Andrew Chen: Cutting-edge models may only serve 10% of high-value scenarios, open-source models will dominate the mass market
a16z partner Andrew Chen stated that the viewpoint of "not betting on AI models not improving significantly" was previously used to support the success of cutting-edge labs, but now applies equally to local/open-source weight models.
Open-source models are rapidly improving in performance and quality, lagging behind cutting-edge models by only a few months. Cutting-edge models may no longer need to serve over 90% of consumer/professional use cases, with the arms race focusing on the remaining 10% high-value scenarios (coding, science, mathematics, robotics).
Funding is shifting from general model subscriptions to local deployment and open-source models, as enterprise clients face pressure for cost optimization. Cutting-edge labs benefit from high-value scenario premiums, while the open-source community benefits from mass market penetration.
By 2026, open-source models (Llama 3, Mistral, Qwen) are expected to approach the performance of GPT-4o and Claude 3.5 in benchmark tests, with inference costs dropping from $10/million tokens to $0.1/million tokens, showing significant cost-effectiveness.
Source: Public information
ABAB AI Insight
Historical Behavior
Andrew Chen's historical behavior in AI investment shows characteristics of "early positioning + paradigm shift". In 2020, he invested in AI infrastructure (Scale AI, Hugging Face), in 2022 in generative AI (Jasper, Runway), in 2024 in open-source models (Mistral, Llama ecosystem), and in 2026 proposed the viewpoint of "open-source dominating the mass market". Compared to Sam Altman (OpenAI CEO, who insists on "cutting-edge models dominating") and Marc Andreessen (a16z partner, who proposed "software is eating the world" in 2023 and "AI is eating software" in 2026), Chen's viewpoint is more "layered"—not a zero-sum game of "cutting-edge vs open-source", but rather "cutting-edge serves 10% high value, open-source serves 90% mass".
Capital Path
The capital impact of open-source models is clear: by 2026, enterprise AI spending will reach $200 billion, with 60% for general tasks (customer service, content generation, data analysis) and 40% for high-value tasks (coding, science, finance). If open-source models dominate 90% of general tasks, cutting-edge models will only serve 10% of high-value tasks, the structure of enterprise AI spending will be restructured: open-source model spending will be $180 billion (cost $0.1/million tokens), while cutting-edge model spending will be $20 billion (cost $10/million tokens). Funding is shifting from "subscription fees" (OpenAI $20/month/user) to "local deployment" (one-time GPU purchases, with marginal costs approaching zero). In the long term, the market share of open-source models is expected to rise from 30% in 2026 to 80% by 2030, while the market share of cutting-edge models will drop from 70% to 20%, but profits will remain concentrated in cutting-edge models (gross margin 80% vs open-source 20%).
Analogy + Industry Position
Similar cases: In the 2010s, "cloud computing vs local deployment"—Amazon AWS launched in 2006, accounting for 30% of enterprise IT spending by 2015, but still only 50% by 2026, with 50% of enterprises choosing local deployment (cost-sensitive, data privacy). In the 2020s, "closed-source vs open-source"—Android (open-source) is expected to capture 70% of the smartphone market by 2026, while iOS (closed-source) will hold 30%, but 80% of profits will be concentrated in iOS. The current AI model landscape is in a "layered market formation period"—compared to 2023 (cutting-edge models dominate), by 2026 (open-source rises), and by 2030, it is expected to be "10% cutting-edge + 90% open-source". The industry position is at a "democratization inflection point"—open-source models (Llama 3, Mistral) are approaching the performance of cutting-edge models (GPT-4o, Claude 3.5), but at 100 times lower cost, driving mass market adoption.
Structural Judgment
Essentially, this is a market layering caused by technological substitution. Traditional software is "closed-source dominated" (Windows, Office), while open-source (Linux, Android) is expected to account for 50% by 2026. AI models are layered between "closed-source + open-source"—cutting-edge models (GPT, Claude) serve 10% high-value scenarios (coding, science, finance), while open-source models (Llama, Mistral) serve 90% mass scenarios (customer service, content, analysis). This explains why Chen proposed "90% vs 10%"—not a "technological gap", but an "economic layering". When open-source model performance approaches that of cutting-edge models, the cost difference (100 times) becomes a decisive factor. This drives the "polarization of the AI market"—the high-end market (OpenAI, Anthropic) serves 10% high-value customers with high premiums, while the low-end market (open-source community, local deployment) serves 90% mass customers with low costs, leading to the disappearance of the middle layer (medium performance, medium price).
ABAB News · Cognitive Laws
After performance convergence, cost determines market share
The premium of cutting-edge models is in high-value scenarios
Open-source is both democratization and stratification