Founders Must Separate Trillion-Dollar Ambition From Revenue Reality

Elad Gil argues that the recent run from near-zero to trillion-dollar valuations has distorted expectations about what a normal technology cycle looks like. In roughly five years, he says, Anthropic, OpenAI, and SpaceX each moved into a valuation range that historically took companies closer to 15 or 20 years to reach. Google’s arc began in the 1990s; SpaceX started in the early 2000s. The compressed timing is unusual enough that investors and founders may be treating it as a durable baseline rather than an exceptional inflection.

Gil’s concern is not that AI, robotics, materials, or other emerging fields cannot support major companies. Over a 20-year horizon, he expects many. Nor is it that a $100 billion outcome is unimpressive. The issue is the combination required for a multi-trillion-dollar company: a single business needs roughly $50 billion to $100 billion in revenue, good margins, and the capacity to reach that scale quickly enough to justify current expectations.

That is a different proposition from identifying a large total addressable market. Gil thinks investors are often conflating market size with the speed at which a company can capture it. For a physical-goods business in energy, robotics, or a related category, the constraint is not only demand. It is whether the company can build the manufacturing, deployment, and operating footprint needed to turn demand into revenue on a three- to five-year timetable.

He describes technology history as a kind of punctuated equilibrium. Social, SaaS, cloud, security, and crypto each produced bursts of company formation and value creation, followed by consolidation. AI will likely have further discontinuities—new jumps in model capability could create another opening for startups—but the first wave has already produced some consolidators. The near-term question is not whether technology will continue producing large companies. It is how many businesses can meet the revenue, margin, and speed requirements of the highest tier before today’s leaders deepen their positions.

Sarah Guo objects to the tendency to anchor markets to their closest historical proxy. She sees an underappreciated investing skill in rethinking what a market becomes when AI changes the commercial model. Application companies in law, medicine, and other professional domains are still often evaluated as if they will charge by seat, lawyer, or doctor. A company delivering a service or an outcome, she argues, may capture a different portion of the economic value.

Guo points to coding as evidence that a market can expand far beyond conventional software spending. The relevant comparison is not necessarily the existing budget for developer tools; it may be the larger value created through higher consumption, better outputs, and greater productivity. Gil agrees that coding is much bigger than many investors expected. His distinction is that a market becoming much bigger does not automatically make it a trillion-dollar market on a five-year timetable.

Elad Gil sees two opposing errors in the startup market. At mid- and late-stage, investors may be pricing companies as though many can reach trillion-dollar scale at extraordinary velocity. At the earliest stage, he thinks some unusually strong founders are avoiding ambitious opportunities because they fear that the major AI labs will enter and win them.

He names markets pursued by companies such as Harvey, OpenEvidence, Decagon, Sierra, and Cognition as examples of large areas that could plausibly have appeared on a frontier lab’s roadmap. Yet those companies went after them anyway. More recently, he sees a tendency to seek refuge in hardware, narrow vertical applications, or products whose underlying capability might ultimately be supplied by an inference cloud. The defensive logic is that the labs will not build this particular hardware product or niche application, leaving a safe space for the startup.

Gil does not claim that labs will fail to absorb any adjacent market. Some markets, he says, they will “just eat naturally.” But he believes founders are increasingly staying away even from areas where a startup can build a better product, distribute it better, or develop an experience that a general-purpose lab is unlikely to prioritize. His concern is specifically about what the strongest new founders choose to do, not the median company entering the market.

Sarah Guo says she, too, is often disappointed by how little ambition founders show relative to what may be possible. But she resists a blanket conclusion: companies in her portfolio are taking on central premises and difficult competitive territory. The point of agreement is that fear of the labs should not become a substitute for strategic judgment. A founder needs to distinguish between a market structurally owned by a model provider and one where the model provider’s existence is simply part of the competitive environment.

Elad Gil proposes that companies schedule a standing board discussion, ideally every six months in the current AI environment, on whether they should consider an exit in the following six-month period. The meeting should be set in advance, rather than triggered by founder exhaustion or investor pressure. Its purpose is to make an otherwise emotional decision available for deliberate analysis.

A small group of companies should not sell in the near term, in Gil’s view; he places Anthropic and OpenAI in that category. Most companies should at least consider it. He believes many businesses have a 12- to 18-month window in which a sale may be their best possible outcome, and that treating independence as an unquestioned virtue can obscure that window.

The case for reviewing the question more often is the speed of AI change. Gil characterizes one year of AI time as roughly three to four years of a normal technology cycle. Model capabilities, vertical applications, infrastructure, and AI roll-ups all look radically different from three years ago. If the underlying facts change at that pace, a company’s assumptions about its own position need regular revision.

That math includes eventual revenue, valuation multiple, dilution, the years required to reach the outcome, and the likelihood that growth slows before the company achieves its intended scale. Gil’s central claim is that the largest opportunity cost is often the founder’s time. A founder can take a good exit, retain credibility, and start another ambitious company—or remain committed for years to an overcapitalized business that is not working while a major platform shift occurs elsewhere.

He points to companies founded or heavily financed in 2020 and 2021 that are still operating five or six years later despite weak trajectories. Their runway may allow them to continue, but continued operation is not necessarily a good use of a strong founder’s most productive years. Secondaries can relieve personal pressure without resolving that larger question, he argues.

Sarah Guo agrees that companies need situational awareness, and that pride in never considering a sale is “nonsense.” Her additional test is whether the company is capturing value as costs fall and capabilities rise. A business on the wrong side of that secular change—and unable to explain how it reaches the right side—should seriously consider selling.

Guo also stresses a competing risk: a founder can be right about the future and still be unable to finance the path there. Gil expects giant recent outcomes to return substantial capital to venture firms, producing larger funds seeking the next trillion-dollar businesses and making financing easier over the next year or two. Guo’s objection is that private markets can remain irrational for long periods. A founder building against an unpopular technical or market thesis may face a practical financing constraint even if the thesis is ultimately right.

She compares financing risk to avoiding a margin call. Founders have concentrated exposure and must navigate prevailing beliefs, not merely have a correct model of the future. Gil’s reply is that the relevant variable may be not actual scale but perceived scale. Both views point to the same discipline: align the financing structure with the thesis horizon, and do not let available capital erase the need to assess whether the company can survive the market’s narrative about it.

Elad Gil describes a widespread belief inside frontier AI labs that code may be effectively solved within months, followed by some form of light recursive self-improvement by the end of the following year. In that view, models would begin by improving post-training and eventually contribute to larger parts of their own development. The belief creates an intense atmosphere: if a researcher thinks productive human work has only a year or year and a half left before displacement, each week can feel disproportionately consequential.

Sarah Guo considers self-improving training systems plausible in principle, particularly as an extension of current progress in code and mathematics. But she is skeptical of the standard 18-month forecast. She says smart, self-aware researchers have perceived a near-term “knee in the curve” toward recursive self-improvement or artificial superintelligence repeatedly over the last five years. The extension from coding to training code and data-pipeline work is easier to believe than the full timeline. Harder questions include how models obtain data in complex, less verifiable domains and whether physical compute access becomes the binding constraint.

Gil thinks compute scarcity creates a counterintuitive competitive effect: it reinforces an oligopoly. If compute is distributed roughly in proportion to the largest labs, constraints put a ceiling on how far any one lab can outrun the others. The labs remain closer competitors than they might be in a world of unconstrained access.

The scarce resource is also changing how labs think about people. Gil says that in many fields, a relatively small number of people generate a disproportionately large share of meaningful results, and he sees the same power law in AI research. When compute is scarce, the cost of a researcher is not merely salary; it is the token budget associated with that researcher’s work. Some labs, he says, have slowed research hiring unless a candidate clears an exceptionally high bar for that reason.

Gil calls the emerging measure “return on invested tokens,” analogous to return on investment. Rather than allowing everyone to experiment freely with AI tools, organizations will increasingly ask which people and projects deserve disproportionate token allocations, and what return those allocations create.

That is why, in Gil’s view, the death of SaaS is overstated. A company may not want to spend expensive model capacity replacing a relatively cheap SaaS product if those same tokens can produce a new core product, a major margin improvement, or another higher-value outcome. He expects organizations to move from broad experimentation, to spend measurement and open-source substitution, and then to explicit choices about where scarce tokens earn the highest return.

Guo sees an implication for researchers who fall outside a lab’s highest-priority allocation. A researcher without an outsized token budget may have more promising places to apply their comparative advantage: physical supply-chain bottlenecks, biology, and other fields where AI diffusion could be valuable. Gil agrees that a researcher outside a top lab’s small, highly provisioned group can still be exceptional—and that there are many valuable places for that person to go.

Sarah Guo asks whether anything could materially disrupt the current AI landscape: a retreat from capital expenditures amid concern over debt and returns, or a technical break from transformers. She expects the industry to consume whatever compute and power is available regardless of architecture. Pressure to find more memory- and energy-efficient approaches will increase, but catching up to transformers’ scale and hardware fit remains difficult.

Elad Gil sees the high-probability outcome as diffusion: a new architecture is discovered, copied, and adapted by the existing large players. A lower-probability alternative is that a neo-lab develops a breakthrough, keeps it secret long enough to scale it, and establishes a decisive model advantage. But Gil’s view of the experience so far is that knowledge tends to spread as people move among organizations.

The more consequential question for him is regulatory. Guo raises the prospect of restrictions on models already available, including open-source models, as a way to slow progress. Gil imagines a scenario in which safety burdens become high enough to constrain smaller companies while leading labs maintain internal progress. If that happened, he argues, the gap could become a structural advantage for the labs because AI progress is moving so quickly.

Gil frames the issue as risk relative to benefit. He invokes an argument he attributes to Janssen, the pharmaceutical developer: that drug regulation became slower and more expensive because regulators focused too heavily on safety and insufficiently on benefit. Gil explicitly does not present that view as settled fact. He uses it to identify a failure mode in which institutions weigh only downside, regardless of the prospective upside foregone.

Guo notes that many policymakers will be uncomfortable with a technologist’s request to permit experimentation and observe the consequences. Gil responds that the tension accompanies major technologies generally: biotechnology can produce dangerous pathogens and cancer therapies; nuclear technology can produce weapons as well as abundant energy.

The source displays figures indicating that France derives about 70% of its electricity from nuclear power and that U.S. nuclear generation accounted for 18% of electrical output in 2024. Gil argues that U.S. safety concerns foreclosed abundant clean energy, saying that a safety lobby in the 1970s “basically kill[ed]” that path. Guo replies that reactors are being built now, but that the country needs many more.

Gil’s point is not that AI risk is nonexistent. He repeatedly says safeguards are necessary. His concern is where society sets the balance among safety, risk, and outcome—and whether, in a hypothetical high-compliance regime, protective policy could favor organizations with the resources to bear the burden.

That concern also informs their discussion of California’s proposed billionaire tax and a possible founder migration. Guo says it is unclear whether regulators have thought through execution and compliance, but predicts that the immediate effect would be to encourage people creating new value to leave the state. Gil says the bill’s breadth, together with talk of a future exit tax, could intensify that dynamic.

Neither treats migration as simply a lifestyle choice. They see ecosystem formation as a matter of critical mass: capable people working on related problems in one place. Guo points to technology and energy activity in Texas as a response to regulatory conditions and demand. Gil adds that energy and hardware activity are growing there as well. In their account, regulation helps shape where innovation clusters form.