AI Security, Revenue Claims and Wearable Economics: A 20VC Investor Roundtable

Harry Stebbings opens with claims circulating about a purported unreleased Anthropic model, described as Claude Mythos. He says leaked assets suggest a ten-trillion-parameter model with unusual capabilities. The panel does not verify the model’s existence, size or performance; Rory O’Driscoll later says he has seen only screenshots and that the model is not publicly available.

Jason Lemkin treats the reported disclosure as an example of a broader trade-off: teams shipping faster can expose code, databases or personal information by mistake. He expects autonomous agents to increase the volume of work and potentially the number of security mistakes, even if their error rate per task is lower. His arithmetic is an illustration, not a measured forecast.

O’Driscoll says the reported model was described as particularly strong in cybersecurity, making an apparent leak ironic. He speculates that a publishing error may have exposed staged material, while acknowledging that the underlying information is limited. He says the reported claims include greater capability and cost; none of those model characteristics is established by the panel.

O’Driscoll says the alleged leaked material is difficult to track down and that claims about the model remain untested in public. He associates the reported leak with a decline in security stocks, but the panel has not established causation. His caution about limited access is important when considering the discussion’s subsequent assertions about capabilities.

Lemkin recalls working at Adobe when moving source code to GitHub required an exception because the code was regarded as a crown jewel. The older on-premises process, he says, could take a month to release software. Faster releases offer clear benefits, but he argues that the increase in speed comes with security trade-offs.

Lemkin is excited by the prospect of assistants that continue working in the background, sleep and resume without a new prompt. He expects always-on agents to consume far more tokens and become part of daily decisions, while recognizing the privacy and autonomy people may surrender. The features he mentions come from material whose claims the panel has not verified.

Stebbings contrasts Anthropic’s momentum with OpenAI’s reported decision to shut down Sora and questions whether OpenAI’s advertising strategy is working. O’Driscoll calls the contrast overstated: abandoning a product can be a sensible correction, and he thinks it is too early to judge advertising. Lemkin sees the Sora decision as evidence of a substantial strategic reversal.

Lemkin cites reports suggesting Sora had low revenue relative to its computing demands, but his own larger cost estimate is conjecture. O’Driscoll argues that scarce compute should be directed toward uses with stronger economics, such as coding. Both treat the reported shutdown as a choice to focus resources, rather than proof that the wider company has failed.

On consumer advertising, O’Driscoll offers rough estimates of paid conversion and revenues and argues that OpenAI needs either much higher consumer conversion or a large advertising business. His figures and suggested future advertising targets are his own assumptions, not verified guidance. He also believes OpenAI needs a competitive coding and enterprise offering.

The panel discusses reporting about past leadership tensions at OpenAI. Lemkin says managing conflicts among talented executives can consume a CEO’s attention, while O’Driscoll regards persistent senior turnover as a warning sign a board should investigate. Their impressions of internal dynamics are based on outside reporting and observation, not access to the company’s boardroom.

Stebbings floats a hypothetical acquisition of Sierra and a different division of leadership responsibilities at OpenAI. O’Driscoll responds in general terms that a board should ask questions when senior turnover is persistent. The suggestion is the host’s proposal, not a reported transaction or an inside account of board deliberations.

Stebbings raises a reported $40 billion SoftBank bridge loan to purchase OpenAI stock. O’Driscoll estimates substantial leverage at the holding-company level and illustrates how a sufficiently large decline in assets could erase equity. His leverage figures and simplified downside calculation are not independently verified, and the discussion is not an assessment of actual loan terms.

Lemkin notes that real estate funds routinely use debt and wonders whether venture investors would use more if it were available. O’Driscoll replies that real estate debt is supported by relatively predictable cash flows, while he considers SoftBank’s investment portfolio more exposed to volatile technology assets. Both frame the leverage as aggressive, with amplified gains and losses.

Stebbings asks whether reported declines in cybersecurity stocks following the alleged Anthropic leak were justified. O’Driscoll differentiates finding flaws during code development from real-time perimeter defense and identity authentication. In his view, a new code-review capability could threaten some vendors, but a broad sell-off would ignore important differences among security products.

Lemkin sees an opposing force: more applications and more autonomous agents may produce more threats and create demand for security products. He discloses that he is considering another security investment and says he is not a cybersecurity expert. His view that the sector could benefit is an investment thesis, not a prediction of any company’s performance.

O’Driscoll agrees that agents operating inside organizations could create new risks, but says the right defenses are not yet clear. He expects established security companies to acquire emerging solutions before a winner is certain. He also interprets the reported early release of a model to security officers as marketing; the panel does not establish that release plan.

Lemkin argues that security companies with established relationships may find it easier to get meetings when they offer products for emerging agent risks. O’Driscoll thinks both incumbents and newer firms could participate, though he acknowledges that specific solutions remain unsettled. Neither offers evidence that every existing security company will benefit.

Stebbings turns to the difficulty of interpreting AI revenue. Lemkin describes an unnamed portfolio company that sends him three different revenue measures, with annual recurring revenue the smallest, and says he cannot reconcile them. The example illustrates his concern that impressive charts can obscure what each number actually measures.

O’Driscoll describes what he understands as a trailing four-week revenue run rate for Anthropic and OpenAI, annualized over thirteen four-week periods. He estimates roughly $19 billion and $25 billion respectively. These are his reported approximations, not audited recurring revenue or a guarantee that the previous four weeks will repeat.

O’Driscoll also describes an alleged accounting difference in partner channels: OpenAI reporting its share net of a partner’s portion and Anthropic reporting gross sales with the partner’s share as a cost. He presents a hypothetical $100 sale with $20 paid to a partner. The panel does not inspect either company’s statements or verify that these treatments apply.

Lemkin asks how many times the same underlying model tokens are counted as revenue as intermediaries resell them. The participants joke that a chain of resellers could each show a large revenue figure while earning little gross margin. O’Driscoll’s point is that revenue throughout a supply chain cannot simply be added up as end-customer spending.

The conversation turns to Emergent’s public claim of reaching $100 million in eight months. Lemkin says scrutiny is fair when a company prominently promotes a number, but explicitly says he does not know what happened or whether alleged double counting occurred. He is discussing questions about the claim, not presenting verified findings about its accounting.

Lemkin dislikes free trials that convert to paid subscriptions and considers it potentially misleading to count the future annualized fee before payment. He speculates that some startups may do so, while saying he is not asserting that Emergent used that method. In his own informal tests, he found Emergent’s product strong; that impression does not verify its financial claims.

In his own trial of Emergent’s app, Lemkin says it performed well on his test of rebuilding a home page. He contrasts this positive product experience with discomfort about trial billing. He cautions that similar practices may occur elsewhere in the industry; the product test itself says nothing conclusive about reported ARR.

Stebbings criticizes funding announcements that combine investments made at different valuations into one round and advertise only the higher headline. He believes some prominent investors get access at lower prices than later participants. O’Driscoll sets out a hypothetical structure to explain the economics rather than establishing the terms of any named deal.

In O’Driscoll’s example, an investor seeking an average basis of $600 million might buy some shares at a $250 million pre-money valuation and others at $1 billion. The company gets a $1 billion headline, while the investor gets the blended price it wanted. He warns the headline can make the next financing appear to be a down round if progress falls short.

The example becomes more contentious when a lead investor gets access to both prices while followers can buy only at the higher price. O’Driscoll says a follower must decide whether the premium is justified by access or a different view of the company’s value. The illustration does not imply that any actual investor or transaction used those exact numbers.

O’Driscoll says a follower paying the higher price might believe the company is worth more than the blended valuation paid by a lead, or might accept a lower expected return to gain access. He treats this as a trade-off the investor must evaluate, while criticizing the optics of calling the higher price the price of the entire round.

Lemkin thinks some founders have turned headline valuations into a game, moving from one milestone to the next without fully considering the consequences. A large valuation remains a private-market bet, he says, and he is less impressed by it when the associated ARR is uncertain. Strong products and durable customer economics matter more than a fundraising headline.

Stebbings raises the possibility of an Oura IPO and says Whoop reportedly raised $500 million at a $10 billion valuation. These are reported or prospective developments in the discussion, not independently confirmed transactions. The investors see growing consumer interest in health data but debate how durable wearable subscriptions will prove.

Lemkin says customers can love a wearable yet switch when another device offers a better experience. He compares the risk with the change in demand for Peloton and questions whether consumer subscriptions deserve the same assumptions as long enterprise software contracts. He sees a large market, but is unsure what valuation multiple the recurring revenue should receive.

O’Driscoll discloses that his firm has a small Oura investment through the acquisition of a portfolio company and declines to discuss Oura’s numbers. He argues that many valuable consumer businesses must win customers repeatedly without long-term lock-in. In his view, the relevant challenge is continuing to deliver products people prefer.

O’Driscoll offers an alternative interpretation of Peloton: pandemic demand may have arrived in a sudden surge, encouraging the company to expand before demand subsequently fell. Lemkin returns to his investment concern that intense competition can constrain profits, even in a large market. Their explanations are interpretations rather than a verified analysis of Peloton’s results.

Lemkin’s broader venture question is whether a market without monopoly-like protection can deliver the returns investors seek. O’Driscoll responds that large markets and fast growth can create significant value even when customers can leave. The disagreement reflects different preferences about investment risk, not a conclusion about either wearable company’s future.

The investors agree a monopoly would be attractive, but O’Driscoll says a competitive consumer business can still create substantial value. He points to market saturation as another risk for hardware products. Lemkin prefers the predictability of long-term enterprise contracts but acknowledges the accomplishment of building a successful consumer hardware business.

Comparing consumer hardware with enterprise software, Lemkin says he would prefer a long contract with high gross margins, all else equal. But he also asks whether a market that has repriced software would value a fast-growing consumer subscription more highly. Their comparison is a framework for analysis rather than a recommendation to buy or sell a security.

Lemkin questions whether investors still value revenue durability as they once did, noting that software valuations can change. O’Driscoll cautions founders against reshaping a business solely to satisfy a favored investor metric. Sell in the way customers want to buy, he says, and be honest about the volatility and ongoing product development that the model requires.

The group cites a reported Allbirds sale as another example of changing consumer-company fortunes. O’Driscoll avoids predicting Oura’s performance and argues that building a wearable with sensors involves different technical challenges from making a shoe. His comparison concerns defensibility, not an assessment of the clinical reliability of any device.

Stebbings brings up reported layoffs at Epic Games and asks about labor-market effects. The panel contrasts a previous guest’s claim that technology companies had overhired with Epic’s stated business circumstances as they understand them. O’Driscoll says falling demand and the need to reduce costs can explain a layoff without attributing it to AI.

O’Driscoll stresses the human cost: people who lose jobs must find new ones in a difficult market. Lemkin suggests entertainment employment faces changes from social media, production economics and new technology, while O’Driscoll emphasizes competition for attention. Their discussion does not establish a precise causal breakdown for the reported Epic layoffs.

The panel contrasts game demand with broader shifts in entertainment. Lemkin sees the industry adapting to social platforms and technology, while O’Driscoll stresses the fickleness of consumer attention and competition among games. Their perspectives leave the future of entertainment jobs uncertain and do not imply all layoffs were caused by the same factor.

Stebbings raises reports that two Manus founders faced restrictions in China after Meta acquired the company. O’Driscoll describes his understanding of the company’s move from China to Singapore and of the transaction’s status, while acknowledging uncertainty about the latter. The conversation does not independently verify the founders’ circumstances or the reasons for any restrictions.

The investors speculate that reports like these could make founders and investors more cautious about cross-border deals and the personal risks involved. Lemkin says investors can rationalize unfamiliar risks when growth appears exceptional; O’Driscoll notes that consequences can extend beyond a financial loss. These are opinions about possible future behavior, not established outcomes.

Asked what Meta might do, the panel considers the importance of the acquired team alongside the product. Lemkin says he has limited information and believes the application appeared to be operating in the short term; O’Driscoll hopes for an amicable resolution if the company can help. Neither claims to know the founders’ current situation or Meta’s response.

O’Driscoll points to political tensions between the United States and China as part of the context for cross-border deals. The panel also discusses the difficulty of valuing an acquisition when founders are essential to a product’s development. Their comments do not substantiate the alleged circumstances affecting Manus personnel or establish the intentions of either government.

The panel discusses an investor’s reported move from California and whether high taxes encourage wealthy residents to relocate. O’Driscoll says mobile taxpayers may weigh large capital gains against a potential state tax, while Lemkin worries that departures could reduce future revenue. They present a policy argument and predictions, not evidence that a particular tax would necessarily reduce receipts.

Lemkin cites a Washington tax development and estimates of potential Californian tax effects, but the discussion does not establish the policies’ final terms or verify their revenue forecasts. O’Driscoll argues that a policy intended to raise money could harm public services if it instead drove away enough taxable income. That outcome depends on contested behavioral and fiscal assumptions.

O’Driscoll argues that if a tax raises less revenue than expected, the loss may affect social services. His example assumes both a substantial taxpayer response and a particular budget response; neither is demonstrated in the conversation. Lemkin similarly frames relocation as a concern, while acknowledging the separate question of how much wealthy people ought to pay.

O’Driscoll says he would not personally move simply to avoid recurring income tax because his life and friends are in California. He then uses a hypothetical $2 billion realized gain and a 13% difference to illustrate why someone might consider relocation: the arithmetic yields $260 million. The example is illustrative, not personal tax advice or a claim about any individual’s actual gain.

Lemkin recounts differing public recollections by investor Ron Conway and Cloudflare CEO Matthew Prince about past help. He sees the episode as a reminder that investors may remember their contribution more vividly than founders do. The panel does not establish the underlying history, and O’Driscoll suggests both parties’ perspectives can be understandable.

O’Driscoll observes that founders juggle many problems, so a single helpful introduction can fade into the background even when an investor remembers it clearly. He says investors’ most consequential roles tend to be providing capital, participating in CEO decisions and agreeing strategic direction. Other assistance can be valuable without making the investor the central figure in the company’s story.

O’Driscoll says business biographies often give venture investors only a small role compared with founders and employees. In his view, a decisive investment or board decision can matter enormously, but many calls and introductions are only assists. Lemkin adds that even emergency funding is not always remembered in the same way by both sides.

The group closes by saying that founders are most likely to remember investors who supported the company when financing was difficult, although Lemkin jokes that even that is sometimes forgotten. O’Driscoll describes venture investors as well-paid supporting players in a story driven primarily by the people building the company.