AI Spending, Layoffs and IPO Speculation: Jason Lemkin and Rory O’Driscoll on the Economics of Agents

Harry Stebbings opens a roundtable with SaaStr’s Jason Lemkin and Scale Venture Partners’ Rory O’Driscoll on the economics of the AI boom. Their discussion ranges from chip spending and layoffs to purported IPO filings, AI infrastructure startups and the trade-off between buying tokens and employing people. Many financial figures and prospective deals are discussed as reports or estimates, not independently verified facts.

Stebbings says Nvidia reported $81.6 billion in quarterly revenue, guided to $91 billion for the next quarter and announced an $80 billion buyback, yet its stock barely moved. O’Driscoll highlights the much larger profit figure he says accompanied the revenue, while cautioning that some items were nonrecurring. He argues that a strong result need not move a stock much when investors have already priced in extraordinary growth.

O’Driscoll recalls that investors had already repriced Nvidia as they absorbed the scale of potential AI capital expenditure. Lemkin says that in a market expecting companies to beat, raise guidance and accelerate, merely holding steady after results can itself be a strong outcome. Both stress that market expectations, not just the size of the reported numbers, shape the response.

The panel makes rough calculations linking hyperscaler capital budgets, the GPU share of spending and Nvidia’s annualized revenue. Stebbings cites Jensen Huang’s projection of $3 trillion to $4 trillion of AI infrastructure expenditure by 2030. O’Driscoll treats the extrapolation as possible rather than certain: the central question is whether the next increments of investment will produce enough economic return.

Lemkin points to comments attributed to Uber’s COO and reports about Microsoft reconsidering expensive models as examples of customers questioning marginal returns. He distinguishes these claims from proof that AI has no value: companies can be getting real benefits while finding the additional cost difficult to justify or measure.

The panel returns to the reported example of Uber consuming its annual AI budget in just four months. The immediate question is not whether employees used the tools, but whether management can identify the productivity gained and the right level of future spending. Stebbings pushes back on treating a measurement problem as conclusive evidence of zero productivity.

O’Driscoll argues that businesses must evaluate AI on the value of a completed task, not only the price of an individual token. If a model is cheaper per token but needs more inference or fails more often, the result may not be cheaper. The useful calculation compares the total cost and quality of accomplishing a job against the alternatives.

The participants debate whether customers will continue paying a premium for frontier models as less expensive alternatives improve. Lemkin worries that model costs can consume software margins, particularly when customers expect AI features without paying enough extra. O’Driscoll sees scope for both cheaper inputs and customers spending more on harder tasks as model capabilities improve.

O’Driscoll refers to research on the changing price and performance of AI models. His reading is that lower unit prices can be offset by users choosing more capable, computationally intensive systems. A declining price per token therefore does not by itself settle how much a company should spend or how much work it will undertake.

Stebbings asks how model providers’ own economics fit with demand. The participants cite figures they say they have heard or seen reported for OpenAI and Anthropic and make forward-looking comparisons of their growth and potential profitability. These numbers and forecasts are not established by the conversation, and the panel acknowledges limited visibility into private companies’ finances.

The panel considers whether enterprise budgets for people and AI tokens are starting to compete. Lemkin’s concern is that spending heavily on premium models does not automatically make a business more efficient; leaders must establish what customers and employees get for the expense. O’Driscoll frames the question as an allocation decision between human work, tools and the quality or volume of output.

The discussion moves from the cost of a model to the economics of the company using it. A tool can generate impressive activity without creating enough measurable business value, while a more expensive system could be worthwhile if it completes a task reliably. The speakers differ on how quickly organizations will be able to distinguish those cases.

Stebbings and the guests discuss whether heavy demand for frontier AI justifies the infrastructure being built. Their answers remain conditional on sustainable customer value, pricing and margins, not simply on the pace of model adoption. Strong current usage is one input into that calculation rather than a guarantee of future returns.

Lemkin returns to the pressure on software companies that must pay for inference while maintaining gross margins. A company can pass some expense to customers, route work to cheaper models or redesign the task, but each choice affects the experience. The panel treats model selection as an economic decision, not just a contest over benchmark performance.

The conversation turns to reported layoffs at ClickUp and Cloudflare and the explanations offered for reductions. Stebbings asks how much is attributable to AI. The panel distinguishes companies’ statements and the guests’ interpretations from a demonstrated causal breakdown; hiring during the pandemic, organizational restructuring and AI adoption can overlap.

Lemkin argues that executives are reconsidering how many people they need as software and agents take on portions of work. O’Driscoll is more cautious about attributing every cut to AI, noting that earlier overhiring and other management decisions matter too. A layoff announcement alone does not establish how much work an AI system actually replaced.

The guests discuss the human consequences: a company-wide efficiency story looks very different to someone who has lost a job. They explore whether routine attrition could reduce headcount without mass layoffs, but their arithmetic is an illustration of what might happen over time, not evidence that any particular company followed that path.

Lemkin describes his own preference for a small team and discusses concentrating compensation on people who can produce unusually high value. That is his management perspective, not a general conclusion about employees who have been laid off. The panel also recognizes that workers face the practical challenge of acquiring skills for changing roles.

The question is not only whether AI can perform a task but who captures the benefits. The panel debates higher pay for a smaller number of highly productive employees, the fate of other workers and the limits of treating staffing as an abstract optimization problem. Their projections are uncertain and do not establish what will happen across the economy.

Stebbings raises reports that OpenAI had confidentially filed for an IPO and asks what that might mean for Anthropic. The panel discusses a possible timetable, valuations and fundraising needs, but a confidential filing is not independently established here and does not guarantee a public listing. None of the transaction details they discuss should be treated as confirmed.

O’Driscoll compares figures circulating about the two companies and sketches scenarios for growth and profitability. He explicitly acknowledges that outsiders cannot see the full numbers. His prediction that Anthropic could overtake OpenAI or become profitable is a forecast based on incomplete information, not a verified financial result.

Lemkin says he would like exposure to both companies, while the guests debate what the market might pay for a leading model provider. The discussion separates the appeal of a rapidly growing business from the valuation risk investors would take at a potential offering. Timing, pricing and the companies’ future economics remain unknown.

The panel considers why a company might go public while still investing heavily and whether an IPO by one provider would increase pressure on another. Access to capital and public scrutiny could both matter, but the participants do not claim to know the companies’ actual plans. Their comparisons rely on reported or speculative financial data.

The conversation shifts to what the guests describe as a SpaceX filing and a proposed transaction involving xAI. They discuss purported revenue, a possible valuation and AI-related commitments, including a claimed monthly arrangement involving Anthropic. The transcript provides no independent confirmation of the filing, acquisition, contract or terms; the subsequent valuation discussion is conditional on those claims.

The speakers examine what a very large proposed valuation would imply relative to trailing sales. Lemkin warns that even a business with compelling technology can be a poor investment at an excessive entry price. O’Driscoll says he admires the ambition of the scenario while being skeptical of buying at the valuation being discussed.

The guests debate whether a combined space, connectivity and AI business would justify the optimistic assumptions they attribute to the alleged filing. They compare the situation with past high-priced technology deals and stress that an investor’s view of the business and an investor’s view of its price can diverge.

Orbital data centers become a thought experiment. The panel considers energy, launch costs, cooling, connectivity and the difficulty of building infrastructure in space. The speakers disagree about eventual feasibility and economics; their speculation does not demonstrate that a proposed system works or can compete with terrestrial facilities.

The potential role of Starlink leads to a broader discussion of the network and AI infrastructure a space-based business might assemble. The guests see possible strategic connections, but also substantial technical, financing and execution questions. Their estimates of future scale and return are opinions rather than established outcomes.

Stebbings asks the guests to weigh the ambition against the price. Lemkin says he can support a founder’s long-term vision while questioning a valuation based on a very high multiple of current sales. O’Driscoll similarly distinguishes enjoying an optimistic narrative from deciding to own a share at the proposed price.

The panel returns to the scale of projected AI spending and the possibility that different companies want the same limited supply of compute. Claims about future contracts, capacity and market size are treated as scenarios: neither demand forecasts nor a purported prospectus establish what a business will ultimately earn.

The participants discuss infrastructure investment as a bet on sustained AI demand, with the warning that spectacular forecasts may be reflected in prices long before profits materialize. Even if the underlying technology succeeds, the returns to a particular investor depend on entry valuation, dilution, competition and execution.

The discussion moves to Polsia, which Stebbings describes as a one-founder AI startup that has attracted attention with unsolicited email and a funding round. He says he met the founder during the fundraiser and cites a reported valuation of around $250 million, while being unsure of the exact amount raised. Those figures are his account, not independently established here.

Lemkin jokes about the valuation and compares the hype around the product with other AI phenomena, but acknowledges that he does not know enough to judge the business conclusively. He then tries its onboarding during the recording and says it does a good job researching his business and specifying a proposed inbound sales tool.

His enthusiasm stops at the payment screen: he says the product requests a $49 monthly subscription before showing him the resulting value. Lemkin’s experience is that tools he keeps using let him see value first, even if that requires the company to spend a little on tokens. This is his individual test and buying preference, not a finding about every customer’s experience.

Stebbings raises reported funding rounds for Exa and OpenRouter and mentions a possible deal involving Manus’s founders; the panel does not establish those transaction terms independently. O’Driscoll sees the first two as examples of infrastructure being built below the foundation-model layer, where developers need search, model access and other components to make agents useful.

O’Driscoll explains why an agent may need current web information in addition to the material encoded in a model. He describes Exa and Parallel as tools for structured search, and OpenRouter as a way to choose among models. In his view, routing can help customers decide when a cheaper model is sufficient rather than paying for the most expensive model on every query.

Lemkin discloses that he is a small investor in Exa and says he used it before investing. He considers agent-native search especially interesting because autonomous workflows need tools designed for machines to retrieve current information, not necessarily the same interfaces people use. His claim about future demand is an investor’s thesis, not a guarantee of market growth.

Stebbings asks what the upside could be at the reported Exa valuation and whether one company would dominate. O’Driscoll thinks developer tools may support more than one significant supplier because individual developers choose them and the market lacks some of the network effects associated with consumer advertising. He presents that as a possibility, not a settled market structure.

For O’Driscoll, the investment case depends on many agents needing frequent access to structured information. He speculates that search could expand into curated or internal information and uses a hypothetical revenue multiple to reason about valuation. The size of the opportunity depends on adoption and pricing that the panel cannot know in advance.

The panel praises Benchmark’s investing and discusses rapid markups in AI companies. Lemkin discloses that he is a small shareholder in Monaco and describes a reported valuation increase. He argues that the best moment to invest may be just as early traction becomes obvious, although a paper markup is not a realized return.

O’Driscoll pushes back on dismissing seed investors and says early product-market fit can reduce risk considerably relative to a company with no revenue. He observes that some AI companies appear to move from early traction to visible scale in weeks or months rather than years. That speed creates pressure to make decisions with less evidence.

Lemkin points out the missing data: a young company may have little renewal history, and a powerful incumbent could enter its market. O’Driscoll agrees that investors may have to pay more with less information if they want to participate early. Both describe a trade-off between speed and uncertainty rather than a dependable investment formula.

The guests ask whether agent infrastructure can compensate for lower revenue per use with much larger volume. Lemkin cites a figure he attributes to Databricks’ Neon—that more than 90% of its databases are created by agents—and argues that tools will need widespread, repeated use for this model to work. The figure and the prediction are his account, not independently verified in the interview.

O’Driscoll suggests that developers may eventually settle on a small set of widely needed components, such as databases, search and observability. Products needed by only a narrow subset of agents could have a much smaller market despite attractive early adoption. Search and model routing are candidates in his view, but he says he could be wrong.

For a final round, Stebbings relays an unnamed CEO’s claims: a custom-built CRM replaced a $600,000 Salesforce contract in three weeks; the company expects to eliminate 80% of its SaaS; and it would not change its Anthropic usage if prices doubled. These are claims about one company, not independently checked evidence or a general forecast.

Lemkin calls the CRM story provocative but says the real comparison must include maintenance, integrations, collaboration and the value of an established ecosystem. His own organization uses Salesforce headlessly with its own agents on top. A small, highly specialized company might rationally build its own system, he says, but he does not see this example as a leading threat to established CRM vendors.

O’Driscoll challenges the pricing claim from a different angle: if the return on an AI project would remain attractive at twice the price, management might have other valuable tasks it should fund. That is a marginal-return argument, not proof that doubling AI expenditure would always be profitable. Lemkin notes that the company Stebbings describes is unusually efficient, according to the figures cited in the conversation.

Lemkin says his own team’s agents can produce more ideas than its people can review and implement. For him the constraint is not simply the cost of another token but human attention and judgment. Even useful suggestions create a queue: someone has to decide which ones matter and carry them through.

O’Driscoll offers an optimistic possibility: if AI produces more genuinely valuable ideas than one employee can handle, adding another person could raise the return on human work. Lemkin agrees in principle but says he wants an unusually capable person for his particular needs. Their closing theme is the same question running through the discussion: how should a business allocate resources between people and AI, and what follows from that choice?

The participants end with a light exchange about their different personalities and social habits. The substantive question remains unresolved: costs, capabilities and jobs will vary by company, and the value of either an agent or an employee has to be assessed in context.