AI Compute Deals, Cerebras’s IPO and the Cost of Building: A 20VC Roundtable

Harry Stebbings, SaaStr founder Jason Lemkin and Scale Venture Partners’ Rory O’Driscoll discuss AI compute deals, software competition, IPOs and the demands of building a company. Financial figures, deal terms and corporate plans are presented as reports or the speakers’ accounts; their valuation, legal and market predictions are opinions rather than established outcomes.

Stebbings says Anthropic has required board approval for secondary sales and special-purpose vehicles, and cites a reported fall in secondary-market pricing. The panel discusses the reported policy and its possible consequences without independently verifying the announcement, particular transfers or the quoted price range.

O’Driscoll distinguishes a legitimate special-purpose vehicle that holds shares directly from an arrangement through which an existing shareholder tries to transfer the economic value of shares without an approved transfer. A company may control changes to its capitalization table, he says, although the rights and obligations depend on the actual documents and applicable law.

Using himself and Lemkin as hypothetical parties, O’Driscoll explains that a buyer of a contractual economic interest might not appear on a company’s share register and could face counterparty risk. He speculates that disputes could arise if a seller lacked the shares or made conflicting promises. This is an illustrative scenario, not evidence that a named party did so.

O’Driscoll suggests the reported policy could reduce uncertainty ahead of a potential IPO, while acknowledging he does not know what approvals Anthropic granted in the past. Lemkin calls the development less novel, saying transfer restrictions and earlier warnings about SPVs already existed. Their descriptions of company motives and particular prior behavior are interpretations, not verified findings.

The discussion turns to a reported arrangement for Anthropic to obtain compute capacity from an Elon Musk-linked business. O’Driscoll sees a possible economic rationale when one party has available capacity and another wants it. His estimates of utilization, revenue and competitive positioning, and his claim about the companies’ motives, are not established by the interview.

Lemkin argues that competitors can also be suppliers, comparing the possibility to companies that compete in finished products while selling components to each other. He thinks separate business units may have incentives to use idle infrastructure productively. The panel’s discussion of the arrangement and its financial effects is based on its account of reported terms, not a verified contract.

The guests describe a fast-changing race for compute and speculate that Anthropic may seek capacity from multiple providers. O’Driscoll says any prospective SpaceX offering would require investors to understand recent changes in its business mix. His revenue examples and references to other possible transactions are hypothetical or reported, not independently confirmed.

Stebbings cites a reported $200 billion Anthropic commitment to Google over five years. The panel discusses why a cloud provider might supply compute to a model company that competes with its own Gemini products. The headline amount, timing and conditions are not independently established in the conversation and should not be treated as realized revenue.

O’Driscoll sees a trade-off: infrastructure revenue may be attractive, but selling capacity also helps a model competitor grow. He is unsure whether the greatest long-term value will sit with model providers or with capital-intensive infrastructure owners. Lemkin adds that a provider can benefit from both its own model and a customer’s growth.

Lemkin cites reported enterprise usage shares for Gemini and Claude and says organizations can use multiple models, so the figures need not sum to 100%. The panel does not inspect the underlying survey, definitions or methodology. His interpretation is that serving more than one growing model provider may still be valuable to Google.

Stebbings raises a Goldman Sachs projection that agents will increase token consumption 24-fold by 2030. Lemkin thinks that may be conservative, proposing that parallel agents and adoption outside technology could produce much greater usage. His alternative multiple is a conjecture he says needs better modeling, not a measured forecast.

Lemkin gives coding as an example: parallel agents might generate several implementations, after which a model and a human select or combine the best. He sees potential for more work to happen simultaneously but acknowledges that not every workflow benefits from many agents. The amount of useful additional work and its cost would depend on the task.

O’Driscoll cautions that token volumes alone can mislead. Hardware improvements and software optimizations may reduce the cost per token, while more sophisticated tasks may use many more tokens. He says forecasting revenue requires estimating both forces and their interaction, and admits that precise forecasts are difficult.

Lemkin offers a counterargument to unlimited token growth: some experienced engineering leaders tell him teams generate more code and consume more tokens than they can productively use. He distinguishes experimentation and waste from work that reaches production. These are reported conversations and his own judgment, not a measured industry-wide rate of waste.

Stebbings contrasts that concern with Atlassian CEO Mike Cannon-Brookes’ argument that demand for software and features may be effectively unlimited. Lemkin asks whether an average engineer can use very high monthly token spending productively. He thinks organizations will eventually distinguish useful output from activity that merely looks impressive.

O’Driscoll invokes Goodhart’s law: making token use a target can change employee behavior and weaken it as a measure of productivity. He cites reports of workers using tokens to satisfy targets, but the conversation does not verify specific incidents. Companies need a measure of value delivered, rather than assuming high consumption is itself success.

O’Driscoll also asks whether the best engineers might use more tokens effectively because they can direct complex work. Lemkin agrees that highly capable people may benefit from broad access, but worries about low-value output elsewhere. O’Driscoll says his team surveyed engineering leaders about spending and still lacks a clear, objective measure of productive results.

Stebbings asks whether model providers will enter legal work, financial analysis and other application categories. He refers to Anthropic templates and a purported future legal product; neither its launch nor its competitive effect is established here. Lemkin says some startups he invested in were affected by new model features, but believes specialized applications may retain important advantages.

For legal work, Lemkin argues that reliability, domain-specific workflows and integrations may matter more than saving on a general model subscription. He points to the risk of submitting legal documents containing hallucinations. His view is a product argument, not a guarantee that any named legal tool is accurate or appropriate for a particular matter.

Lemkin thinks model companies may offer pieces of customer-support or legal workflows without committing to complete enterprise applications, though he says he could be wrong. A model feature that makes a simple application unnecessary can nevertheless hurt a startup. He reports that one of his investments had to change direction after functionality appeared in Claude.

O’Driscoll compares the question to earlier platform shifts. Microsoft dominated broad productivity software while many enterprise application companies built on its platform; later, investors debated whether cloud providers would absorb independent infrastructure products. He uses these historical examples to argue that platform ownership does not automatically eliminate all specialized firms.

O’Driscoll expects broad intelligence tools and enterprise applications to coexist, particularly when customers need coordinated workflows, customization and relationships. He discloses that GCAI, a legal AI company serving corporate customers, is in his portfolio. His prediction that specialized providers will retain a role remains a view, not a guarantee.

Lemkin separately predicts that traditional marketing automation could become less relevant if agents no longer need interfaces designed for people to compose messages. O’Driscoll distinguishes two questions: whether models absorb new applications, and whether agents make older software obsolete. They agree that new products can also age quickly as underlying models improve.

The panel says startups close to the model layer face particularly rapid competitive pressure: an improvement by a model supplier can change the value of a separate application. Their discussion is a strategic warning rather than evidence that any particular product will disappear or that all legacy software faces the same timeline.

Stebbings moves to public markets, citing declines at HubSpot and AppLovin and a Cloudflare layoff announcement. Lemkin compares Monday.com with HubSpot: he says both were slowing, but Monday raised guidance while HubSpot lowered it. He interprets the different stock reactions through growth expectations; the figures and causation are his account.

O’Driscoll says investors must consider two separate factors: whether the business is strengthening or weakening, and the price already paid for the shares. He speculates that a large cost reduction can also raise questions about a company’s outlook. Reported layoffs do not alone establish the reasons for cuts or prove how much work AI replaced.

A company trading at a high multiple can disappoint even after a solid quarter if investors expected more, O’Driscoll argues. A company already priced for trouble may rebound on modest positive news. His examples are interpretations of market behavior, not a formula for future share prices.

Lemkin describes ZoomInfo as an incumbent in sales intelligence and argues that Clay and similar tools captured growth it might otherwise have enjoyed. He says SaaStr uses Clay and likes it, but admits its revenue is hard to know. His claim about the precise source of ZoomInfo’s lost growth is an interpretation, not a demonstrated causal calculation.

O’Driscoll emphasizes Clay’s ability to compare multiple data providers in a workflow, which he thinks can commoditize individual data sources. He also praises the addition of agents. This competitive explanation reflects his analysis of the products; it does not establish that all changes in a rival’s revenue were caused by Clay.

The guests discuss the possibility of a private-equity purchase for a slow-growing public software company. O’Driscoll warns that a buyer and an attractive price are not guaranteed. He uses the metrics cited in the conversation to illustrate a potential transaction, not to announce an offer or predict a sale.

O’Driscoll argues that a mature SaaS business may earn a better valuation if it demonstrates growth, profitability and a credible future. His illustrative multiples and growth targets are views about what investors might accept, not assurances about a particular company or its stock.

Stebbings says Cerebras has raised the proposed IPO price range to $150–$160 and cites reported oversubscription, a $4.8 billion raise and a $48 billion fully diluted valuation. The conversation takes place before the offering; these figures and the IPO outcome are not independently established and could change.

O’Driscoll expects strong initial trading because, in his experience, a large increase in the proposed range suggests demand. He distinguishes a short-term first-day pop from the harder question of how the business will perform over two years. Even his confident prediction is a prediction, not knowledge of the eventual market price.

Lemkin agrees that indicated demand appears strong but says chip competition, capacity hedging and the early stage of deployment make longer-term outcomes difficult to predict. He points to reported support or commitments from OpenAI and Amazon as part of the story, without independently verifying the terms or their ultimate value.

O’Driscoll says Cerebras historically depended heavily on a small number of customers and that future plans involve a different customer mix. He discloses that one of his portfolio companies, Tabula, uses Cerebras for inference. He sees potential demand for fast responses, while acknowledging that the medium-term business case remains a bet.

Stebbings compares Cerebras’s proposed valuation with Nvidia’s much larger market capitalization and asks about a long-term upside case. O’Driscoll says the hypothetical comparison can be attractive if Cerebras becomes a meaningful competitor, but explicitly calls it risky. Neither an assumed probability nor a comparison of market caps establishes future returns.

Stebbings raises a claimed 20% Benchmark stake. O’Driscoll corrects him, saying the S-1 lists Foundation, Benchmark and Eclipse at roughly 8%–9% each. The percentages are his reading of the filing and remain subject to the document’s definitions and timing; they are not a realized profit calculation.

Lemkin praises Foundation for incubating Cerebras and supporting an uncertain, capital-intensive idea early. O’Driscoll credits the investors and founder for persisting through years when the opportunity was less obvious. Their discussion of paper value and potential fund returns is not evidence that the investors have sold or realized gains.

The guests move to fintech: Stebbings mentions a reported Ramp valuation near $40 billion, Parker’s Chapter 7 filing and Gusto reaching $1 billion. O’Driscoll contrasts Ramp’s broad corporate-card market and software expansion with the narrower economics he attributes to Parker. These are the panel’s descriptions of different businesses, not a full account of either company’s financial condition.

O’Driscoll says card interchange alone may leave only modest contribution margins after customer incentives, so software, payments and purchasing tools can expand the opportunity. He likes the idea of agents helping automate procurement. Lemkin says SaaStr may compare Ramp and Brex when rebuilding its financial stack; he has not yet decided to switch.

The guests separate product strategy from price. O’Driscoll cites approximately $1 billion in Ramp revenue and a reported $40 billion valuation, then says he probably would not invest at that price without a clear growth case. Lemkin questions whether investors sufficiently adjust revenue multiples for margin quality. These figures and opinions are not independently verified investment guidance.

O’Driscoll illustrates valuation risk by asking how many years of rapid growth would be needed to reach a more conventional revenue multiple. He notes that financing preferences and existing investors’ earlier entry prices can affect their overall economics. The scenario is mathematical illustration, not a forecast of Ramp’s future growth or realized returns.

Stebbings mentions reports that Lime is preparing for an IPO and praises its survival and growth. O’Driscoll says he would like to see the financial numbers. The discussion does not establish an offering date or financial results, and a company preparing for a listing may not complete one.

The panel briefly discusses litigation involving Musk and OpenAI. O’Driscoll offers his understanding of the proceedings and predicts that OpenAI’s arrangements will survive, while acknowledging it is his gut judgment. His account of procedure, likely ruling and the parties’ positions is not an independently verified legal assessment.

Lemkin describes changing SaaStr’s event format toward workshops and demonstrations instead of conventional talks, arguing that podcasts already make many interviews accessible. O’Driscoll says practical demonstrations of how agents are built and secured could make an in-person event more valuable. This is their assessment of a particular format, not a general claim that talks have no value.

Asked about memory-chip stocks, O’Driscoll says a prior increase alone is not enough to decide whether a stock is expensive. He would compare current earnings with the duration of the capital-spending boom and the time needed to add manufacturing capacity. He says he does not have a developed conclusion, and later states that he does not hold a long position in Nebius.

Stebbings raises a statement he attributes to MrBeast and his own experience that major success involved sacrificing mental health. O’Driscoll agrees demanding work often involves serious sacrifices, but says severe stress can impair decisions and that people should seek ways to regain perspective. These are personal experiences and opinions, not proof that mental-health harm is necessary for success.

Lemkin contrasts his first startup, which he says sold after about a year, with a later five-year journey through financial strain and near failures. He describes the lasting change he felt from prolonged intensity and says a short break did not restore his previous outlook. His account is personal and should not be generalized into a clinical claim about every founder.

Lemkin believes years of intense founder work can change relationships, perspective and the ability to switch off. He supports seeking mental-health help while saying his own experience was not simply solved by a vacation. Stebbings recalls a UiPath founder’s description of loneliness, and Lemkin relates the sentiment to his own interpretation of sustained pressure.

The guests discuss whether exceptional company outcomes select for unusually intense founders willing to reject acquisition offers. Lemkin says he often tells founders to consider selling when the economics work for them, because continuing can require years of uncertain effort. These are their investment and personal judgments, not a rule for founders or a reliable assessment of anyone else’s mental state.

O’Driscoll agrees intensity can matter but emphasizes the need to sleep, maintain health and find ways to clear one’s head. He recounts leadership advice about staying functional under pressure and says exhaustion can weaken decisions. The participants close acknowledging both the sacrifices they see and the value of perspective.