Thinking Machines, AI Infrastructure and the Venture Portfolio Tradeoff

In this October 2025 roundtable, Harry Stebbings talks with Roger Ehrenberg, Jason Lemkin and Rory O’Driscoll about the sale of a venture asset manager, the reported departure of a Thinking Machines co-founder for Meta, AI infrastructure and portfolio construction. Their figures, reports and predictions reflect the discussion at the time and are not independently verified here.

O’Driscoll recalls investing alongside venture pioneer Arthur Rock in the 2000s. He remembers Rock as a formidable and direct boardroom presence. The group jokes about the changing expectations placed on investors at different stages of their careers.

The investors discuss the challenge of raising another fund after financial success. Limited partners may ask for more distributions when a fund is young, then question whether a successful manager remains motivated once those distributions arrive. Ehrenberg jokes about his own return to institutional fundraising.

Stebbings introduces the reported acquisition of Industry Ventures by Goldman Sachs. He cites an initial price of $665 million, a performance-dependent additional payment and approximately $7 billion under management. The guests congratulate founder Hans Swildens on building the business over many years.

Lemkin asks how to value a fund-of-funds business and why assets under management do not translate directly into proceeds for its owner. O’Driscoll separates two questions: the market value of an asset manager and the founder’s personal choice between selling a business and keeping its future income stream.

O’Driscoll estimates the initial price at roughly 10% of assets under management and compares that with various public asset managers. He stresses that the percentage varies with the business model. His comparisons are valuation heuristics rather than verified transaction multiples or forecasts of returns.

Ehrenberg, who previously worked in financial-institutions M&A, agrees that fund-of-funds economics differ from taking a fixed share of all assets. He sketches a hypothetical combination of carried interest and management fees. He discloses that Industry Ventures was an early institutional limited partner in IA Ventures funds.

The guests see distribution as a possible rationale for Goldman Sachs: its private-wealth channel could offer private-market products to clients, while Industry Ventures brings institutional relationships. These are their interpretations of the deal’s strategy, not Goldman’s confirmed financial projections.

O’Driscoll argues that a secondary and fund-of-funds platform can be more transferable than a small venture partnership whose value depends heavily on a few individual investors. The buyer must still retain people and preserve the capabilities behind the track record. The group treats saleability as a spectrum rather than a guarantee.

Ehrenberg describes Industry Ventures as an asset-management business built around an institutional platform and fee stream. O’Driscoll contrasts that model with small partnerships built around individual investment decisions, while suggesting that a media or founder-services platform may have more transferable value.

Stebbings raises reports that Thinking Machines co-founder Andrew Tulloch left for Meta after the lab raised $2 billion, citing a reported $3.5 billion figure for the move. The compensation details and circumstances are not established in the conversation. The panel uses the report to debate founder responsibility and investor risk.

Lemkin says he finds it difficult to reconcile a founder’s early departure with the commitment made to colleagues and backers. He asks whether investors must now treat such departures as a more explicit risk even when a company has raised an exceptionally large seed round.

O’Driscoll counters that an extraordinarily large offer changes the decision in ways most founders never face. Ehrenberg still objects to leaving people who relied on a founder’s commitment, saying that maximizing immediate personal gain is not the only value at stake. These are their ethical judgments, not verified accounts of Tulloch’s motives.

Ehrenberg compares some AI talent transactions with acquiring selected assets rather than an entire business, though he sees a difference between those arrangements and the founder departure under discussion. The guests debate how expensive talent competition can change the relationship between a company and the people building it.

O’Driscoll asks what founders and investors can do prospectively. He suggests examining vesting, cliffs and repurchase rights so that co-founders have discussed the consequences if someone leaves. He does not claim to know the actual contractual terms of this case.

Stebbings worries that fundraising is becoming more transactional, with price and speed crowding out personal relationships. Lemkin says his own downside protection as a seed investor has often been confidence that a founder will persist, and that a legal preference alone cannot replace that judgment.

O’Driscoll notes that a research lab investment may depend on a group of technical people, not just the chief executive. He suggests an engineer’s career incentives can differ from those of a founder determined to run a company. Ehrenberg responds that accepting a founder role also creates responsibilities to colleagues and investors.

The guests compare a hypothetical valuable but illiquid stake in a startup with a reported large package of more liquid public-company stock. O’Driscoll stresses liquidity and risk; Ehrenberg notes the startup stake could also have significant upside. The actual package and personal finances remain unverified.

O’Driscoll wonders how deeply investors in a fast-moving round could have met and assessed every co-founder. Ehrenberg argues that a large investor should meet co-founders. Both are discussing due-diligence standards in general; neither establishes what any specific investor actually did.

O’Driscoll argues that an unusually large one-time payoff can weaken the reputational incentives that normally shape repeated business relationships. Ehrenberg agrees that this can increase the range of outcomes. They regard concentration and unexpected talent departures as risks investors need to plan for, not risks with an easy contractual fix.

Stebbings cites reports that SoftBank was arranging a $5 billion margin loan against Arm shares to invest in OpenAI. The guests describe Masayoshi Son’s willingness to make concentrated bets and debate the use of leverage. They present the loan as reported, not independently confirmed transaction terms.

Lemkin suggests a modest loan against a much larger stake may be manageable, while O’Driscoll reminds him that extreme declines in share prices are possible and a margin call cannot be ruled out. The risk depends on loan terms, collateral value and volatility, none of which are fully established in the discussion.

Stebbings asks whether enormous data-center construction and demand for compute make this investment cycle different from a bubble. O’Driscoll questions a simple comparison between data centers and office buildings, since demand for each is shaped by different trends, including remote work.

O’Driscoll describes AI researchers’ confidence in the historical relationship between compute and model performance and their expectation of very large infrastructure investment. He treats that as a view expressed by researchers, then asks the separate question: will enough capital be available, and will it earn an economic return?

Lemkin points to his own experience using AI coding tools to build applications quickly. He sees this as evidence of real utility, but personal usage does not establish that every infrastructure project will be profitable. The panel distinguishes technological progress from the economics of financing it.

O’Driscoll worries that even if model performance keeps improving, marginal economic returns could eventually fall below the cost of additional capital. Ehrenberg asks whether efficiency gains could reduce computing requirements. These are competing hypotheses about future infrastructure demand, not settled outcomes.

Lemkin expects better tools to encourage people to build more rather than simply reduce total compute use. He argues that extremely rapid product creation makes seed investing harder: founders can build convincing software so quickly that technical capability alone may no longer distinguish a company.

O’Driscoll says the window between an uncertain early company and an obviously fast-growing one may be shrinking. Investors can be forced to choose between acute uncertainty and a much higher price. The example of rapid company growth and valuation is illustrative and relies on figures cited in the conversation.

Ehrenberg says his new fund focuses in part on areas where legal, regulatory or intellectual-property complexity creates additional barriers. He is comfortable with early uncertainty when it fits a deeply researched thesis. His comments also disclose his interest as a fund manager in the strategy being discussed.

O’Driscoll, referring to an idea from Box’s Aaron Levie, says AI adoption will diffuse at different rates across markets. Simple application creation may spread quickly, while regulated or complex enterprise uses may require longer sales and validation cycles. Investors should align expectations with the market concerned.

Stebbings raises the reported financings of prediction-market companies Polymarket and Kalshi and asks whether capital can make a winner. Ehrenberg attributes much of the opportunity to differences in regulatory treatment relative to traditional sports betting. His characterization is an opinion, not a legal finding.

Lemkin points to the example of a solo founder building Polymarket during lockdown and argues that new founders can still emerge from unlikely places. He sees the distribution of capital as important, but the story also shows why the next company cannot always be picked in advance.

O’Driscoll argues that financing both competitors does not necessarily anoint either winner. He distinguishes a powerful investor endorsement that influences enterprise customers from a consumer marketplace where users primarily care whether a service works and pays out. His claims about each platform’s business mix are unverified estimates.

Ehrenberg uses a broader definition of kingmaking: large amounts of capital can help a small group fund marketing, distribution and customer incentives. O’Driscoll agrees that capital is necessary in such markets but questions whether a particular venture firm can determine the eventual winner.

The speakers point to political connections around prediction markets and offer critical interpretations of the regulatory environment. These are their observations and opinions; the discussion does not establish that any person engaged in improper conduct or that any particular regulatory decision was improperly influenced.

Ehrenberg argues that tax and regulatory differences may move betting activity between jurisdictions or toward offshore services. O’Driscoll expresses a general preference for less economic regulation. These are policy arguments, not an analysis of all consumer-protection or enforcement tradeoffs.

The conversation shifts to portfolio management. Stebbings compares a concentrated growth-fund strategy with making many smaller AI investments. Lemkin suggests diversification can buy time when an investor believes the opportunity is real but does not yet know which companies will prevail.

O’Driscoll says diversification reduces both the chance of an exceptionally high outcome and exposure to an individual failure. A manager with more conviction and the ability to pick winners may choose to concentrate; different stages and fund objectives justify different degrees of concentration.

Ehrenberg describes his new fund’s intended approach: begin with roughly 20 to 25 companies, then concentrate second and third checks where conviction in the team, execution and market strengthens. Historically, he says, three to five companies could account for around 75% of invested capital. This is his strategy and experience, not a promise of future returns.

Ehrenberg gives an example of an initial $1.5 million investment at a $10 million post-money valuation and explains that he may invest more if the company performs. Stebbings and Lemkin discuss how quickly competing investors can raise prices and make the follow-on decision more difficult.

O’Driscoll says later exits may require more revenue and longer holding periods than before. For his firm, that suggests holding a somewhat broader portfolio initially, then concentrating in promising companies as evidence arrives. His estimates about exit thresholds are observations and expectations, not market rules.

Stebbings challenges the idea that winners can be identified early. Looking back at his own fund, he says some early standouts did not become the strongest long-term outcomes, while other eventual leaders were less obvious at first. The group agrees that early momentum is an imperfect signal.

Ehrenberg describes The Trade Desk’s difficult early journey alongside Wise’s comparatively steady growth. Both became significant investments, but they required different judgments at different moments. He argues that a multi-stage process lets an investor respond to new evidence rather than demanding perfect foresight at the first check.

Ehrenberg says IA Ventures held meaningful stakes in The Trade Desk and Wise at their IPOs, and a smaller stake in Datadog that still produced a valuable return. These are his own reported portfolio examples and carry the perspective of an investor in those companies; they do not imply similar results elsewhere.

O’Driscoll reframes the argument: an existing investor may have better information than an outsider, even if that information is far from certainty. He says his firm’s internal analysis found stronger outcomes when companies met their first two years of revenue expectations. The figures he cites describe that analysis, not a universal probability.

Lemkin reflects on his more concentrated seed approach. To invest a large share of a fund early, he says, he must seek a high hit rate and may turn away unusual opportunities where he lacks conviction. Concentration can improve exposure to chosen winners while excluding others.

Ehrenberg explains that he often starts with about 1–2% of a fund and may commit much more in a subsequent round. He recalls adding to DigitalOcean after its progress and an outside investor’s participation increased his confidence. The example illustrates his allocation process, not a forecast.

Stebbings asks whether this still works when a company’s next round jumps to a much higher valuation. Ehrenberg says each check must stand independently: even a large markup can be attractive if the risk-adjusted future return is sufficient. The prior entry price should not be the sole anchor.

Ehrenberg recalls an investment in The Trade Desk at a substantially higher valuation than his initial checks, which he says still produced a strong multiple. Lemkin counters that rapid valuation increases may reduce the effect of a follow-on on ownership and personal economics. Their views depend on assumptions about future exits.

O’Driscoll identifies the central risk: a follow-on price can be too high relative to the ultimate exit value. If so, an investor can decline the round and retain the earlier position. He sees the option to invest later as useful even when some options expire without being exercised.

Stebbings considers the opportunity cost of holding cash for a follow-on versus funding a new company. Ehrenberg says the decision should compare the risk-adjusted value of each opportunity, incorporating both new operating information and price. A reputable outside lead can be a useful signal, O’Driscoll adds, but it is not infallible.

Lemkin warns that allocating up to 10% of a fund to several early winners can use reserves quickly. If an investor remains on a board for years, he believes continued engagement with the company is an ethical obligation, even when later checks are smaller.

Ehrenberg describes seeking alignment among limited partners across funds so that a later fund can invest in an existing portfolio company with fewer conflicts. He recalls an earlier case where the LP bases differed, the issue went to the advisory committee, and his firm ultimately declined to make the investment.

O’Driscoll returns to the tradeoff between reserving capital for marginal companies and investing aggressively in the best opportunities. Ehrenberg says he will support companies as he can but will not optimize the entire allocation around avoiding difficult future conversations. He also mentions recycling capital as a way to fund additional investments.

The group closes with a lighthearted recollection of one such uncomfortable conversation and thanks Ehrenberg for joining. Their discussion presents individual investment approaches and past experiences, not guarantees or personalized financial advice.