🔍 Read the full analysis: Scenario 2027: Could A Compute Crunch Follow AI’s Rapid Advances? 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 systems could accelerate AI development in 2027, then run into limits on chips, data centres and power. Public evidence shows rapid gains on a narrow software-task benchmark, but the trend depends on the period measured, the benchmark has limits, and the scenario’s proposed sixfold acceleration has not been established.
Thorsten Meyer published a 2027 AI compute-crunch scenario on November 18, arguing that AI systems could speed up the work of developing future systems, compressing progress before hitting constraints in chips, data-centre construction and electricity supply. The article presents this as a scenario to test, not a confirmed forecast: available evidence shows faster performance on some software tasks, but does not establish that progress will accelerate sixfold or that a physical bottleneck will arrive in 2027.
The scenario has four steps: AI tools take on more model-building work; that work speeds up subsequent development; capability gains arrive faster than they can be matched by new computing infrastructure; and, later, automation in physical settings such as factories and laboratories helps expand capacity. Meyer calls this last stage embodied self-improvement. Each step could fail, and the article does not present the proposed timeline as a measured result.
A central indicator is METR’s 80% time horizon, an estimate of how long a software task an AI agent can complete with four-in-five reliability. METR’s recorded results rose from 0.89 minutes for GPT-4 in March 2023 to 185.9 minutes for an early Claude Mythos Preview in April 2026. These are benchmark measurements, not projections of general workplace performance. METR says the tasks are mainly in software engineering, machine learning and cybersecurity, and are cleaner and more specified than economically valuable work in real settings.
The article’s own curve fits produce very different dates. Fitting an exponential trend to METR points from 2024 gives a doubling time of 3.4 months and crosses a 16-hour task horizon in January 2027; fitting from 2019 gives a 6.4-month doubling time and crosses it in June 2028. Meyer also reports that a hyperbolic fit reaches infinity in November 2027, but says this was the worst-performing of five fits in backtests. METR’s public page was marked no longer actively updated on September 8, 2026, and the organization cautions that measurements above 16 hours are unreliable with its current task suite.
The Gap Between Software and Power
The question matters because software gains and physical capacity move on different schedules. AI systems may help design models, write code or run experiments, but those advances do not by themselves create additional chips, data centres or connected electricity supply. If demand for computing grows faster than infrastructure, the result could be slower deployment or competition for scarce capacity rather than an immediate halt to technical progress.
There are signs of substantial infrastructure investment, but they do not confirm a coming crunch. The source article reports that U.S. 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 could be the first scaling constraint and that new power supply can take three to five years after a decision. Those figures and estimates describe pressures and timelines; they do not show how much capacity will be available in 2027 or how quickly demand will rise.
For workers and businesses, a further distinction is between capability and adoption. The article cites estimates 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. A faster benchmark result would not, on its own, prove widespread deployment or job losses. Investment, reliability, integration and the availability of compute all affect how quickly capabilities translate into operational change.
Popular right now

Ninja Foodi Smart 6-in-1 Indoor Grill and Air Fryer, 4 Qt, FG551

Bose QuietComfort Headphones (2nd Gen), Wireless Noise-Cancelling Over-Ear Headphones, White Smoke

Vitamix 5200 Blender with 64 oz Container, Black
As an affiliate, we earn on qualifying purchases.
What the Benchmark Can Show
METR defines its time horizon as the task duration, measured by human expert completion time, at which an AI agent is predicted to succeed at a specified reliability. Meyer focuses on the 80% reliability measure; METR’s better-known 50% horizon is longer for the same model. The metric is useful for tracking a particular kind of agent performance, but it is not a general measure of intelligence or a direct forecast of economic impact.
METR has identified limits that matter to forecasts based on the series. Its tasks mostly concern software and related technical work, and it says a two-hour task is closer to work a new hire or freelance contractor could complete than to the work of an experienced colleague. The organization also warns that its current task suite cannot reliably measure horizons above 16 hours. 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 also compares acceleration claims with METR’s assessment of AI-assisted model development. According to Meyer’s account, METR estimates Anthropic is improving its models on well-specified tasks at about twice its earlier rate, while saying it would not be surprised by a rate between 1.3 and 3 times. That is distinct from the scenario’s sixfold acceleration, which would be needed for a year of progress at the old pace to arrive in about two months. The two figures should not be treated as equivalent evidence.
“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
The Unproven Steps to 2027
The largest uncertainty is whether improvement on METR’s software tasks predicts a broad acceleration in AI research and development. The benchmark covers a limited, relatively well-specified set of work, and its longer horizons are difficult to measure. The article’s different curve fits also show that the projected date changes substantially with the chosen starting period; none establishes that a particular threshold will be reached in 2027.
It is also unclear whether AI-assisted development is accelerating at the rate required by Meyer’s scenario. The reported METR estimate of roughly twice the earlier rate is below the proposed sixfold pace, and the article does not establish that this estimate will persist or apply across all stages of model development. Likewise, data-centre construction and expected power lead times do not reveal how much usable compute will be online, where it will be available, or whether efficiency gains could offset demand.
The source 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 an attributed company classification, not independent proof that AI systems can rapidly and autonomously build substantially more capable successors. The article also discloses that it was researched and drafted with AI assistance, including Anthropic’s Claude, and says Anthropic, OpenAI and METR did not review or sponsor it.
Signals to Watch in 2027
The scenario can be tested against future benchmark results, but readers should watch for transparent, repeatable measurements and clear explanations of task selection, reliability and human-time estimates. METR’s public results page was no longer being actively updated as of September 8, 2026, according to Meyer, so whether comparable data will continue to be published is an open question.
Other signals include evidence that AI systems are taking on more of the research and engineering work behind subsequent models, and whether independent evaluations show that this work is producing faster capability gains. On the infrastructure side, data-centre construction, power connections and new electricity supply will help show whether computing capacity is keeping pace with demand. These indicators could clarify the scenario’s separate links, but none alone would prove a 2027 turning point.
For now, the appropriate reading is a conditional scenario rather than a prediction. Meyer’s proposed outcome depends on software progress accelerating well beyond the roughly twofold rate attributed to METR, while physical infrastructure becomes a limiting factor and later expands. Which, if any, of those conditions arrives on the suggested timeline remains to be seen.
Key Questions
Is a compute crunch in 2027 confirmed?
No. The 2027 crunch is a scenario published by Thorsten Meyer, not a confirmed event or established forecast. The article itself sets out evidence against the scenario and identifies multiple steps that could fail.
What does METR’s 80% time horizon measure?
It estimates the duration of a task an AI agent can complete with 80% reliability, based on tasks timed by human experts. METR’s suite focuses mainly on software engineering, machine learning and cybersecurity, so the measure does not represent all work or general AI capability.
Why do the article’s projected dates differ?
Meyer’s fits use different periods of METR data. The fit starting in 2024 gives a 3.4-month doubling time and reaches a 16-hour horizon in January 2027; the fit starting in 2019 gives a 6.4-month doubling time and reaches it in June 2028. These are extrapolations, not observed outcomes.
What could limit AI development even if models improve quickly?
Physical infrastructure could constrain how quickly additional AI capability can be deployed, including chips, data-centre space and electricity supply. The source article cites a three-to-five-year timeline for new power supply after a decision, but that estimate does not establish a specific shortage in 2027.
Source: ThorstenMeyerAI.com
