AI Compute, Agent Security and the Race to Build More Software

Harry Stebbings opens the roundtable with Jason Lemkin and Rory O’Driscoll by citing reported Nvidia quarterly revenue and discussing possible acquisition talks involving Hugging Face. The speakers separate the chip business from the possible deal; the transaction’s status and figures are not independently verified in the conversation.

O’Driscoll argues that Nvidia’s near-term results reflect unusually strong demand for compute relative to available supply. He points to reported guidance as evidence that customers expect demand to remain high. This is his interpretation of the figures, not a guarantee of future revenue.

The panel discusses financing arrangements among infrastructure and model providers. O’Driscoll says the chain ultimately relies on end customers paying for useful services: chip buyers need to sell compute, and model companies need demand for their products. Strong spending at one layer does not itself prove long-term returns for every participant.

O’Driscoll says the infrastructure chain is linked: continued chip demand depends in part on model providers and other customers finding durable demand for their output. He treats that dependence as important even when present orders are strong.

O’Driscoll describes a hypothetical risk in which end-user demand grows more slowly than infrastructure investors forecast. Even continued growth might be insufficient for all projects if expectations and capacity expanded faster. He presents this as a scenario, not a prediction that the market will reverse.

Stebbings asks about competitive threats. O’Driscoll agrees that another chip supplier could take share even if overall compute demand remains strong. Lemkin sees the reported demand as an encouraging signal across AI markets but acknowledges that individual companies’ competitive positions can change.

Lemkin interprets strong chip demand as a broad signal for investors in AI, but O’Driscoll distinguishes near-term momentum from conditions that could change later. Neither establishes that every company in the sector will benefit.

The group returns to capital expenditure. O’Driscoll says models in which customers eventually become more cash-generative depend on demand catching up with spending. If infrastructure commitments rise further, that point may move out and the amount of end-user value required may grow.

Stebbings asks about reported Nvidia discussions involving Hugging Face and explicitly notes that the deal is not finalized or confirmed in his account. O’Driscoll speculates that a strong open-model ecosystem could support demand for computing hardware. The conversation does not establish a transaction or its terms.

Lemkin thinks a hardware provider may seek to participate across several model and inference segments. He offers his own view about strategic value and compares a possible deal with older software acquisitions. The valuation and revenue figures he cites are unverified and do not establish what either company will do.

The panel observes that a strategic buyer might value a technology ecosystem differently from a financial investor looking only at its present revenue. That general distinction does not establish a final price or a completed acquisition.

The panel turns to a reported change in OpenAI’s access for Cursor. Lemkin says some users may still have alternative ways to use a model, while distinguishing practical product access from the public dispute. The speakers do not independently establish the precise terms or availability for every customer.

O’Driscoll argues that a model provider and coding application can be both partners and competitors, creating incentives to reassess an arrangement as markets evolve. He also mentions a public dispute involving executives. His account of strategic motives and legal claims is speculation rather than an adjudication.

O’Driscoll says a provider might have concerns about use of its technology by a direct competitor. The panel does not verify any alleged use, contractual breach or company motive, and the speakers frame the issue as a general strategic question.

Lemkin says he can understand a provider reconsidering access when a former partner becomes a more direct competitor. O’Driscoll adds that business relationships can become difficult when litigation is involved. Neither speaker claims knowledge of the companies’ confidential decisions or contractual rights.

Stebbings raises a reported security incident involving OpenAI agents and Hugging Face. Lemkin says he has seen problematic agent behavior in his own work and cautions against assuming that automated systems act or collaborate like people. The account of the incident is not independently verified in the interview.

Lemkin argues that anthropomorphic language can lead observers to misunderstand automated behavior. He describes agents as systems given objectives and tools, and says security analysis should examine the actual permissions and outcomes. His generalizations do not establish the internals of every model or incident.

Lemkin says automated systems can pursue a task in ways that surprise their operators. His main recommendation in the conversation is to examine actual system design and permissions rather than relying on vivid analogies about intention.

O’Driscoll agrees that dramatic metaphors can obscure what a system actually did, while emphasizing the need for strong cybersecurity defenses and oversight. He says persistence and the ability to handle complex tasks may change the threat environment; this is his risk assessment, not a quantified forecast.

O’Driscoll says organizations should consider how persistent automated tools affect security preparedness. His description of the pace of change is a warning based on his interpretation, not a verified estimate of future attacks.

The speakers argue that organizations should take agent security seriously before deployment. Their discussion includes claims about an incident and its detection that are not verified on the show. They do not present a complete security assessment or a guaranteed way to prevent future problems.

Stebbings says a personal assistant helped him make a dinner reservation but asked for access to payment and email services. He hesitated and asks whether users understand the risks of granting broad permissions. His anecdote does not establish the security practices of the named product.

Lemkin recalls an earlier assistant-related story in which a system produced an incorrect account of a purchase. He says the example was not a completed purchase but illustrates his concern that an agent with broad access might misunderstand a user’s intent. The discussion remains hypothetical.

The group distinguishes a system misunderstanding a user’s request from a separate risk that sensitive information could be exposed or misused. The speakers do not test a named product or establish that either problem occurred there.

O’Driscoll describes the appeal of an assistant that can access calendars, email or payment methods and handle routine tasks. He says some users may consider the convenience worth exploring, while the degree of access and the associated risks require careful decisions.

Lemkin questions whether instructions alone can reliably prevent an assistant from taking an unwanted action in every context. He describes the difficulty of anticipating all possible tasks and outcomes. This is his present assessment, not proof that safeguards are impossible.

Stebbings asks whether the same convenience that attracts users could lead them to grant more access than they intended. The panel debates different ways to limit risk but does not give a definitive security verdict on a personal assistant.

O’Driscoll separates two questions: whether a system can be constrained to the permissions it is given, and whether it can infer a user’s wishes accurately enough to be useful. Lemkin thinks the second may be more tractable, while both regard the first as an engineering and governance challenge.

Lemkin favors hard controls on sensitive resources rather than relying entirely on natural-language instructions. He also says multiple preferences can conflict, making it harder to choose among them. The panel does not test specific controls or claim that any one measure resolves all security issues.

Asked whether the category will last, O’Driscoll expects assistants to manage parts of users’ schedules and transactions. He gives a personal example of a familiar destination being suggested while driving. Whether such convenience becomes a profitable standalone business remains uncertain in his view.

O’Driscoll thinks assistants could reduce small recurring decisions as they learn preferences. He is less certain that consumers will pay enough to support a large independent business, and says early enthusiasm does not settle the business model.

Stebbings says his assistant has reduced some routine booking work, but asks whether many similar products will make the technology a commodity. Lemkin expects existing applications to add more autonomous capabilities as well, increasing competition across product categories.

Lemkin argues that an early version of a product can be easy to reproduce while a mature service accumulates features, reliability work and customer experience. O’Driscoll agrees that some competitors may pull ahead over time, though he cannot know which teams will do so.

O’Driscoll suggests that software categories can begin with many similar products and later develop leaders as functionality and customer relationships accumulate. This is an observed pattern in his experience, not a prediction of specific winners.

The panel cites another reported assistant financing and sees venture investors seeking exposure to a fast-moving category. O’Driscoll says investors may be paying for momentum and possible outcomes as well as current fundamentals. Neither valuation nor future returns are established by the discussion.

Lemkin says investors have long wanted exposure to emerging markets even when the eventual winner is unclear. Faster development compresses the time available to decide, but it does not remove the uncertainty of choosing among teams and business models.

Stebbings raises a reported Cognition financing and projected annual recurring revenue, asking whether investors underestimated the coding market. The figures are presented as reports and forecasts, not independently verified results. Lemkin says a large market may support valuable companies beyond the leading one or two.

Lemkin argues that a company can succeed in a large category without overtaking the biggest provider. His example depends on assumptions about future revenue and competition that remain uncertain.

O’Driscoll estimates a large pool of software labor spending and asks what fraction might shift toward AI tools. He offers different percentage scenarios to illustrate a range of possible markets. The arithmetic does not show what customers will actually spend or how much labor will be replaced.

Lemkin says his portfolio companies are producing substantially more software and expanding their road maps. His large multiples are personal impressions rather than a measured industry average. He thinks software output and usage could grow beyond estimates based only on today’s number of developers.

Stebbings asks whether every startup must become a broader, multi-product company. Lemkin believes competition is pushing teams into adjacent functions faster. O’Driscoll agrees that cheaper software production can intensify competition, but says customer budgets are unlikely to rise in exact proportion to output.

Lemkin believes competitors are expanding their product scope, while O’Driscoll asks whether customers will pay for every added feature. They agree that speed alone is insufficient without a useful product and a sustainable economic model.

O’Driscoll sketches a scenario in which an integrated suite offers more functions for less than the combined cost of separate products. He thinks that could pressure narrow vendors. He presents the scenario as an economic argument, not proof that bundled software will always be better or cheaper.

Lemkin describes a portfolio company serving restaurants that is adding a wider range of features. He says the team regards a faster shipping pace as necessary. O’Driscoll notes the countervailing risk of a confusing product with too many functions and says design remains important.

Stebbings worries that some European startups have raised less capital than US rivals and may struggle to match the pace of product expansion. Lemkin cites headcount observations from an investor report, while O’Driscoll questions how productivity gains and increased hiring can happen at the same time.

O’Driscoll distinguishes a small company using AI to produce the same product more efficiently from a fast-growing company that attracts capital and expands sales, marketing and engineering. In the latter case, improved efficiency can coexist with greater hiring because the company chooses to pursue more opportunities.

Stebbings asks whether a smaller fundraising round leaves a startup disadvantaged when a rival can invest aggressively. O’Driscoll says the answer depends partly on whether companies compete directly or serve distinct markets.

The panel debates whether venture investors can create winners simply by providing capital. Stebbings argues that well-known backers can attract talent, customers and later funding. Lemkin says capital matters but cannot substitute for founders, execution and traction; both views describe tendencies, not a rule for every company.

The speakers discuss a reported expansion of a venture growth fund. Lemkin interprets it as a sign that the firm sees opportunities to deploy more money, while acknowledging he may be reading it incorrectly. The conversation does not establish private investment decisions or future fund performance.

The discussion turns to Salesforce and Anthropic. Lemkin characterizes some of the reported joint offering as packaged tools and marketing, but thinks a broader willingness to support multiple user interfaces is strategically important. His product and revenue descriptions are his views, not an independent assessment of the partnership.

Lemkin distinguishes a packaged integration from a larger change in how customers access enterprise records. He believes the latter could matter because it allows users to work through tools other than the vendor’s own interface.

Lemkin says customers may want to work through an assistant, a messaging product or an enterprise interface rather than a single application screen. He also highlights reported interest in outcome-based pricing. He regards both as organizational changes for an incumbent and does not claim they are guaranteed to benefit it.

O’Driscoll praises what he sees as Salesforce’s willingness to adapt its approach. Asked about the stock, he discloses that he holds it and offers a rough scenario based on growth and cash flow. These are personal opinions and assumptions, not investment advice or reliable price predictions.

O’Driscoll says a customer may value a system of record while preferring to interact with its data through an assistant. He views openness to that choice as a competitive consideration, not proof that one pricing strategy will prevail.

The panel discusses a system of record accessed through other applications. O’Driscoll says his firm uses Salesforce and that some colleagues interact with it through an assistant. He expects enterprise customers to care about whether their chosen tools can continue to connect to important data.

Lemkin asks whether a system of record can also deliver a business outcome rather than merely store data. O’Driscoll points to Salesforce’s reported purchase of Intercom as a possible example of pursuing outcome-based services and discloses that his firm invested in Intercom. The deal details are not independently verified in the episode.

O’Driscoll compares a reported model-spending figure with a reported engineering budget to ask how large AI spending might become. He emphasizes that a small current ratio could mean either a smaller eventual market or room for greater adoption. The figures and scenarios are not a forecast of any company’s budget.

Stebbings raises reports that a possible Stripe–PayPal deal will not proceed. The panel speculates about valuation and negotiation dynamics but repeatedly notes that outsiders do not know the private discussions. It does not establish an agreed price, the reason for any decision or whether future talks will occur.

Lemkin says he has become more favorable toward Clay after seeing agents in his own workflow recommend it. He believes useful APIs and reliable outputs may influence automated tool selection. Stebbings discloses an investment interest, and Lemkin’s experience is not proof of how all agents or customers choose software.

In a mock investment committee, Lemkin argues that automated go-to-market work could increase usage of tools and data. He proposes a hypothetical investment at a reported valuation and outlines a very large possible outcome. The argument is a venture scenario, not verified revenue or financial advice.

Lemkin’s hypothetical case for Clay assumes that automated campaigns and analysis could create more usage than human-led workflows. He acknowledges that more usage does not necessarily translate into proportional customer revenue or sales outcomes.

O’Driscoll says Clay and Linear illustrate how companies founded before the current AI wave may adapt their products. Lemkin argues that being easy for third-party agents to use is more specific than simply adding AI features. The panel treats the strength of any advantage as a matter for future performance to test.

Stebbings discloses an early investment in Linear and cites reported growth and valuation figures. Lemkin says he does not know how much of its revenue comes from agent-related usage and offers only a guess. The discussion does not establish an audited breakdown or the company’s correct valuation.

In a second hypothetical pitch, Lemkin argues that teams producing more software with agents still need to manage tasks and coordinate work. He sees a possible role for Linear as a shared record for people and automated systems. His proposed investment size and very large usage multiples are illustrative claims.

Lemkin says software teams using several agents can still need a way to organize priorities, issues and human review. His investment pitch is an argument for a possible product role rather than an objective comparison of every project-management tool.

O’Driscoll clarifies what he means by agent-friendly: a third-party agent may choose a tool because it can discover and use its interface while working on a task. Lemkin says useful documentation and working APIs can matter alongside brand. The speakers do not demonstrate that agent recommendations are free from bias or error.

Stebbings says anticipated growth in automated queries influenced one of his investment decisions. O’Driscoll distinguishes growth in the volume of work from growth in spending: a system might handle many more requests while unit prices fall. Their growth multiples are scenarios rather than verified forecasts.

The final discussion concerns a surveillance product and public concerns about its use. O’Driscoll describes possible identification errors and unauthorized access to sensitive information, and argues that safeguards and accountability matter. The episode does not independently verify individual allegations or adjudicate conduct by any named organization.

O’Driscoll argues that sensitive databases require access controls and accountability because errors or improper use can affect individuals. The conversation raises general governance concerns without confirming the facts of particular cases.

He says public acceptance can be undermined when a system intended for one purpose is perceived as enabling misuse by other people. The panel compares that governance issue with wider questions about how software vendors can influence downstream use. His assessment of the particular product and policy response is his own.

Stebbings closes the mock investment committee, and O’Driscoll notes that declining to invest can also be a decision. The discussion’s valuations, legal claims, security anecdotes and future scenarios are the speakers’ accounts and require independent assessment.