Apple, Google and the Race to Fund AI Growth

In this July 2025 roundtable, Harry Stebbings, Jason Lemkin and Rory O’Driscoll discuss venture partnerships, AI coding tools, competition among large technology companies and Figma’s planned IPO. Their financial figures, personnel reports and predictions reflect the conversation at that time and are not independently verified here.

Stebbings raises a reported partner departure from Benchmark and asks what it means for venture firms. Lemkin says mobility has increased and that successful investors may prefer to start their own funds. The guests discuss industry incentives, not the private reasons behind any individual departure.

O’Driscoll praises Benchmark’s established partnership model but sees a broader change: a talented investor may now be able to raise capital independently rather than stay at a leading firm. He expects Benchmark to adapt, which is his forecast, not a statement of its hiring plans.

Lemkin says he might have stayed at a prestigious firm because its brand can help win access to founders. He and Stebbings each run their own branded investment businesses, making their experiences relevant but not universal. Independence offers autonomy and economics while requiring a manager to build and maintain a reputation.

Stebbings says he would value the freedom to invest across stages and retain economics in an independent fund. He offers his interpretation of the investor’s choices without claiming to know the complete reasons for the departure. Lemkin counters that the value of an established institutional brand is easy to underestimate.

The guests discuss Elad Gil and other investor-led platforms. O’Driscoll says limited partners have sometimes relaxed traditional preferences for small, focused teams when an individual has a strong record and access. He warns that a successful recent record does not remove the need to test a model through a downturn.

O’Driscoll describes the tradeoff between changing practices in response to new evidence and abandoning rules that may have protected investors across cycles. He sees value in both perspectives and says the durability of newer fund structures remains to be tested.

Stebbings worries that deploying money quickly can create visibility and access, while a slower fund may appear less relevant. O’Driscoll says fast deployment can produce both strong opportunities and short-term paper markups, but the outcome depends on the quality of the investments and eventual realized returns.

O’Driscoll contrasts investors with durable records with firms that gained visibility by deploying rapidly during a hot market. He argues that deal volume alone is not proof of successful selection. His examples and assessments are investment opinions, not verified fund-level performance data.

The panel compares early-stage partnerships with firms led by a dominant investor. O’Driscoll says early-stage work often requires several partners capable of supporting companies, while a concentrated later-stage strategy can involve fewer, larger decisions. He presents this as a framework, not a rule that every firm follows.

Asked whether large investor-led funds make him question Scale’s model, O’Driscoll says his firm’s approach reflects experience through downturns. He sees correlated valuation risk in late-stage portfolios. Lemkin jokes about liquidation preferences; O’Driscoll explicitly notes that they are not a guarantee against losses.

Stebbings argues that some criticized large portfolios may still contain excellent companies. O’Driscoll distinguishes a fund’s overall record from the experience of limited partners who entered at different times. Lemkin adds that broad investing in a buoyant market can include both strong and weak choices.

O’Driscoll observes that some successful multi-stage firms earned access through early investments and later deployed more capital into established winners. He believes early relationships can be an advantage, although follow-on investments still require judgment about price and risk.

Stebbings cites reports of a possible Anthropic fundraising valuation rising from around $100 billion toward a higher range. O’Driscoll discusses reported revenue acceleration and says the math is compelling if the figures are correct. The round, valuation and private-company figures are treated as reported or estimated, not independently confirmed.

Lemkin describes his own heavy use of AI coding tools and says his spending could reach thousands of dollars a month. He predicts that some technology companies may eventually allocate far more per developer than conventional software budgets. This is his extrapolation from personal experience, not evidence of a standard industry budget.

The guests discuss how longer tasks and parallel agents could increase token consumption. O’Driscoll asks whether customers will continue to accept high prices and how tool providers will manage inference costs. Falling unit prices and rising usage could pull total spending in opposite directions.

Lemkin sees scope for a much larger developer-tools market as users run more tasks at once. O’Driscoll says a very large company may need applications beyond developers to justify the highest valuations. Both acknowledge that product demand, cost and future market size are uncertain.

The discussion turns to the scarcity of technical talent. The guests say AI startups face difficult hiring even when they have cash. Their comments about individual companies and compensation reflect their experience and estimates, not a comprehensive labor-market survey.

Stebbings asks how Microsoft, Google and Amazon compare in AI infrastructure. The guests discuss distribution, chips, cloud capacity and model partnerships. Their rankings are opinions as of the recording, and O’Driscoll notes that some quarterly results were still to be reported.

The panel asks whether Microsoft missed an opportunity with GitHub Copilot as competitors grew. O’Driscoll argues that an early distribution advantage does not guarantee continued leadership and wonders whether Microsoft has focused enough on the core product. Revenue figures he uses for competitors are estimates rather than verified comparable results.

Stebbings contrasts Google’s own models and distribution with Microsoft’s and Amazon’s partnerships. O’Driscoll favors Google in a hypothetical comparison and says AWS appeared to be growing more slowly in the cloud segment. He qualifies the comparison because cloud accounting and reporting are not identical.

Lemkin says customer demand for compute looked strong, and O’Driscoll agrees that available capacity was a constraint in the information they had seen. Neither establishes that demand is literally unlimited or will remain so. Their hypothetical stock preferences are not recommendations.

O’Driscoll worries that infrastructure spending and depreciation could eventually outpace the revenue companies earn from AI, even while demand grows. His figures are broad estimates with differing scopes. He says he cannot time a slowdown and acknowledges that spending has continued to rise despite earlier concerns.

Stebbings asks whether large incumbents have become too powerful. The guests debate regulation, competition and oligopoly; O’Driscoll prefers a lighter regulatory approach, while Lemkin notes that a few competitors can still compete vigorously. These are policy and market-structure opinions, not legal determinations.

O’Driscoll argues that in the newer model market, startups have challenged established technology companies despite the incumbents’ resources. He views Google as having executed well relative to some peers. The assessment reflects the panel’s 2025 view and does not establish the outcome of the competition.

Asked what Apple should do, O’Driscoll says its board should assess whether management can respond effectively to AI. He discloses that Apple is his largest individual position and credits its leadership with past performance. His criticism and any suggestion of a management change are opinions, not a report that a change is planned.

Lemkin pushes back that Apple’s hardware and App Store could continue to benefit even if it does not lead model development. The financial percentages he cites are his estimates and should not be treated as verified segment disclosure. O’Driscoll agrees the hardware business may be resilient but questions where future growth will come from.

The guests discuss whether Apple could rely on distribution and partnerships while others supply AI. That strategy might protect existing economics, they say, but could leave another provider owning more of the customer relationship. They do not know the parties’ future agreements or whether either scenario will occur.

The panel considers Meta’s reported spending on AI talent. O’Driscoll says a company committing heavily to infrastructure may see expensive specialists as worthwhile, though the marginal value of another hire can fall. The headcount, packages and eventual return are discussed as reports and estimates, not independently verified contracts.

Lemkin argues that a highly profitable incumbent has room to absorb unsuccessful bets. O’Driscoll adds an important condition: the core business must remain strong while the company experiments. Both are discussing a possible financial path, not asserting that any specific expense can simply be ignored.

They say founder control can give some public companies more latitude to invest through periods of lower margins. O’Driscoll contrasts this with management teams under more direct shareholder pressure. The discussion is about governance incentives and does not guarantee immunity from market or business risk.

Stebbings raises Figma’s upcoming IPO and asks whether the market is paying it enough attention. O’Driscoll describes the then-reported filing range and predicts strong demand and a possible first-day increase. The episode was recorded before the result, so its pricing comments are forecasts rather than reported outcomes.

Lemkin recalls how much attention Figma’s earlier proposed Adobe sale received, contrasting it with the current focus on AI applications. O’Driscoll says an investment can still produce a valuable exit even after the technology conversation has moved on. Neither equates an indicated IPO valuation with cash already distributed.

O’Driscoll notes that a decade-long path from startup to IPO may span a technology cycle. The guests discuss how Figma, Canva and AI coding tools might compete or integrate in prototyping. Their product comparisons and future competitive claims are opinions, with no settled winner.

Lemkin says he wants prototypes to behave like working products, not merely static designs. Stebbings distinguishes Canva’s more consumer-oriented audience from professional design use cases. The panel expects competition and partnership to change as products develop.

In a quick-fire round, the guests forecast future revenue for AI coding tools and discuss Lovable. Stebbings has an investment interest in Lovable; Lemkin describes himself as a user of coding tools. The numerical targets and margin estimates are speculative and are not independently verified financial results.

Asked whether OpenAI might exceed an $800 billion valuation by the end of the following year, the guests disagree. They debate product progress, funding demand and market sentiment. These are dated predictions based on reported fundraising, not evidence of a completed financing or future price.

O’Driscoll compares estimated AI infrastructure spending with earlier technology and railway booms. The guests also discuss historical space-program spending but disagree over the figures and units. Those historical comparisons are unverified in the episode and do not establish a spending ceiling.

The panel distinguishes a durable technology opportunity from the possibility of a temporary break in investment. Lemkin speculates that major model companies could seek capital from a range of large investors. O’Driscoll says financing at a private valuation is different from proving the same price in public markets.

The conversation closes with the guests reflecting on how their views changed during the discussion. Their assessments of company management, valuations and future spending are opinions made in July 2025, not verified subsequent outcomes.