🔍 Read the full analysis: Scenario 2027: AI Advances Quickly, Until Compute Runs Short? on ThorstenMeyerAI.com
Get the latest gadgets delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
Thorsten Meyer’s November 18 scenario argues that AI could accelerate the development of future AI systems in 2027, then run into limits in chips, data-centre construction and power. The article presents this as a possibility, not a forecast: benchmark trends depend on the period measured, METR’s public chart is no longer actively updated, and current evidence covers only a narrow slice of work.
Thorsten Meyer published a scenario analysis on November 18 arguing that AI systems could help build their successors quickly enough to compress years of progress into months in 2027, before limits on chips, data centres and power slow the cycle. The piece is explicitly a scenario, not a forecast presented as fact, and its central claims depend on a software-development feedback loop that has not been shown to produce the projected pace.
The argument has four links: AI systems do more work on research and development; that work accelerates later systems; the resulting progress reaches a physical compute ceiling; and a slower, more embodied cycle involving factories, robots and automated laboratories eventually expands capacity. Meyer calls that later phase “embodied self-improvement.” He says the scenario could lead to wider effects on labour and businesses, but that outcome depends on several steps working as described.
The article cites METR’s 80% time horizon, an estimate of the length of software tasks an AI agent can complete with 80% reliability. In METR’s results, the measure rose from 0.89 minutes for GPT-4 in March 2023 to 185.9 minutes for an early Claude Mythos Preview in April 2026. Those are measured task results, not evidence that AI can perform all economically valuable work for the same duration.
Meyer’s own curve fits produce different timelines. Using METR points from 2024 onward, an exponential fit doubles the measured horizon every 3.4 months and reaches 16 hours in January 2027; fitting from 2019 gives a 6.4-month doubling time and reaches that mark in June 2028. The article says its hyperbolic fit, which reaches infinity in November 2027, performed worst in backtests. These are modelled extrapolations, not METR forecasts.
The Limits Behind Faster AI Research
If AI systems materially speed up the work involved in developing later systems, the interval between major capability gains could shrink. That would matter to companies, workers and policymakers trying to plan for changes in software development and other tasks. But the article’s evidence does not establish that a year of progress will arrive in two months: Meyer’s scenario implies roughly sixfold acceleration, while the METR estimate cited in the article puts Anthropic’s improvement rate on well-specified tasks at around twice its earlier rate, with a range of 1.3 to 3 times described as plausible.
The proposed compute ceiling shifts attention from software to infrastructure. The source reports that US data-centre construction ran at an annual rate of $96.5 billion in September 2026, compared with $50.7 billion a year earlier. It also cites Epoch AI’s view that power may become the first constraint on scaling, and that adding supply can take three to five years after a decision. A rapid increase in demand could therefore meet slower construction and energy timelines, although the timing and severity of any shortage are not established here.
Even a substantial rise in model capability would not automatically mean widespread adoption or immediate job losses. The source says 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 different outcomes: adoption and reported cuts. They do not establish that future deployment will follow the same pattern or that capability gains will translate directly into employment effects.
Products Worth Considering

Bose QuietComfort Ultra Earbuds (2nd Gen), Wireless Noise-Cancelling Earbuds, Black

Samsung Galaxy Tab A11+ Kids Edition, 128GB, Royal Blue

Bose QuietComfort Wireless Noise Cancelling Headphones – Twilight Blue Limited Edition
As an affiliate, we earn on qualifying purchases.
What METR’s Benchmark Can Show
METR evaluates frontier AI agents using tasks with estimated human completion times. Its 80% horizon is the task duration at which an agent is predicted to succeed four times out of five. The measure is useful for tracking progress on certain tasks, but its scope matters: METR’s task suite primarily covers software engineering, machine learning and cybersecurity, and the organization says those tasks are cleaner than real economically valuable work.
METR also cautions that longer tasks are hard to assess. Meyer reports that measurements above 16 hours are unreliable with the current task suite. Of 31 tasks estimated to take humans at least eight hours, METR measured human completion times for only five; the rest are estimates. The article further says METR marked its public page no longer actively updated on September 8, 2026. That makes the listed record measurements informative historical evidence, but not a continuously updated indicator for testing what happens next.
The source also points to a claimed change in the development loop: it says OpenAI’s GPT-6.2 Astra was the first model the company released externally to meet its “High” threshold for AI self-improvement. The threshold is OpenAI’s own classification as reported by Meyer; the article does not provide enough detail here to treat it as an independent measure of broad, autonomous AI research capability.
“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
Where the 2027 Projection Could Fail
The central uncertainty is whether AI-assisted development will accelerate rapidly enough to produce the scenario’s proposed sixfold pace. The source’s benchmark measures performance on selected software tasks, not the full research process or general economic work. Its curve fits also change substantially depending on which years are included, and the most dramatic fit performed poorly in the article’s backtests. The projection to 2027 is therefore sensitive to assumptions and starting dates.
Other links are also unresolved. The article does not establish how much AI currently contributes to building future models, how quickly that contribution could grow, or whether gains on benchmark tasks translate into faster model development. It also does not quantify when available compute would become binding or how much new energy and data-centre capacity could be added. The reported construction rate is not itself proof of an imminent shortage.
The source identifies a conflict of interest disclosure: Meyer says the article was researched and drafted with AI assistance, including Anthropic’s Claude. He says Anthropic, OpenAI and METR did not review or sponsor it. No independent review of the article’s projections is reported in the supplied material.
Signals to Watch Through 2027
The scenario can be tested against future measurements of agent performance, especially whether task horizons keep rising and whether the pace holds across a broader range of tasks. Because METR’s public page was marked as no longer actively updated, readers will need to distinguish new, comparable benchmark results from projections based on its existing data. A sustained rise in measured task duration would support one part of the scenario, but would not by itself prove the full development loop.
Other signals include evidence that AI is materially speeding model research, changes in data-centre construction and power availability, and whether firms expand AI use beyond current reported levels. The article gives no confirmed milestone showing that the proposed 2027 acceleration or compute ceiling has arrived. The next step is evidence, not a settled timetable: future results will show whether the feedback loop strengthens, slows or is overtaken by infrastructure limits.
Key Questions
Is Meyer predicting that artificial superintelligence will arrive in 2027?
He presents a possible scenario in which it arrives in 2027 or shortly after, but explicitly says the article is not a forecast presented as fact. The term “artificial superintelligence” also has no agreed definition.
What does METR’s 80% time horizon measure?
It estimates the duration of a task, measured by human expert completion time, at which an AI agent is predicted to succeed four times out of five. The cited tasks are mainly in software engineering, machine learning and cybersecurity.
Does the benchmark prove that AI progress will accelerate sixfold?
No. The sixfold figure is part of Meyer’s scenario, while his article cites a METR estimate of about twice the earlier rate for Anthropic’s improvement on well-specified tasks. The benchmark does not establish a sixfold acceleration in overall AI research.
What could limit AI progress in this scenario?
Meyer identifies physical infrastructure, including chips, data-centre buildings and power. The timing and severity of any constraint remain unclear, and the article does not establish that a shortage is imminent.
Source: ThorstenMeyerAI.com
