Investor Elad Gil: Most AI Companies Should Consider Selling Within the Next 18 Months
Entrepreneur and investor Elad Gil noted in a series of random thoughts on AI that the current revenue sizes of OpenAI and Anthropic are approximately close to 0.1% of the US GDP. He predicts that AI-related revenues could soon reach 1%-2% of the US GDP, with the core issue shifting to "how this productivity is accounted for in statistics." He also emphasized that AI has pushed top researchers into a state akin to a "collective IPO," where large tech companies like Meta offer extremely high salaries and equity to match other labs' offers, allowing this group to achieve wealth leaps in a very short time, potentially changing their behavior patterns and risk preferences.
At the structural level, Elad Gil proposed that "computing power is the new currency," believing that computing power budgets are becoming the core resource constraint for hiring engineers, prioritizing teams, and assessing project importance. The existing computing power ceiling has temporarily reinforced an oligopoly in the model layer. He predicts that many companies have outsourced layoffs and stopped hiring, leading to a form of "hidden layoffs" where total employee numbers first stabilize and then gradually decline. During the "Slop Era," AI will first tackle closed-loop tasks with high economic value, such as software development and customer service.
In terms of product and market structure, Elad Gil pointed out that the true stickiness of AI tools is shifting from "model differences" to "harness and workflow"—the usage environment, interface, and brand built around the model, which may even have short-term defensive power exceeding that of the model itself. He further posits that the essence of the AI era is "selling labor units rather than software seats," with many AI markets having potential scales 10 to 100 times that of traditional software seat models. In this context, he offers a rather radical conclusion: most AI companies should consider selling within the next 12-18 months, reiterating in podcasts and media interviews that "the valuation peak window for AI companies is only about 12 months."
Source: Public Information
ABAB AI Insight
Elad Gil's entire judgment pulls AI back from a "technical narrative" to a "macroeconomic and asset pricing structure." When the revenue sizes of labs like OpenAI and Anthropic begin to be measured as a percentage of GDP, AI is no longer a marginal industry but has transformed into a new infrastructure investment cycle similar to the combination of "cloud computing + internet." This implies that the productivity gains driven by AI are likely to replay the statistical mismatches of the internet era—real efficiency improvements are obscured within free services, quality enhancements, and new business forms, making it difficult to fully reflect in traditional GDP and productivity metrics in the short term, thus causing a structural misalignment of "perception has changed, but data lags."
The concepts of "computing power as the new currency" and "top researchers in a collective IPO" reveal the new distribution of capital and labor in this cycle: in the previous internet cycle, equity incentives primarily followed the IPO rhythm of companies, whereas this time, AI researchers have already achieved wealth revaluation akin to IPO levels through substantial salaries and equity while companies remain private or unlisted. This will change their risk tolerance and liquidity decisions—some may turn to independent entrepreneurship, establish research organizations, or take on dual roles of "capital + technology," transforming this group from "high-skilled labor" to "technology-capital hybrids," accelerating the concentration of technology and funds among a very small population.
For corporate organizational structures, "hidden layoffs + stable headcount" means that AI has not immediately led to large-scale visible unemployment, but rather gradually reduces the number of employees corresponding to unit revenue through hiring freezes, outsourcing transfers, and natural attrition. This makes AI more like a "slow squeeze" technology: companies no longer expand their workforce while revenues continue to grow, leading to increased profit margins, with labor shares gradually absorbed by capital and a high-skilled minority. The first to be impacted are often developing country enterprises engaged in standardized outsourced services like customer service and IT support, which are in replaceable segments of the global value chain. AI prioritizing "eating closed-loop" tasks will first create localized negative GDP impacts and employment pressures in these countries.
In terms of industry competition structure, Elad Gil's emphasis on "harness" actually reiterates an old rule in software history: true defensibility often lies not in underlying capabilities, but in the usage environment, integration depth, and mindshare. The model layer is unlikely to fully differentiate in the short term under the computing power ceiling and oligopoly, so whoever can occupy the default entry point for developers and enterprise users through workflows, interfaces, and brands will find it easier to maintain bargaining power when subsequent models become homogenized. This is similar to the logic of "controlling the console and ecosystem" in the cloud computing era: what appears to be sold is basic capability, but the essential competitive point is the ecosystem and lock-in.
The most controversial point is his judgment on the "valuation window for AI companies": most AI companies should consider selling within 12-18 months. The structural logic behind this view is that as foundational models rapidly expand capability boundaries, many AI applications that currently seem like "product companies" may be directly absorbed by upstream models or large platforms in the form of functions, leading to rapid disappearance of differentiation and valuation compression. In other words, the current high valuations reflect more of a scarcity premium of "not yet being swallowed by platforms" rather than long-term independent value. This short window period and sharp decline in peak valuations closely resemble the fates of some companies (like portals and tool applications) during the early internet acquisition wave, only this time the speed is faster and the concentration is higher.