🔍 Read the full analysis: Scenario 2027: How Far Can AI Progress Before Compute Runs Out? on ThorstenMeyerAI.com
Get hardware and tech essentials 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, 2026 article lays out a scenario in which AI helps build more capable AI, compressing a year of progress into about two months before compute constraints slow the cycle. METR measurements show rapid gains in software-task duration, but the benchmark is limited, its public page is no longer actively updated, and the evidence does not establish that a 2027 acceleration will occur.
Thorsten Meyer published a 2027 AI-progress scenario on November 18, arguing that AI systems could help accelerate the development of later systems until limits on chips, data-centre construction and power slow the cycle. The article presents this as a scenario, not a forecast established by current evidence; its central claim—that a year of progress could arrive in roughly two months—remains uncertain.
The scenario describes a chain of events: AI systems take on more of the coding, testing and research involved in building future models; that work speeds up development; and eventually expansion runs into physical limits because new compute requires chips, buildings and electricity. Meyer then proposes a slower phase of what he calls embodied self-improvement, in which automation in factories, robotics and laboratories could help expand capacity. Each link is conditional, and the article does not show that the full chain is already happening.
As evidence of improving software-agent capability, Meyer cites METR’s 80% time horizon: the duration of software tasks, measured by estimated human completion time, that an AI agent is predicted to finish successfully four times out of five. In the selected measurements, the horizon rose from 0.89 minutes for GPT-4 in March 2023 to 185.9 minutes—about 3 hours and 6 minutes—for an early Claude Mythos Preview in April 2026. These are task-benchmark results, not measurements of overall AI ability or the pace of economic change.
The article’s fitted trend lines produce very different dates. Meyer’s fit to measurements since 2024 reaches a 16-hour horizon in January 2027, while a fit using data from 2019 reaches that point in June 2028. A hyperbolic fit reaches infinity in November 2027, but Meyer says it performed worst in his backtests, having already projected an infinite horizon by January 2, 2026 when fitted to earlier models. METR marked its public page as no longer actively updated on September 8, 2026, adding uncertainty to the availability of comparable future measurements.
Why Compute Could Set the Pace
The scenario matters because it links two questions that are often discussed separately: whether AI can speed up the work of developing AI, and whether the physical infrastructure can keep up. If software-based research accelerates faster than data centres and electricity supply expand, capability gains could encounter a bottleneck that software alone cannot remove. If automation also helps build or operate that infrastructure, the constraint could shift over time—but the article treats that as a possibility, not an established outcome.
The timing also affects how readers interpret rapid benchmark gains. A longer software-task horizon could indicate that agents handle more extended work, but it does not by itself show that they can independently build frontier models, replace workers at scale or produce a sudden economy-wide transformation. Meyer cites figures 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 business use and reported job reductions over that period; they do not settle how adoption may change.
Products Worth Considering

Logitech MX Mechanical Wireless Illuminated Keyboard, Tactile Quiet, Graphite

Ledger Flex Crypto Wallet – Neptune Blue

Logitech MX Creative Console with 9 Customizable Keys, Graphite
As an affiliate, we earn on qualifying purchases.
What the Benchmark Measures
METR is a nonprofit that evaluates frontier AI agents. Its time-horizon measure estimates how long a task would take a human expert, then estimates the task duration at which an agent succeeds at a specified reliability level. The 80% measure used in Meyer’s comparisons is stricter than a 50% success horizon, which can yield longer task durations for the same model.
METR’s task suite primarily covers software engineering, machine learning and cybersecurity. METR says these tasks are cleaner and more clearly specified than real economically valuable work, and that a two-hour benchmark task is more like work assigned to a new hire or contractor than to an experienced colleague. It also warns that measurements above 16 hours are unreliable with its current suite. Of 31 tasks estimated to take humans at least eight hours, the source material says human completion times were measured for only five.
Meyer also cites a METR estimate that Anthropic is improving models on well-specified tasks at around twice the earlier rate, while METR said it would not be surprised by a rate between 1.3 and 3 times as fast. A sixfold acceleration would be needed to turn a year of progress at the earlier pace into two months. These figures are not equivalent: the cited estimate does not establish a sixfold acceleration. The article also reports that OpenAI’s GPT-6.2 Astra became the first model released externally by the company to meet its “High” threshold for AI self-improvement; that is OpenAI’s stated classification, not independent evidence that recursive self-improvement is under way.
“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
Evidence Behind the 2027 Timing
The timing depends on which historical data points are used and how they are fitted. Meyer’s article says the faster exponential fit reaches 16 hours in January 2027, while a longer-run fit points to June 2028; neither date is a measured result. The hyperbolic curve’s November 2027 endpoint is particularly weak evidence because Meyer reports that the same method performed poorly in backtesting.
METR’s measure covers a limited set of software tasks, includes estimated human task durations and has stated reliability limits beyond 16 hours. Its public page was no longer actively updated as of September 8, 2026, according to the article. The source material does not establish whether later comparable measurements will be released, how much AI systems currently contribute to frontier-model research, or whether reported gains translate into faster model development.
Infrastructure is another open question. Meyer cites US data-centre construction running at a September 2026 annual rate of $96.5 billion, compared with $50.7 billion a year earlier, and says Epoch AI expects power to be the first scaling constraint, with new power supply taking three to five years from a decision. The article does not specify a common measure of compute demand against that construction spending, nor does it establish when or where electricity limits will bind.
Signals to Watch in 2027
The scenario can be tested against future evidence rather than its headline date alone. Relevant signals include whether new, comparable benchmark measurements continue to extend agents’ reliable task duration; whether independent assessments find that AI systems are doing more of the research and engineering used to develop later models; and whether the measured pace of improvement approaches the roughly sixfold acceleration implied by the two-month claim.
Readers can also watch for data on data-centre construction, chip availability and electricity connections, alongside evidence of expanded automation in laboratories and factories. Meyer’s article does not provide a confirmed timetable for such developments. Until those signals are documented, the 2027 timeline remains a scenario to test, not an established prediction.
Key Questions
Is the article predicting that artificial superintelligence will arrive in 2027?
No. Meyer describes a scenario and personal bet, not a forecast established by current evidence. The article also notes that artificial superintelligence has no agreed definition and focuses on the proposed mechanism rather than the label.
What does METR’s 80% time horizon measure?
It estimates the length of software tasks, measured by human completion time, that an AI agent is predicted to complete successfully four times out of five. METR’s task suite is limited and does not directly measure general economic productivity.
Does the benchmark prove that AI is improving itself?
No. The benchmark tracks agent performance on selected tasks. Evidence that agents can complete longer software tasks does not by itself establish that they are independently developing future AI systems or accelerating that process at the rate proposed in the scenario.
What could constrain AI progress even if models improve quickly?
The scenario points to chips, data-centre buildings and power supply. Meyer cites an estimate that new power supply can take three to five years from a decision, but the timing and severity of any constraint remain uncertain.
Source: ThorstenMeyerAI.com
