permissions involved. He says fund documents can leave limited partners with little ability to sell without approval. This is one participant’s account of a specific process, not a finding about the unnamed parties.

O’Driscoll agrees that more liquidity can benefit investors and employees when companies remain private for much longer than expected. Secondary sales can help clean up an old fund or give employees access to value, though they may involve discounts, transaction costs and restrictions.

Lemkin says liquidity can be scarce outside a handful of highly sought-after companies. O’Driscoll adds that liquidity can disappear when investors become fearful even if capital still exists. Both see public markets as one possible source of liquidity, with their own tradeoffs.

The panel discusses Snyk, citing an estimated roughly $300 million in annual recurring revenue and 26% growth, alongside rumors of private-equity interest. They ask whether the company could go public or sell; neither the operating figures nor potential buyer interest is independently confirmed in the conversation.

O’Driscoll cites his review of 2025 IPOs, including a median revenue run rate near $931 million, and says the reported Snyk figures would sit near the lower end of the range he observed. The number is his stated sample and metric, not a universal IPO threshold.

The guests use rough security-company comparisons to speculate about possible valuation, including private-equity and public-market scenarios. Lemkin explicitly calls his comparison approximate. Their figures are hypothetical estimates, not an offer, an appraisal or evidence that an IPO is impossible.

O’Driscoll argues a company can also continue compounding, make acquisitions and build scale before seeking liquidity. The choice depends on growth, competitive durability and whether management and investors can tolerate a longer holding period. A company can be valuable even if it does not match an earlier private valuation.

Lemkin says several of his own portfolio companies with strong operating metrics had not received private-equity offers, despite his expectation that they might. The anecdote highlights his uncertainty about demand; it is not a claim that no buyers exist or that any specific company has failed.

O’Driscoll recommends that management teams take control of their options instead of assuming a buyer will arrive. One tool he describes is a fresh equity grant tied to growth, intended to motivate a founder or executive whose earlier equity has already vested. Its suitability depends on the company and the terms.

The guests discuss rebuilding growth through product expansion, relevant AI offerings and acquisitions. Each requires operational execution and may bring its own risks. A new incentive plan or a strategic narrative is not a substitute for finding customers and delivering sustained growth.

Stebbings asks about founders who may not want to continue the long journey. O’Driscoll and Lemkin discuss supporting founders with experienced executives and recognizing that the decision to continue or sell can depend on personal circumstances, incentives and the actual offer.

O’Driscoll rejects asking an early-stage founder whether they would accept a hypothetical $500 million offer. He argues that answers given before a real offer exists reveal little: people who say they will never sell can change their minds, and people expected to sell may choose to continue.

Stebbings says very strong or troubled investments can sometimes become apparent in the first two months, while the middle remains difficult to judge. O’Driscoll agrees there can be useful early signals, but frames them as imperfect judgment rather than certainty about outcomes.

Stebbings turns to Vercel’s reported $300 million raise at a $9.3 billion valuation and asks whether a high-priced round with a comparatively modest amount of new capital creates dangerous expectations. The guests treat the financing figures as reported and debate the tradeoff rather than declaring a verdict.

Lemkin links Vercel and Supabase to the growth of app development: developers need hosting and databases, and both companies had become important choices in his experience. He considers the underlying trend compelling while acknowledging that the price of the investment remains a separate issue.

O’Driscoll agrees that both businesses were positioned before the generative AI boom and have benefited from a new wave of application building. He suggests a valuation can rise rapidly while the revenue multiple remains similar if revenue rises just as quickly; this is an illustrative framework, not verified current financial data.

Stebbings asks whether the same growth rate at a much larger scale still supports the same multiple. O’Driscoll says the risk is a market-size ceiling or sudden deceleration. A company can appear to maintain its growth trajectory and then discover that the addressable opportunity is smaller than investors assumed.

O’Driscoll disputes calling every high-priced round a suicide round. He says the most dangerous combination is rapid deceleration plus enough spending to force a return to market at a lower valuation. Cash already raised and the company’s actual burn matter alongside the headline price.

Stebbings argues that large early rounds can help investors crown leaders, especially in smaller markets. O’Driscoll challenges the size distinction, suggesting that the largest AI markets also display strong capital advantages. They disagree about how much credit belongs to the investor versus the founding team.

Lemkin says capital-intensive categories may have room for only a limited number of companies capable of raising enough to compete. O’Driscoll responds that a strong product and early traction usually come first, after which investor brand, financing and momentum can reinforce an existing lead.

Stebbings pushes back that substantial rounds can occur before meaningful execution, when investors are backing an exceptional founder or insight. O’Driscoll concedes the phenomenon but emphasizes a common pattern of early customers and growth followed by successive rounds from prominent firms.

The guests discuss legal AI competitors, including Harvey and Legora, as examples of companies attracting substantial capital. They debate whether investor reputation and a large balance sheet can influence hiring, enterprise customers and later funding. They do not establish that any company has definitively won its market.

O’Driscoll points out that some AI businesses have high variable costs as usage grows, which makes the need for capital different from traditional high-gross-margin software. Stebbings sees funding as an offensive advantage; the panel also recognizes that spending heavily can create pressure if growth slows.

Asked whether founders should accept a large offer of capital, O’Driscoll says that in a competitive market he would generally lean toward taking attractive funding, while considering what rivals may do. Lemkin says founders vary considerably in their appetite for spending and risk, and investor advice cannot change every team’s nature.

O’Driscoll brings up a newly discussed SPAC and says its sponsor economics appear better aligned than some earlier structures. He describes a payout tied to a higher stock price but remains skeptical that the structure will replace a well-run IPO. His description concerns the deal as discussed, not all SPACs.

The guests discuss incentives and promotional risks in public-market transactions. Claims and accusations raised in the conversation are not established as wrongdoing. O’Driscoll’s comments about disclosure rules are a simplified opinion in the episode, not a complete or current statement of securities law.

Lemkin raises the reported $2 billion investment in Polymarket by Intercontinental Exchange, owner of the New York Stock Exchange, at a reported $9 billion valuation. The panel discusses the changing regulatory and political environment; its claims about particular relationships and legal status are opinions and reports, not findings of misconduct.

O’Driscoll asks whether prediction markets can build durable non-sports businesses as well as sports-related activity. His figures about activity mix are uncertain estimates. He speculates that a strategic investor may value data, distribution or market access, but the discussion does not establish the deal’s complete terms.

In a quick-fire question, Stebbings asks whether Tim Cook might leave Apple within the year. The guests regard it as unlikely while acknowledging unconfirmed succession speculation. O’Driscoll says he owns Apple shares and has trimmed his position, a relevant financial interest in his comments.

Stebbings asks whether Replit and Lovable could exceed $250 million in annual recurring revenue by year-end, citing then-current figures that are not independently confirmed. Lemkin predicts they could, but says casual users may churn more quickly than committed users. He describes his own extensive use of Replit.

Lemkin discusses a third-party traffic analysis and suggests interest may have temporarily flattened, though traffic is not the same measure as paid revenue or retention. He sees improving products and a more committed customer base as possible supports for growth, while treating his revenue call as a prediction.

The panel closes by noting how quickly AI coding platforms are improving and how difficult it is to extrapolate growth from early adoption. Their year-end revenue forecasts and market comparisons are opinions made in October 2025, not reported subsequent results.