Ethan Mollick posted on X that 'we are quite short of compute,' warning of a market-splitting dynamic where complex AI agents become expensive while chatbots get cheaper. The observation captures a structural divergence in AI economics that most vendor press releases elide.
Key facts
- Agentic workflows require 10-50x more compute than single-turn queries
- OpenAI o1 agentic tasks cost $10-100 vs $0.01-0.10 for single GPT-4o
- Anthropic Claude Opus charges $15 per million input tokens
- Microsoft Copilot costs $30/user/month; agents could be 10-20x that
- Startups Adept, Imbue, Cognition pre-purchased compute capacity
Ethan Mollick posted on X that 'we are quite short of compute, and that is going to result in compute becoming very expensive for complex agentic workflows even as single-turn chatbots get cheaper.' [According to @emollick] The post frames a bifurcation: 'the richest companies & most pressing use cases will use AI agents & everyone else will be stuck with chatbots.'
Why agents cost more
The key mechanism is that agentic workflows require multiple inference calls per task, often 10-50x more compute than a single-turn query. OpenAI's o1 model, for example, uses chain-of-thought reasoning that can cost $10-100 per task for complex agentic loops, versus $0.01-0.10 for a single GPT-4o completion. [Per public pricing pages] Anthropic's Claude Opus similarly charges $15 per million input tokens; an agent that iterates through 50 tool calls burns through 500K+ tokens per task, pushing per-task cost into dollars.
The structural divergence
The divergence is structural: inference efficiency gains (KV-cache optimizations, speculative decoding) apply disproportionately to single-turn, stateless queries. Agentic workflows demand stateful context, multi-step planning, and tool integration — none of which benefit equally from batch-size scaling or prompt caching. [As previously reported in AI literature] The result is a compute-cost wedge that widens as model capability improves.
Market implications
Mollick's framing echoes a pattern visible in enterprise deployments: Microsoft Copilot (single-turn) costs $30/user/month; an equivalent agentic assistant with multi-step workflow automation would cost 10-20x that in compute alone. [Per Microsoft pricing] Startups building agentic products — Adept, Imbue, Cognition — have raised large rounds partly to pre-purchase compute capacity, as previously reported. The market is already sorting by budget.
What to watch
Watch for GPU cluster booking lead times from AWS, Azure, and GCP in Q2 2026 earnings calls — if lead times extend beyond 6 months for H100/B200 instances, the agent-vs-chatbot price wedge will widen further. Also track whether OpenAI or Anthropic launch agent-specific pricing tiers that formalize the split.
