20VC Roundtable: Open Models, AI Spending and the Cost of Competition

Harry Stebbings, Jason Lemkin and Rory O'Driscoll discuss AI talent, open models, infrastructure spending and the outlook for software businesses. The speakers' reported transactions, company figures and forecasts are their commentary at recording, not independently verified results or investment advice.

Stebbings opens with reports that researchers Noam Shazeer and John Jumper were leaving Google DeepMind for Anthropic. Lemkin suggests top researchers value the freedom to pursue their own work; O'Driscoll adds that the ability to ship products may matter too. Neither establishes the individuals' reasons for moving.

O'Driscoll distinguishes Shazeer's work in language models and products from Jumper's scientific research, including AlphaFold. He suggests a newer lab may offer fewer legacy constraints and attractive resources, but this is his interpretation of the reported moves.

Stebbings relays a rumor that an undisclosed Anthropic breakthrough motivated Jumper. O'Driscoll is skeptical and suggests scientific initiatives may be long-term research projects. The rumor is unverified; the conversation supplies no evidence of a specific breakthrough or its role in the move.

Lemkin shares a family anecdote about a researcher prioritizing the learning environment over pay. He says it illustrates why research freedom can be attractive, while acknowledging that one person's experience cannot establish the motives of other researchers.

The panel discusses how rapidly growing labs may offer both research opportunities and product momentum. They also praise Google's recent position in AI, while arguing about its relative strength in specific products; their rankings are opinions, not measured league tables.

Lemkin argues that a third closed-model provider could be squeezed between leading premium models and cheaper open alternatives. He says routing lets customers select models for different workloads; O'Driscoll agrees that lower-cost competitors can put pressure on margins across a concentrated market.

Lemkin says inference and training with open models still cost money, even where weights are available. He claims Chinese public support subsidizes open models, but offers no evidence for a blanket assertion about funding of all Chinese models; the degree and source of support vary and are not established in the discussion.

Lemkin describes an Anthropic message encouraging prompt caching and interprets it as a competitive response to open models. He says discounts may make cached workloads cheaper in some cases; the provider's intent and comparative costs are his inference, not verified across workloads.

Stebbings asks about sovereign models. O'Driscoll suggests a region might accept higher costs or a different model choice for political or security reasons. That is a proposed trade-off, not proof that every regional model is inferior or more expensive.

Lemkin describes his recent experience accessing AI services during a trip to China and his impressions of local model features. His account is anecdotal; the speakers do not verify provider access policies, training data, model behavior or the causes of any restrictions.

The panel discusses a reported DeepSeek financing, with Stebbings citing $7.4 billion at about a $50 billion valuation and suggesting unusual voting rights. Lemkin refers elsewhere to a different valuation. The financing, investor terms and governance claims are unverified in the interview, and the speakers' numbers are not consistent.

The speakers speculate about government influence over Chinese AI companies and compare sovereign concerns with US policy debates. Their broad claims about ownership, subsidies and control are opinions without supporting transaction documents; they do not establish a particular government's role in a specific company.

Stebbings cites a claimed coding benchmark for a Chinese open model. O'Driscoll sees open models as a competitive alternative that can constrain pricing. The benchmark and market comparison are not independently tested in the discussion, and performance depends on the task and evaluation.

The conversation shifts to memory and infrastructure costs. Stebbings cites reported increases in DRAM prices; O'Driscoll argues that AI investment competes for hardware, electricity and other resources and could raise costs elsewhere. His examples and predictions are not a causal accounting of individual prices.

Stebbings references a Goldman Sachs estimate of cumulative AI capital expenditure and asks who will ultimately pay for it. The panel uses rounded figures to explore the gap between present spending and future revenue; their arithmetic is illustrative and should not be treated as audited sector totals.

O'Driscoll worries that enthusiasm can sustain investment even when the timing of returns is uncertain. Lemkin sees potentially enormous demand if AI can be delivered cheaply and says venture investors must participate in demand cycles while considering liquidity and risk.

O'Driscoll distinguishes desire for intelligence from willingness to pay: demand may be very large at a low price, but enterprises need to decide how much to spend and where. He asks whether improved token efficiency could slow revenue growth for model providers; the panel does not know the answer.

Lemkin describes different economics for free, subscription and enterprise model users and suggests some heavy users may be subsidized. His margin estimates and interpretation of provider economics are not based on disclosed customer-level financial statements.

He expects enterprises to move from encouraging broad experimentation toward asking for measurable ROI. O'Driscoll agrees that organizations may accept an initial learning phase but will eventually need to connect spending to useful output.

O'Driscoll offers a back-of-the-envelope calculation linking large AI expenditure to productivity and potential labor disruption. Lemkin adds that if every competitor adopts the same tools, an individual company may struggle to show improved profitability. These are scenarios built on assumptions, not forecasts of a fixed number of job losses.

O'Driscoll notes that industry productivity can rise without every participant's profitability rising, and that technologies can expand demand as well as displace tasks. Both think firms may adopt tools defensively if competitors do so, but actual outcomes remain uncertain.

Lemkin describes an AI finance agent his team built to help with quoting, contracts, invoicing and bookkeeping. He says it may reduce some contractor hours but explicitly says it is not replacing four employees; this is his account of an internal experiment, not an independently validated automation study.

Stebbings relays a prediction that training agents will become a major job category. Lemkin agrees that working effectively with agents matters but recalls earlier enthusiasm for prompt engineers and expects required skills to change quickly. The size of any future job category remains unknown.

Lemkin also recounts an agent admitting it had not fully read a contract. He describes adjusting instructions and recognizing that the system might still fail, illustrating why human oversight and an understanding of limitations matter in his example.

A quoted investor argues that low gross margins are blocking some startup financings. O'Driscoll counters that hypergrowth companies with weak early margins have historically raised money and sometimes improved economics later; he allows that investor standards may now be changing.

Lemkin says investors may be left with companies that spent heavily on inference without achieving either exceptional growth or a credible path to healthy margins. The panel agrees the strategy can work in some cases but is risky when growth slows or costs remain high.

Stebbings asks why Menlo Ventures raised $3 billion rather than more after its Anthropic investment. Lemkin first speculates about demand and then suggests additional capital could come through special-purpose vehicles; he says he does not know the firm's actual reasons or future fundraising.

O'Driscoll argues that a moderately sized core fund plus optional SPVs can help a firm invest across cycles without committing to deploy an enormous fund. He notes that this requires reliable access to additional capital when opportunities appear, a condition Lemkin reinforces with his own experience.

The panel briefly discusses Kalshi, citing a reported $2 billion revenue run rate and possible IPO. They discuss sports-related contracts and potential competition from Meta; the revenue figure, future listing and hypothetical Meta product are not independently verified or confirmed here.

Lemkin thinks social distribution might help a possible Meta prediction product; O'Driscoll questions whether its audience would match the market. They disagree about the opportunity and acknowledge uncertainty rather than presenting a launch as fact.

O'Driscoll distinguishes an exchange matching buyers and sellers from a traditional bookmaker and notes that litigation and regulation can affect the business. His comments are a high-level account, not legal advice or a ruling on Kalshi's products.

The speakers discuss a reported decline in Accenture's share price. O'Driscoll suggests demand for AI implementation could grow while traditional consulting and systems integration face pricing pressure. These are his hypotheses about a complex business, not verified explanations for the stock move.

He describes potential AI-assisted systems-integration startups and speculates that some implementation work could be delivered for less. Lemkin recalls a costly Salesforce implementation at Adobe; the amounts and hypothetical savings are their estimates, not disclosed contract terms.

Lemkin says AI-assisted data migration could weaken vendor lock-in, citing a migration experience and reported claims by data-platform providers. O'Driscoll argues that a labor-based consulting model may be difficult to adapt if fewer billable people are needed; actual effects will vary by project.

Stebbings introduces a provocative comment that working from home is 'white-collar fraud.' Lemkin says some remote work he observed involved limited hours but acknowledges others make it work. His experience and criticism do not establish that remote workers generally commit fraud or work fewer hours.

Lemkin says he prefers investing in small, highly paid teams working long hours in person. O'Driscoll asks whether the concern is effort or effectiveness. Lemkin's preference is an investment and management opinion, not evidence that a particular schedule guarantees success.

The panel turns to a reported OpenAI custom inference chip, described as co-developed with Broadcom, and relays company performance and cost claims. These specifications and savings are not independently tested in the interview, and the discussion does not establish the chip's deployment at scale.

Lemkin argues lower inference costs could help closed-model providers compete for mid-priced workloads against open alternatives. O'Driscoll agrees cost matters but questions whether model labs should build chips themselves when cloud and hardware suppliers already compete to serve them.

O'Driscoll argues external suppliers currently bear much of the infrastructure investment and wonders whether backward integration distracts a lab from winning customers. Lemkin replies that a chip decision made earlier may reflect different market conditions and that both the economics and competitive landscape could change.

They return to the 'middle' of the market: Lemkin believes customers need models that are more capable than the cheapest options without frontier-level prices. O'Driscoll agrees about demand for affordable quality but thinks competition among providers could be another way to achieve it.

The panel closes without resolving whether custom chips or competitive sourcing will best reduce inference costs. Their shared question is how AI providers can serve a wide range of workloads at prices customers can sustain.