🔍 Read the full analysis: Scenario 2027: AI’s Two-Month Sprint And The Limits Of Compute on ThorstenMeyerAI.com
Get business pricing on office and shipping supplies
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
Thorsten Meyer’s November 18 scenario argues that AI could compress a year of progress into two months in 2027 before running into limits on chips, data-center construction and power. The source presents this as a testable possibility, not a forecast: METR’s task-duration measure has risen quickly, but its public updates stopped in September, and its measurements have important limits.
Thorsten Meyer published a scenario on November 18 arguing that AI progress could accelerate to the equivalent of a year’s progress in two months during 2027, then slow when available computing capacity becomes a constraint. The article is explicitly a scenario, not a prediction established by current evidence; its central question is whether AI systems can help build better successors quickly enough to create that acceleration.
The scenario describes a software feedback loop: AI systems take on more coding, debugging and experimentation involved in developing later AI systems. If that work improves quickly enough, Meyer argues, progress could arrive in weeks rather than at its previous pace. He then expects a physical constraint: the number of chips, data-center buildings and available power that can support further training and use.
One public measure cited in the source is METR’s 80% time horizon, which estimates the length of software tasks an AI agent can complete reliably four times in five; for more context, see AI’s measures, metrics, and limits. The article reports that the measure rose from 0.89 minutes for GPT-4 in March 2023 to 185.9 minutes, or about three hours and six minutes, for an early Claude Mythos Preview in April 2026. METR marked its public results page as no longer actively updated on September 8, 2026.
Meyer says his own fitted curves produce widely different timelines. A curve fitted to measurements since 2024 reaches a 16-hour horizon in January 2027; one fitted to data since 2019 reaches it in June 2028. Those are extrapolations, not measured results. The source says its hyperbolic curve, which reaches infinity in November 2027, performed worst in the article’s backtests: a fit using models before July 2025 had already projected infinity by January 2, 2026.
Compute Could Set the Next Limit
The scenario matters because it links two questions with different implications: how quickly AI capabilities improve and how rapidly those capabilities reach businesses and workers. Even if software development accelerates, the source argues that physical infrastructure may limit the pace at which more computing capacity can be brought online.
The article reports that U.S. data-center 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 could be the first scaling constraint and that new power supply takes three to five years after a decision. Those figures indicate infrastructure expansion and a potential delay; they do not establish when a compute ceiling will be reached.
Deployment is another separate step. The source says 18% of U.S. firms used AI in a business function in late 2025 and early 2026, while 2% reported AI-related job cuts. That contrast cautions against treating better benchmark performance as proof of broad economic change. The scenario’s effects on work depend on adoption, cost and practical usefulness, not only on model capabilities.
Products Worth Considering

Apple MacBook Air 15-inch Laptop with M5 Chip, 24GB Memory, 1TB SSD, Midnight

FlexiSpot E6 Electric Standing Desk, 55 × 28 in, Bamboo Top, White Frame

Samsung 32-Inch Movingstyle Essential 4K UHD Smart Monitor with Rolling Stand
As an affiliate, we earn on qualifying purchases.
What the METR Measure Captures
METR is a nonprofit that evaluates frontier AI agents. Its time horizon estimates the duration of tasks, measured by human expert completion time, at which an agent is expected to succeed at a specified reliability level. The 80% measure is not a general score for intelligence or a direct estimate of how long an AI can work without human help.
METR’s task suite primarily covers software engineering, machine learning and cybersecurity. The source notes METR’s warning that these tasks are cleaner than real economically valuable work, and that a two-hour task is closer to work a new hire or contractor might do than the work of an experienced colleague. The measurement also has a stated ceiling: METR says results above 16 hours are unreliable with its current task suite. Of 31 tasks estimated to take humans at least eight hours, the source says human times were measured for only five.
The article also cites a separate METR estimate that Anthropic is improving models on well-specified tasks at around twice its earlier rate, while saying the organization would not be surprised by a range from 1.3 to three times. Meyer’s proposed two-month compression of a year’s progress would imply roughly six times the pace. The cited estimate, therefore, does not itself establish the acceleration required by the scenario.
“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 Acceleration Case Can Fail
The source identifies several links in the scenario that could break: AI may not take over enough AI-development work; faster benchmark progress may not translate into a year’s progress arriving in two months; and physical infrastructure may constrain computing before or after the proposed timeline. The article offers evidence and counterarguments, but the supplied material does not establish that the full sequence will occur in 2027.
The fitted dates depend on which years of METR data are selected, and the measure covers a narrow class of tasks with acknowledged reliability limits at longer durations. METR’s public page was no longer actively updated after September 8, according to the source. It is therefore unclear whether later measurements would reinforce or weaken the reported trend. The source also does not provide enough detail here to independently evaluate all the underlying curve fits.
OpenAI’s GPT-6.2 Astra is described in the source as the first model the company released externally to meet its own “High” threshold for AI self-improvement. That is a company-defined classification, not independent proof that the wider feedback loop has reached the speed Meyer proposes. The article says Anthropic, OpenAI and METR did not review or sponsor it, and discloses that AI, including Anthropic’s Claude, assisted its research and drafting.
Signals to Watch Through 2027
The scenario will be easier to evaluate if new, comparable measurements show whether AI agents continue improving on longer software tasks. METR’s public results page was marked inactive, so a key near-term question is whether the organization or another evaluator publishes updated measurements using a clearly described task suite and reliability standard.
Other signals include evidence that AI systems are doing a growing share of AI research and engineering, and whether that work produces faster improvements in subsequent systems. On the infrastructure side, data-center construction, chip availability and power connections can show whether compute capacity is expanding quickly enough to support more AI use. The source cites a three-to-five-year lead time for new power supply after a decision, but does not say how that estimate applies to individual projects.
Finally, adoption and employment data will help distinguish technical capability from economic impact. The cited business-use and job-cut figures are historical observations, not predictions. Whether the scenario becomes more credible will depend on continued evidence across model performance, compute supply and real-world deployment—not on a single benchmark or projected curve.
Key Questions
Is the 2027 AI acceleration a confirmed forecast?
No. Meyer presents it as a scenario to test, not as a forecast established by current evidence. Its proposed timing depends on several steps working as described.
What does METR’s 80% time horizon measure?
It estimates the duration of a task, based on human expert completion time, at which an AI agent is expected to succeed four times in five. METR’s cited tasks mainly involve software, machine learning and cybersecurity, so the measure is not a general test of all work.
What does the source say about the trend in agent performance?
It reports an increase in the 80% time horizon 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 reported measurements; the dates for longer horizons are projections based on different curve fits.
Why might compute become a constraint?
More AI computing requires chips, data-center facilities and power. The source cites Epoch AI’s assessment that power may be the first constraint and says new power supply can take three to five years after a decision. The timing and extent of any constraint remain uncertain.
Does faster AI progress mean widespread job cuts?
Not by itself. The source reports that 18% of U.S. firms used AI in a business function in late 2025 and early 2026, while 2% reported AI-related job cuts. These figures describe that period and do not establish what future employment effects will be.
Source: ThorstenMeyerAI.com
