TL;DR
A scenario published November 18 argues that AI systems could speed up AI development in 2027 before limits on chips, buildings and power slow progress. Public benchmark data shows rapid gains in software task length, but the projected timing depends on which trend line is used, and the scenario is not a forecast.
Thorsten Meyer published a 2027 AI scenario on November 18 in which AI systems speed up the development of later systems, then encounter limits in chips, data centres and power. The piece presents this as a scenario rather than a forecast; public benchmark data shows progress in software task performance, but does not establish that a sixfold acceleration will occur in 2027.
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
The scenario’s central claim is that AI could take on a growing share of the coding, debugging and experimentation used to build future AI systems. If that feedback loop tightens, Meyer argues, progress that previously took a year could arrive in roughly two months. That pace would amount to about six times the earlier rate. The article calls the later, physical phase “embodied self-improvement,” in which automated labs, factories and robots help expand capacity. These are elements of Meyer’s proposed scenario, not measured outcomes.
As evidence on software capability, the article cites METR’s 80% time horizon: the length of software task an AI agent completes successfully four times out of five. METR’s measured result rose from under one minute for GPT-4 in March 2023 to 185.9 minutes for an early Claude Mythos Preview in April 2026. METR marked its public page as no longer actively updated on September 8, 2026, according to the article.
Meyer says his own fits to METR’s data produce different timelines. An exponential fit using measurements since 2024 doubles task length about every 3.4 months and reaches 16 hours in January 2027; a fit using data since 2019 doubles about every 6.4 months and reaches that mark in June 2028. The benchmark does not measure beyond 16 hours, the article says. It also reports that a hyperbolic fit—one that rises without limit—performed worst in the author’s backtests, reaching infinity by January 2, 2026 when fitted to models available before July 2025.

The Physical Limits on Faster AI
If AI systems materially speed up the work of developing later systems, the pace of capability gains could affect company planning, investment and demand for infrastructure. But the scenario’s own proposed bottleneck is physical: additional software capability cannot by itself supply chips, buildings or electricity. Meyer cites US data-centre construction at a September 2026 annual rate of $96.5 billion, compared with $50.7 billion a year earlier. That figure signals construction activity; it does not show how much usable computing capacity is online or prove that demand will outstrip supply.
The distinction matters for employment and business impact, too. The article cites estimates that 18% of US firms used AI in a business function in late 2025 and early 2026, while 2% reported AI-related job cuts. Those figures describe reported adoption and job cuts in that period. They do not establish that broader job losses will follow from faster model development, or when any such effects might occur.
Products Worth Considering

Logitech MX Brio Ultra HD 4K Webcam

Kindle Paperwhite Signature Edition 32GB (2024 Model) – Metallic Black

DJI Mini 4K Drone Combo with 4K Camera, 3-Axis Gimbal, and Two Batteries
As an affiliate, we earn on qualifying purchases.
What the Benchmarks Measure
METR’s time horizon measures performance on a particular set of software tasks, using the threshold that an agent must succeed four times out of five. It offers one way to track agents’ ability to handle longer tasks, but it is not a direct measure of general intelligence, economic productivity or the ability to build a new AI system without human input. The article says that above 16 hours, the task suite no longer measures the projected range.
The article also reports that METR estimates Anthropic is improving its models on well-specified tasks at around twice its earlier rate. METR is quoted as saying it “would not be surprised” by a rate between 1.3 and 3 times the earlier pace. That estimate is well below the scenario’s proposed sixfold acceleration. Separately, the article says OpenAI’s GPT-6.2 Astra is the first model it has released externally to meet the company’s “High” threshold for AI self-improvement. That is OpenAI’s classification; the source does not provide a shared industry standard for the threshold.
Infrastructure adds a different timescale. The article cites Epoch AI’s expectation that power will be the first constraint on scaling, and says new power supply can take three to five years from the decision to build. That forecast helps explain why software progress and the physical expansion of computing capacity may move at different speeds.
““We reach ASI either in 2027 or shortly after. The effect: in two months we see the progress of a whole year, and then further progress until we hit a ceiling in available compute.””
— Thorsten Meyer
Which Growth Curve Fits?
The available measurements do not determine which extrapolation is right. The article’s projected dates shift from January 2027 to June 2028 depending on the starting period used for the fit, while the hyperbolic curve’s failed backtest illustrates the risk of treating a steep line as a reliable forecast. METR’s public results page was no longer actively updated as of September 8, 2026, according to the source, leaving the article without a continuing series of updates from that benchmark.
It is also unclear whether software task gains will translate into faster end-to-end AI research, whether additional data-centre and power capacity will arrive on the required schedule, or whether the proposed automated labs and factories will become capable and widespread. The source does not establish a date for a compute ceiling or quantify its capacity. Its chart depicts a ceiling as an illustration, not a measured value. The impact on jobs and business adoption remains uncertain as well.
Signals to Watch in 2027
The scenario can be assessed against future measurements of AI agents’ task performance and evidence that AI systems are contributing more of the work used to develop later systems. A fresh benchmark series would help show whether the recent pace continues, slows or changes; any comparison should specify the task suite, success threshold and measurement period.
Infrastructure figures can test the proposed ceiling: data-centre construction, completed capacity, access to chips and new power supply would show whether physical expansion is keeping pace with demand. Business adoption and reported employment effects can show how capability gains spread beyond labs. Until those signals are available, 2027 remains the scenario’s proposed timeframe, not an established outcome.
Key Questions
Is the article predicting that AI will reach artificial superintelligence in 2027?
No. Meyer presents 2027 or shortly after as his scenario. The source says artificial superintelligence has no agreed definition and does not establish that the milestone will occur.
What does METR’s 80% time horizon measure?
It measures the length of software task an AI agent completes successfully four times out of five. It is a benchmark measure, not a direct measure of general intelligence or workplace productivity.
Does the benchmark data support six times faster AI progress?
Not on its own. The article reports METR’s estimate of about twice the earlier rate for Anthropic’s improvements on well-specified tasks, with a possible range of 1.3 to 3 times. The sixfold figure belongs to Meyer’s scenario.
What could limit AI development in this scenario?
The proposed constraints are chips, data-centre buildings and electricity. The article cites Epoch AI’s expectation that power may constrain scaling first, but it gives no measured date or capacity for a compute ceiling.
Source: Thorsten Meyer AI
