The AI Platform Shift: From Capex Explosion to Commodity Intelligence

Generative AI is driving a massive, capital-intensive platform shift that is transforming tech companies from asset-light software firms into infrastructure-heavy utilities. While LLMs are currently being commoditized, the real opportunity lies in moving up the stack to create specialized applications that turn raw intelligence into daily essential tools. Ultimately, AI functions as a tool for automating tasks, forcing businesses to redefine their value propositions around human judgment and unique workflows.

Key Points

Generative AI represents a 10-15 year platform shift that is fundamentally resetting the tech industry's capital structure and investment cycles.
The current AI landscape is defined by a massive capex explosion where tech giants are spending hundreds of billions on infrastructure, despite models currently appearing to be low-margin commodities.
AI adoption is widespread but shallow; while hundreds of millions use these tools, they have not yet become a daily essential for the vast majority of the workforce.
AI automates 'tasks' rather than entire 'jobs,' acting as a force for price elasticity that allows for more output but requires human 'taste' and 'opinion' to provide real value.
Historical precedents suggest that infrastructure providers rarely capture the full value of a platform shift; the real winners will be those who build the applications and workflows on top of the models.

Sentiment

The community is broadly sympathetic to Evans' analysis and appreciates his historically-grounded, non-hyperbolic approach. However, there is meaningful pushback on his commoditization thesis, with several commenters arguing that frontier models will consolidate into a duopoly rather than commoditize like telecoms. The discussion is substantive and informed, with Evans himself actively engaging and defending his positions. A minority of commenters are hostile, questioning Evans' credibility by pointing to his past crypto commentary, though this is mostly dismissed by others.

In Agreement

Models are showing clear signs of commoditization, with multiple commenters noting that a dozen models now exist at roughly frontier-level performance from six months ago at a fraction of the cost
Value will likely move up-stack to apps, workflows, proprietary data, and product — not remain at the model layer, echoing Evans' core thesis
We are still very early in AI adoption, with usage 'a mile wide and an inch deep,' and most industries like legal showing very low AI usage rates
Historical analogies are valuable but predictions about new technology are inherently unreliable — we are probably asking the wrong questions about AI just as people asked the wrong questions about the early internet
Open-source models like DeepSeek are raising the competitive bar and accelerating commoditization of the model layer
Chatbots are barely products and AI labs have had limited success delivering products above the chat layer, with coding agents being a notable exception

Opposed

Frontier models may follow semiconductor fab dynamics (Rock's Law) rather than telecom dynamics, with escalating training costs creating a monopoly/duopoly rather than a commodity market
Unlike the telecom bubble, AI compute demand actually outstrips supply — all providers are desperately capacity-constrained despite massive capex, suggesting this is fundamentally different
Low usage frequency may reflect compute constraints and deliberate throttling by AI labs rather than a genuine lack of use cases
Chatbots' simple, flexible interface is actually their strength — AI labs are using them to surface valuable use cases before building specialized agents on top
Current models may be hiding massive inefficiency, and more compact representations could emerge that make today's datacenter investments obsolete — similar to the mainframe-to-PC transition
The comparison of LLM tokens to telecom bits is flawed because token quality varies dramatically between models, unlike standardized data bits