Gokul Rajaram on Software Moats, AI Products and Venture Returns

Harry Stebbings asks Gokul Rajaram how operating at Google, Facebook, Square and DoorDash shaped his investing. Rajaram says Google taught him to look for a remarkable core product: distribution cannot rescue a weak value proposition indefinitely, in his view. He recalls Gmail’s unusually large initial storage allowance as an example of a product that surprised users.

At Facebook, Rajaram says, he learned the power of distribution and products that become more useful when people use them together. He cites Figma as an example of collaborative sharing that can aid adoption and defensibility. He has invested in Figma, so his favorable assessment is also connected to a financial interest.

Square taught him about adjacent products, Rajaram says. He recalls the company growing from payments to around 11 products with over $50 million in revenue each, and says its key metric became the median number used by a merchant. Those numbers and the relationship between product usage and retention are his recollections, not independently verified results.

He describes Square Capital as a product that grew naturally from payment flows and merchant data. In his account, some products contribute profit while others support retention, and teams should know which objective they are serving. He regards a multi-product portfolio as important, though his prescription is a view rather than a rule for every company.

Rajaram calls DoorDash the most operational of the companies where he worked. He recalls that when COVID disrupted restaurants, DoorDash waived its revenue share for a month despite a short-term financial cost. He regards the decision and the combination of physical operations and software as lessons in long-term execution; the account is his own.

Asked whether software-stock volatility is justified by AI, Rajaram says he considers the broad sell-off an overreaction. He believes software companies differ substantially in how defensible they are. Both speakers mention exposure to software investments, and his market assessment is an opinion rather than a prediction of any stock’s performance.

Rajaram introduces eight possible moats, adapting ideas from Hamilton Helmer. The first is proprietary data that a competitor cannot easily obtain; the second is deeply embedded workflow. He calls workflow alone potentially weak and contrasts the depth of an enterprise resource planning system with a lighter support workflow.

The third is regulatory protection, including licenses, capital requirements or procurement relationships. Rajaram points to Coinbase and discloses that he is on its board. His claim that the requirements create a strong advantage is his commercial assessment, not proof that customers have no other lawful custody options.

The fourth is distribution. Rajaram recalls wanting to use Xero for accounting but switching to QuickBooks because his accountant preferred it; he views accountant relationships as a potential Intuit advantage. He does not know the terms of those relationships, and a single anecdote cannot establish exclusive distribution.

The fifth moat is an ecosystem of outside developers and applications, which he illustrates with Shopify. The sixth is a network effect such as a marketplace’s density of customers, merchants and couriers, illustrated by DoorDash. Rajaram previously worked at DoorDash; these examples express his view of the difficulty of replication.

The final two are physical infrastructure and scale-based cost advantages. He offers Amazon and TSMC as examples and predicts physical assets may remain difficult to replicate even as AI improves. He suggests assigning one point per moat and regards four or more as strong; the scoring is an informal heuristic, not a validated measurement or guarantee.

Stebbings applies the framework to Atlassian and monday.com after their reported share-price declines. Rajaram tentatively gives Atlassian three points for data, workflow and ecosystem and monday.com one for workflow. These are quick judgments based on incomplete discussion, not independent security assessments or recommendations on either stock.

On Klaviyo and Shopify, Rajaram thinks the relationship and the strength of any preferred distribution matter more than a simple assumption that Shopify will build the same product. He notes that Shopify is an investor, but says he does not know how strong its promotion is. His view of Shopify’s plans is speculation, not inside confirmation.

Stebbings challenges Rajaram’s exclusion of brand, arguing that brand may matter more when technology is commoditized. Rajaram thinks enterprise buyers may be less attached to brands than consumers and predicts data portability and product replication will reduce switching costs. That forecast is uncertain; he does not establish that all migration will become easy or costless.

Using Salesforce as another example, Rajaram gives it an approximate score of three and questions whether pure software production still produces a scale moat. He thinks systems of record remain relatively attractive but must decide whether value lies in data or workflows and develop complementary products accordingly. His suggested pricing and strategy are hypotheses, not company plans.

They discuss share buybacks as potential confidence signals. Rajaram believes company and founder purchases can communicate conviction, while Stebbings questions the significance of a particular executive’s reported purchase. The discussion does not establish why any company or executive traded, or whether a buyback predicts subsequent returns.

Asked what distinguishes an effective AI add-on, Rajaram says a company should reframe the user experience, not merely attach a model to an existing feature. He cites Notion as a product he likes and suggests its agents should improve through use and model tuning. His product assessment and assumptions about competitors are not independently tested.

He uses document processing to illustrate the opportunity to change a workflow: a product might extract useful information while a document is uploaded instead of simply storing it. He argues teams should revisit interactions as model capabilities change. His claims about reliability are broad impressions, not measured accuracy for legal or other sensitive documents.

Rajaram says Marathon invests heavily in fintech and believes moving money can create defensibility through data, workflows and regulatory requirements. That is an investment thesis with a direct commercial interest. For early software startups, he looks for data that improves over time and genuinely embedded workflows, while acknowledging that many other moats are too early to judge.

Stebbings questions narrowly targeted voice or support agents. Rajaram says they may become viable businesses but predicts a single function within one vertical will rarely support the largest outcomes. He favors teams with the ambition to serve more of a vertical’s workflow and uses ServiceTitan as an example; the cited product count and valuation are approximate.

They compare broad platforms with vertical products. Rajaram cites revenue streams at Robinhood and Coinbase and discloses his Coinbase board role; those company figures are his estimates. He argues that a vertical company seeking a very large valuation may need multiple products and a wider share of customer spending.

Asked whether vertical software remains investable for venture funds, Rajaram says yes, particularly if it addresses spending on outsourced services and human labor rather than only existing software budgets. He describes potential expansion into services as a thesis, not evidence that a specific company will capture those budgets or achieve a particular valuation.

Rajaram predicts businesses may first replace outsourced business-process spending, then leave some departing employees’ roles unfilled, and only later consider layoffs. He claims AI services can sometimes be faster or cheaper, but gives no comparative study. The sequence is his observation and forecast, not a universal description of employers or a measured labor-market effect.

Stebbings raises the problem of private software companies with high previous valuations and slower growth, citing an example using his own estimated revenue and growth numbers. Rajaram predicts some may struggle, merge or sell to private equity, while others may rebuild around AI. Neither the company-specific figures nor the proposed outcomes are verified in the conversation.

Rajaram points to Intercom and Podium as companies he believes created new AI-led products, citing approximate revenue growth. He argues leaders may need to migrate existing customers even if the new product is cheaper, rather than protect sunk costs. His examples and claimed growth are unverified and the approach carries risks not fully analyzed in the interview.

Stebbings challenges whether paying a high price today already assumes several years of exceptional growth. Rajaram replies that his optimistic case depends on taking a much larger share of service and software spending. This is a disputed investment assumption, not evidence that any particular company will achieve the required revenue or valuation.

Rajaram does not expect seat-based pricing to disappear because it gives buyers predictable access costs. He distinguishes access products from products that perform work and suggests output-based pricing can fit the latter. He explicitly says he does not know Harvey’s actual pricing, so his contract-processing example is hypothetical.

On venture investors trying to “king-make” a company with a large early round, Rajaram says funding can send a signal, but the company still must execute. He notes that large funds and smaller funds play different games. Both speakers are investors, so their views on funding and fund competition reflect their own professional interests.

As rapid growth becomes more common, Rajaram says durability matters more than a spectacular initial revenue trajectory. He wants to see customer or gross retention and net revenue retention, and says he would favor strong retention over faster growth with poor retention. Examples of companies’ past revenue changes are his rough recollections, not audited figures.

Stebbings asks whether Granola’s strong retention would withstand a competing product from a large model provider. Rajaram agrees that retention should be evaluated against the competition and shocks a company has actually faced. He also says an unusually useful product can persuade people to buy a separate tool they previously would not have considered.

The pair cite Granola and Gamma as examples of potentially creating demand outside an existing bundled tool. Rajaram argues such companies still need to grow beyond a single product, suggesting adjacent documents, presentations or websites as possibilities. Those are illustrative ideas, not announced roadmaps or proof that either company will succeed.

On AI-company margins, Rajaram expects lower inference costs could help, but says the more important long-term question is whether a valuable product gives a company pricing power. He cites recollections about PayPal and changing marketplace economics. His past-price examples and forecasts are not independently verified, and stronger pricing power does not guarantee higher margins.

Stebbings pushes back that it is hard to forecast margins three to five years ahead in a changing market. Rajaram agrees the potential for defensibility and future pricing power matters more to him than a precise early margin. They use consumer subscriptions and luxury goods as illustrations, not as comparable financial analyses.

Asked whether businesses can serve less wealthy consumers profitably, Rajaram says scale and an inexpensive or free entry product can be important because revenue per user is lower. He cites Robinhood and discloses an investment in Atlas, which he describes as serving wealthy customers. His observations are business models to consider, not evidence that any specific financial product is suitable.

Rajaram still performs bottom-up market sizing despite the difficulty of predicting new demand. He suggests segmenting customers, estimating willingness to pay, understanding the problem and speaking directly with potential buyers about budgets. He says a large existing customer base often contains several distinct segments rather than one uniform market.

He recalls underestimating Shopify because he focused on the number of existing e-commerce merchants. In retrospect, he says, the platform enabled more people to become sellers. He calls these non-consumption markets: a new product changes behavior rather than merely winning existing spending. His recollection illustrates a missed judgment, not a valuation forecast.

Rajaram says a new behavior can produce an unusually large outcome if a bet works, but can also leave a company without a market if it does not. Discussing Facebook, he credits real identity as a differentiator and recalls the company resisting pressure to abandon that approach in Japan. These are his historical interpretations, not independently verified causal findings.

Stebbings asks how to avoid letting past losses distort later decisions. Rajaram says investors must revisit each opportunity from first principles and cites a venture investor backing Instacart after a costly Webvan investment. The specific historical loss is his recollection; his lesson is about resisting the tendency to dismiss a whole category.

On entry price, Rajaram says that for seed and Series A investments, a truly exceptional company can sometimes overcome a high price. He cites his investment in Fair and estimates a 100–200 times return. This is his own approximate account, without dates, realization status or verification; price and dilution still affect actual returns.

He says later-stage pricing can be more damaging even when a company grows, illustrating the point with an unnamed business whose revenue rose while its valuation stayed roughly flat. His example describes the possibility of a weak investor return despite operating progress. The amounts and ultimate outcome are not independently established.

Stebbings describes Series A rounds priced at hundreds of millions of dollars and asks how a smaller fund can preserve meaningful ownership. Rajaram says a fund cannot build its whole strategy around those prices; it may need to invest earlier, accept less visible traction and be patient. His proposed double-digit ownership target is a fund strategy, not a guarantee.

Stebbings questions whether a fund with limited investable capital and sizable checks can achieve those ownership goals. Rajaram says Marathon keeps roughly 35% in reserves and uses a mix of inception, seed and Series A investments. He cites Lead Bank and founder Jackie Reses as an example involving his colleagues and investment relationships.

Stebbings is skeptical of repeated claims of proprietary founder access. Rajaram says an investor must offer something founders value, such as advice, distribution, customers or hiring assistance. He agrees the best founders may not need much help, but thinks a small number of consequential decisions can make a difference.

For limited partners choosing a manager, Rajaram recommends speaking directly with founders about why they selected a particular investor and what alternatives they had. He suggests examining recent investments rather than accepting generic claims about networks. This is his due-diligence preference as a fund manager, not a verified method for predicting returns.

Stebbings challenges Marathon’s 35% reserve allocation and asks whether more initial investments would improve diversification. Rajaram contrasts two investors in The Trade Desk, where he serves on the board: one focused on initial checks, while another followed on. He says one approach can maximize a multiple and the other produce more dollars; his deal details and return descriptions are approximate.

Rajaram favors selectively doubling down because an existing investor may know a company and founder better than a new investor does. He acknowledges the danger of excessive concentration and frames the choice as exploration versus exploitation: more initial bets can improve discovery, while follow-ons increase exposure to identified opportunities.

Stebbings says his first fund could at different times have been named for Hopin or Clubhouse and now he expects Linear to be a major driver. Both hosts have interests in the portfolio and the outcome remains uncertain. His example challenges the ease of identifying winners early enough to commit reserves.

Rajaram responds that follow-on decisions need a thesis about the company’s competitors, why it is differentiated and whether that advantage can last over a venture time horizon. He argues a smaller portfolio creates time to study and help each business rather than forcing a partnership to make a new investment every month.

He praises a concentrated approach at Greenoaks and offers approximate figures about its portfolio and outcomes. The numbers are his claims, not independently checked fund results. Stebbings reflects on learning from successful investors; Rajaram says he learns most from entrepreneurs who are actively building and serving customers.

Rajaram says he speaks with people at model companies to understand changing capabilities and potential competitive overlap. He compares a large company’s roadmap to a powerful vehicle that can overtake a startup directly in its path, while leaving adjacent areas open. This is a metaphor for competitive research, not privileged confirmation of future product releases.

He regrets passing on Vanta after meeting its founder because he had already committed to another company in the same space. He says he now tries to meet multiple businesses in a category before deciding. His claim about the eventual winner is a personal judgment; Stebbings also recalls the opportunity as a missed investment.

On selling, Rajaram says he held his Figma investment for roughly 13 years until its IPO, even though earlier liquidity opportunities existed. He notes subsequent price challenges. The timeline and his experience are personal investment recollections, not advice about the stock or a verified calculation of realized returns.

He distinguishes multiple of invested capital from internal rate of return, which incorporates timing. In his view, a large multiple earned over a very long period can yield a less compelling annualized return. He says a fund should assess its prospective return at each liquidity opportunity against its obligations and expectations for limited partners.

Rajaram suggests selling at least part of a holding when the prospective return no longer supports the fund’s goals, especially if a sale could return a meaningful share of the fund. He notes that secondary sales can be subject to company rights and liquidity constraints. This is his portfolio-management view, dependent on assumptions rather than a universal rule.

He describes Quince as a missed opportunity: he dismissed it as a direct-to-consumer company rather than investigating its product and repeat purchasing. The valuation and 35–40% repeat-purchase figure he cites are unverified and lack a defined measurement period. His lesson is to examine individual company differentiation instead of rejecting an entire category.

Stebbings counters that exceptional founders can pivot from an unattractive initial idea, so a thesis-led investor can miss them. Rajaram agrees that a broad seed portfolio may primarily back people, while a concentrated portfolio requires greater conviction in both founder and business. Both approaches involve different risks and assumptions.

Rajaram describes repeat founders, first-time consumer founders and AI-lab researchers as different founder archetypes. He says an early investor may take many small bets on exceptional people, while his own fund cannot justify ownership economics in frontier labs already priced at billions. These are his investment preferences, not an assessment of any named lab’s future.

He says large venture funds can make many early investments and concentrate later in companies they believe are strongest. He also describes cases in which a partner leaves and a portfolio company loses its internal advocate. Those anecdotes and his prediction of more investor spin-outs are not evidence that every large fund or founder faces that problem.

In the quick-fire round, Rajaram says he changed his mind about fully remote early-stage teams after observing founders struggle to align and iterate. He now prefers some in-person work, suggesting roughly three days a week. This is an experience-based preference, not proof that remote work causes startup failure.

He advises new graduates to spend two or three years learning at a strong company before founding one. Asked to choose funds, he names First Round Capital, Benchmark and Greylock and discloses that he has been a limited partner in the first two. His choices reflect personal relationships and interests rather than independent recommendations.

Rajaram says leaving Google to start a company was among his hardest career decisions. He attributes different strengths to leaders he has worked with: technical vision, growth, design and physical-world operations. These are his subjective assessments of colleagues and leaders, not a comparative performance study.

He says one of his biggest misses was failing to imagine the eventual scale of Facebook and Google, recalling much lower valuation discussions years earlier. He also estimates that Fitbit was between 500 and 1,000 times his investment at its IPO, before its later decline. The return is his unverified estimate at a specified point, not necessarily realized proceeds.

Rajaram says ambitious founders applying AI to difficult problems make him optimistic about the coming decade. He sees more young people using AI tools and says he has recently invested in more college dropouts as an angel. He also cautions that leaving school is not necessarily right for most of them; his view of their future success is speculative.

Stebbings thanks Rajaram and closes the conversation. Their wide-ranging assessments reflect the perspectives of investors with commercial interests in the sector; the interview does not independently verify financial figures or establish future returns.