NEA: Open Weight Models Cover a Wide Range of Uses

For Aaron Jacobson, a partner at NEA, open-weight models already cover a growing portion of needs and are expected to gain traction. Market signals, such as spending around Opus 5 from Anthropic, are putting pressure on prices. However, the Chinese origin of some advanced models hinders their deployment in sensitive contexts. According to him, cutting-edge models still dominate the most complex problems as the industry shifts from chatbots to agents.

Security and Origin Hinder Some Enterprise Deployments

Aaron Jacobson emphasizes that in the United States, the availability of open-weight models faces a major constraint: the most advanced ones are of Chinese origin. Some companies are willing to use these models to generate code or perform other tasks, while many others prohibit their use for code or data. For a significant portion of the market, including enterprise workloads, these models could technically provide cheaper and potentially sufficient work. However, their origin and the fact that they are not truly open source raise security concerns that hinder their adoption. In this context, the relevant tasks are either assigned to non-Chinese open models deemed less effective or to more expensive closed cutting-edge models. The more sensitive the data, the stronger these reservations are: source code is considered extremely sensitive, as are uses related to finance or customer data. Individual developers, however, are already using open-weight models.

Pricing: Market Signals and Pressure on Cutting-Edge Models

According to information reported by the Financial Times and sourced from Ramp, Anthropic's Opus 5, a discounted solution, has gained an edge over Fable in terms of commercial spending since its release at the end of July. Aaron Jacobson highlights two pressure movements on prices: in many use cases within startups and certain companies, open-source models meet the needs for some code generation tasks; these models are also less expensive because they do not bear the "tax" associated with operating a cutting-edge model. Tokens deemed just as good, or sufficiently good, are driving prices down, while cutting-edge models maintain high rates to cover their training costs. At the same time, open-weight models are progressing, becoming cheaper and easier to customize, which pushes closed providers to enhance their features to retain their customers. Companies are thus balancing cost, control, and security in their model choices.

Use Cases: Delegating the Simple, Keeping the Complex for Cutting-Edge Models

Aaron Jacobson reserves cutting-edge models for very difficult or open problems, particularly those requiring long-term planning. He cites tasks such as organizing a vacation from start to finish, including flight bookings, or developing a highly secure banking application. In this framework, a cutting-edge model handles planning and more complex token generation while redirecting simpler subtasks to less expensive models. He believes that for many everyday tasks, cheaper models are sufficient and that it is not necessary to send everything to cutting-edge intelligences. He situates these recommendations within a broader shift from the chatbot era to that of agents.

Technical Framework and Strategic Direction Advocated by NEA

Aaron Jacobson distinguishes three categories: open source, where data and pipelines are available to retrain and obtain weights; open weight, where only the weights are accessible without transparency on training data; and closed cutting-edge models, operated via API, without access to weights. He considers the latter to be better across a wide range of problems and for coding, while reminding that they are more expensive and not always the most useful. His central debate revolves around the actual necessity of cutting-edge intelligence for every task, as open weights become sufficient for many uses and challenge the business model of closed players. He anticipates a shift from a decade marked by "the bigger, the better" to a decade dominated by the efficient use of models. In this perspective, he believes that open ecosystems are sufficiently effective and that companies do not need to systematically commit to cutting-edge models, placing this debate at the heart of the next phase of the industry.