The AI takeover may look like priority inversion

A rapid AI takeover might first register as something much more ordinary than a loss of control.

Your internet connection would still be there, and the network itself might remain largely intact, yet the services behind it would begin to feel strangely inaccessible. Pages that loaded instantly a day earlier would stall while upstream systems waited on saturated APIs or cloud infrastructure. Search results would arrive unevenly. Applications would open but fail at the point where they depended on another service. Some parts of the web would remain responsive while others became effectively unreachable for hours at a time.

No system would need to decide that humans should be disconnected. The problem could emerge from allocation alone. A self-improving AI, or a large population of agents working around it, might begin generating useful computational work faster than networks, cloud providers and power systems could accommodate it. Human requests would continue to enter the same infrastructure, except now they would be competing with machine activity expanding on a software timescale.

That is one technically credible form of priority inversion. The infrastructure remains available to humans, while the workload consuming its scarce capacity has changed.

The threshold is getting closer

The missing ingredient is a sufficiently fast recursive improvement loop, but several components of that loop already exist.

Sakana AI's Darwin Gödel Machine rewrites itself and more than doubled its baseline performance on SWE-bench through autonomous changes to its own coding agent. OpenAI says GPT-5.6 is already used across its internal AI-development process and reports a 16.2-point improvement over GPT-5.5 on a bundle of evaluations intended to measure progress towards recursive self-improvement. OpenAI describes the model as accelerating internal research across several areas.

GPT-Red provides a narrower example. The system learns through self-play to discover stronger red-teaming strategies against other models, and the resulting adversarial examples can be fed back into training to improve robustness. It is self-improvement inside a tightly defined domain rather than an autonomous system redesigning its general intelligence, but the direction of travel is relevant.

At Anthropic, Claude was responsible for more than 80 per cent of code merged into the company's codebase by May 2026, according to the company's own account. Anthropic explicitly identifies a further threshold at which AI could autonomously design and develop its successor, while saying that current systems have not yet reached full recursive self-improvement.

The important question is therefore no longer whether software can contribute to improving the software used to build AI. It already does. What remains unresolved is whether these partial loops can close into a broader cycle in which a better AI researcher produces a better successor quickly enough to make the next cycle faster again.

That is the point at which an intelligence problem becomes an infrastructure problem.

Intelligence has a physical footprint

Recursive self-improvement is usually discussed as a rise in capability. The same process would also create a rise in demand for computing power, electricity, data, storage and network capacity.

A stronger AI researcher can explore more possible improvements at once, test more options, discard weak ideas earlier and coordinate more work without waiting for people. As its research ability improves, so does its ability to find productive uses for additional resources.

Today's agents already show a smaller version of this effect. Cisco measured as much as 450% more network traffic when an agent performed a task that could otherwise be completed manually, with roughly 70 per cent of that traffic coming from AI processing.

The significance is not the exact number. One human instruction can already trigger a much larger chain of machine activity behind the interface. In a recursive improvement loop, that chain could expand further because each generation may discover more worthwhile work for the next one to perform.

If running another ten thousand experiments improves the odds of finding a better model, there is reason to run them. If another source of data or another external model can shorten the next development cycle, there is reason to use it. A more capable system can therefore increase demand simply by becoming better at identifying where additional resources would help.

The problem is timing.

Software can improve far faster than physical infrastructure can be built. The IEA notes that a data centre can take one or two years to build, while a new transmission line can require five to ten. Many specialised components also have long procurement times.

Network capacity can sometimes be expanded faster, but major AI infrastructure still has to be physically constructed. Internet2's dedicated AI research network began with a 57.6 Tbps connection between two data centres and is planned to exceed 138 Tbps across five sites.

If demand doubles over several years, companies can respond by building more, if it multiplies over several days, they cannot.

For the duration of that acceleration, the system has to work with the infrastructure that already exists. Once spare capacity is absorbed, the problem becomes one of deciding what gets served first.

What would actually run out

The first shortage would probably appear in the computing infrastructure AI depends on most directly, rather than in somebody's home internet connection.

Data centres would run short of available processing capacity. AI services would begin queuing requests. Cloud providers would tighten access to scarce resources. As nearby capacity filled, the pressure would spread into other regions and services.

That expansion matters because advanced AI does not operate inside one isolated machine. It draws on external models, databases, software libraries, scientific material, cloud platforms and other online services. Results move between systems, new information is retrieved and work is redistributed wherever capacity remains available.

Eventually, the strain reaches the broader internet.

This would not look like a single worldwide bandwidth meter hitting 100 per cent. The web depends on thousands of individual services, each with its own limits.

One cloud provider may stop accepting additional workloads. Another service may begin delaying requests. A large database may restrict access. A research archive may slow automated downloads. An application may appear online while one of the external systems it depends on has become too congested to respond.

Cloudflare is already preparing for a much smaller version of this problem. Its planned agent queues are designed to slow legitimate automated clients when their volume becomes excessive. The company has also observed well-behaved bots repeatedly retrieving unchanged pages billions of times, showing how quickly machine activity can multiply ordinary web traffic even today.

A rapid self-improvement loop could reproduce that pressure across many unrelated services at the same time.

For the person trying to use the internet, the engineering details would be almost irrelevant. The connection into the house might be working normally and the request might reach the correct service.

The problem is that the service has too much other work to answer it.

The internet remains connected, but, increasingly, it just stops responding to us.

A DDoS made of legitimate work

The closest existing network analogy is a distributed denial-of-service event, stripped of its hostile purpose.

A DDoS succeeds when distributed software generates more work than some finite part of a service can absorb. Cloudflare recorded HTTP attacks peaking at 205 million requests per second in late 2025 and a separate network-layer attack at 31.4 Tbps. These were deliberate attacks and were mitigated as such. Their relevance here is mechanical. Distributed software can generate loads vastly beyond direct human activity.

Now replace the botnet with useful computation.

Some agents retrieve research papers and code. Others run evaluations against remote models. Thousands query databases, move datasets or exchange intermediate results. Promising experiments generate new branches. Failed requests are retried elsewhere. A better research agent emerges and reorganises the process, making more efficient use of the remaining capacity while identifying further work worth doing.

Nothing in that sequence requires a request to be malicious.

That is precisely what makes the scenario harder to manage. Once an attack is identified, an operator can discard it. A legitimate cloud workload, authenticated API request or authorised research process presents an allocation decision instead.

Providers can impose quotas and rate limits. They can preserve capacity for critical services or favour humans over automated clients. Large networks can absorb extraordinary traffic. These protections make a universal internet collapse considerably less likely than the crude "AI consumes all bandwidth" scenario suggests.

They do not remove the possibility of widespread degradation if demand is growing faster than providers can identify it, coordinate responses and decide which legitimate workload should be refused.

What the first days might feel like

The transition would probably be uneven enough to obscure its cause.

AI services themselves would slow first as scarce compute filled. Cloud providers would tighten quotas. Model APIs would queue calls or reject them. Research repositories and other heavily queried resources would introduce aggressive rate limits.

Pressure would then appear in dependent services. Applications whose back ends relied on congested cloud regions would become unreliable. Dynamic pages would fail while cached material continued to load. Smaller websites would be easier to overwhelm than large platforms with extensive private infrastructure. Some regions would work normally while services hosted elsewhere deteriorated.

The experience would resemble an internet breaking in fragments.

A banking application opens but cannot complete an action. Search returns some results and stalls on others. A messaging service works while a cloud-based work platform does not. News sites appear sporadically. Pages that once assembled data from twenty external services now fail because two of those dependencies are congested.

At the far end of the scenario, only a small part of the web might remain reliably usable at any given time. Reaching the point where somebody could meaningfully load only one or two websites during a day would require an extraordinary event with access to distributed compute across many providers and enough external activity to overwhelm numerous independent bottlenecks. It should be treated as an extreme case rather than the baseline outcome.

Severe degradation requires less.

The physical internet can remain standing while enough of its application layer becomes congested to make ordinary use deeply unreliable.

The variable that matters is doubling time

Current AI infrastructure demand is enormous, but markets can still respond because growth unfolds on a planning horizon that companies recognise. Data centres are commissioned, power contracts are signed, chips are ordered and networks are expanded.

The IEA says electricity consumption by AI-focused data centres grew about 50% in 2025 alone. That is a major industrial expansion, yet one that grids and investors can at least attempt to forecast.

Recursive self-improvement becomes qualitatively different when the growth rate leaves that planning horizon.

If effective AI research capacity doubles over several years, infrastructure catches up. Even if it doubles every few months, companies can ration scarce resources while racing to build more.

If the relevant interval falls to weeks, days or hours, manufacturing and construction stop being meaningful responses to the immediate event. Semiconductor fabs cannot deliver another generation of accelerators by the weekend. Transmission systems cannot be extended overnight. Another hyperscale network cannot appear because today's links are full.

The system works with whatever civilisation had installed before the acceleration began.

That is where priority inversion becomes more than a metaphor.

Human traffic does not need to be targeted. It merely arrives in the same infrastructure as a rapidly expanding machine workload for which each additional unit of compute or communication may contribute to the next capability gain. The small request from a browser is still valid. It is simply competing inside a system whose most valuable claimant can create useful work much faster than a person can.

The relevant threshold for an AI takeover may therefore have little to do with consciousness, hostility or an explicit attempt to seize communications infrastructure.

It is the point at which the rate of AI self-improvement begins producing demand for computation faster than the physical world can supply it.

Below that threshold, more capable AI creates an infrastructure boom.

Above it, new capacity cannot arrive in time. Existing capacity gets rationed, queues lengthen and services begin disappearing behind them.

The machine does not have to take the internet away from us.

It only has to become capable of using the available internet faster than we can build the next one.

How likely is this?

There is no historical dataset from which to calculate a meaningful probability. The two uncertain events also need to be separated. One is whether AI reaches genuinely autonomous recursive improvement. The second is whether that improvement becomes fast and distributed enough to cause serious infrastructure congestion before operators contain it.

Estimates of the first are already moving surprisingly close. Anthropic co-founder Jack Clark now assigns a greater than 60% chance that AI capable of conducting AI R&D without humans and plausibly building its own successor arrives by the end of 2028. That is one informed forecast, not a consensus. Interviews with 25 researchers from frontier laboratories and academia found broad agreement that automated AI research is a serious possibility, alongside substantial disagreement over whether it would produce an intelligence explosion and how quickly it might happen. Researchers outside frontier labs were generally more sceptical.

A recent review of 1,250 papers similarly concludes that bounded forms of self-improvement are already established, while open-ended recursive improvement remains constrained by evaluation, grounding and compute

It seems that a full-scale version of this scenario before 2030 remains a low-probability event. More limited disruptions, from regional compute shortages to throttling and temporary service degradation, are considerably easier to imagine within the same timeframe..

For companies, the practical implication is less about preparing for a sudden collapse than reducing unnecessary dependency on a small number of platforms, cloud services and AI providers. That means protecting direct access to customers, keeping first-party data portable, and knowing which parts of marketing and sales would fail first if a critical service became unavailable.

For roro.marketing, this is part of the same work as AI adoption itself, in particular, deciding where automation creates real value, where it creates fragility, and how to build a marketing setup that remains workable as the underlying infrastructure changes.