AIThis post was created with the assistance of artificial intelligence (AI).

🔍 Read the full analysis: AI’s 2027 Speed-Up: How Long Before Compute Runs Out? on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on office and shipping supplies

  • Business-only prices and quantity discounts
  • Tax-exempt purchasing
  • Multiple users, one account, clear invoices
As an affiliate, we earn on qualifying purchases.

TL;DR

Thorsten Meyer’s November 18 scenario argues that AI systems could help build better AI quickly enough to compress a year of progress into about two months in 2027, before compute constraints slow the cycle. That is a scenario, not an established forecast: benchmark trends depend on the period measured, the key public benchmark is no longer actively updated, and power and deployment remain major uncertainties.

Thorsten Meyer published a scenario analysis on November 18 arguing that AI progress could accelerate sharply in 2027 as AI systems take on more of the work involved in developing future systems, then slow when chips, buildings or electricity become constraints. The article presents this as a possible scenario, not a forecast established by current evidence; its central claims depend on how benchmark trends continue and whether technical progress translates into broader deployment.

Meyer’s proposed sequence has four links: AI systems help with coding and experimentation for new models; that work speeds up successive development; available computing capacity becomes a bottleneck; and later improvements in physical infrastructure and automation allow deployment to expand again. Meyer calls the proposed physical-world feedback loop embodied self-improvement. Each link could fail, and the source does not establish that the full sequence will occur in 2027.

The article uses METR’s 80% time horizon as a public measure of agent performance on software tasks. The measure estimates the length of a task, in human expert completion time, at which an AI agent succeeds four times in five. Selected record-setting measurements in the source rise from 0.89 minutes for GPT-4 in March 2023 to 185.9 minutes, or about three hours, for an early Claude Mythos Preview in April 2026. These are benchmark observations, not projections of general workplace capability.

Meyer’s own curve fits produce sharply different dates. An exponential fit using measurements since 2024 doubles about every 3.4 months and reaches a 16-hour horizon in January 2027; a fit using data from 2019 doubles about every 6.4 months and reaches that point in June 2028. The article says the more dramatic hyperbolic fit performed worst in its backtests, reaching infinity by January 2, 2026, when fitted to models released before July 2025. METR’s public page was marked no longer actively updated on September 8, 2026, according to the source.

At a glance
analysisWhen: Published November 18, 2026; the 2027 s…
The developmentThorsten Meyer published a scenario analysis testing whether AI-driven research could sharply accelerate AI progress in 2027 before physical compute constraints intervene.

Compute Limits Could Slow AI Gains

The scenario matters because it connects two questions often treated separately: how quickly AI capabilities may improve and how quickly the physical infrastructure needed to run them can grow. If AI tools substantially speed up model research, the pace of software progress could put pressure on available chips and data-centre capacity. If power supply and construction cannot keep pace, capability gains may not translate into equivalent near-term increases in usable compute.

The source points to US data-centre construction running 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 expectation that power could be the first scaling constraint, with new power supply taking three to five years from a decision. Those figures indicate infrastructure investment and a potential lead time; they do not show that a specific shortage will occur in 2027.

There is also a difference between technical capability and economic effect. The source reports 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 measures do not settle how AI will affect productivity, employment or adoption in 2027. They do caution against treating a fast-moving benchmark as proof that the same pace of change is already reaching most businesses.

Products Worth Considering

Apple MacBook Neo 13-inch Laptop with A18 Pro Chip, 8GB Memory, 512GB SSD, Liquid Retina Display, and Touch ID – Blush

Apple MacBook Neo 13-inch Laptop with A18 Pro Chip, 8GB Memory, 512GB SSD, Liquid Retina Display, and Touch ID – Blush

Canon PIXMA MegaTank G3270 Wireless All-in-One Color Inkjet Printer

Canon PIXMA MegaTank G3270 Wireless All-in-One Color Inkjet Printer

As an affiliate, we earn on qualifying purchases.

What METR’s Benchmark Tracks

METR is a nonprofit that evaluates frontier AI agents. Its time-horizon measure estimates the human-expert duration of tasks an agent can complete at a specified reliability. The tasks in the cited benchmark primarily involve software engineering, machine learning and cybersecurity, rather than a broad sample of paid work. METR also cautions that its tasks are cleaner than economically valuable work in real settings and that a two-hour task is closer to work a new hire or contractor could perform than an experienced colleague’s work.

The benchmark has measurement limits near the longer task durations relevant to the scenario. The source says measurements above 16 hours are unreliable with the current task suite; among 31 tasks taking humans at least eight hours, METR measured human completion times for only five and estimated the rest. The benchmark can show changes on its task set, but does not directly measure the time needed to build a new frontier model or predict when AI will reach a general intelligence milestone.

Meyer says METR estimates Anthropic is improving its models on well-specified tasks at about twice the earlier rate, while allowing that a rate between 1.3 and three times would not surprise it. The article contrasts that estimate with the scenario’s proposed sixfold acceleration: delivering a year of progress at the old pace in two months. That gap is a key uncertainty, not evidence that the sixfold rate has been achieved.

Where the 2027 Case Could Break

No agreed definition of artificial superintelligence is provided, and Meyer says the scenario does not depend on resolving that label. More directly, the data cited do not establish that AI development will accelerate sixfold, that AI systems will soon do enough research work to drive that acceleration, or that the resulting gains will arrive on the proposed timetable.

The projected dates depend on which benchmark points are included and which curve is fitted. The longer-history fit and the post-2024 fit yield different dates, while the hyperbolic model’s poor backtest argues against treating its apparent 2027 endpoint as reliable. METR’s discontinued active updates also leave open what comparable, public measurements will be available to test the trend. The source does not explain how its benchmark projections would map to training runs, compute demand or performance across tasks outside software work.

Infrastructure and adoption are uncertain too. The cited construction rate does not specify how much new capacity will be available to AI systems, and a three-to-five-year power-supply lead time is not a prediction of a particular shortage. The article does not establish whether embodied automation will become capable enough to speed up factories or research labs, or whether businesses will adopt new systems quickly. The timing and scale of any compute ceiling remain unknown.

Signals to Watch Through 2027

The scenario will be easier to assess if new, comparable evaluations show whether agents can reliably complete longer and more varied tasks. Readers should distinguish measured benchmark results from extrapolated curves, check what tasks and reliability thresholds each result covers, and note whether the underlying measurement method has changed. A return of regular public updates from METR or another evaluator could help establish whether gains persist beyond the data cited here.

Other signals include the pace of new data-centre construction, access to chips and connected power, and evidence that AI systems are materially contributing to model research rather than simply assisting with selected tasks. Business adoption and workforce data can help show whether capability changes are affecting operations. The source sets out no confirmed milestone that would prove the 2027 scenario; the next test is whether the claimed acceleration appears in new evidence while infrastructure and real-world use keep pace.

Key Questions

Is AI expected to run out of compute in 2027?

No. The 2027 date is a scenario advanced by Thorsten Meyer, not a confirmed prediction. The source identifies chips, data-centre capacity and electricity as possible constraints but does not establish when or whether a shortage will occur.

What does METR’s 80% time horizon measure?

It estimates the duration, measured by human expert completion time, of tasks an AI agent can finish with 80% reliability. The cited tasks mainly cover software engineering, machine learning and cybersecurity, so the measure is not a direct test of general workplace capability.

Why do the article’s projections give different dates?

Meyer fitted different growth curves to different portions of METR’s historical measurements. The fit using data since 2024 reaches a 16-hour horizon in January 2027, while the fit using data since 2019 reaches it in June 2028. These are model-based projections, not measured future results.

What evidence would support the scenario?

Comparable evaluations showing sustained gains on longer tasks, evidence that AI systems are speeding up model research, and data on available chips, data-centre capacity and power would help test the proposed chain. Benchmark growth alone would not confirm that broader deployment or a compute ceiling is imminent.

Source: ThorstenMeyerAI.com

This content is for general information only and is not financial, tax or legal advice. Consult a qualified professional for decisions about your money.
You May Also Like

When AI Builds Itself: Inside Anthropic’s Evidence on Recursive Self-Improvement

Anthropic presents data suggesting AI is increasingly capable of automating its own development, raising questions about recursive self-improvement.

Pltr Stock

Palantir’s stock (PLTR) increased significantly following its latest earnings release, signaling investor confidence despite mixed financial results.

What Does A 5.7% AI Gain Mean When The Price Doubles?

Fable 5 costs twice as much as Opus 4.8 and scores 5.7% higher on one benchmark, while public productivity evidence remains limited.

AI’s Future In 2027: 10 Things I Expect, With Confidence Ratings

Thorsten Meyer forecasts agents, AI-written code, persistent memory and power limits in 2027, with confidence ratings averaging 72.5%.