AI agents challenge software buying as investors chase larger bets

Three investors debate concentrated venture funding, shrinking software customer growth and the economics of agents.

JLJason Lemkinゲスト · SaaStrRORory O’Driscollゲスト · Scale Venture Partners
チャンネル
20VC with Harry Stebbings
公開日
2026年11月19日
動画
1:15:12
読了時間
13 分
原文
英語
YouTube で開く
スクロールすると動画は右下に小さくなって再生を続けます。字幕は閲覧言語に合わせて切り替わります。
閲覧言語2 言語で読めます
この記事はまだ日本語に翻訳されていないため、英語の原文を表示しています。
要約

30 秒でわかる

Businesses may delay buying conventional software while paying for agents that handle urgent work, argues investor Jason Lemkin. He points to slowing new-customer growth at HubSpot and Monday and describes his own team’s use of an assistant for deal screening, follow-ups and content. Rory O’Driscoll cautions that mature markets were already saturating and questions whether low-price agents can sustain their computing costs. The investors also debate the proposed Anthropic IPO, Nvidia’s leadership uncertainty and the risks behind rapid venture funding.

  1. 1Nvidia uncertainty could change competition among AI modelsO’Driscoll says an open-weight alternative needs substantial financing and thinks a weaker Nvidia push could benefit closed-model providers.
  2. 2A small workforce at a large AI company sparks a jobs debateLemkin compares Anthropic’s projected employee count at a large revenue run rate with earlier technology companies. O’Driscoll cautions that compute-heavy AI labs may not represent every employer.
  3. 3Fast deployment can compress returnsO’Driscoll says investing quickly after large fundraising cycles can produce weaker returns; Lemkin adds that the pressure to deploy may itself drive valuation increases.
  4. 4A founder’s return may not restore an incumbent’s old growth rateLemkin points to a decline in net new HubSpot customers, while O’Driscoll says even a successful turnaround must overcome a large, slower-growing base.
  5. 5Small businesses may defer software instead of replacing itLemkin thinks customers can use spreadsheets and general AI tools longer while buying agents that solve an immediate problem, such as answering calls.
  6. 6Connected agent workflows can make switching costlyLemkin describes an assistant screening deals and editing content, but worries that independent vendors may struggle to fund continuous use.

ポイント

  1. Lemkin thinks small businesses may postpone new software purchases rather than build complete replacements themselves.
  2. Test an agent on work that is not getting done, such as customer follow-ups, and evaluate the result.
  3. O’Driscoll says a mature software company can improve its growth without returning to startup-like rates.
  4. Rapid fund deployment and rising entry prices can reduce future investment returns, O’Driscoll argues.
  5. Do not infer that a particular outage was an AI attack without evidence; the panel distinguishes causes from future security risks.
全文

チャプターごとに整えた全文を読む

文字起こしのみ
01

Nvidia’s board faces a difficult founder decision

The investors disagree about whether Jensen Huang will return and how the company’s board might weigh unresolved issues. They consider the effect on Nvidia’s deals and competitors.

Harry Stebbings introduces a week of news, including debate over Jensen Huang’s possible return to Nvidia, an anticipated Anthropic listing, Microsoft service outages and a SpaceX Starship catch. Investors Jason Lemkin and Rory O’Driscoll join him to consider the implications for companies and their backers.

Stebbings asks whether Nvidia’s board was right to let its founder leave after a sharp stock decline and public appeals from investors and employees. The panel separates the question of whether he should return from a prediction about whether a return is likely.

O’Driscoll says that as a board member he would have sought time to investigate rather than make an irrevocable decision in haste. A leave of absence and an interim chief executive, he suggests, could have allowed the board to establish the facts.

He says there are different strands of allegations and questions about chip export controls, and outsiders do not know what the board knew or when. That uncertainty makes it difficult to judge the process, even if he would have preferred more time.

O’Driscoll notes that several directors had backed Huang for decades through Nvidia’s rise. He thinks a board that chose to part with such a long-standing leader may find it difficult to reverse itself within a week, although shareholders could in theory replace directors.

He also sees unresolved questions around exports as a potential obstacle to any return. He stresses that this is a guess rather than knowledge of the board’s deliberations and allows that events could prove him wrong within days.

Lemkin considers a return more likely. He argues that if Nvidia’s arrangements with overseas data centers become the central issue, responsibility would extend beyond Huang to other senior executives, including those who signed financial statements.

The panel notes that the board has formed an export-related risk committee. O’Driscoll says that if he owned the company outright and there were no threat of criminal proceedings, Huang would be the person he would most want to run it; the question is whether a public company can resolve the surrounding issues.

The investors disagree on whether Huang’s return would restore the stock immediately. One concern is that lost China-related revenue and a change in appetite for the aggressive financing deals Nvidia had pursued could affect the business even if the founder returned.

O’Driscoll argues that a weaker Nvidia open-model push could benefit Anthropic. In his account, building a competitive open-weight US model requires substantial resources; without Nvidia’s support, closed-model providers might face less competition.

Lemkin describes the offshore data-center arrangement as having been legal under the prior rules, while saying model makers might regard it as a way for competitors to access and copy their work. The group treats these assessments as uncertain and turns to Anthropic’s prospective IPO.

02

A prospective IPO changes venture expectations

Anthropic’s projected headcount provokes a debate about employment and compute. The group also considers exceptional returns, larger funds and the danger of investing too quickly.

Lemkin focuses less on the proposed valuation than on the number of employees needed to reach a large revenue run rate. He contrasts a projection of fewer than 6,000 Anthropic employees at $100 billion in revenue with much larger headcounts at Google and Meta at a comparable scale.

He says compute expense is part of the explanation but believes the contrast has serious implications for jobs. O’Driscoll agrees the model uses more computing power per employee, yet cautions that other businesses may not be able to substitute compute for labor to the same degree.

The investors disagree over how broadly that example will carry across the economy. Lemkin points to changes in front- and back-office work; O’Driscoll expects compute-intensive companies to be at the leading edge rather than a universal template.

Stebbings asks what an Anthropic IPO could mean for venture funds. He anticipates that investors receiving distributions could redeploy capital and that early employees might become angel investors themselves.

O’Driscoll says reaching a trillion-dollar valuation in roughly five years would expand investors’ sense of what is possible. Early backers could achieve exceptional outcomes, but he describes venture returns as concentrated in a small number of unusually successful investments.

Lemkin cites Carta data for the third quarter: the number of deals fell while dollars invested rose sharply. He sees that as a shift toward concentrating much more money in fewer companies and repeatedly financing perceived winners.

He expects fund investors to ask how many portfolio companies can reach valuations above $20 billion, rather than simply count $1 billion unicorns. Getting into such companies before their valuations rise remains important to the return.

O’Driscoll says funds with both early-stage and growth investing capabilities can pursue those rare outcomes at different points. The earlier the entry, the greater the uncertainty over which companies will become very large; investing later can still make sense if the price and outcome work.

Stebbings observes that large funds are preparing to raise again after a short deployment cycle. O’Driscoll says money put to work quickly in an enthusiastic market may earn lower returns than checks written when investors were more fearful.

Lemkin adds that rapid fundraising and spending can help drive consecutive valuation increases. Even when revenue growth supports part of a markup, capital moving quickly may compress future returns, though investors may accept lower multiples on much larger sums.

03

Three late-stage deals demand different risk assessments

Kalshi, Mercor and Fractile offer distinct market opportunities at high prices. The investors compare regulation, customers, technology and valuation.

The panel considers three hypothetical investments: prediction-market operator Kalshi, data company Mercor and inference-chip developer Fractile. The exercise compares their market opportunities with late-stage prices rather than assuming a successful business is automatically a good investment.

Lemkin is uneasy about Kalshi’s regulatory and litigation risks at the quoted valuation and cannot confidently assess a pre-revenue chip startup. He chooses Mercor, while saying he would have preferred to invest at an earlier, lower price.

Stebbings makes the case for all three: a large US betting market for Kalshi, demand from model developers for Mercor’s data, and specialized inference hardware for Fractile. These are his investment arguments rather than agreed forecasts.

O’Driscoll says investors must move beyond calling each company impressive and compare the price with the size of its market. For Fractile, he wants to understand its technological advantage relative to a public inference-chip company already shipping products and recording revenue.

On Mercor, he says the customer base for training data may be small, but individual AI labs can spend very large amounts. On Kalshi, he would investigate the eventual size of sports betting and prediction markets and the effect of legal restrictions before making a decision.

The group settles on a hypothetical Mercor investment. Lemkin says even if the return is not exceptional, a stake could provide the fund with useful knowledge of the market, though the participants remain conscious of the entry valuation.

04

Fewer large buyers could weaken venture exits

Lemkin questions assumptions about acquisition exits and liquidation preferences. The panel discusses Nvidia’s potential transactions and capital-intensive startups.

Lemkin returns to Nvidia’s role as a buyer and financier of AI startups. He argues that investors who assumed a large company would always acquire or rescue a venture should reconsider what their liquidation preferences actually protect if that buyer becomes less active.

The group discusses the uncertainty around a proposed Hugging Face transaction and potential regulatory scrutiny. O’Driscoll says the price may be manageable relative to Nvidia’s size, but a prolonged regulatory process could complicate approval.

Lemkin says a year with fewer acquisitions by both Nvidia and Salesforce would look different for venture-backed companies seeking liquidity. He cites Poolside as an example of a company whose decision to sell reflected the immense capital needed to keep building.

05

Can returning founders reverse a software slowdown?

HubSpot’s new-customer growth frames a discussion of founder-led turnarounds. The investors distinguish a better trajectory from a return to earlier growth.

Stebbings turns to HubSpot co-founder Brian Halligan’s return. Lemkin says the concern is not only share price: HubSpot had consistently added about 10,000 net customers per quarter, he says, but the recent additions and guidance had fallen substantially.

He contrasts that pattern with another software company’s decline in larger new accounts. Price increases and expansion sales can sustain revenue growth for a while, he says, even as a weakening flow of new customers makes long-term growth harder.

Lemkin expects a returning founder to bring product knowledge and authority to make difficult changes, but asks whether that will be enough. He sees a real difference between improving a company’s direction and restoring its former growth rate.

He points to Intercom: its founder returned, changed pricing and pushed into AI agents, producing a substantial new business. But its later sale means investors did not see whether that product could grow into a much larger standalone business.

O’Driscoll agrees that changing the trajectory of an established company with a large revenue base is hard. A founder might improve customer growth and produce a healthy public company without matching a much smaller startup growing several times a year.

He notes that several boards have brought founders back in difficult moments. At the same time, an incumbent application may be optimizing an older product when customers need a substantially different product built around AI.

HubSpot’s repositioning around a customer-relationship system is an attempt to respond, O’Driscoll says. He is unsure whether its typical small or medium-sized customer cares about the technical terminology, and notes that a faster-growing new product still sits inside a much larger existing business.

Lemkin says founders may return with freedom to change teams and spending, yet even a talented leader may face a ceiling on growth. The practical goal may be to improve the outcome available to the business rather than recreate its earliest expansion.

The panel also considers market saturation. O’Driscoll says many potential customers already have a customer-management system, so a mature software category cannot add buyers indefinitely; AI is one pressure alongside that finite market.

06

Small businesses may wait to buy software

Lemkin argues that customers may continue with existing tools while spending on agents that automate urgent work. O’Driscoll and Stebbings debate how far custom-built alternatives have spread.

Lemkin regards the recent decline in new customers at HubSpot and Monday as a sharper change in small-business buying. He does not think a mass movement of ordinary small businesses building their own replacement systems fully explains it.

O’Driscoll describes investors creating custom interfaces with AI for an existing Salesforce system. Stebbings questions whether that behavior among technically sophisticated workers generalizes to an ordinary small business without similar skills.

Lemkin’s explanation falls between those positions: many businesses may defer buying another application, continue using spreadsheets, or use a general AI tool for a limited task. They need not have built a full custom replacement for the software vendor to lose a sale.

He says portfolio companies selling agents that answer calls, reschedule appointments or handle billing find clearer demand. A business that struggles to keep a receptionist may immediately understand the value of a service that covers that work around the clock.

In his view, the shift is about a customer’s immediate return rather than simply switching to a competing software brand. A small company might postpone a conventional subscription for a year or two while paying for automation that solves an urgent staffing problem.

07

Critical systems and third-party tools face reliability risks

The panel discusses service outages without asserting an unconfirmed cause. It considers cascading failures, cybersecurity and the resources needed to protect critical services.

Stebbings raises outages affecting Microsoft services and US air traffic. O’Driscoll says the initial explanations he has seen point to ordinary infrastructure or configuration problems rather than evidence of an AI-driven cyberattack; he does not claim to know all the causes.

Lemkin argues that AI agents could make it easier to probe systems for weaknesses or overwhelm them with traffic. He separates that concern about future attacks from the unresolved causes of the particular outages under discussion.

He describes the danger of a cascading overload: when some servers fail, their traffic moves to fewer remaining machines, which can fail in turn. Fixing the original problem can become harder while the cascade is still spreading.

O’Driscoll says critical services need stronger operational and security practices. An outage of email is inconvenient, but an extended failure of air-traffic systems would have wider economic and safety consequences, in his view.

He recalls advising a company as it grew to invest in security expertise and reliability. The panel argues that systems supporting important public and business functions need enough maintenance and skilled people to handle increasingly demanding conditions.

Stebbings describes the disruption to flights as frightening. The panel uses it to discuss a trade-off for startups: small teams can build products more quickly, but their services may become dependencies for customers with substantial security requirements.

Lemkin says an insecure third-party plug-in can create a risk for many organizations that use it. He discusses recent vendor incidents as examples of how a problem in one supplier can affect multiple customers, while cautioning that exposure can occur in a neglected corner of an application.

The group notes that customers have increasingly named vendors when reporting incidents. Lemkin sees a tension between celebrating very small startups and expecting the extensive security capability needed when their products become critical infrastructure.

08

General agents gain value from accumulated workflows

Lemkin explains how his team uses a connected assistant and why switching would be disruptive. The group weighs permissions, compute costs, competition and staffing.

Stebbings asks whether OpenAI’s purchase of a travel-focused assistant company signals a shift from specialized assistants to general-purpose agents. Lemkin answers by describing his own use of Meta’s Muse across his business.

He says the agent handles tasks including outbound work, content editing, analysis of website traffic and checking his team’s workflows. Its usefulness across many separate tasks makes it harder for him to switch to a competing assistant.

Lemkin says his agent reviews incoming venture opportunities, ranks them and reminds him about promising items he has not read. It can communicate with another team member’s agent; he describes these as workflows his team has built rather than universal product capabilities.

Asked about permissions, he says he has given it access to his email but not a credit card. If purchasing became necessary, he would consider a separate account with a limited balance rather than unrestricted access.

The investors discuss whether successful agents become sticky once they accumulate useful workflows. O’Driscoll says users may switch when another product can do a valuable task the first one cannot, but changing a functioning collection of automations has a cost.

Lemkin gives the example of an agent interacting with his WordPress site, a task he says other agents had not handled as well. A specific capability could justify adding or changing products even when the user is otherwise settled.

They compare what heavy users pay with the computing resources their agents consume. O’Driscoll says the viability of a low-price offering depends partly on how many users operate it as intensively as the investors do.

Lemkin suggests an assistant connected to a database could identify and invite relevant founders to an event, potentially delaying a purchase of marketing automation software. He presents this as an example from the workflows his team has already tested.

O’Driscoll asks whether a consumer company’s general agent could become a small-business tool, and whether a specialist would offer enough improvement to displace it. Large distribution and the ability to subsidize compute could give an incumbent an advantage.

Lemkin suggests a practical starting point: identify work that the team is not getting done and ask the agent to attempt it. He mentions follow-up emails after an event as the kind of task that can be tested against the results a human would produce.

He is enthusiastic about the product’s browser, memory and task-handling capabilities, but worries about the economics for independent competitors. An agent available continuously may consume more infrastructure than a startup can afford to include at a low subscription price.

O’Driscoll says a specialist could still compete if most users need less computing power than heavy business users and if it can move quickly. He also acknowledges the difficulty of competing with a large platform that has distribution and deep resources.

The discussion moves to staffing. Lemkin expects businesses increasingly to ask an agent to do a task first, then bring in people for gaps; he gives collections work as an area where his team has reduced outside spending.

He expects some workforce reductions to happen through not replacing departing employees rather than immediate layoffs. Stebbings says workers also need to improve as capabilities change; the group notes that supervising agents can itself take many hours.

09

Starship’s achievement and the compute investment case

O’Driscoll praises the engineering while examining the role of future compute demand in SpaceX valuation. The speakers close with opposing predictions on Nvidia leadership.

The investors briefly turn to SpaceX’s Starship catch. O’Driscoll calls it a substantial technical achievement that could reduce launch costs, while separating his assessment of the engineering from his view of the company’s valuation.

He says a large part of the investment case depends on future demand for computing, not only rockets or Starlink. Whether customers buy more compute or cancel orders is therefore central to how he would evaluate the stock.

The group closes by returning to Huang’s possible return. Lemkin predicts it could happen within days and argues the opportunity may narrow as the news cycle moves on; O’Driscoll remains skeptical. Stebbings ends the conversation.

同じチャンネルの記事

1:15:58EN
20VC with Harry Stebbings· 2026年11月14日ポッドキャスト · 2 人Fractile’s Walter Goodwin wants to break the AI memory bottleneckゲスト Walter Goodwin · ホスト Harry StebbingsGoodwin explains why faster inference needs both memory bandwidth and capacity, and what must happen before a new chip can reach scale.SemiconductorsInferenceインフラとデプロイ+2
50:17EN
20VC with Harry Stebbings· 2026年11月9日ポッドキャスト · 2 人Mistral’s Arthur Mensch on sovereign AI and the compute bottleneckゲスト Arthur Mensch · ホスト Harry StebbingsThe CEO says customers need model choice and continuity, while European infrastructure must scale ahead of demand.AI Strategyエンタープライズ AIインフラとデプロイ+2
57:17EN
20VC with Harry Stebbings· 2026年11月7日ポッドキャスト · 2 人Scott Wilson on WashU’s Endowment, Concentration and Co-Investingゲスト Scott Wilson · ホスト Harry StebbingsWashU’s CIO explains its manager relationships, direct investments, liquidity constraints and views on private-market valuations.InvestingVenture CapitalDecision Making+2
1:08:33EN
20VC with Harry Stebbings· 2026年11月5日ポッドキャスト · 3 人Harry Stebbings, Jason Lemkin and Rory O’Driscoll on AI valuations, agent economics and venture exitsゲスト Jason Lemkin, Rory O’Driscoll · ホスト Harry StebbingsThe investors debate pricing pressure on model providers, the threat agents pose to platforms, and how venture returns depend on large acquisitions.Venture CapitalAI StrategyInvesting
1:02:12EN
20VC with Harry Stebbings· 2026年11月2日ポッドキャスト · 2 人Modal CEO Erik Bernhardsson on Inference, Compute and Building an AI Platformゲスト Erik Bernhardsson · ホスト Harry StebbingsErik Bernhardsson discusses GPU capacity, open models, sandbox growth and why Modal is expanding beyond a single infrastructure product.InferenceAI platform engineeringインフラとデプロイ+2
1:20:42EN
20VC with Harry Stebbings· 2026年10月31日ポッドキャスト · 2 人David Tisch on Seed Investing, Ownership and Finding Outlier Companiesゲスト David Tisch · ホスト Harry StebbingsThe BoxGroup founder discusses portfolio construction, competition for seed deals and the relationships behind venture investing.Venture CapitalInvestingFounders+2