20VC Roundtable: AI Spending, Model Competition and Software Acquisitions

Harry Stebbings, Jason Lemkin and Rory O'Driscoll discuss enterprise AI spending, model competition, software acquisitions and venture financing. Their company figures, market views and predictions are commentary from the recording, not verified financial results or investment advice.

Stebbings opens with Brian Armstrong's account that Coinbase cut AI spending by 50% while usage increased, including by using open models. The panel treats this as an example of cost optimization; the discussion does not independently audit Coinbase's figures.

Lemkin says he values the data but is tired of executives presenting AI activity without demonstrating business growth. O'Driscoll disagrees with dismissing the example: he thinks a non-frontier company showing how it managed a rapidly rising bill is especially relevant to ordinary enterprises.

O'Driscoll interprets the reported reduction as a response to a sharp spending increase after coding tools took off. He expects chief financial officers to ask their own teams to examine similar costs, while Lemkin says lower inference spending alone will not necessarily restore a company's revenue growth.

The panel asks what widespread optimization could mean for frontier-model revenue. O'Driscoll explicitly says one company's 50% reduction does not imply that a model provider's total revenue will fall by half; he expects the competitive environment could nevertheless affect growth.

Lemkin says the larger issue is whether increased token spend produces measurable product velocity, revenue or productivity. He cites portfolio companies that want to invest in AI but cannot yet clearly connect additional spending to a corresponding improvement.

O'Driscoll argues the expected benefit varies by business: a software company may hope for faster product development and revenue, while a manufacturer might look mainly for savings. If neither appears, he expects management to question the expenditure.

Lemkin recalls a portfolio company's board asking for a return-on-investment case before approving a large increase in token spending. He describes this as a transition from unconstrained experimentation toward closer financial scrutiny, based on his experience rather than a survey.

The panel separates the number of tokens from revenue: cheaper open models could serve much of the volume while higher-priced frontier models still capture substantial revenue. They speculate that infrastructure spending creates risk if a provider needs exceptionally large future revenue to recover it.

Stebbings asks what Lemkin wants CEOs to show. Lemkin says he wants evidence that AI contributes to revenue or margins, while recognizing that Coinbase is exposed to volatile crypto markets. O'Driscoll agrees that software businesses face pressure to show acceleration but says the financial business should not automatically be expected to gain AI-driven revenue.

Stebbings introduces Anthropic's allegations that Chinese model developers used its outputs for distillation. O'Driscoll describes the allegation as repeated prompting and use of answers for training. The speakers discuss allegations and possible legal theories; their discussion does not establish that a named company committed theft or breached a contract.

O'Driscoll notes the tension between model providers asserting rights over outputs and disputes about training on others' copyrighted works. He distinguishes a possible terms-of-service dispute from other potential legal claims, but the discussion is opinion, not a determination of liability.

The panel speculates that a US provider might seek government restrictions on models allegedly distilled from its outputs. O'Driscoll and Lemkin attribute possible strategic motives to Anthropic; neither provides evidence establishing an agreement, an actual ban or the company's private intentions.

Stebbings suggests restrictions could become likely, while O'Driscoll distinguishes intellectual-property allegations from national-security questions. The panel considers hypothetical tariffs, bans and enterprise procurement concerns; these scenarios are not descriptions of enacted rules.

O'Driscoll argues that running open weights domestically can reduce certain concerns and says he would not support a ban on that basis. His sweeping reassurance about security is his opinion: open weights or local hosting alone do not establish that every security risk is absent.

Asked directly whether they would ban the models, O'Driscoll says no and argues alleged misconduct should be addressed separately from security. Lemkin thinks the US economy and investors are heavily exposed to AI and speculates that this could encourage political support for the sector.

O'Driscoll counters that protecting model prices could make AI more expensive for the rest of the economy. He compares it with a hypothetical restriction on computer clones and argues that competition can spread the benefits of cheaper technology.

Lemkin describes a transition from a concentrated model market toward stronger price competition. The two disagree over whether a concentrated market can support innovation: Lemkin sees a possible short-term benefit from feature competition, while O'Driscoll favors open competition.

Stebbings discloses that he owns Microsoft shares and asks about a recent decline. O'Driscoll argues that the company faces questions about its own AI products and reliance on partners; these are his competitive opinions, not a recommendation to buy or sell the stock.

Lemkin says he has no complete theory of Microsoft's complex business but is concerned by guidance for slower Azure growth. He argues public markets demand strong growth and asks whether cloud demand should accelerate with AI; his interpretation of market expectations is not a forecast.

O'Driscoll contrasts Microsoft's cloud role and investment in OpenAI with Google's own model offerings. The speakers discuss Microsoft's ownership and commercial arrangements in broad terms, without establishing the current contractual details or their future value.

The panel speculates about volatility if Anthropic were to list publicly. No IPO date is established in the conversation, and the discussion of timing and trading behavior is hypothetical.

Stebbings raises a reported Kalshi financing at a $40 billion valuation and cites a $2 billion revenue figure. The panel discusses these as reported or stated figures, not independently verified transaction terms; no IPO is confirmed by the discussion.

Lemkin views prediction markets partly through demand for sports betting, while O'Driscoll questions whether the company can reach a much larger valuation quickly. O'Driscoll sees possible growth in sports or financial products but notes that both market size and regulation are uncertain.

O'Driscoll argues that betting on sports or financial outcomes could reach a broader audience than election predictions. The panel's discussion is a view of potential markets, not a statement that the products are equivalent or that any particular activity is legally classified as gambling.

Stebbings discloses an investment in FOMO and describes a product for taking positions on stock-price direction. His enthusiasm reflects that interest, and the panel's discussion of potential market expansion is speculative.

The speakers ask whether SpaceX volatility could affect other AI-related IPO plans. They think market conditions might influence timing but do not know whether any company will delay or proceed. Their discussion reflects expectations at recording, not confirmed listing plans.

O'Driscoll says Bending Spoons is expected to list around the time of the recording, citing its portfolio of acquired consumer software. He and Stebbings discuss a possible $20 billion valuation and revenue multiples as estimates; the conversation does not verify the eventual offering or price.

O'Driscoll questions the premium for a collection of mature consumer products, describing the company's strategy as price increases and cost optimization. Lemkin is more optimistic: he says a long list of potential acquisitions could support growth if execution continues. Both assessments are conditional.

Lemkin predicts a B2B software consolidator could buy mature products with sticky customers and improve retention and growth. O'Driscoll agrees there may be acquisition opportunities, but both acknowledge that the thesis depends on finding suitable assets and executing well.

Lemkin criticizes his experience as a customer of Marketo, citing reliability, service and price concerns, and imagines a buyer improving the product. These are his personal allegations and hypothetical proposals, not independently verified claims about the company or its revenue.

O'Driscoll suggests examples in search optimization and incident management where an acquirer could add AI-enabled capabilities. His discussion of specific transactions, market shares and valuations is presented as his understanding, not confirmed deal data.

Lemkin argues that acquired software companies can become complacent and need accountable managers who act quickly. He bases that view on his experience with portfolio companies and customers; it is not an assessment of every private-equity-backed business.

O'Driscoll adds an important distinction: a B2B buyer may need to develop new AI-enabled revenue, not merely cut costs or increase prices. That could make the operating task harder than optimizing a mature consumer product.

They debate acquisition prices and financing for public software companies. The proposed bids are examples, not offers, and their comments about board duties are informal opinions rather than legal conclusions about any particular transaction.

The speakers mention PagerDuty and Asana as possible hypothetical targets, while acknowledging product and competitive challenges. No acquisition is proposed or announced by the companies in this discussion.

Stebbings says Chamath Palihapitiya raised $135 million for AI software company 8090 and is its CEO. O'Driscoll welcomes the decision to build in a competitive market. The funding and company description are presented as the host's account, not independently verified here.

Lemkin says he is skeptical of wealthy investors running startups while pursuing many other activities, based on unsuccessful investments he has made. He explicitly allows exceptions; his doubts about Palihapitiya's time commitment are speculation, not an established fact.

The panel turns to Stebbings' deleted post about declining to invest in a company projected to grow annual recurring revenue from $1.5 million to $5 million. He says he deleted it because engagement was low, and asks whether his assessment was wrong.

O'Driscoll says such growth can be harder to finance in a market rewarding extreme AI growth, though it is not impossible. He distinguishes a judgment about current fundraising conditions from a judgment about the founder's worth and urges investors to communicate with care.

Lemkin says founders benefit from honest feedback before a fundraising process, rather than encouragement from investors who already know they would not invest. O'Driscoll agrees that boards should discuss options such as profitability or a smaller raise instead of sending founders into an avoidable surprise.

Lemkin argues that $1.5 million to $5 million in revenue can still lead to a significant company, especially with efficient spending, even if investors may be difficult to reach. His examples of fast-growing portfolio companies and their revenue are his own claims, not audited comparisons.

O'Driscoll agrees that slower early growth does not make a large outcome impossible and cites Procore as an example. Founders should understand the current funding environment and plan accordingly, he says; the panel's views are not a universal fundraising rule.

The final topic is a proposed Claude integration in Slack. Lemkin says he tried but could not access it under his plan, so he has not evaluated it firsthand. He outlines a hypothetical autonomous cross-application agent that could challenge traditional software, while emphasizing there is not yet evidence for that outcome.

O'Driscoll considers Slack an interesting entry point because conversations contain context about how people actually work. He agrees that capturing that context could aid automation, but calls it an entry point rather than proof of a threat to Salesforce or other applications.

Lemkin speculates that some application opportunities might be small relative to Anthropic's broader ambitions and therefore not receive sustained attention. His revenue and materiality comparisons are illustrative opinions, not confirmed company plans or financial guidance.

The speakers close by observing that very large companies can affect smaller ones without necessarily focusing on them. Their conversation returns to a recurring uncertainty: whether AI adoption creates measurable value and how competition changes the economics.