INNOVATIONWeeks to result

The AI Capability Diffusion Model

Cutting-edge AI spreads fast; what was unique becomes ubiquitous

Problem it solves

stagnant innovation

Best for

Technology leaders, strategists, and entrepreneurs assessing how quickly AI capabilities commoditize and planning their adoption strategies accordingly.

Not ideal for

People in industries with minimal technology exposure or those looking for specific tactical AI implementation guides rather than strategic mental models.

Overview

Why this framework exists

The AI Capability Diffusion Model describes how breakthrough AI capabilities rapidly spread from a single provider to many competing systems, fundamentally changing the competitive landscape. In late 2023, GPT-4 was the only cutting-edge publicly available model. Within months, six to ten comparable models emerged from different companies and countries, including open-weight versions anyone could run on personal hardware. This pattern reveals that building capable AI did not involve some magical formula only one company possessed.

This framework matters because it challenges the assumption that any single AI advantage is durable. When open models can run on gaming PCs and approach smartphone compatibility, the barriers to accessing frontier AI capabilities collapse. Leaders who understand this diffusion pattern can make better decisions about build-vs-buy, timing of adoption, and where genuine competitive advantages lie: not in access to AI, but in how it is applied.

The model suggests that the window to shape how AI transforms any given field is narrow and exists primarily during the fluid early period, not after capabilities become widespread and commoditized.

Core principles

4 total
  1. What was exclusive AI capability yesterday becomes commodity tomorrow.
  2. Building capable AI did not involve some magical formula only one company had.
  3. The opportunity to shape how AI transforms your field exists now, when the situation is fluid.
  4. The remarkable aspect is not individual breakthroughs but the pace and breadth of change.

Steps

3 steps
  1. Map the Current Capability Frontier
    Identify which AI capabilities are currently cutting-edge and which have already diffused to multiple providers. Track not just what models can do, but how many organizations can deliver similar results. This mapping reveals where genuine scarcity still exists versus where capabilities have commoditized, helping you avoid overpaying for what is becoming freely available.
    Pro tipFollow AI benchmark leaderboards and open-source model releases to track diffusion speed in real time.
  2. Identify Your Application Window
    Determine where in your specific field or industry AI capabilities have not yet been widely applied, even if the underlying technology is available. The competitive advantage lies not in having access to AI which diffuses rapidly but in being among the first to apply it thoughtfully to domain-specific problems. Map the gap between what AI can now do and what your industry has actually implemented.
    Pro tipTalk to domain experts who are not AI specialists as they often see application opportunities that technologists miss.
    WarningDo not wait for the perfect model. By the time you have evaluated every option, the application window may have closed.
  3. Build on Commodity Capabilities
    Design your strategy around the assumption that any AI capability you use today will be widely available within 12-18 months. Build your moat through proprietary data, domain expertise, workflow integration, and user relationships rather than through exclusive access to any particular model. Use open-weight models where possible to reduce dependency on any single provider and maintain flexibility.
    Pro tipTest your strategy by asking: if everyone had access to the same AI models tomorrow, would our approach still be valuable?

Checklist

Saved in your browser

Examples

2 cases
Meta Llama Running on Personal Computers

Meta released Llama 3.3 as an open-weight model with GPT-4 level performance that could run entirely offline on consumer gaming PCs. This single release demolished the assumption that frontier AI required massive cloud infrastructure and expensive API access, making cutting-edge capabilities available to anyone with a decent computer.

OutcomeGPT-4 level AI became accessible to individual developers and small companies without cloud costs or API dependencies.
Ethan Mollick, One Useful Thing, 2023
Chinese AI Labs Matching Western Models

Three Chinese companies, Alibaba (Qwen), DeepSeek, and Yi, released open multilingual models with GPT-4 level performance within the same timeframe as Western competitors, demonstrating that AI capability development was not confined to Silicon Valley or any single research tradition.

OutcomeThe geographic monopoly on frontier AI development was broken, with competitive models emerging from multiple countries simultaneously.
Ethan Mollick, One Useful Thing, 2023

Common mistakes

2 traps
Assuming AI Advantages Are Durable
Many organizations invest heavily in partnerships with a single AI provider, believing that access to their model is a lasting competitive advantage. The diffusion pattern shows that what one company builds, others replicate quickly, often within months. Building strategy around exclusive AI access is building on sand.
Waiting for Stability Before Acting
The pace of AI change tempts many leaders to wait until things settle down before making decisions. But the fluid early period is precisely when the opportunity to shape outcomes exists. By the time the landscape stabilizes, the winners will already be established and the cost of entry rises dramatically.

Origin story

How this framework came to be

Ethan Mollick, a professor at Wharton, documented this pattern in his newsletter One Useful Thing in late 2023, observing a month of unprecedented AI releases. He noted that at the end of 2022, only OpenAI GPT-4 existed as a cutting-edge model. By December 2023, companies including Anthropic, Google, Meta, Chinese firms like DeepSeek and Alibaba, and France Mistral had all released comparable models. The sheer speed of this diffusion from monopoly to abundance in roughly one year prompted Mollick to articulate the pattern as a fundamental feature of AI development rather than an anomaly.

Source

Traced to primary
Source · ESSAY
What Just Happened
Ethan Mollick · 2023
Open source →

Related frameworks

Browse all Innovation →