Anthropic, Statsig and the Economics of AI Growth: Cliff Obrecht and 20VC

In this September 2025 roundtable, Harry Stebbings, Jason Lemkin and Rory O’Driscoll are joined by Canva co-founder and COO Cliff Obrecht. They discuss Anthropic’s financing, OpenAI’s reported acquisition of Statsig, AI application valuations, public software companies and the costs of serving AI. Figures and forecasts reflect the discussion at the time and are not independently verified here.

Stebbings introduces Anthropic’s $13 billion financing at a reported $183 billion post-money valuation. O’Driscoll says the absolute price is high but its reasonableness depends on future growth. He starts from reported revenue run rates and models what a much larger business might earn the following year.

O’Driscoll illustrates how a company moving from roughly $1 billion to $9 billion in annualized run rate over a year might produce a much lower recognized revenue figure for that year. He then assumes growth to $30 billion in run rate the following year to estimate a forward multiple. Both the averaging shortcut and future growth are assumptions, not reported GAAP results.

The investors emphasize the central uncertainty: exceptional current growth may slow. A valuation that looks modest against a projected future revenue number becomes expensive if that forecast is missed. Obrecht says demand for the round was strong; the amount of oversubscription is his account, not independently established here.

The panel compares Anthropic with reported valuations for Databricks and Canva. Lemkin says forward multiples can appear more reasonable when growth persists. Obrecht speaks as Canva’s co-founder and COO, with an interest in the company’s valuation and financing decisions.

Asked how a heavily oversubscribed company chooses investors, Obrecht says Canva considered long-term relationships and potential IPO shareholders. He reports that Canva expected to close the year near $4 billion in revenue with growth close to 40%, and says Fidelity anchored a secondary transaction. These are management statements and expectations, not subsequently verified results.

Obrecht says Canva had more than $1 billion in cash and had been profitable for years, so the transaction was intended to give employees and existing investors liquidity rather than raise primary capital. He says demand from would-be sellers changed after Figma’s IPO, illustrating how public valuations can affect a private secondary.

When employees weigh whether to sell, Obrecht says Canva shows them a range of public-company comparisons rather than anchoring on a single hot stock. He also describes Canva’s earlier valuation decline and subsequent growth. The figures illustrate his account of market changes, not a promise about future share prices.

O’Driscoll says public and private investors can apply dramatically different multiples over time. Sustained operating growth can eventually make an earlier valuation look more reasonable, but it cannot guarantee that every high entry price will work. Obrecht says Canva turned down some even higher prices in a prior financing.

Obrecht worries that some AI products may pull forward revenue from enthusiastic early adopters without crossing into the mainstream. He says reaching a much wider audience requires distribution, and moving from tens of millions to billions in revenue is a substantial leap.

O’Driscoll asks whether AI application renewals will hold up after the initial excitement. Obrecht says Canva aims to build useful features into existing workflows rather than chase gimmicks. He expects customers to experiment broadly and consolidate tools over time; that expectation is not an observed industry-wide renewal rate.

Obrecht says Canva lets employees try several security-approved AI and coding tools and plans to narrow the selection as winners emerge. He describes the company’s own purchasing policy, including its willingness to spend for productivity. Tool adoption and cost effectiveness will still need to be measured.

Stebbings introduces OpenAI’s reported acquisition of Statsig for $1.1 billion in stock and cites approximately $75 million of annual recurring revenue. He discloses that a partner at 20VC was an angel investor. The panel debates the price and structure using reported terms, not independently confirmed financial statements.

O’Driscoll suggests a late-stage Statsig investor may value receiving OpenAI shares even if the nominal sale price matches the prior round. Lemkin offers a similar interpretation of different investors’ incentives. Their speculation about individual investors’ access, preferences and motivations is not established by the deal announcement.

Obrecht says he would not choose an acquisition for Canva in the hypothetical posed by Stebbings and prefers to follow its own course. The exchange highlights that an all-stock offer can have different implications for a founder, early investor and later investor.

The conversation turns to Meta’s reported multibillion-dollar investment in Scale AI and changes to its AI organization. O’Driscoll says combining highly paid talent can create management challenges and questions whether the acquired asset will meet Meta’s needs. These are opinions about a developing situation, not proof of a failed transaction.

O’Driscoll notes that data-labeling needs can shift from simpler labels to expert tasks and that customers have several potential suppliers. He suggests Meta may choose vendors on performance even where it has an investment. Assertions about Scale’s internal condition and any future write-down remain speculation.

Lemkin argues that highly paid external hires can sometimes serve a useful purpose even if they differ from a company’s original culture. The group does not know how Meta’s hiring strategy will perform. Obrecht then says AI contributed an estimated 20% of Canva’s reacceleration, with core execution and international expansion also important.

Stebbings cites reported potential financings for Lovable at $4 billion and Vercel at $9 billion and asks whether rapid markups reflect progress or excess capital. Stebbings has an investment interest in Lovable, which is relevant to his comments. The potential rounds and revenue projections are not independently confirmed in the discussion.

Obrecht sees both real opportunity and fear of missing out. He worries that early traction may not automatically translate into mainstream adoption, while saying Lovable appears well positioned if it executes. His comments about a possible market cycle are predictions rather than established outcomes.

Lemkin questions whether a large markup weeks after a round reflects new information or growth investors should already have modeled. Stebbings says Lovable was exceeding its plan and offers a year-end revenue forecast. These are participant claims and projections, not verified accounts of the company’s finances.

O’Driscoll offers three explanations for a rapid second round: new operating information, an earlier or later mispricing, or validation from a prestigious investor. More than one can be true. Obrecht adds that if a company compounds into a much larger market leader, a difference in entry price may feel less decisive.

O’Driscoll cautions that if no meaningful information changed, a sharply higher price may mean one of the rounds was mispriced. If performance exceeded the prior plan, however, a substantial repricing could reflect real new evidence. The panel cannot determine which explanation fits every reported transaction.

Lemkin compares reported Lovable and Replit prices and questions whether similar companies can be valued very differently within days. His comparison uses his own estimates of their businesses and revenues. A superficially similar revenue figure does not prove the two companies have identical economics or risk.

Asked whether raising a lot of money harms a company, O’Driscoll says a larger balance sheet can help through a downturn if management remains disciplined. Obrecht says Canva deliberately raised relatively little to limit dilution, but describes that approach as risky and says he would now advise founders to be somewhat overcapitalized.

The panel shifts to quarterly results from public software businesses. Lemkin says some established companies appear to be benefiting from AI, while others have not yet seen a comparable effect. He describes the improvement as uneven and often modest rather than a universal return to exceptional growth.

O’Driscoll says a stock can rise sharply when results exceed depressed expectations, even if the company has not returned to a dramatically higher growth rate. He argues that a strong existing business and improved execution can matter alongside AI features. Market reaction is not the same as proof of a permanent acceleration.

Obrecht says a large private company with sophisticated investors already faces significant reporting and performance expectations. Canva wants to be ready for an IPO, he says, but has not chosen when to list. The company’s employee-liquidity needs are a major part of its evaluation.

Obrecht says private secondary transactions offer some liquidity but can be awkward or restricted in certain jurisdictions. He wants employees and, eventually, customers to be able to share in Canva’s success. This is the company’s stated objective, not an announced listing date.

Stebbings raises the possibility of a direct listing. Obrecht says it is one option but questions the short-term performance of prior examples and values relationships with long-term shareholders. O’Driscoll counters that a smaller first-day price jump can reflect a more efficient opening price, not necessarily a failed listing.

The guests discuss how lockup design, allocation and investor relationships can affect a listing. Their descriptions are general observations; a company’s actual options depend on its transaction and rules. Canva has not committed to a specific route in this conversation.

Stebbings cites Jensen Huang’s estimate of a $3–4 trillion AI infrastructure opportunity over five years and asks whether AI revenue can support that spending. O’Driscoll uses an illustrative required return to show how much profit would be needed. His arithmetic is a scenario, not evidence that the investment or profit has occurred.

Lemkin argues that AI penetration remains early and could expand dramatically across existing software customers. Obrecht says Canva already sees billions of AI uses per month, but notes that model distillation and on-device processing may reduce unit costs. The potential scale of usage and the cost per use are separate variables.

O’Driscoll cites a decline in Notion’s gross margin from roughly 90% to 80% as an illustration of AI-related costs, then asks whether Canva could spend a comparable share of revenue on AI. The Notion comparison is his interpretation and does not establish that the entire change was caused by AI.

Obrecht says Canva’s AI costs can reach that order of magnitude, especially while it trains models as well as serves features, and expects the proportion to fall over time. The panel distinguishes model training from inference and recognizes that the forecast depends on cost reductions and usage growth.

Obrecht says the company expects model costs to decline and evaluates which tasks require the strongest models versus smaller or more efficient ones. Lemkin counters that falling prices can encourage people to use much more AI. Both effects can occur at once; neither establishes Canva’s future margin.

O’Driscoll questions whether software companies will willingly transfer enough value to infrastructure and model providers to justify very large capital expenditure. Lemkin thinks new uses and increased token consumption could expand the opportunity. They offer competing economic scenarios, not established market totals.

Obrecht says Canva must model the costs of new AI features across its large user base. He describes a unified credit approach: plans include an allowance, with additional usage-based charges for heavy users. The aim is to let people use the features while managing exposure to compute costs.

Asked about equipping engineers with expensive coding agents, Obrecht says Canva encourages security-approved tools and is not narrowly price-sensitive when they improve productivity. He does not say the company budgets $10,000 per engineer per month; the group discusses that amount as a hypothetical benchmark.

Obrecht says one marketer using AI might create and distribute vastly more content than before. A flat per-seat price may not cover heavy compute use, so Canva is considering a hybrid of seat and consumption pricing. The example is his description of a possible model, not proof of a completed pricing change across every product.

Stebbings asks whether search traffic will shrink as users turn to AI assistants. Obrecht says most Canva acquisition is organic, with search only one portion, and claims its SEO traffic is growing. He also cites referrals and image uploads associated with ChatGPT; these are Canva’s reported metrics, with different denominators.

Obrecht says Canva has a team working on visibility in AI-generated answers. O’Driscoll distinguishes a product company that can receive a referral from a publisher whose answer may be summarized without a visit. The effect of AI search depends on the business model rather than a single claim that SEO is dead.

Asked to choose an investment among OpenAI, Anthropic and xAI at the valuations suggested by Stebbings, Obrecht declines to pick one and praises the category. He speculates that AI demand could accelerate investment in lower-emission energy. That is his forecast, not a demonstrated environmental outcome.

Stebbings describes the discomfort of investing later in a company he passed on early. O’Driscoll says he has done exactly that, including with Box: if founders deliver against an earlier plan, an investor gains information that can justify revisiting the decision even at a higher price.

O’Driscoll says sustained execution across multiple observations can be valuable, while acknowledging that AI products can move in and out of product-market fit. In his view, a founder who repeatedly adapts to a changing market may provide a useful signal. Past execution does not guarantee the next stage.

Obrecht says some successful early Canva investors found ways to maintain or increase their exposure in later rounds. O’Driscoll separates whether the later investment is attractive from whether a fund has the structure and reserves to make it. Their discussion reflects the founder’s and investors’ interests, not a blanket instruction to follow on.

Stebbings cautions that a higher valuation can sometimes reflect a price jump without an equivalent improvement in the company. O’Driscoll says a strong outside-led round and positive operating data can be informative, but the investor still has to judge the price and distinguish durable winners from overvalued companies. The guests close on that uncertainty.