AI Coding Valuations, NVIDIA Risk and the Future of Private Markets

Harry Stebbings welcomes Tomasz Tunguz of Theory Ventures alongside Jason Lemkin and Rory O’Driscoll. Recorded in November 2025, the roundtable discusses AI coding companies, NVIDIA, credit risk and the changing market for private-company shares. Its valuations and predictions reflect the participants’ views at that time.

Stebbings introduces a reported Cursor financing of $2.3 billion at a $29.3 billion valuation. Tunguz says coding is a strong AI use case and points to rapid growth, but his cited productivity gains and model-speed comparisons are observations rather than independently verified benchmarks.

Tunguz argues that a small team and growth could make the investment attractive, while warning that customer retention and enterprise adoption need examination. The reported revenue, employee counts and proposed return are not independently verified financial disclosures.

Lemkin sees AI coding tools becoming a standard developer expense rather than an optional productivity boost. The panel then constructs a very large market-size estimate by multiplying assumed developer counts and annual spending; those inputs are speculative and do not establish actual demand or Cursor’s future revenue.

Tunguz describes paying for coding assistance and sometimes exhausting a subscription allowance. He uses this personal experience to ask how much high-usage customers might pay, not to establish average willingness to pay or future pricing across all developers.

O’Driscoll challenges the trillion-dollar estimate while agreeing that a narrower professional-developer market could still be large. He identifies profitability and durability against competition as the main questions at a high entry price.

The group discusses a platform risk: an application may pay a model supplier that also sells a competing coding product. Lemkin asks how a vendor can sustain gross margins; the percentages mentioned in the exchange are estimates, not audited company results.

Tunguz says developers may switch when a model improves substantially but stay when workflows, configuration and memory create enough friction. His five-year retention scenario is a guess, and his impressions of specific models are not independent comparisons.

Tunguz describes a small internal model-distillation experiment and says there may be opportunities to lower costs. He does not know whether coding companies will reach traditional software gross margins; the experiment is not a benchmark for the industry.

The investors make hypothetical market-share picks among Cursor, Microsoft, Anthropic and other coding products. These percentages are personal forecasts, not current share measurements, and the panel does not know which products will lead in five years.

Lemkin describes his recent Replit experience and the potential for multiple agents to review work. He remains open to new competitors because the technology is changing quickly; his product observations and claims about context are his own usage impressions, not verified technical specifications.

He says quality assurance remains a bottleneck and imagines an agent that could handle more of it. O’Driscoll counters that enterprise standardization and switching costs may eventually stabilize market shares, while allowing that a major technical change could disrupt them again.

Tunguz frames the disagreement as a question of when a rapidly moving market settles. The speakers use different historical analogies to make the case for continuing disruption or eventual stability; neither analogy proves the future path.

Lemkin says his organization transferred a prompt from another agent into Salesforce Agentforce and made it work after iteration. His anecdote suggests some work may be portable, but does not measure switching costs for other customers or validate the products generally.

Tunguz asks whether commoditization could reduce prices. O’Driscoll distinguishes getting more tokens for the same expenditure from an outright price war in which customers pay less; the latter is a scenario the group considers, not an observed universal market trend.

The guests compare cheap alternatives to enterprise software with the high costs of replacing deeply integrated products. Their central question is whether AI agents will be embedded in workflows like enterprise applications or become easier to swap like commodity components.

Lemkin says his team interacts with sales agents on top of Salesforce rather than logging into the system as often. He argues that value could migrate to the agent layer even if the underlying customer remains; this is a single organization’s experience and Salesforce is among the products he uses.

He reports high starting prices for some agents and forward-deployed support, and speculates that a later price war could compress revenue. These prices are examples from his experience, not a representative price survey or a forecast guaranteed to occur.

O’Driscoll says integrations can make a product harder to remove, while a portable prompt and loosely coupled data could make substitution easier. Tunguz asks what happens if customers control and route those prompts across suppliers; the discussion is a scenario, not evidence that all implementations are interchangeable.

Stebbings contrasts early-stage company building with repeated late-stage valuation increases. The panel cites reported Cursor and Ramp financing rounds and their own observations of newly valued startups; a private markup is not necessarily a realized gain or a liquid market price.

Lemkin points to established firms participating in later rounds, while O’Driscoll says large managers may find those investments attractive when they can deploy substantial capital. Their comments about named firms’ internal risk assessments are interpretations, not statements from the firms.

O’Driscoll warns that private-market trading can look easy on the way up but lacks comparable liquidity during a downturn. A later mark does not guarantee an investor can sell a position at that price, especially when buyers retreat.

Stebbings asks whether reported NVIDIA sales by Peter Thiel and SoftBank signal a market top. The guests offer possible alternative explanations and explicitly do not know the sellers’ motives. A trade by one investor is not proof of the direction of the overall market.

Tunguz turns to Oracle credit-default-swap spreads as one indicator he is watching. He says the relative change may suggest higher perceived risk even if the absolute default probability remains small; a CDS spread is a market price, not a direct measurement or prediction of default.

The group discusses Oracle’s reported data-center commitments and borrowing as a possible source of investor concern. Their comments on contracts, credit markets and company motives are interpretations and do not establish that Oracle cannot meet its obligations.

Tunguz identifies customer concentration as another risk for NVIDIA while noting that many major buyers generate cash. The percentages and company comparisons are his estimates from available information, not verified revenue attribution in this discussion.

He proposes that a slowdown in inference demand or an underused data center could affect expectations. O’Driscoll says the episode was recorded before an NVIDIA earnings release, so their guesses could be overtaken by information available when viewers watched.

Tunguz worries that leverage could amplify a correction if infrastructure demand weakens. O’Driscoll suggests limited electricity supply might slow construction, but presents that as an upside scenario, not proof that overcapacity or a correction will occur.

Lemkin recalls sharp software-stock declines during earlier cycles and says volatility can occur even in a growing market. The participants’ suggested magnitudes and timing are speculation rather than a forecast investors can rely on.

O’Driscoll notes that long-run market charts can hide a long recovery after a severe decline. His comments about portfolio allocation reflect his personal risk tolerance, not individualized financial advice.

Stebbings describes exuberance around a startup accelerator batch and asks about fundraising behavior. O’Driscoll says bargaining power changes with the supply of capital, while arguing that both founders and investors should still treat one another respectfully.

Tunguz asks whether the US venture market could reach $500 billion by 2030. The group debates the definition of venture capital and the concentration of money in a few large companies; the historical totals cited differ by source and are not reconciled in the episode.

O’Driscoll reminds the group that venture funding has previously contracted sharply, so extrapolating a rising line can miss downturns. Lemkin observes that several unusually large financings can dominate aggregate totals even if funding for other startups is weaker.

The investors reason that future industry inflows may depend heavily on returns from a small number of large private companies. The concentration could magnify good or bad outcomes; their scenario is not a verified allocation or guarantee of future fundraising.

Stebbings asks O’Driscoll about Scale Venture Partners’ investment in GC AI, a legal tool for in-house teams. O’Driscoll explicitly discloses the investment and cites customer references, adoption and capital efficiency as his reasons; company performance claims are not independently verified.

O’Driscoll says he values cash-efficient growth when paying a high entry price, particularly if later financing could be difficult. He argues that customers ultimately choose software for its usefulness rather than solely because a famous venture firm invested in it.

Tunguz discusses a reported Stripe tender and compares the direct cost of a late private financing with the cost and recurring obligations of listing. The specific costs are rough examples; actual transaction fees, disclosure, liquidity and investor rights differ substantially.

The guests discuss an apparent private-market “access premium” for a limited number of sought-after companies. O’Driscoll cautions that this may not apply to ordinary businesses without strong demand for their private shares.

Tunguz hypothesizes that retail money could reach venture through funds and intermediaries. The conversation treats regulatory changes and fund structures as possibilities, not as confirmation that any specific retirement-account or exchange-traded product is available or suitable.

O’Driscoll says investors may keep supplying capital while returns appear strong, but eventual losses can change behavior. Stebbings notes that retail inflows could arrive before long-term private-market performance is clear; both timing estimates are speculative.

The group raises liquidity mismatch and infrequent valuation of private assets as risks for funds promising access to retail investors. They discuss other asset classes as analogies, not proof that a named venture fund has a particular redemption problem.

O’Driscoll reflects on the responsibility of telling investors when they have lost money. He warns that fees earned by managers and long lockups can leave investors with disappointing outcomes, even if fundraising initially looked successful.

Lemkin says the IPO market feels weaker than he had hoped. Tunguz argues that liquidity should also be measured through mergers and secondary sales, citing estimates of their growing share; the numbers are not reconciled against a complete dataset.

Tunguz imagines a broader secondary market for smaller private companies, while Lemkin questions how much liquidity would exist for weaker names. O’Driscoll agrees that a market-clearing price may exist but expects many of the largest eventual exits still to go public.

The panel compares fees in private and public investment structures. O’Driscoll sees higher aggregate costs in many private arrangements; Tunguz predicts competitive pressure could lower some fees. Their examples are simplified and do not describe every fund or product.

Tunguz proposes that early investors could sell portions of a stake over multiple private rounds rather than wait for an IPO. O’Driscoll accepts the analogy to partial public-market sales but stresses the different transaction costs and uncertain downside liquidity.

O’Driscoll says the trend toward larger private markets is visible, but its durability in a significant downturn remains untested in his view. He highlights the risk of being unable to sell private holdings when liquidity is most needed.

The group considers whether continued acquisitions of public companies and private secondaries could reduce the number of listed software businesses. This is a forecast dependent on future financing, regulation and investor behavior, not a demonstrated inevitability.

The panel’s quick-fire segment asks for hypothetical choices between coding companies and legal AI businesses. The picks are personal judgments at the time, not investment advice; O’Driscoll’s Scale investment in GC AI is relevant context for his comments on the sector.

Asked about an OpenAI IPO, the guests offer different quarter and year guesses. They are predictions made in November 2025, not company announcements or confirmed financing plans.

The guests discuss how ownership and dilution can influence a CEO’s incentives. Lemkin cites an unnamed executive with anti-dilution terms as an anecdote; it does not establish the terms or motivations of any named company leader.

The episode closes with jokes about forming a growth fund and shifting away from early-stage investing. The remarks are banter, not an announcement of a new fund or an investment mandate.