AI Funding, IPO Expectations and the Politics of Layoffs: Jason Lemkin and Rory O’Driscoll

Harry Stebbings joins SaaStr founder Jason Lemkin and Scale Venture Partners investor Rory O’Driscoll to discuss AI financing, public software companies, chip infrastructure, prospective IPOs and the politics of job losses. Their company figures and transaction details are presented as claims or reports from the recording, while their market and employment forecasts are opinions, not established outcomes.

Stebbings opens with reports that Anthropic is in talks to raise $30 billion at a valuation above $900 billion, compared with a reported $380 billion valuation in February. He also raises a reported move by Andrej Karpathy to Anthropic. The guests discuss the financing as a possibility; the conversation does not establish its completion, terms or his employment independently.

Lemkin asks whether annual recurring revenue multiples still matter at this scale, citing an estimated 18 times revenue for the reported price. O’Driscoll contrasts that with the higher multiples sometimes paid for smaller private companies. Their comparison depends on the accuracy and comparability of the reported revenue, growth and valuation figures.

Lemkin says discounted future cash flow must ultimately matter, but observes that growth investors often use revenue multiples without fully adjusting for weaker gross margins. If the 18-times figure and improving margins were appropriate, he argues, the proposed Anthropic round might look attractive against earlier-stage deals. That is a conditional investor judgment, not a verified assessment of value.

O’Driscoll argues that a large, fast-growing company can look cheaper on a revenue multiple than a young startup despite its size. He treats a future IPO as likely, but a potential listing remains uncertain and would not eliminate valuation, financing or execution risk. Stebbings asks why Anthropic would accept the suggested price if investors view it favorably.

O’Driscoll estimates that raising $30 billion at a $900 billion valuation would be a small fraction of ownership and could strengthen the balance sheet against heavy compute commitments. He says the company may need further capital, though he is unsure whether another round would precede a possible listing. Both the dilution calculation and future financing needs are his assumptions.

Lemkin offers his interpretation of how Sam Altman and Dario Amodei approach fundraising: he thinks one pushes harder on price while the other may prefer a fast, relatively straightforward deal. These are his impressions of personal motives, not independently verified accounts of either executive’s intentions or any financing negotiation.

The guests compare the financing structures they understand Anthropic and OpenAI to be considering, with O’Driscoll suggesting that cash timing and conditions can make headline amounts misleading. Their descriptions of the terms are conversation claims, not confirmation of closed investments. He favors understanding how much cash is actually available and when.

Stebbings cites Salesforce CEO Marc Benioff’s reported statement that Salesforce would spend $300 million on Anthropic tokens, mostly for coding. Lemkin estimates that this works out to roughly $15,000–$20,000 per engineer annually and a modest share of total employment cost. The calculation depends on the quoted spend and headcount, which the panel does not independently verify.

O’Driscoll says his team surveyed around 40 portfolio and outside companies and found average AI spending of roughly $1,200–$1,300 per developer per month, with a lower median. He treats the Salesforce estimate as broadly comparable, while noting variation among heavy users. The small survey is an example, not a representative estimate for all companies.

O’Driscoll argues that model-provider valuations depend in part on how much of engineering and knowledge-worker spending shifts to tokens. Using a hypothetical $1 trillion in future token revenue, he estimates that it would require substantial spending relative to wage bills. His math is a rough scenario based on assumptions, not a forecast guaranteed by Salesforce’s current spend.

He sketches two possibilities: if spending remains near current levels, some projected markets may have been overestimated; if companies multiply usage, tokens could absorb a much larger share of operating costs. Lemkin notes a countercase: better models and more efficient workflows could reduce the number of tokens needed for a given task.

Lemkin says SaaStr has 21 agents, three autonomous, and that its direct AI costs for two people are about $2,000 a month. He stresses that figure excludes third-party applications and tokens embedded in Salesforce, so it is not the entire cost of the workflow. His own small team is one example, not a market-wide benchmark.

Lemkin describes hearing Klaviyo CEO Andrew Bialecki explain that product-adjacent employees must commit code and use AI or agents. He says the company uses its own framework and recounts Bialecki’s close estimate of the approximately $257 direct monthly cost of two SaaStr agents. These are Lemkin’s accounts of specific practices and costs, not independently verified company-wide results.

The guests return to scale: O’Driscoll says a small token bill may be barely visible relative to R&D, whereas the much higher share required by some large market projections would materially affect budgets. He frames the comparison as a hypothetical; it does not demonstrate that spending at that level would cause a specific layoff.

Asked about Klaviyo, Lemkin says he is more optimistic about established software companies than he was 90 days earlier. He points to reported growth at Figma, Atlassian and Twilio as reasons to think some incumbents still have time to develop agent products. His assessment of particular companies and future stock performance remains an opinion.

Lemkin nevertheless wants evidence that internal AI adoption turns into products customers will buy. In his view a company can be advanced internally yet lag startups externally. He speculates that strong leaders could improve their position by 2027, but says the growth must show up in results rather than solely in a narrative.

O’Driscoll agrees that a rebound does not mean software stocks will regain the exceptionally high multiples of 2021. He argues that mature companies tend to be valued increasingly on growth and cash flow. His valuation ranges are illustrative opinions, not a rule or prediction that any particular stock will reach a given multiple.

Stebbings cites Datadog’s reported first $1 billion revenue quarter and annual recurring revenue above $4 billion. The panel treats the performance as an example of a software company growing despite a difficult sector narrative, while distinguishing operational results from the stock price and earlier exuberant valuations.

Turning to Figma, Stebbings cites reported accelerating growth and 139% net dollar retention; Lemkin says he sold his holdings at the end of the preceding year. Lemkin remains critical of Figma’s vibe-coding product but acknowledges that he underestimated how a broader increase in software creation could benefit the company.

Lemkin notes that he and Stebbings have invested in WorkOS and RevenueCat, companies he sees as beneficiaries of more software being built. He argues that Figma can also add value to existing users by applying AI inside design workflows, even if he believes its initial move into application generation was slow. These are his product and investment views.

O’Driscoll asks whether competitors such as Lovable could draw users into a different design flow, where a working prototype replaces an initial mockup. Lemkin thinks professional design capabilities remain limited but argues Figma missed an opportunity to turn its designs directly into production prototypes. His estimate of revenue left on the table is speculation.

Stebbings contrasts those results with Wix, whose share price he says fell 45% after a buyback, and cites a reported $150 million ARR for Base44. O’Driscoll interprets the market as questioning growth in the legacy business despite the new product. His interpretation and the cited metrics are not independently verified in the discussion.

Lemkin argues that website builders face pressure from AI-assisted site creation and from Shopify in commerce. He thinks Wix and Squarespace remain useful for people who prefer them, but sees stronger competition for other users. His description of competitors’ prospects is a personal assessment rather than proof that their businesses are terminal.

O’Driscoll says a newer product can initially be too small to move a much larger company’s aggregate growth rate, even if the product grows quickly. Over time, sustained growth could change the combined picture. That possibility depends on retention, adoption and execution, and does not establish a turnaround.

On Wix’s buyback, O’Driscoll argues that buying shares before a further decline may consume capital that could have supported future acquisitions. He contrasts retaining strategic flexibility with trying to support the stock in the short run. This is his retrospective opinion based on the price movement he cites, not knowledge of management’s motives.

The guests debate whether buybacks should offset dilution from acquisitions or stock compensation. O’Driscoll argues that price and alternative uses of cash should determine the decision. Lemkin speculates that management teams may also try to address activist pressure; he does not establish that as the reason for any named company’s decision.

O’Driscoll outlines two broad strategies for a mature software business: optimize an established, slower-growth product for cash, or preserve resources for reinvention and acquisitions. He notes that even a buyback investors once supported may be criticized if the shares later fall. Management, in his view, must weigh both price and strategic options.

Asked whether the stock under discussion will be higher in a year, O’Driscoll tentatively says yes but admits he has not examined churn closely. He reasons from an estimated revenue multiple and possible operating margins, then acknowledges that both revenue and the multiple could fall further. This is a speculative scenario, not an investment recommendation.

Lemkin predicts that at least one beaten-down, founder-led software company will successfully combine an existing customer base with a strong AI product. He imagines a focused team building something competitive with a startup and distributing it to existing customers. He cannot identify in advance which company would succeed, and the outcome is uncertain.

Stebbings cites rapid growth at Nebius and asks whether it reflects compute scarcity or excessive enthusiasm. O’Driscoll says the economic answer depends on whether demand stays ahead of supply: scarce capacity can command strong returns, while plentiful capacity can turn the business into a commodity. Neither outcome is certain.

O’Driscoll relays an argument attributed to investor Gavin Baker that slow data-center permitting may prevent too much capacity from arriving at once. He sets out a hypothetical mismatch between capital expenditure and token revenue, then considers how slower construction might keep compute scarce. His numbers are scenarios, not observed proof that capacity will be balanced.

Lemkin observes that hardware, memory and networking companies appear to be benefiting alongside AI infrastructure while traditional software has lagged. He says shorting a boom years before supply catches up is difficult and sees no evidence of a near-term crash in his own assessment. That is a time-specific view, not a guarantee about markets.

O’Driscoll argues that sustained infrastructure spending eventually requires customers to buy enough useful AI services, directly or through software companies. If application businesses fail to grow, he thinks demand for model providers could eventually weaken. The relationship is his economic thesis; future spending and revenue remain uncertain.

Lemkin asks whether the guests would invest in a new compute-capacity startup at seed stage. O’Driscoll describes a team he liked but passed on because he worried about adding capacity several years into a capital-intensive boom. The founders’ response made him reconsider; he still does not know whether the decision was right.

Asked about feedback to founders, O’Driscoll says a clear short refusal is usually sufficient after one meeting, while more detailed reasoning may be useful after substantial work on a deal. He also writes down his own conclusions so he can compare them with later outcomes and learn from opportunities he declined or nearly pursued.

Lemkin agrees that detailed feedback is most useful after investors have gone deep enough to articulate a real reason. After only a brief meeting, a long explanation may prompt an unproductive argument. The value of the practice, for both sides, depends on how much the investor actually understands the business.

Stebbings says Cerebras priced its IPO at $185 after increasing the range and rose 68% on its first day. He asks whether this will open the market to more offerings. Lemkin thinks it helps companies with similarly strong investor demand, but does not assume that a single successful transaction opens the window for every smaller software company.

O’Driscoll calls Cerebras a distinctive case: a semiconductor and inference company that, he says, secured OpenAI as a major customer just as demand for AI compute rose. He notes that an earlier attempt at an IPO did not proceed. His interpretation of demand and the company’s prospects should not be generalized into a guaranteed outcome for other listings.

Asked whether he would buy at $300, O’Driscoll says probably not. He cites his understanding of historical returns after first-day IPO pops and distinguishes a company’s technological potential from the price an investor pays. He presents the historical pattern as his own claim; it is not independently analyzed in the discussion.

Stebbings introduces reports of a possible SpaceX listing on June 12 at a $1.75 trillion valuation with a $75 billion raise. O’Driscoll says the panel is recording before the expected S-1 and that he has not seen it. Date, valuation, offering size and the transaction itself therefore remain reported possibilities, not confirmed outcomes in this conversation.

O’Driscoll expects a prospectus to contain historical financials and wonders how much light it would shed on more recent AI-related transactions. He refers to reported xAI, Anthropic and Cursor arrangements, but the panel does not independently establish their status or terms. His description of what might appear in a future filing is a prediction, not a reading of a document.

He says a company can be appealing while investors still face substantial uncertainty about its current financial profile and price. He expects enthusiasm could support a listing in a favorable market, but cautions that a change in investor preference toward cash flow could alter the story. Neither trading demand nor future disclosure is known.

Lemkin anticipates strong retail interest in a potential SpaceX offering and asks whether that could produce a dramatic price rise. O’Driscoll notes that the reported starting valuation and offering size would make a tenfold rise much harder than for a small, thinly traded stock. Both are discussing hypothetical market behavior, not an actual offering result.

O’Driscoll describes how institutional investors might sell if an IPO rapidly reaches their price targets. He mentions a rumored institutional allocation but explicitly says he does not know the facts, and uses it only as an example. He also recalls Facebook’s weak initial trading to illustrate that an admired company can fall below its offering price.

Lemkin speculates about a rise toward $5 trillion, while Stebbings suggests $3 trillion. O’Driscoll cautions that a decline is also possible and later says he would take the below-$3-trillion side at the end of the first month. These are the guests’ competing bets, not facts or promises of a future valuation.

O’Driscoll returns to Facebook’s IPO as a warning against assuming a popular story must trade upward. He thinks positive trading is more likely in his scenario but stresses that high pricing leaves room for disappointment. The panel’s disagreement illustrates uncertainty about sentiment, supply of shares and fundamentals.

Stebbings describes a reported offer of $2 million in OpenAI tokens to each startup in a Y Combinator batch in exchange for equity. The guests discuss its possible impact on startup valuations and financing, without independently confirming the offer’s complete terms or whether credits are transferable.

O’Driscoll sees the reported offer as a way to attract developers and assumes, without confirming, that the tokens cannot be transferred. He thinks the benefit could reduce some costs but may not replace the need for cash to hire people. Lemkin argues that the credits could reduce risk for early startups and create upward pressure on valuations.

The guests discuss whether a reported reference price could anchor later financing rounds and whether startups might raise less cash or offer investors a smaller ownership stake. These are hypothetical effects: the value to each company would depend on the actual terms, its token usage and what it would otherwise spend.

Lemkin suggests that a startup could use token credits to offer customers generous free trials, treating inference as a marketing expense. O’Driscoll agrees that this differs from using the credits solely for engineering or paying customers. The economics still depend on usage, conversion and any restrictions attached to the offer.

O’Driscoll notes an opportunity cost if a model provider gives away capacity that could otherwise be sold. Lemkin agrees that the trade-off changes if capacity would otherwise be idle. The guests speculate about relative compute availability at OpenAI and Anthropic, but their conversation provides no direct evidence of either company’s spare capacity.

The discussion turns to the reported outcome of litigation involving Elon Musk and OpenAI. The guests offer interpretations of a limitations issue and the jury’s reasoning, but their reconstruction of private deliberations is speculative. Their predictions about an appeal and views of the parties’ motives are opinions, not verified legal findings.

The panel discusses possible questions about Sam Altman’s indirect business interests and OpenAI’s organizational structure. Lemkin advances claims and theories about a venture fund, board knowledge and personal motives; O’Driscoll offers a different interpretation. The interview does not independently establish those allegations, the reasons for personnel decisions or the individuals’ intentions.

O’Driscoll worries that public unease about AI may grow as industry leaders discuss risks, job displacement and rising costs. He offers his own political interpretation of public sentiment, not polling evidence or a verified account of any jury’s views. Lemkin takes a more favorable view of Altman’s public approach.

O’Driscoll contrasts a past graduation audience’s enthusiasm for ChatGPT with a reported negative reception for Eric Schmidt at another event. He treats the anecdotes as signs of changing sentiment, not a representative measurement. The guests cite several reported technology layoffs, and explicitly note that LinkedIn said its cuts were a realignment rather than caused by AI.

Lemkin argues that technology companies may eventually have a social obligation to hire more people if AI displaces large numbers of workers. O’Driscoll qualifies the argument: the outcome depends on how many jobs AI can actually replace, and executives may overestimate the technology or cut too deeply. Their positions are predictions and policy opinions, not established labor-market results.

O’Driscoll says that if very large white-collar job losses occurred, a substantial social response might be needed, but he personally doubts the most extreme projections. He warns that announcing layoffs in the name of AI could have political consequences. Lemkin worries that displaced workers may struggle to find new jobs; that is his forecast, not a finding about every worker.

Near the end, Stebbings relays a headline suggesting OpenAI might file for an IPO as soon as Friday. O’Driscoll explicitly says he does not know whether the report is true and is reacting in real time. He speculates about what a confidential filing might mean, while the group agrees they would need to wait for further information.

The participants end by noting that a reported filing would only begin a process and would not necessarily reveal public financial details immediately. The headline, the timing of any listing and the trading outcomes discussed throughout the episode remain unconfirmed at the time of their conversation.