20VC Roundtable on Open Models, AI Economics and Payments Deals

Harry Stebbings asks investors Jason Lemkin and Rory O’Driscoll about recent Chinese open-weight model releases, including Kimi and Alibaba's Qwen. Lemkin cautions that a benchmark is only a benchmark: he wants to see performance in real use before drawing broad conclusions.

Lemkin says Chinese-developed models were already important in OpenRouter traffic and distinguishes model origin from where a model is hosted. His usage figure describes his understanding of one platform, not a complete count of global inference.

O’Driscoll sees continued progress by well-funded teams rather than a sudden change in direction. He says models differ greatly in size and compute needs, so comparisons of price and capability depend on the model and task. His assessment of the gap with frontier models is a dated judgment, not a universal ranking.

The speakers discuss a public comment by an OpenAI policy executive and the reaction to it. O’Driscoll argues that policy proposals favoring restrictions will be scrutinized when a person's employer could benefit commercially; this is his view of incentives, not proof of anyone's private intent.

Lemkin expects some US government and regulated customers to be cautious about Chinese-developed models, while noting uncertainty about how extensive restrictions might become. He also cites an example of increasing token use in a regulated company and expects pressure for lower-cost alternatives.

O’Driscoll argues against a blanket ban and distinguishes sensitive government use from lower-risk commercial workloads. He refers to a reported policy discussion but does not establish a settled government rule; his conclusion is a policy preference.

Lemkin says he remains concerned about potential data leakage and would not accept a simple assurance that local hosting eliminates every risk. His concern is not evidence of a backdoor or misconduct by any named model provider.

O’Driscoll distinguishes open weights from full access to training data and asks what technical controls can limit an application's access to information. He thinks some risks can be managed but says political and organizational trust are separate questions. Neither speaker demonstrates a specific compromise.

The panel asks whether a low-cost American open-weight model can sustain a business. O’Driscoll outlines possible demand for cheaper intelligence and discusses Thinking Machines as an emerging effort, while emphasizing uncertainty about both costs and the business model.

Stebbings says he received an investment offer involving Kimi and cites a valuation in that context. The anecdote is not an independently verified company valuation or a recommendation. The guests continue to ask why more US labs have not pursued the same segment.

The discussion turns to a report that OpenRouter is in acquisition talks and to competing routing products. Lemkin sees a possible window to sell as more platforms add similar functions, while O’Driscoll considers why a cloud provider might value a model-agnostic gateway.

O’Driscoll says strategic value to an acquirer may differ from a company's standalone value, particularly when a product helps a buyer win customers. Both participants discuss a hypothetical sale; the interview does not confirm a buyer, offer, leak or transaction.

Asked what a board should do with a substantial offer, O’Driscoll says founders should take the opportunity seriously and weigh their goals and confidential concerns. He argues that investors should support an informed founder decision, rather than treating a multiple as the only consideration.

Lemkin contrasts an investor's calculation with a founder's personal risk: a life-changing cash offer may be worth more than the uncertain prospect of doubling it after years of work. His hypothetical ownership and payout figures illustrate his view, not OpenRouter's actual ownership or an actual offer.

O’Driscoll agrees that the financial comparison is risk-adjusted but adds that some founders may prefer to continue building their life's work. The right decision, in his view, depends on personal goals as well as money.

Lemkin worries that routing functionality could become a feature included by larger platforms. O’Driscoll resists treating every product as an identical commodity, but agrees that inclusion by other vendors can change competitive dynamics.

Stebbings discusses Fireworks AI's reported funding, valuation, revenue run rate and token volume and discloses 20VC's participation in the round. These figures and forecasts are presented by the host and have not been independently audited in the interview.

O’Driscoll links demand for independent inference providers to organizations wanting to run different open-weight models. He says strong demand can improve near-term economics, while provider concentration, compute costs and competition remain important questions.

Stebbings relays Fireworks management's description of margins in the mid-30% range and an ambition to move further into data-center infrastructure. O’Driscoll notes that owning more of the stack could offer control but also require considerable capital; neither future margins nor returns are guaranteed.

The panel contrasts the large sums flowing into AI infrastructure with the smaller revenue base it observes in applications. The market totals and company figures cited are rough speaker estimates, not a reconciled accounting of AI spending.

O’Driscoll divides the ecosystem into infrastructure, foundation models and applications, and asks when application revenue will support the infrastructure being built. Lemkin says he is still waiting for a broader application-layer boom, though both acknowledge individual growing products.

Lemkin describes building an agentic recruiting application and labeling examples to improve its results. He says this helped a particular workflow and distinguishes his own rules and reasoning layer from a separately trained foundation model.

He believes companies will increasingly customize intelligence with domain-specific data. His enthusiastic claims about the improvement from a small number of expert examples reflect his experience and expectations, not a controlled study or evidence applicable to every use case.

Stebbings discloses an investment in Mercor while asking about data-labeling businesses. O’Driscoll says the market outlook depends partly on whether enterprise model development adds demand or instead slows growth at the frontier labs that buy much of the data.

O’Driscoll says he does not know how much open models will affect frontier lab growth over the next one to two years. He outlines a possible reallocation of compute demand and financial commitments if growth slows; this is a scenario, not a forecast of insolvency or a known market outcome.

Lemkin argues that price competition may pressure frontier providers, but they also need to recover the costs of training and serving models. O’Driscoll adds that rapid growth can mask difficult unit economics, whereas falling prices and margins would change the calculation.

The panel contrasts inference's physical serving costs with traditional software bundling. Their figures for model pricing, growth and margins are illustrative estimates, and they acknowledge that competitive outcomes are uncertain.

Stebbings raises the reported Stripe and Advent proposal involving PayPal. O’Driscoll compares the companies' payments volumes and discusses potential strategic benefits and operational challenges. The speakers are discussing an offer and possible transaction, not a completed acquisition.

He cautions that Stripe and PayPal report revenue differently, so gross and net figures cannot simply be compared. The numbers he gives are approximate and not independently verified in the conversation; integration and financing would require detailed analysis.

Lemkin worries that combining a faster-growing business with a slower-growing one could reduce the combined growth rate. O’Driscoll stresses the potential scale benefit but recognizes the operational work and financing complexity of any deal.

The guests speculate about how PayPal's board might evaluate an offer, future performance and alternatives. Their comments on directors' obligations and potential negotiations are general opinions, not legal advice or evidence of board deliberations.

Both say they think a deal could happen, and Lemkin speculates about a higher price and negotiation process. The interview does not establish the bidders' strategy, the board's decision, a final price or completion; their timing and outcome predictions remain uncertain.

O’Driscoll notes how heavily AI companies influence the value of the private technology market and sees related demand in semiconductors and energy. He mentions Databricks among large private companies, but reported financing and valuations remain claims in the discussion.

The panel compares early-stage and later-stage AI investment pricing. Stebbings describes examples where a later company had far more revenue for a valuation only several times higher; he frames the choice as a risk-adjusted question rather than a guaranteed return.

Lemkin says earlier investors can own more of a winner but must be able to identify it under intense competition. O’Driscoll says growth investing has benefited from recent valuation increases while warning that favorable conditions need not last.

They discuss funding rounds split into tranches at different prices. Lemkin questions whether charging investors different prices in a short period is worth the relationship and employee-option consequences, though he acknowledges that a large capital raise can change the calculation.

O’Driscoll says founders can calculate the blended valuation but should understand future incentives and deal terms. Both guests describe tradeoffs and possible signaling benefits; their remarks are not a substitute for reviewing the terms of any particular financing.

The conversation moves to nuclear energy and a reported fundraising by Valor Atomics. O’Driscoll sees continued technical activity alongside significant risk and questions why some early nuclear companies are public while much larger mature technology companies remain private.

Lemkin discusses a data-center company's public offering and contrasts its reported growth and valuation with more exuberant AI stories. The participants acknowledge that debt and enterprise value matter; their figures and assessment are not independently verified.

O’Driscoll compares long-term relationships among semiconductor suppliers with what he describes as more volatile memory pricing. These are his interpretations of supply-chain behavior and company economics, not verified evidence of any supplier's private intentions.

He suggests that demand, investment and memory costs can affect infrastructure businesses, and describes several possible paths for Nvidia depending on future capital spending. Stebbings notes his own Nvidia position; neither speaker's scenario is a guaranteed share-price outcome.

The guests return to the uncertainty surrounding future frontier-model growth and its effects across the technology supply chain. They close with a light exchange, having distinguished current demand from the many financial and policy outcomes that remain unknown.