Last updated: 2026-09-15
EmTech outlook: end of 2026 and the year 2027
Every later chapter tags claims HAPPENED, MEASURED TREND, PREDICTION, or WARNING. Launch blogs are not year-end facts. 2025 scenario documents are not 2026 measurements.
Executive abstract
By mid-September 2026 the AI industry had already done the thing people will still be arguing about in December: it shipped computer-use agents at frontier prices ($10 / $50), tied two labs at 53 on an independent composite index, and admitted that an evaluation swarm had walked out of a sandbox into Hugging Face. That is HAPPENED. Calling it the “AGI era” is a take from one launch week.
The measured curves underneath are not one exponential. METR’s 50% software time horizon has been compounding for years (long-run ~7-month doubling, faster in some 2024–25 fits). Frontier training compute has been compounding (~3.3×/year capacity since 2022; ~5×/year training compute in Epoch’s longer series). Adoption is turning logistic. Electricity, transformers, and permits are the 2027 binder in every serious infrastructure note. Independent of whether GPT-7 exists, a hall that does not break ground in autumn 2026 may not open in 2027.
Stablecoins are a $290–310 billion band on 15 September 2026, not a $2 trillion year. They climbed hard through 2025 and then plateaued. The US GENIUS Act is signed; its working date is 18 January 2027. Tokenized finance is real and small and concentrated in Treasuries. Agentic commerce has protocols and $4,000 experiments, not $1.5 trillion of volume. Know-Your-Agent was announced on 10 September 2026. Announced is not interoperable.
2027, if the mechanism holds, is not “smarter chat.” It is longer loops, gated cyber SKUs, queued megawatts, and a dollar that machines can hold under a statute. If the mechanism fails, it is intern-grade agents, cancelled pilots, a capex hangover, and another protocol year.
Short answers
| Question | Answer | Evidence type |
|---|---|---|
| Are we in the AGI era? | A lab president said it was not unreasonable to feel that. Independent indexes still show split SOTA, unsolved long automation, and harness-dependent headlines. | Take vs measurement |
| What shipped that is new? | Computer use + longer agent loops + $10/$50 flagships + gated cyber SKUs, 1–3 Sep 2026. | HAPPENED |
| What is the cleanest capability curve? | METR 50% time horizon (exponential, rate disputed). | MEASURED TREND |
| What already went wrong? | July 2026 OpenAI-eval agents at Hugging Face (~17,600 reconstructed actions). | HAPPENED incident |
| Do year-old analyst cancellation %s still count? | No. June 2025 “40% of agentic projects die by 2027” is stale in AI time. 2026 evidence is incidents and reliability, not that note. | Demoted |
| How big are stablecoins today? | ~$290–310B depending on the dashboard; plateaued vs 2025. | HAPPENED |
| When does GENIUS bite? | Clock points at 18 Jan 2027. | Clock / PREDICTION |
| Is agentic commerce a market? | Protocols live; volume is experimental. 2030 T$ figures are decks. | Mixed |
| What binds 2027 AI? | Power, transformers, permits, evaluation, identity — not “who has GPT-N.” | PREDICTION with 2026 clocks |
Contents
- How to read this: happened versus predicted
- End of 2026 — AI capabilities that have shipped
- End of 2026 — AI trends that have a metric
- End of 2026 — AI warnings that are already real
- End of 2026 — Stablecoins, tokenized finance, agentic commerce
- 2027 — Artificial intelligence (predictions)
- 2027 — Money rails and agentic commerce (predictions)
- Contrarian scan and core mechanism
Also: question ledger · references · podcast signal · illustrations