20VC on AI Financing, Token Budgets and the Cost of Startup Intensity

Harry Stebbings, Jason Lemkin and Rory O’Driscoll discuss reported AI-company financings and listing plans, public software valuations, token budgets and intensive work cultures. Their company figures and predictions are statements made in the episode, not independently verified filings, financial results or investment advice.

Stebbings begins with a report that Anthropic has filed to go public and asks whether a listing would help the broader ecosystem. Lemkin argues that exceptionally fast growth and very large potential valuations may reset expectations for founders, investors and employees. The captions do not supply a filing or establish that a listing took place.

Lemkin wonders why investors or employees would settle for smaller outcomes when a few companies appear capable of reaching extraordinary scale quickly. He sees a possible psychological effect: people may feel pressure to join the hottest companies or disregard businesses that could otherwise be valuable.

O’Driscoll pushes back on building a strategy around finding another extreme outlier. In his view, investors should acknowledge a missed opportunity and continue making good investments rather than assume a comparable outcome will recur on demand. His comments about company values and frequency are judgments, not verified forecasts.

O’Driscoll admits that he has reflected on missing an opportunity to meet Anthropic early. He says that an eventual public listing could remove some mystery by making financial information visible and returning liquidity to investors, though the timing and amount of any distribution are unknown.

Lemkin explains that, at his current career stage, he is increasingly interested in investments that could produce a billion-dollar position for his fund. He distinguishes that from a billion-dollar company outcome and acknowledges that he would have applied a different threshold earlier in his career.

Stebbings challenges the premise: many investors say they underestimated the eventual market and outcome of their best investments. Lemkin answers that he cannot know the final size in advance but can eliminate opportunities where, in his judgment, the team or ambition creates a clear barrier to the upside he seeks.

O’Driscoll corrects the size of his fund to $900 million and asks how many billion-dollar positions are realistically available. He prefers underwriting a plausible base case while identifying a credible, uncapped upside. A strong return from a smaller outcome can still matter to a fund, he says.

Stebbings says he is more willing to take risk when an investment can, rather than will, produce a huge outcome. O’Driscoll recalls investing through the dot-com collapse and warns that investors can become especially brave after a long period of rising markets. His recollections are a caution about risk, not a prediction of a specific crash.

Lemkin worries that high expectations can make it harder for founders of solid, less spectacular businesses even to get meetings. O’Driscoll agrees that unusually large outcomes can anchor expectations, but believes public disclosure could eventually make a prominent company seem less mysterious.

The guests discuss reports about Anthropic and OpenAI pursuing listings, while acknowledging uncertainty about the precise filing language and sequence. O’Driscoll argues that companies with large AI infrastructure needs have an incentive to seek capital sooner. The episode does not independently establish either company’s filing status or listing date.

O’Driscoll also cites a reported Google capital raise and estimates that several large companies could collectively seek hundreds of billions of dollars. He describes this as a rush to secure financing for capital-intensive AI. The proposed amounts, investors and timing are his account of reports, not verified transactions.

On the choice between equity and debt, O’Driscoll argues that a company might use both to preserve balance-sheet flexibility for uncertain future spending. Lemkin asks why management would accept dilution if investment payback were rapid. The guests do not know the company’s internal plans or eventual returns.

O’Driscoll describes the shift from asset-light cash generation toward heavy capital expenditure as a potential change in the investment profile of technology companies. He says payback periods are uncertain; his examples of infrastructure economics are illustrative rather than a demonstrated result for any specific deal.

Stebbings asks whether stronger reported earnings at Snowflake, MongoDB and Salesforce mean the software downturn is over. O’Driscoll says some companies may have recovered because growth reaccelerated or investors saw a connection to AI spending, but treats individual results and valuations as distinct cases.

O’Driscoll discloses that he bought a cloud-software fund after its decline and describes its subsequent rebound. He says a recovery from an oversold level does not establish that the same return will repeat. The performance figures are his own account for the recording period, not current prices.

Lemkin agrees that the panic may have eased while underlying concerns remain. He points to pressure on per-seat software and the possibility that larger AI budgets compete with other spending. The statistics and multiples he cites are his or reported estimates, and a budget trade-off is not a universal rule.

The guests distinguish software products that might benefit as AI agents generate more activity from products sold chiefly through human seats. Lemkin cites Twilio, Okta and Datadog as examples he believes have benefited, though his reported growth and stock returns are not independently verified in the episode.

Lemkin argues that some buyers may reduce unused seats to fund token consumption. O’Driscoll adds that companies need careful product decisions to connect existing products to new demand. Both speak as investors offering interpretations rather than establishing the future trajectory of every software company.

Stebbings cites a reported $1 billion Cognition financing at a $26 billion valuation and about $492 million of annual recurring revenue for Devin. These figures come from the discussion and are not supported by transaction documents or audited results in the captions.

Lemkin recalls that two strong technical leaders in his portfolio experimented with Devin early. He finds the ambition of an autonomous software engineer compelling: a developer could assign a task and receive a proposed code change. He acknowledges that earlier versions were imperfect and does not provide a controlled evaluation of present capability.

Lemkin says he hopes autonomous agents can handle work that human engineers do not want to do, while retaining strong engineers. O’Driscoll stresses that the market is highly competitive and larger companies could challenge specialized providers. Their predictions do not establish that any role can safely or fully be automated.

Stebbings compares uncontrolled token spending to giving every employee an unlimited company card. O’Driscoll hypothesizes that recent model improvements and pricing changes encouraged heavy use before finance teams recognized the cost. His account of widespread corporate spending is an inference, not a survey.

O’Driscoll cites a reported Uber budget of $1,500 per engineer per month as an example of a possible cap. He corrects himself when calling that the right amount, saying instead it is slightly above an average he had seen. The report and comparison are not independently established in the discussion.

He expects older models to become cheaper as newer frontier models arrive, although he says frontier usage itself may not become cheaper. If the code produced is useful, he thinks companies could slow spending growth rather than abandon AI. The price curve and adoption path are his forecasts.

Lemkin says two fast-growing companies in his portfolio had already exhausted their token budgets, while slower-growing companies had not told him the same. He explicitly does not know whether the connection is causal. He expects organizations to become more deliberate about testing, feature development and waste.

Lemkin describes how Replit can combine models for complex work, using one to build and another to review. He sees consumer-facing products under pressure to manage costs because modest price increases can cause customer churn. The product behavior is his account, not an independently tested comparison.

Multi-agent workflows may increase demand even as teams try to control costs, Lemkin says. He believes developers often prefer the strongest available model over cheaper older ones, which can limit the savings from falling prices. He acknowledges that different use cases can require different models.

O’Driscoll interprets the emergence of budget scrutiny as evidence that AI has become a meaningful spending category, while giving a speculative estimate of its potential market size. He asks whether companies will route simpler work to cheaper or open models instead of using the most expensive system for every task.

Lemkin distinguishes models used inside customer-facing applications from those used by developers. He believes cost optimization may be easier for high-volume application requests, while developers under pressure may resist restrictions on the tools they prefer. This is his view of likely behavior, not evidence that restrictions cannot work.

Lemkin predicts that some leaders will choose greater token budgets over additional engineering or product headcount. He describes a hypothetical fixed budget in which managers decide how to divide spending between people and AI. This is his forecast and preferred management choice, not a finding that AI has caused specific layoffs.

He speculates that quality assurance and customer-success roles could be particularly exposed to such decisions. O’Driscoll challenges him to quantify engineering costs separately from application spending, and asks whether a smaller team could still ship complete, reliable products rather than merely produce more code.

The guests work through a hypothetical: replace 100 roles in a 400-person organization and redirect the equivalent budget to tokens. O’Driscoll calculates that this could imply a substantial token allocation per remaining employee and contrasts it with the lower reported per-engineer cap. Neither number measures actual savings or verified productivity.

O’Driscoll says the ratio of model spending to engineering compensation is a key unknown for estimating the market. He is not convinced by the higher scenario and admits he does not know the state of the art. Larger market-size and valuation outcomes he describes depend on assumptions, not observed future results.

Lemkin regards per-person caps as potentially temporary and predicts that departmental leaders will eventually choose between staff and token spending. O’Driscoll agrees that leaders face a difficult trade-off: individual workers benefit from more computing while finance teams must manage the total cost.

O’Driscoll suggests there may be demand for tools that let engineers work effectively while giving finance teams confidence in spending. The right budget could change as products and models improve. He frames this as an opportunity and an unresolved management problem, rather than a fixed formula.

Lemkin describes how he would respond in a hypothetical leadership role if a smaller team could deliver substantially more. Stebbings says the speed with which Lemkin identifies potential cuts is unsettling. O’Driscoll asks whether engineering leaders can actually promise higher end-to-end output after such reductions.

Lemkin says he knows leaders who believe it is possible, particularly at fast-growing startups. O’Driscoll accepts that small startups can do a great deal with few people but questions whether similar results can be reproduced at a large company. They agree that the claimed productivity gain still needs to be demonstrated.

Stebbings relays an executive’s claim that a company spends more on tokens than engineering salaries, but the participants are unsure about its engineering headcount. O’Driscoll says such examples are useful only if the size and output of the team are understood. He cautions against treating a single anecdote as a general rule.

Lemkin adds a countervailing possibility: if teams can launch many more products, they may need more people to manage and sell them even if each product takes fewer engineers. He expects some organizations to grow again for that reason, though the eventual size and number of products are uncertain.

O’Driscoll agrees that faster coding does not automatically translate into company-wide productivity. Products also need packaging, pricing, marketing, sales and training. He cites the idea that the slowest link in an organization can limit the benefits of acceleration elsewhere; his GDP projection remains his own view.

Stebbings raises a report that law firm Kirkland & Ellis plans to spend $500 million over five years building AI tools, potentially using proprietary data and workflows. The captions provide no independent budget documentation. He asks what the reported plan could mean for specialist legal-AI vendors.

Lemkin says the reported annual spend should be judged against the firm’s size and need not rule out buying outside products as well. In his view, organizations should experiment with building where they see value, and vendors must justify their offering. His revenue comparison and assessment are estimates.

O’Driscoll says building sophisticated technology within a partnership can be difficult, even if announcing a plan attracts attention. He distinguishes experimenting with an internal tool from demonstrating that it can replace a specialized product. The outcome, he says, remains to be seen.

O’Driscoll examines a broader build-versus-buy question. A firm may happily buy a common administrative tool but hesitate to share workflows it considers central to its competitive advantage. He says he does not know what particular vendors intend or how any firm’s confidential information is handled.

The guests also consider whether an AI vendor might eventually compete with its customers by supplying a complete service. O’Driscoll treats that as a strategic concern and explicitly says he is not asserting that the named legal-software vendors plan to become law firms.

O’Driscoll predicts that AI could expand access to some legal information and services for individuals and small businesses who cannot afford traditional assistance. He also expects high-stakes transactions to continue requiring trusted human judgment. These are possibilities, not evidence that an AI system provides reliable legal advice in any particular case.

The guests discuss why high-value legal work may still command a premium: clients want accountability and experience when errors are consequential. O’Driscoll recalls an evaluation in which a lawyer rejected a tool despite an estimated 98% accuracy, because the remaining errors were unacceptable for her work. It is an anecdote, not a benchmark.

Lemkin turns to the possibility of AI agents helping people manage their finances. He imagines a service that gathers a person’s circumstances, goals and risk tolerance and helps them understand choices. He describes a desired product, not proof that an agent can deliver suitable advice or protect every user from loss.

O’Driscoll discloses an investment in Range, a company operating in financial planning. He distinguishes assessing a client’s overall allocation and goals from selecting individual stocks to outperform the market. He sees potential for AI in the first task but says the record for consistent trading outperformance is not established.

Lemkin suggests applications should help customers understand their domain and decisions, whether through education or agent action. Stebbings asks how much control a user should delegate. O’Driscoll reiterates that explaining a portfolio is different from promising an agent can reliably beat a market index.

The panel discusses an agent that could explain risks and historical comparisons before a user trades. Lemkin likes the idea of making those trade-offs clearer; O’Driscoll agrees information can be helpful while questioning whether predictions would beat simpler alternatives. Their examples are not personalized financial recommendations.

Stebbings cites a warning attributed to Apollo about private-equity software returns and mentions a reported allocation to private assets at Harvard. O’Driscoll discusses the risk that lower software valuations and leverage can impair equity after debt is paid. The cited warning and allocation are not independently verified here.

O’Driscoll illustrates the arithmetic with a hypothetical company bought at a high revenue multiple using debt, followed by slower growth and a lower market multiple. The example is not an actual Salesforce acquisition. He says owners might need years of operating work or acquisitions to recover value, with no guarantee of a particular return.

The guests disagree about how to respond to difficult funds. Lemkin suggests limited partners with large gains elsewhere may move on; O’Driscoll says managers still have a responsibility to work through challenged assets. He thinks the incentives and personal capital committed by managers can affect that effort.

Stebbings asks how unusually large potential distributions might affect retention at venture firms. The carry figures he mentions are projections, not confirmed payouts. Lemkin and O’Driscoll say some people may continue investing because they enjoy it, while others may choose a different life or retire.

O’Driscoll says he would continue working even after a hypothetical very large payout, while acknowledging others might pursue different goals. Lemkin describes his own threshold for continuing to invest and argues that firms need not last forever. Their comments reflect personal preferences rather than verified details of other firms’ decisions.

The guests consider whether a successful investor who wants more economic exposure could invest personal capital alongside a fund. This is a hypothetical discussion of incentives and risk; the examples cited do not establish the holdings or future returns of any named investor.

Stebbings turns to the “996” work ethic and describes a reported intense culture at Corgi, asking whether leading companies are working differently from earlier cycles. He cites a valuation as context, but the conversation does not establish a causal link between long hours and company success.

Lemkin recalls working weekends early in his career and as a founder. He argues that intense work is not new, and that the important questions are how far expectations extend through an organization and for how long. He says early employees asked to make that trade should have meaningful compensation or equity upside.

O’Driscoll agrees that demanding roles have long existed, but says such intensity is not sustainable across an entire lifetime. He distinguishes a founder’s voluntary commitment from expecting the same schedule of a much larger workforce. These are the guests’ views, not a claim that extreme hours are necessary or appropriate for every worker.

O’Driscoll recalls working intensely on his own unsuccessful startup without making money, and warns that effort does not guarantee a large outcome. Lemkin says employees promised substantial upside should at least have a meaningful chance of sharing in success. Their examples do not establish likely individual compensation.

O’Driscoll warns that sustained stress can impair judgment even as hours increase. For work requiring judgment, stepping back and protecting psychological health can matter more than performing busyness. Lemkin agrees that company building is a marathon, though he continues to predict difficult staffing choices.

The panel notices a tension: people predicting extensive automation also report working exceptionally hard and struggling to hire. O’Driscoll is skeptical of forecasts of near-term mass unemployment; Lemkin is less confident about the future for workers he considers less exceptional. Both are making contested predictions, not presenting verified employment outcomes.

O’Driscoll closes with a personal anecdote and argues that people displaced from highly paid technology roles may find other work, even if compensation or conditions differ. Lemkin remains uncertain. The discussion ends without resolving how AI will affect overall employment, wages or job quality.