The 2027 Outlook

Predictions are easy. Confident predictions about AI in 2027 are harder. Here is what feels load-bearing vs what feels like extrapolation.

Five near-certain things

• Inference cost continues to fall ~3x/year for the same quality.
• Open-weight models stay within 6 months of frontier closed.
• Reasoning models become standard for hard tasks.
• Long context windows reach 10M tokens; effective recall lags behind.
• Multimodal becomes the default; text-only is a special case.

Three likely things

• Generalist computer-use agents become reliable enough for real work.
• On-device 7-13B models match 2024-frontier API quality on most tasks.
• Specialised AI hardware (Cerebras, Groq, Tenstorrent) carves out 10-20% of inference market.

Three wildcards

• A new architecture (state-space, retentive, neuro-symbolic) replaces transformers as the dominant choice.
• Capability hits a wall, scaling stops paying off and progress relies on algorithmic breakthroughs.
• Major regulatory action that meaningfully constrains frontier development in the West.

Strategy

Build with portability assumptions. Don’t lock to one provider. Routing/gateway tooling is essential infrastructure. Skills that compound: evaluation rigour, prompt engineering as a practice, ML operations as a workstream. Skills that depreciate: deep knowledge of any single API.