Nvidia’s AI Spending Outlook, Anduril’s Army Contract and the Venture Math Behind Big Bets

Harry Stebbings opens the roundtable with Jason Lemkin and Rory O’Driscoll by asking what stood out from Nvidia’s GTC event, including discussion of data centers in space. Lemkin says the company projects unusual momentum, pointing to initiatives around agents, open models and partnerships. His characterization of competitors and corporate momentum is an investor’s impression, not a measurement of their performance.

Lemkin sees risks to Nvidia from customers developing or using alternatives such as Google and Amazon chips, but says the breadth of Nvidia’s announcements suggests confidence. He mentions Groq-related technology and the agent platform NemoClaw as examples. Whether those initiatives achieve the effects he expects remains to be seen.

O’Driscoll focuses on the trillion-dollar headline. He says Nvidia’s demand figure covers multiple years and, in his reading, is broadly consistent with existing analyst forecasts, rather than a trillion dollars of revenue in one year. His explanation for a relatively muted share-price response is an interpretation of market expectations, not a verified account of every investor’s reaction.

He contrasts Nvidia’s rapid historical growth with the growth implied by future forecasts. In his view, those forecasts require customers to keep investing extraordinary sums in infrastructure for several years. Past growth makes the scenario conceivable, but does not establish that spending will continue on that path.

Stebbings asks whether Nvidia could become a $10 trillion company by 2030. Lemkin shifts to a different measure, speculating about $10 trillion of cumulative revenue over a longer period, and the guests clarify that Nvidia’s current trillion-dollar headline is cumulative. These are separate concepts: market capitalization, revenue and bookings cannot be interchanged.

The conversation turns to inference demand. Lemkin expects agents to run continuously and generate far more token usage, suggesting this could support hardware demand. O’Driscoll notes that higher volume does not automatically mean higher revenue: if the price per token falls faster than usage rises, revenue can decline. Neither gives a verified forecast for either variable.

O’Driscoll compares the expected capital spending to previous infrastructure booms and calls a sustained expansion a demanding assumption. He offers rough illustrative ratios between Nvidia revenue and wider capital expenditure, then suggests there is a meaningful chance the projected path does not materialize. His percentages and spending figures are estimates, not a formal model.

Lemkin reads Nvidia’s message as confidence that many model developers will still need substantial data-center and inference infrastructure. He thinks several possible industry paths could benefit the company, while acknowledging competition around its position. His further multiyear cumulative-bookings prediction is speculative and should not be confused with guidance for a single year.

Stebbings turns to Atlassian’s announced reduction of 1,600 positions and reports of a potential 20% workforce reduction at Meta. He describes the Meta figure as speculative; the panel has no confirmation that such a reduction occurred. He asks how much layoffs reflect earlier hiring, slower growth or AI.

Lemkin argues that some companies are reassessing the skills and roles they need rather than simply responding to immediate cash shortages. He predicts significant organizational change as AI tools improve, including changes to engineering and go-to-market work. His suggestion that large portions of teams may become unnecessary is a forecast and judgment, not evidence about any individual company’s actual plans.

Lemkin quotes an engineering leader’s account of changing software work and argues that AI may alter both writing and reviewing code. He also expects sales and marketing teams to rethink work previously done through sheer headcount. The conversation does not establish that all those activities can already be automated or that all affected roles will disappear.

O’Driscoll proposes several distinct explanations for layoffs. One is correcting earlier overhiring; another is slower growth pushing a company to seek profitability. A third is genuine productivity improvement from AI, which he thinks may apply to coding but should not automatically be assumed to apply equally across every function.

His fourth category is shifting resources toward compute and capital expenditure; he applies that interpretation to Meta. This is his inference about spending pressures, not confirmation of the company’s motives for any proposed layoffs. A fifth possibility, prompted by Lemkin, is replacing some existing roles with different skills and potentially hiring again.

Stebbings asks what kind of employee companies should seek. Lemkin’s answer is practical AI fluency: he wants people who can explain how they use tools to improve their own work. His examples span engineering, sales, marketing, product and quality assurance, and he presents his estimates of how many candidates meet that standard as personal observations.

Lemkin disputes a simple account that all 2021 hiring was wasteful. In his view, high-growth companies needed many people under the workflows and growth assumptions of the time; both assumptions have since changed. His broader argument is that managers should reassess what work still needs people, while recognizing that future outcomes are uncertain.

Asked how to test AI fluency, Lemkin suggests asking candidates what useful commercial AI product they have introduced or seriously evaluated recently, and how they assessed its value. He proposes a demanding monthly cadence as his own hiring heuristic. He later clarifies that a candidate need not have purchased a tool if they can explain a deep evaluation and its limitations.

Lemkin recounts a problem he says SaaStr encountered with an unnamed AI company’s pricing agent. In his account, the agent suggested a much larger bill because it had not been properly trained on pricing. The anecdote illustrates his concern about implementation and testing, but the company and circumstances are not independently verified in the interview.

He describes the job as deploying useful agentic products and understanding how they fit an organization’s workflow, rather than merely experimenting with prompts. Lemkin says he sees some companies in which he has invested embracing that approach and others lagging. This is his management advice and assessment of his portfolio, not a proven hiring rule.

O’Driscoll stresses that willingness to learn matters more than age and says investors should spend time using the tools they assess. Both guests frame adaptation as a continuing responsibility, including for themselves. Their advice reflects a view about current technology, not a conclusion that every role requires the same tools.

Lemkin recalls finding early AI systems too technical for his own use and contrasts them with newer software he believes generalists can deploy. He describes the familiar elements of workflow design and decision trees, while saying agents still require training. His experience is encouraging to nontechnical users, but does not demonstrate that all deployments are simple or risk-free.

He argues that someone who has successfully deployed enterprise software can learn to deploy an AI agent, and describes buying, configuring and training an AI sales tool. He predicts that employees who do not adapt could face employment consequences. That is his forceful personal warning, not a finding that any individual will be laid off.

Stebbings introduces Anduril’s Army agreement, describing a $20 billion contract with a five-year base and five-year option that consolidates more than 120 separate actions. The number refers to the discussed headline contract amount over its term, not $20 billion of annual revenue or guaranteed spending. The interview does not establish the amount the Army will ultimately order or pay.

O’Driscoll interprets the arrangement as an important procurement consolidation rather than simply a wholly new program. In his account, fewer separate contracting steps can make it easier for units to acquire products. His description of customer concentration and the agreement’s practical effect is an investor’s interpretation, not an audit of the contract.

He identifies Lattice, Anduril’s software connectivity system, as central to the discussion and explains why connecting different hardware systems quickly could matter in defense. He believes the arrangement signals a stronger role for Anduril as a supplier. The panel does not establish exclusive status, a monopoly or operational performance for any particular system.

The size of the agreement prompts Lemkin to question his own seed-investing approach. He says he previously backed strong founders in relatively small markets on the expectation that the market could expand, but now struggles to pursue opportunities he cannot imagine becoming enormous. He explicitly acknowledges that this shift in his thinking may be a flaw.

Lemkin worries that early investors chasing only the largest outcomes will pay high entry prices for several competing companies and suffer many losses. O’Driscoll separates the appeal of a large market from the danger of blindly backing the third or fourth entrant. Neither claims that a large market guarantees an individual startup will win.

O’Driscoll disputes the idea that defense necessarily offers unlimited room for new entrants. He discusses the constraints of the defense budget and speculates that only a handful of companies may become very large winners. The percentages and winner count are his rough judgments, not verified budget analysis or a forecast with supporting calculations.

The guests discuss power-law returns and fund size. O’Driscoll notes that a 10% stake in a $1 billion outcome would return about $100 million before relevant adjustments: roughly one times the size of a $100 million fund. Larger funds therefore need larger absolute outcomes, but excessive focus on outliers can encourage investors to overpay or imitate recent successes.

O’Driscoll argues that market size alone is not enough: growth, the ability to lead a market and entry price matter too. Lemkin agrees that a mid-sized market can create a good company, but says an expensive seed valuation may make the resulting return inadequate. Their disagreement is about investment economics rather than the intrinsic value of smaller businesses.

Stebbings asks what is large enough, citing AI sales and support applications. Lemkin says growth investors can wait for evidence that a product and market are working, although they may pay a higher price. He mentions Intercom’s acceleration in a discussion that includes O’Driscoll’s investment connection; performance claims are the speakers’ assessments.

Lemkin considers a counterargument to his own position: AI agents could make separate sales and support markets converge, expanding what looked like a narrow opportunity. That expansion is a hypothesis, not an established market size. The hard seed-stage decision is whether to invest before convincing evidence of expansion exists.

Stebbings points to early coding products as examples of apparently small categories that might become large. Lemkin admits that he might not have recognized the opportunity in Replit before the current AI wave, while describing how investors could have seen a broader future for web-based software development. His account is retrospective, not evidence that the outcome was obvious at the time.

Lemkin says he is more skeptical of narrow tools that fix one small weakness in a coding product, particularly if model providers could add the feature themselves. A startup can describe a huge future market, he says, but he wants evidence that its product can occupy a defensible and meaningful position. This is a screening preference, not a claim about any named startup.

O’Driscoll pushes back on the idea that moving to growth investing solves the problem. When capital floods an attractive area, investors may bid up prices and reduce future returns. He contrasts periods when growth capital was scarce with periods when late-stage valuations rose sharply, emphasizing that a good business can still be a poor investment at the wrong price.

His firm’s approach, O’Driscoll says, is to look for both early evidence of product-market fit and a credible larger trend. He uses Anduril’s early border-watchtower work as an example of near-term traction paired with a possible broader defense opportunity. Whether any investor could identify the eventual outcome in advance is a separate question.

Stebbings moves to Travis Kalanick’s return with Atoms, described as a rebranding and expansion of his earlier CloudKitchens and City Storage Systems work into robotics and industrial use cases. The guests discuss an interview in which Kalanick favored practical machines over an exclusive focus on humanoid robots. The company’s future commercial success remains uncertain.

O’Driscoll agrees with the view that wheels and task-specific machines can be useful in factories and warehouses, where legs may introduce cost and complexity. He does not rule out humanoids forever; he thinks specialized equipment may be a more practical near-term path for some work. This is his engineering and market judgment, not a universal performance comparison.

Lemkin argues that Uber might have been worth $1 trillion had Kalanick remained its CEO, because he believes the company could have moved faster in autonomy and food delivery. He also acknowledges serious concerns around Kalanick’s earlier tenure. The valuation is a counterfactual opinion, not an observable result or evidence about what would have happened.

O’Driscoll challenges the counterfactual by distinguishing a talented founder from the opportunity and constraints facing a business. He asks whether Uber needed to move toward profitability and public-market access at a time when autonomous driving was still years from substantial deployment. A more financially focused manager, he suggests, may have been suited to that period.

Lemkin replies that professional management might have worked well through an earlier phase but believes a more aggressive product-led approach would now be valuable. O’Driscoll partly agrees that the right management emphasis can change over time, while noting that it remains unproven whether Uber must own autonomous-driving technology to participate successfully.

Asked when a board should replace a founder, O’Driscoll says it is an extremely disruptive step. He describes two possible high-threshold reasons: decisions that threaten the company’s survival and serious systemic behavioral problems. He says he lacks visibility into the specific private deliberations at Uber; the framework does not establish the reasons for any particular decision.

The discussion returns to Atoms as an investment. Lemkin admires Kalanick’s earlier entrepreneurial insight but says, based on the interview, he lacks sufficient conviction to invest from his own fund. O’Driscoll prefers to begin with the market: he finds industrial autonomy plausible, while questioning whether one robot can address many different industrial segments effectively.

Stebbings raises a hypothetical $20 billion post-money valuation based on an AI-generated estimate of what a fundraising round might seek. The figure is unverified and should not be treated as a confirmed financing or target. The guests use it to discuss whether a very large growth fund could size a speculative position differently from a smaller fund.

Stebbings says he might make such a bet from a hypothetical multibillion-dollar growth fund, while saying he would not do so with his own 20VC fund. O’Driscoll cautions against investing merely because a manager needs to deploy capital and admires a founder: if the opportunity and price are unconvincing, he says, a smaller fund may be the better answer.

Lemkin recalls meeting Kalanick during the Red Swoosh era and being struck by his explanation of the future of online video. The story illustrates why founder insight matters to him, but he distinguishes admiration from having sufficient conviction for a particular investment. The guests return to the point that position size and fund size affect how much risk a manager can take.

The final topic is Adobe. O’Driscoll corrects the sequence after Stebbings initially suggests a stock decline caused the CEO’s departure: the earnings and transition news were announced together, followed by a share-price decline. The panel does not know how much of the market reaction came from either announcement.

Lemkin, who says he previously worked at Adobe, considers the absence of an already named successor unusual. He speculates about possible reasons for the transition but repeatedly acknowledges he does not know what happened. The identity of any eventual successor and the board’s private motivations remain unresolved in the interview.

The guests discuss how difficult it can be to run a public software company when growth slows and markets demand profitability. They speculate that more CEOs may consider leaving under such pressure. These are general views about leadership incentives, not evidence of the private reasons for Adobe’s transition.

O’Driscoll initially treats the announcement as potentially abrupt, then checks and says Shantanu Narayen is staying on as CEO until a successor is found. He revises his interpretation accordingly. The panel considers possible board motivations, including investor pressure, but presents them as conjecture rather than established facts.

Lemkin argues that a large installed base and low customer churn can protect an incumbent without guaranteeing future growth. He says he has not seen convincing evidence in Adobe’s product response that would make him confident in growth. This is his investment assessment, not a claim that Adobe cannot grow.

O’Driscoll compares Adobe with Intuit. In his view, accounting and tax workflows may be comparatively durable, whereas creative work could change more substantially if users adopt new ways to produce content. He predicts significant disruption risk for Adobe over five years, while acknowledging that the future of both companies is uncertain.