Series A Fund Math, SpaceX Ambitions and the AI Valuation Debate

Harry Stebbings opens with a discussion of AI spending data from Ramp. He says Anthropic captured 73% of the new spending in the sample over a recent period, and asks whether OpenAI is losing its grip on enterprise AI. Rory O’Driscoll stresses the distinction between new spending and the existing base: the same data still put OpenAI ahead on total spending. The figures describe Ramp’s sample, not the whole market.

O’Driscoll regards the shift in marginal spending as a leading indicator, but warns that the dataset has a particular customer mix. Tech companies and fast-moving startups are overrepresented, so he would not extrapolate the proportions directly to every enterprise. Even with that qualification, the panel sees meaningful evidence of Anthropic’s recent momentum.

Jason Lemkin agrees that the picture looks different inside the technology ecosystem. In his experience, developers and startups are increasingly enthusiastic about Claude. That does not establish the same pattern among mainstream consumers or every business customer, and a lead in one group does not by itself settle the broader competition.

The panel reflects on how quickly apparent invincibility can disappear in AI. Stebbings recalls a time when OpenAI seemed impossible to displace; the guests argue that leadership is less permanent than it looks when products and model performance can change within months. Strong current demand is an advantage, not a guarantee of a lasting monopoly.

Lemkin cautions against treating a temporary aura of leadership as a permanent condition. He points to the speed of new releases and organizational change as reasons an investor should remain alert to reversals, even when one company looks dominant today. The group discusses OpenAI’s product and organizational responses in that context.

O’Driscoll thinks a large company can recover ground by focusing on a few essential problems rather than trying to win every contest at once. He sees signs that OpenAI is addressing important issues, even if the response may have arrived later than he would have preferred. His assessment is a view of its strategy, not evidence that the contest is decided.

Lemkin says the improvement in coding tools has been especially striking to people who use them daily. In his view, that experience explains some of the intensity of developer enthusiasm for competing models. He also cautions that an apparent lead can move quickly as providers release stronger versions.

The participants discuss whether token usage provides a reliable map of model competition. OpenRouter’s data offer a useful window into one distribution channel, but they do not represent all direct API and consumer use. Comparisons based on token counts also need to account for differences in how many tokens models use to perform a task.

Lemkin says switching models inside an application is not always as easy as changing a name in a configuration file. Teams must test outputs, revise prompts and scaffolding, and make sure existing workflows still behave as expected. For his own SaaStr agents, he says that work makes him reluctant to switch without a substantial improvement.

Habits also matter. Lemkin describes using Claude for some work while continuing to reach for ChatGPT for certain research tasks. He treats that as an illustration of learned behavior rather than a market-wide measurement: the best model on one benchmark is not automatically the tool a customer chooses for every job.

The panel returns to the challenge of comparing model economics. The guests mention very different estimates of token consumption and efficiency, but acknowledge that a rigorous comparison would need equivalent tasks, output quality and complete cost data. A raw token count alone cannot establish which product is cheapest or most useful.

Stebbings moves to Elon Musk’s proposed Terafab chip factory and the possibility of SpaceX computing infrastructure in orbit. The discussion links potential demand from Tesla and space-based data centers, but those future uses and their scale remain hypothetical. The guests also discuss a possible SpaceX IPO at a $2 trillion valuation; it is a scenario, not a completed offering.

O’Driscoll distinguishes an announcement from a working fabrication plant and from a functioning orbital data-center network. A prediction market’s estimated probability for a valuation is a price for a bet, not proof that the valuation or the underlying projects will materialize. The business case depends on execution, cost and timing.

The guests acknowledge Musk’s record of building ambitious businesses while debating how much confidence to place in his schedules. O’Driscoll notes that major projects can take much longer than initial expectations. Lemkin is more willing to credit the long-term ambition, but the group does not establish the feasibility or economics of orbital compute.

They consider the argument for using solar power in space and the obstacles involved in launching, operating and replacing computing infrastructure there. These are competing judgments about a proposed system, not verified engineering or financial conclusions. The valuation discussion turns on what investors believe could eventually be built, as well as the businesses operating today.

The conversation separates the multiple activities investors might value inside SpaceX, including launch services and Starlink, from speculative future businesses. Lemkin questions how to apply conventional cash-flow valuation when a company is also being priced for projects that have not yet produced the projected results. The panel does not resolve what valuation is justified.

Stebbings cites a Wall Street Journal report that Jeff Bezos was seeking a $100 billion investment vehicle focused on AI and industrial businesses. The panel discusses what such an initiative could mean for incumbent industries. The report is presented as a proposal, not evidence that a fund of that size has been completed or deployed.

The guests compare approaches to industrial change: build a new company from scratch, provide software to existing businesses, or buy and modernize an incumbent. Each approach requires different amounts of capital and carries different execution risks. Their speculation about Bezos’s aims should not be mistaken for a statement of his intentions.

An acquisition strategy could give a technology investor existing distribution, customers and assets, but transforming an established company is difficult. The guests use Walmart as a counterexample to the idea that an incumbent cannot adapt: they point to its ability to compete in the internet era. Buying a large business would also mean paying for the value already recognized by the market.

The panel discusses the appeal of building again after a founder has already achieved enormous financial success. They suggest that a large industrial project may offer a different kind of challenge from simply investing. Those are the speakers’ interpretations, not verified accounts of any individual’s private motivations.

The conversation moves to where ambitious industrial companies can be built and managed. The guests exchange observations about founders, regional ecosystems and the practical work required to change an established business. Their broader point is that financing an effort and operating it well are separate challenges.

Stebbings raises the reported Nvidia–Groq arrangement and a headline figure of about $20 billion. The panel debates the value of inference technology and talent to a much larger buyer. They also speculate about the deal’s structure and economics; without the underlying transaction documents, their estimates of proceeds, taxes and rights transferred should not be treated as established facts.

The guests discuss whether a very high purchase price relative to Groq’s reported revenue could be rational for a strategic buyer. O’Driscoll argues that a buyer may value technology, people and speed differently from a standalone public-market investor. The multiple by itself does not prove either that the price was sensible or that the acquired capability will deliver the expected return.

The speakers describe Groq as having demonstrated a product and commercial demand, while distinguishing those achievements from the much larger valuation discussed in reports. Their revenue and founder-proceeds figures are conversational estimates. The important investment question is whether a buyer can put a capability to use at a scale the smaller company could not reach alone.

The panel also discusses whether the form of an arrangement affects taxation and investor returns. Their examples assume particular asset transfers and tax treatment that are not independently established in the interview. The actual economics would depend on the signed agreements, rights, entity structure and the applicable law.

The Nvidia discussion ends with a broader observation about strategic outcomes: an unusual buyer may pay for time, talent or infrastructure rather than a conventional revenue multiple. That possibility matters to founders and investors, but the panel does not suggest that every expensive AI startup can find such a buyer.

Stebbings shifts to Figma after the launch of Google’s Stitch. He discloses that he owns Figma shares and describes a loss on that investment; he also discloses an Amazon holding. The guests disagree about how much a new AI design tool changes Figma’s prospects and whether the stock market has reacted too strongly.

O’Driscoll argues that a proof of concept should not by itself determine the value of an established product. He also notes that markets may be reassessing the durability of software revenue, not just reacting to a single demonstration. In his view, long-term value depends on what a company can eventually earn rather than any one day’s share-price move.

Lemkin says he has tried Stitch and sees impressive capabilities, while cautioning that Google’s willingness to sustain a product matters to customers. His experience is an individual product assessment, not an independent head-to-head study. The panel asks whether Figma can turn its established workflows and customer relationships into an advantage as the tools change.

The investors consider why software shares could fall even before a company’s reported revenue collapses. If investors become less confident that current revenue will persist, they may assign a lower value to the same present earnings. That is their explanation for market behavior, not proof that a particular company will lose its customers.

Lemkin argues that established software businesses face both an opportunity and a difficult transition: they can add AI to trusted products, but they also carry existing features, integrations and technical debt. A successful response requires useful products and a way to demonstrate customer demand. The panel’s judgments about Figma’s products and management are opinions based on their experiences.

The guests debate Figma Make and what evidence would change their minds. Lemkin says an incumbent needs to show that its agentic offering can compete with newer tools; O’Driscoll emphasizes giving teams time to respond while watching actual customer adoption. Neither treats a short product demonstration as a complete test of long-term outcomes.

Lemkin warns that technical debt is not just old code: mature products also have years of accumulated features and integrations. Those obligations make rapid product changes harder, even as they give existing customers reasons to stay. He worries that a company can fall behind if its response does not improve fast enough, while acknowledging that its future execution remains uncertain.

Asked whether he would buy a software stock after a sharp decline, Lemkin says he would rather miss an initial rebound than assume the company has solved its AI challenge. O’Driscoll agrees that a potential short-term bounce is different from a conviction that the business will win over five or ten years. Their comments are personal investment views, not recommendations.

Lemkin proposes charging for AI as one test of product value. If a company cannot persuade customers to pay for a new capability, he questions whether the capability is strong enough to change its business. That is his heuristic; other pricing strategies and bundled products could make the relationship between price and utility more complicated.

For smaller-business software, Lemkin suggests looking for at least a 50% increase in average revenue per user attributable to AI. He believes Notion may have achieved meaningful pricing uplift, based on its public statements and pricing, but acknowledges that it is a private company and his inference depends on incomplete information. The threshold is his rule of thumb, not an industry standard.

O’Driscoll broadens the test: a company should be able to show traction somewhere, whether in pricing, growth acceleration or contract value. The panel discusses Notion’s pricing as an example and Lemkin says he sees potential in AI features in Slack. These assessments rely on their use and interpretation of company information, not verified causal evidence.

Stebbings raises Kleiner Perkins’ announced funds and argues that leading today’s Series A rounds may require a fund around $1 billion. His illustration assumes 20 investments and initial checks of roughly $25 million to $30 million, with additional capital for follow-ons. The conclusion depends on the portfolio size, target ownership, check sizes and reserve policy a firm chooses.

O’Driscoll says his firm’s approximately $900 million fund operates at a meaningful scale for the Series A business, though not every deal requires the same check. He gives an example of an average $20 million initial investment and a reserve equal to 50% of that initial amount, producing $30 million of intended capital per deal. He prefers a diversified portfolio closer to 30 companies given longer waits for exits.

A clarification matters to the arithmetic: a reserve equal to 50% of the initial check is not the same as reserving half of the total capital for follow-ons. In the example, $20 million goes in initially and $10 million is reserved, so the initial investment is two-thirds of the $30 million total. The speakers emphasize that both reserve conventions are used in the industry.

O’Driscoll says the initial Series A or B check is the central investment for his firm, not merely a small option to write a much larger growth check later. A firm with a separate growth fund can pursue a different strategy. The right fund size therefore follows from how much ownership the firm wants, how many companies it backs and how it plans to finance them.

Stebbings asks whether growing rounds are making the economics uncomfortable. O’Driscoll says average round sizes in his firm’s target market have risen over roughly the past year to year and a half, though he still finds deals that fit. At the most capital-intensive end, a $30 million or $40 million check may no longer be competitive, forcing a decision about whether the risk is worth it.

The group compares fund growth and changing valuations with earlier venture cycles. O’Driscoll estimates that the valuations associated with his firm’s checks have roughly doubled compared with 2010, while suggesting the broader industry has risen more. These are his observations about his own portfolio and the market, not a comprehensive market dataset.

Lemkin describes a trade-off between concentrated bets and accepting lower ownership as rounds grow. He mentions a reported exit-ownership example to illustrate how a very large outcome could still work for a seed fund even after dilution, but says he does not know the details behind that ownership figure. The example is a scenario rather than a verified calculation of a particular fund’s returns.

O’Driscoll jokes that the mathematics itself has not changed; what has changed is the risk implied by larger entry prices. Lemkin agrees that higher valuations require higher eventual outcomes to support returns. Both say the pace of changing deal sizes feels unusually fast, although that comparison is their experience rather than a measured historical result.

Lemkin identifies a further risk: a high private valuation can leave a company with few potential acquirers. If a startup is too expensive for established companies to buy and cannot achieve an attractive IPO, a strong business may still produce disappointing investor outcomes. He does not claim acquisitions are impossible; he questions how many buyers exist at the prices investors are underwriting.

O’Driscoll offers a hypothetical example of an AI application valued privately at $10 billion that might be worth $5 billion at an eventual IPO. Depending on the security’s terms, the last-round investors could take a loss or receive contractual protection that shifts the economics. The example is a scenario, not a forecast or a description of a specific company’s financing.

The guests discuss secondary sales as a possible source of liquidity for early investors. They caution that a strong company can still have an interim private valuation above what a buyer or public market will eventually pay. Whether to sell depends on the investment, available terms and expectations; their discussion does not establish a universally correct strategy.

O’Driscoll adds that AI applications marketed as replacing more of a customer’s work may be valued above the older software companies they aim to displace. That can create a strategic paradox: the incumbent that might benefit most from buying a challenger may be unable to afford it. A smaller earlier acquisition could be possible, but it carries its own uncertainty.

Lemkin says he would feel more comfortable with elevated valuations if large cash acquisitions were occurring frequently enough to provide credible exits. He worries that raising money at a high paper valuation can be easier than achieving a meaningful realized exit. That is his assessment of market conditions and liquidity, not a claim that every highly valued company will fail.

O’Driscoll says his firm has participated in parts of the market while avoiding some of its highest valuations. He acknowledges that the restraint has looked costly in a rising market, and credits investors who took bold early positions in Anthropic. He expects discipline may prove valuable across a full cycle, but admits that the outcome is not yet known.

Stebbings says he regrets not being more flexible with his Series A fund’s mandate and using access to participate earlier in popular companies. O’Driscoll translates that into a portfolio decision: a manager could spend several of the same-sized investment slots on later-stage momentum deals instead of initial Series A positions. Such a choice may offer markups, but it changes ownership, concentration and exit exposure.

The panel examines whether later investments in exceptional AI companies could ultimately produce returns comparable to early-stage bets with less risk. O’Driscoll points to Anthropic as an example of the argument, while questioning whether other application companies will have markets large enough to justify their prices. A markup in a later financing is not the same as a realized return.

O’Driscoll observes that momentum investing works in a rising market by definition, making skill difficult to distinguish from exposure to the cycle. Stebbings responds that disciplined investors can appear to lag while others obtain rapid markups or liquidity. They end with the question facing fund managers: when should they stay with a strategy, and when is there a sound reason to break it?