AI Eats the World: Benedict Evans on the Next Platform Shift

Current Adoption & Usage Gaps

• ChatGPT has 800–900 million weekly active users, yet only 5% are paying subscribers.
• Survey data suggests 10–15% of people in the developed world use generative AI daily, with another 20–30% using it weekly.
• A significant usage gap exists: approximately 80% of account holders cannot identify a specific use case for the tool in a given week.
• Usage patterns are bifurcated: high-frequency users are primarily developers, marketers, or those with flexible, open-ended workflows; other professions (e.g., lawyers, accountants) currently lack daily-critical tasks that AI automates efficiently.
• Validation remains a bottleneck for use cases requiring specific, factual accuracy (e.g., data entry, financial analysis) where error rates currently outweigh the efficiency gains of manual verification.

Definition & Perception of AI

• "AI" in general usage functions as a label for "new" technology; once a tool is established (like databases or the web), it is no longer considered AI.
• There is a schism in industry forecasting: some leaders (e.g., Sam Altman) claim PhD-level AI capabilities exist now, while others (e.g., Demis Hassabis) argue this is premature.
• The definition of AGI remains fluid, oscillating between "already here as better software" and "five years away, always five years away."
• Physical and theoretical limits of the technology are unknown, preventing accurate modeling of future capability improvements.

Economic Dynamics & Investment Risks

• The current investment cycle exhibits classic bubble characteristics: circular revenue, heavy leveraging, and synchronized asset appreciation.
• Hyperscalers face a "downside asymmetry" decision: the risk of under-investing is perceived as greater than the risk of over-investing in compute capacity.
• Compute costs are projected to be in the range of $100–250 billion annually for top-tier models, with hyperscalers betting that demand will outpace supply.
• Over-investment risks a "ratchet effect" where capacity remains underutilized if model efficiency improves faster than adoption (e.g., achieving same results with 1/100th the compute).

Platform Shift Comparisons

• Platform shifts are categorized into "sustaining" (value captured by incumbents, e.g., mobile) and "disruptive" (value captured by new entrants, e.g., the internet).
• Mobile shifted the tech landscape but ultimately reinforced incumbent value; generative AI may similarly allow incumbents to absorb AI features without creating new giants immediately.
• Unlike previous shifts (e.g., mobile, internet), the AI shift lacks clear physical limits, making historical analogies less predictive.
• The "Internet" comparison holds significant weight: like the web, AI is a foundational layer that requires new product layers (GUIs, specific workflows) to realize value.

Corporate Strategy & Competitive Landscape

• Google/Alphabet: Viewing AI as a search optimization tool and feature addition; competitive advantage relies on data and existing ad infrastructure.
• Meta: AI is existential; owning the stack is critical to defending the core social content and recommendation loop against disruption.
• Amazon: Focusing on AI as infrastructure (AWS) and a potential revolution in consumer discovery/recommendation (beyond simple retail transactions).
• Microsoft: Betting on the "iPhone of AI" emerging within its ecosystem, potentially unbundling legacy software features into AI agents.
• Apple: Faces a strategic question regarding AI agents; if AI fundamentally changes computing, Apple risks losing its device dominance if the "agent" layer decouples from hardware.
• OpenAI: Currently lacks moats beyond brand and distribution; faces pressure to build ecosystems, infrastructure, and unique features to avoid commoditization by cheaper model providers.

Product Evolution & UX

• Raw chatbots are inefficient for enterprise tasks; value is captured by "unbundling" the model into specialized UIs and workflows (e.g., legal discovery, specific SaaS tools).
• The "GUI" function of AI software is to encode institutional knowledge and decision logic, reducing the need for users to prompt from first principles.
• Successful products will likely define "new jobs to be done" rather than just automating existing ones (e.g., suggesting home insurance rather than just packing tape).
• Early product iterations (like the iPhone 1.0) often failed initially; the market requires time to discover the "killer use case" (similar to the 3G era).

Future Outlook & AGI

• For AI to be deemed "bigger than the internet," it must transition from a probabilistic tool to a reliable, human-level agent capable of consistent, complex reasoning.
• Current systems are viewed as "infinite interns" that require human supervision; the threshold for AGI involves removing the need for such supervision in broad contexts.
• The industry expects multiple winners across specialized categories rather than a single "winner-takes-all" model provider ecosystem.
• Most current questions about the technology's trajectory are considered premature; the "true" questions will only emerge as the technology matures through deployment.