Anthropic, Block and Cursor: A 20VC Roundtable on AI and Software

Harry Stebbings asks Rory O’Driscoll and Jason Lemkin about Anthropic’s dispute with the US Department of Defense. O’Driscoll describes his understanding of a roughly $200 million contract and proposed limits involving mass surveillance and autonomous weapons. He explicitly warns that participants in the negotiations may present self-serving accounts; the panel has not reviewed the contract.

O’Driscoll says the government sought permission for lawful uses and describes a possible supply-chain designation as a source of uncertainty. He says he does not know exactly how a threat would be implemented or what it would mean for other suppliers. His description is a contemporary account, not a legal determination or verified final outcome.

Asked whether Anthropic’s leadership was right to stand by its conditions, Lemkin speculates about the company’s safety commitments and the preferences of its staff. He thinks leadership could no longer assure employees that military use would meet their standards. The interview does not establish the private discussions, motivations or effect of any model on civilian harm.

O’Driscoll agrees that safety has been central to Anthropic’s culture but argues it was naive to enter a defense relationship if the company expected to prescribe use. He believes elected authorities should make defense decisions. That is his normative position; it does not resolve the separate legal and ethical questions about restrictions or oversight.

He compares the dispute with historical tensions between scientists and government decision-makers. O’Driscoll argues that a supplier can choose not to contract, while elected officials carry responsibility for defense. Lemkin’s emphasis on researchers’ concerns and O’Driscoll’s emphasis on democratic accountability frame a disagreement rather than a finding about either party’s legal rights.

O’Driscoll says Anthropic may have gained consumer goodwill from the dispute but also faces a more powerful counterparty. He regards employee alignment as valuable and considers further action by government a business risk. The panel does not establish the eventual legal or commercial effect of those events.

O’Driscoll says the administration should be free to decline a contract without trying to damage the wider company. He worries that a broader government response could increase risk for Anthropic and its investors. His comments describe potential actions and legal uncertainty, not a confirmed outcome or assessment of the government’s motives.

Stebbings asks whether an Anthropic shareholder should be more confident after a reported rise in consumer attention. O’Driscoll says he sees greater uncertainty despite potential brand benefits. Lemkin argues investors may focus primarily on growth and cites an approximate revenue run rate; neither verifies financial data or predicts a shareholder’s actual return.

Lemkin contrasts the bargaining power of sought-after AI researchers with that of employees elsewhere in technology. He believes retention is a central concern for model companies and attributes leadership decisions partly to it. His anecdotes about employment practices and compensation are not independently corroborated, and his comments do not establish what Anthropic employees collectively demanded.

O’Driscoll adds the state as a third actor alongside capital and labor. He argues that government has substantial legal power and that a company should consider that risk before entering a public conflict. His speculation about the use of particular authorities and court outcomes is not legal advice or a determination that such powers were lawfully invoked.

The conversation turns to OpenAI’s reported defense agreement. Lemkin asserts that employees objected and discusses subsequent public changes, while O’Driscoll also sees employee influence. Their accounts of internal reactions and contract language are interpretations without access to negotiations or an independently verified employee survey.

O’Driscoll thinks government, rather than a private company, should ultimately decide lawful military uses, but believes signing the agreement may create internal difficulties for OpenAI. Lemkin sees an opportunity seized during a competitor’s setback; O’Driscoll thinks the decision may also reflect sincerely held principles. Both are speculating about motivations.

Lemkin says rapidly changing model capabilities make it difficult to anticipate what military applications might do in a year. O’Driscoll maintains that constitutional decision-makers should consult experts while retaining accountability. Their different emphasis illustrates the governance challenge, not an agreed solution.

Stebbings and Lemkin question whether political leaders understand AI well enough to make technical decisions. O’Driscoll replies that elected officials should listen to experts and make accountable judgments, even when experts know more about a specific technology. The disagreement concerns governance and expertise, not a conclusion that either institution is infallible.

Lemkin suggests the reported defense and partner contracts may have mattered more when Anthropic’s revenue run rate was smaller. He uses approximate numbers to illustrate the trade-off; O’Driscoll cautions that the parties may have drifted into a disagreement neither anticipated. Their revenue figures and guesses about the negotiation are unverified.

The panel discusses OpenAI’s reported $110 billion financing. Stebbings praises the fundraising scale, while the guests ask how a round of that size compares with private and public capital markets. This is a discussion of reported commitments, not independently verified cash received or final transaction terms.

O’Driscoll says he understands that part of a reported Amazon commitment is conditional on a public listing or an AGI-related event. Lemkin distinguishes funds available now from money subject to future conditions. Their approximate $15 billion upfront and $35 billion conditional breakdown is their account; the interview does not inspect signed agreements or define the conditions.

The guests contrast a commitment subject to conditions with money immediately available to fund operations. They also speculate about whether such commitments might support a later public offering. That proposed financing effect is not confirmed, and the transcript does not supply complete terms or a timetable.

The guests speculate that OpenAI, Anthropic and SpaceX might next raise money through public offerings because of the scale of their financing needs. They acknowledge that another private round is possible. Neither a listing date nor an IPO is established by their discussion.

Stebbings asks whether leaders create a valuation premium. O’Driscoll compares Sam Altman’s fundraising role with Elon Musk’s role in engineering-led companies, then offers hypothetical values if either departed. These are speculative scenarios, not valuations derived from a disclosed model or predictions of an actual event.

The panel considers whether a successor could run OpenAI if its founder left, and whether comparable assumptions would apply to Tesla. Their hypothetical prices and leadership scenarios are subjective, and no departure or acquisition is announced. The exercise reflects their views about founder dependence, not a measured premium.

O’Driscoll guesses an October OpenAI listing at $1.5 trillion, while Lemkin considers the possible effect of conditional funding on a future offering. O’Driscoll cautions that such prices depend heavily on strong markets and future growth. Their figures are predictions and illustrations, not an announced listing or established enterprise value.

O’Driscoll compares the assumptions investors might make about OpenAI’s growth and SpaceX’s future technologies. He cites approximate revenue and growth figures and emphasizes the distance between those fundamentals and very high valuations. The panel does not validate either company’s financials or future technical milestones.

Turning to public software, Lemkin predicts that many companies will report weaker growth through the year. He argues that the problem is not only AI-generated code but also a failure to find new growth. These are broad forecasts from an investor, not reported results for the sector.

Lemkin says backward-looking retention and multiyear contracts may hide deterioration, and predicts some companies will cut staff to protect profitability. His claim that companies will need to reduce teams substantially is speculative. O’Driscoll replies that the price paid for a stock matters alongside its future growth.

O’Driscoll contrasts a highly valued company vulnerable to a small guidance disappointment with lower-valued businesses whose risks might already be reflected in price. He says companies can pursue growth, profitability or a combination. His multiples and possible upside are illustrative opinions, not an analysis of a particular security.

Lemkin argues that incumbent software companies should convert AI demand into products customers will pay for. He says he has used Salesforce’s Agentforce and sees improvement, while acknowledging flaws and not predicting a particular growth rate. His product experience is anecdotal and does not establish performance for all customers.

Lemkin argues incumbents must invest in customer-facing AI products rather than rely only on internal automation. O’Driscoll says that shifting a company toward a new product can require difficult organizational and customer decisions. Their debate distinguishes technical possibility from execution and does not prove any incumbent’s strategy has failed.

O’Driscoll points to the organizational difficulty of changing an established product while serving existing customers, citing Intercom as an example. Lemkin agrees that timing and category matter but believes many firms are moving too slowly. Both are assessing strategy from outside; neither supplies controlled evidence that a particular executive will succeed or fail.

Stebbings asks about Shopify and Spotify. Lemkin says he sees Shopify pursuing agentic commerce and also credits Salesforce with taking action, though he does not know how successful the efforts will be. O’Driscoll resists the idea that all public-company leaders are ignoring AI, and thinks outcomes will vary by industry.

The guests debate whether earlier software valuations adequately reflected slowing growth. They offer rough comparisons of revenue multiples and growth rates across periods and argue that market expectations have changed. These historical averages are their approximations, not a dataset or proof that any decline is permanent.

Stebbings raises Block’s reported 40% workforce reduction and asks whether it reflects AI efficiency or earlier overhiring. Lemkin attributes much of the decision to slower growth and a desire for profitability. That is his interpretation of management motives; the interview does not establish a causal breakdown.

Lemkin cites roughly 3% top-line growth and contrasts it with a higher gross-profit growth figure. Those metrics have different denominators and business definitions, and the panel does not inspect Block’s financial statements. His conclusion that management has abandoned growth is a judgment, not a confirmed company policy.

O’Driscoll agrees that staff levels and profitability matter but cautions that a financial-services company should not be compared mechanically with a software company because pass-through revenue and costs differ. He separates AI as a possible source of new revenue from AI as a way to reduce operating expenses.

He doubts that AI alone explains a 40% reduction at Block and says the company may be relying on some combination of changed staffing and tools to serve customers. Lemkin says the announcement could encourage other executives to make cuts. Neither has evidence sufficient to quantify the share caused by AI or to predict whether the plan will work.

Lemkin predicts that other software companies with slowing growth may follow Block by focusing more heavily on profitability. O’Driscoll asks him to separate possible AI-driven savings from costs that could already have been reduced through ordinary restructuring. Both recognize that the number and timing of future cuts remain uncertain.

O’Driscoll asks what rate of staff reduction AI might justify at an already efficient company. Lemkin first argues that even financial products can face competition from agent-based services, then says executives he speaks with believe they can operate with fewer people. His generalized 40% claim is anecdotal, not a representative survey or a necessary staffing rule.

Stebbings reports speaking with three executives about possible reductions. The panel discusses whether a prominent layoff changes what other leaders regard as acceptable. The sample is small and the future decisions are unknown; reported conversations cannot establish a sector-wide wave or how many jobs AI will replace.

Lemkin offers an example of an AI company expanding its sales team rapidly and questions how many hires will prove exceptional. He acknowledges the human cost of layoffs, while predicting future businesses may start with fewer employees. His comments on individual productivity and future staffing are opinions rather than measured findings.

O’Driscoll says the long-term trend toward valuable companies with fewer employees supports part of Lemkin’s argument, but he does not expect every business to become a one-person company. He cites approximate revenue-per-employee comparisons. Those comparisons depend on company type, period and revenue definition and do not isolate AI’s effect.

The panel distinguishes layoffs becoming more socially acceptable from executives believing cuts are necessary for company economics. O’Driscoll argues slow growth and low cash flow can invite restructuring or acquisition; Lemkin predicts boards will discuss Block’s example. These are hypothetical business scenarios, not evidence of a particular board decision.

Lemkin speculates that future software companies may rely less on labor-intensive sales and marketing, with higher quotas for some teams. Stebbings and O’Driscoll do not establish that such changes are inevitable or beneficial. The forecast should be read alongside the human costs of changes in employment.

Stebbings says people in his network have moved from Cursor to Claude Code, yet cites a report that Cursor grew from $1 billion to $2 billion in annualized revenue in about three months. He also mentions an unconfirmed secondary valuation. The panel has not reviewed the company’s accounts, and an ARR run rate is not completed-year revenue.

Lemkin says similar anecdotes from his portfolio led him to underestimate Cursor, illustrating how a venture portfolio is not a representative sample. He suggests enterprise customers may explain some growth and cites a reported enterprise share, but explicitly says he does not know the full picture. The attribution is a hypothesis, not verified customer data.

Stebbings relays a view that annual subscriptions may delay visible churn. Lemkin says deferred churn is a question for later, while O’Driscoll explains that enterprise procurement, security reviews and contract cycles can slow switching. They agree that strong reported growth does not by itself reveal future retention.

O’Driscoll argues that an enterprise buying cycle can involve security review, procurement and a contract term rather than a daily choice between tools. That may make adoption and switching look different from a startup developer’s experience. The discussion provides a possible explanation, not verified evidence of Cursor’s customer-level retention.

O’Driscoll believes a large expanding market can support multiple companies even when competition intensifies. He notes that he has not seen Cursor’s actual revenue or gross margins. His saturation threshold and expectation of eventual competition are heuristics, not measurements of the coding market.

Lemkin points to reported enterprise deployments and capabilities such as single sign-on and access controls as possible reasons customers choose a tool. He suggests managing autonomous agents safely could become valuable to security leaders. His estimates of buyer preferences and claims about agent risks are opinions; the panel does not test Cursor’s security or prove that every agent leaks data.

O’Driscoll says enterprise relationships and integration can matter alongside technical capability, using cloud providers as an analogy. He expects competition between coding tools to increase as initial adoption spreads. The comparison is a market thesis and does not establish which product will win or what share each holds.

The speakers discuss reports that many Cursor model calls use Anthropic’s API and that revenue may therefore flow through multiple companies in the supply chain. The cited percentage is unverified, and adding the two companies’ revenues would not necessarily represent separate end-customer spending. Neither firm’s gross margin is disclosed.

The panel notes that model providers and application developers can both report growth when an application resells underlying model access. O’Driscoll regards the market as large enough for several participants, while Lemkin expects each to keep adding capabilities. Neither speaker has established audited revenue composition or future competitive shares.

O’Driscoll argues that fast-moving model capabilities may require coding products to change repeatedly. Lemkin wonders how API costs affect Cursor’s margins and compares his own subscription and API experiences. He speculates about discounts and subsidies but acknowledges that he does not know the commercial terms.

Lemkin says his own API usage appears more expensive than a subscription for some tasks, and asks whether an application supplier can sustain margins when it pays for model calls. His inference about negotiated discounts is explicitly speculative; the interview provides no actual Cursor contract or cost schedule.

The panel discusses differences between frontier-model prices and older open models, using approximate cost comparisons. O’Driscoll sees a rapid pace of change; Lemkin predicts it will accelerate. Their prices and forecasts are illustrative and do not establish equivalent model quality, total operating cost or future pricing.

Lemkin recounts using coding tools over several months and says their ability to complete applications improved. He describes a colleague building an internal AI marketing system and is enthusiastic about its output. This is his team’s experience and a company claim, not an independent benchmark or evidence that the system is appropriate for every organization.

He predicts that coordinated agents will make increasingly sophisticated applications easier to build. He also wonders how an investor can distinguish founders at a demo day when many teams can produce a polished prototype quickly. His prediction is uncertain and does not establish that generated software is secure, reliable or commercially viable.

O’Driscoll narrows the claim that software is “uninvestable”: a demo alone may contain less information if a prototype can be produced quickly. He says companies may need additional advantages in distribution, networks, vertical knowledge or execution. The panel treats this as an investing framework, not a categorical conclusion about software investments.

The speakers agree that easier prototyping changes the information an early demonstration conveys. O’Driscoll distinguishes that point from the claim that software has no investment value: a business can still build durable advantages beyond code. Their assessment is a hypothesis about investment selection, not evidence that every demo is equally easy to reproduce.

Lemkin says his founder background once helped him identify unusually capable teams by comparing what they shipped with his own experience. He now finds that signal harder to interpret as tools improve. Stebbings praises his agent-building work, while Lemkin acknowledges that specialists are much further ahead than he is.

Lemkin describes portfolio-company reports of productivity gains and predicts that product-development pace will accelerate substantially. He says two investments generated substantial first-week revenue, but provides no verified financial statements or evidence that those sales will recur. His projections and examples reflect an investor’s own portfolio.

The group concludes that technical ability and deep understanding of a market may remain useful signals even when demos are easier to make. Stebbings mentions a missed investment opportunity, and O’Driscoll promises to return with stock ideas in a later episode. His reported index performance is a personal, time-specific estimate, not a recommendation or verified return.