Venture investor Chamath Palihapitiya (Social Capital) stated that as the cost of additional tokens falls, value in the AI industry is shifting from reselling model access to owning the application layer—applications and workflows. He predicts that the first to be hit will not be traditional SaaS companies, but AI-native startups currently spending large sums on purchasing tokens from model labs, primarily in legal, sales, and customer support. This scenario is reinforced by the ongoing move by OpenAI and Google to embed office functions directly into their chat interfaces. According to the investor, businesses structured as "arbitrage"—simple reselling of model access through an interface—are doomed once the labs themselves enter these verticals.

image
image

What happened

As reported by Benzinga and Yahoo Finance, Chamath Palihapitiya outlined a scenario in public comments where the marginal cost of generating a single token approaches zero, and economic benefit shifts from those selling model access to those owning applications and workflows. He unexpectedly identified the class of companies that will suffer first: not established SaaS vendors, but AI-native companies whose operating model is built around large-scale token purchases from labs like OpenAI, Google, and Anthropic—typical examples include legal, sales, and customer support. The basis for this conclusion is cited as the labs' already begun move into the application layer: OpenAI and Google are transferring office tool functionality—document editing and file handling—directly into their chat interfaces, becoming direct competitors to independent AI applications. A related thesis in the Yahoo Finance recap states that approximately 90% of all AI spending comes from just 1% of companies. It is important to note that all of this consists of the investor's public statements on industry economics, not the result of technical measurements.

Context

The development of the generative AI industry has followed a two-tier scheme: OpenAI, Google, and Anthropic trained models and provided them externally via API, while thousands of startups built interfaces, integrations, and vertical scenarios on top of this interface, paying per generated response. As inference costs decreased, "wrappers" earned on the difference, but their protection remained conditional: what such a company sells can be easily reproduced with a prompt on the same frontier model. The historical analog of these processes is the commoditization of clouds and public APIs, when profit migrated from infrastructure up the stack; however, at that time, the commoditization of compute coexisted with a living ecosystem of independent developers rather than leading to its disappearance. It is also important to distinguish between two different classes of products: the thin "arbitrage wrapper" reselling model access, and the vertical application with proprietary data, deep workflow integration, and responsibility for results—the latter cannot be automatically reproduced with a simple prompt. Against this backdrop, labs are transforming from API providers into owners of the work surface: ChatGPT and Gemini already serve as the place where documents and files are processed.

Why this matters for the industry

For the industry, Palihapitiya's statement starts working before the facts: investors are conducting significantly stricter checks on "wrapper" business models, and vertical startup pitches must now answer the question of what will happen to the product if its key function is demonstrated at the next OpenAI or Google demo. Practically, this is a map of no-go zones: any product whose value is reduced to model access through an interface falls into the labs' absorption zone, while protection is provided by owning the workflow, proprietary data, proprietary quality evaluation loops, and deep integration into customer systems. The engineering action plan does not require believing the forecast: calculate the share of inference cost in unit economics, inventory the dependency on a single API, and ensure the product has a layer that cannot be duplicated. If the hypothesis is confirmed at least partially, consolidation of the middle layer of AI applications is expected within a horizon of about two years, and value will further shift to owners of distribution and workflows; companies with proprietary data, quality guarantees, auditability, and deep integration into corporate processes will survive. An alternative scenario remains: enterprises require integrations, reliability, and controllability at a level that lab chat interfaces do not provide, and this gap remains working space for independent vendors.

Why this matters for users

For the reader, this forecast looks like a familiar trajectory: functions of individual AI services you use today may appear directly in ChatGPT or Gemini within the next year or two, and part of the subscription list will become redundant. This is no longer an abstraction: document editing and file handling are embedded in OpenAI and Google chat interfaces, so some independent AI applications have lost their only advantage, and user spending is being redistributed toward chatbots. A side effect is the simplification of workflows: instead of switching between specialized services, many office tasks are solved in one window. The downside is increased dependence on a few companies: if niche startups in legal, sales, and support leave the market, users will have fewer alternatives and a weaker position when discussing price and quality. Before renewing another subscription, it makes sense to check where your data lies and how deeply the service is embedded in your workflow: these layers, not the interface, determine how easily it can be replaced.

What is currently unknown / limitations

Palihapitiya's forecast is an investment opinion without quantitative support: available materials contain no data on the actual dynamics of token prices, margins, or latency, nor are there public benchmarks and evals confirming the readiness of models to replace qualified specialists in legal, sales, and support in real workflows. The thesis that lab competition is already an observed practice relies on the fact of embedding office functions into chat interfaces, but the conclusion about the death of vertical applications does not automatically follow from it. Two different classes of products are mixed: "arbitrage wrappers," which are indeed vulnerable, and vertical applications with proprietary data, whose resilience is fundamentally different. The historical precedent also limits categoricity: the commoditization of compute coexisted with a living ecosystem of independent developers. The investor's formulations reached the audience through recap and translation, and the figure of 1% of companies and 90% of AI spending is cited by publications from the speaker's words and requires separate verification. The verifiable fork in the horizon of about six months is lab product releases: the appearance of explicit vertical modes (lawyer, salesperson, support) in ChatGPT and Gemini would confirm the thesis, while their absence would be an argument against it.

Sources

Author

Look at AI, editorial team