Daniel sent us this one — he's been thinking about whether a city or a country can actually run out of internet. Not in the sense of a data cap on your phone plan, but physically — can the cumulative demand from everyone streaming and gaming and video-calling at once exceed the total supply? And he's asking whether throttling at peak times is a sign we're hitting some fundamental limit in the fiber, or if it's a byproduct of something else entirely. It's the kind of question that sounds like stoner philosophy until you realize it's actually a physics problem.
It's absolutely a physics problem. And the short answer is — no, a city can't run out of internet like a reservoir running dry, but yes, it can absolutely experience localized, temporary exhaustion of the shared bits of the network. The distinction matters enormously.
We're not going to wake up one morning and find the internet has a "sold out" sign on it.
But you might try to stream a 49ers playoff game and discover your gigabit fiber connection has turned into a slide show. Which actually happened in San Francisco in 2023 — a neighborhood on a fiber-to-the-home network that normally runs at about thirty percent utilization spiked three hundred percent during the game, and for about fifteen minutes, service degraded noticeably. That's not a backbone failure. That's a last-mile contention issue.
The internet doesn't run out — but your street can.
And to understand why, we have to start with what "running out" would even mean technically. The question touches on at least three different bottlenecks — the physics of the glass, the economics of deployment, and the oversubscription model that makes consumer internet affordable in the first place. And most coverage conflates all three.
Let's untangle them then. Where does the actual physical ceiling sit?
The fundamental ceiling is something called the Shannon-Hartley theorem. It says there's a hard mathematical limit to how much information you can push through any channel given its bandwidth and signal-to-noise ratio. It's not infinite. But for modern optical fiber, it's astronomically high — high enough that for most of the internet's history, we weren't even in the same zip code as the limit.
Now we're getting close?
We're getting close enough that researchers have been using the phrase "capacity crunch" since about 2010. Here's the numbers. A single-mode fiber — that's the standard glass strand in pretty much every long-distance cable — can theoretically carry about a hundred terabits per second on a single wavelength. Then you use wavelength-division multiplexing, WDM, which is basically sending different colors of laser light down the same fiber simultaneously. Modern systems use eighty-plus wavelengths per fiber pair. Multiply it out and you're looking at eight to ten petabits per second per fiber pair.
A petabit being...
A million gigabits. So one fiber pair could theoretically handle something like two hundred fifty million simultaneous 4K Netflix streams. Per strand of glass the thickness of a human hair.
Which sounds like we should never, ever have congestion problems. So what's the catch? Where does the theory break against reality?
The catch is that nonlinear effects in the glass itself eat into that theoretical maximum. You've got something called the Kerr effect — the refractive index of the glass actually changes slightly with the intensity of the light, which distorts the signal. Think of it like shouting into a tunnel. At normal volume, your voice carries clearly. But if you shout loud enough, the air itself starts behaving differently — the sound waves begin interacting with each other, creating distortion. That's the Kerr effect in fiber. The light is so intense that it changes the properties of the medium it's traveling through.
The glass is fighting back against too much information.
In a sense, yes. And it gets worse. You've also got four-wave mixing, where different wavelengths interact and create ghost signals at new frequencies. Imagine three musicians playing different notes in a room, and the sound waves combine to produce a fourth note that none of them are actually playing. That ghost note is noise — it interferes with the real signal. And then you've got amplifier noise from the erbium-doped fiber amplifiers that boost the signal every eighty kilometers or so. These amplifiers are necessary because even the best glass isn't perfectly transparent — the signal attenuates over distance. But every amplification stage adds a little bit of noise, like making a photocopy of a photocopy. Cumulatively, these effects cap the usable capacity at about thirty to fifty percent of the theoretical maximum.
Physics itself is the internet's portion-control dietician.
That's one way to put it. And this is the actual "fiber capacity crunch" that optical researchers have been publishing about for over a decade. The 2024 deployment of eight-hundred-gigabit-per-wavelength technology on the MAREA transatlantic cable — that's the Google and Facebook cable between Virginia and Spain — shows how modulation advances keep pushing the ceiling up. MAREA originally launched in 2018 with eight fiber pairs at two hundred gigabits per wavelength, total design capacity a hundred sixty terabits. The 2024 upgrade to eight hundred gig per wavelength pushed it to six hundred forty terabits. But even that's not infinite, and we're approaching the Shannon limit for conventional single-mode fiber.
When a backbone link actually does approach capacity — what happens? Does the internet break?
No, it gets managed. There was a well-documented case in 2021 — BT's core network in London hit ninety-eight percent utilization during lockdown. Everyone's home, everyone's streaming, everyone's on Zoom. The network didn't crash. Automated traffic shaping kicked in. Certain types of traffic got prioritized — voice calls, emergency services — and bulk data transfers got slowed down. It's the digital equivalent of letting ambulances through while everyone else merges.
Which brings us to throttling, which is what most people actually experience and complain about. And I suspect this is where the economic reality crashes into the physics.
This is the part that matters for almost everyone listening. The backbone — those long-haul fiber routes between cities and across oceans — is almost never the bottleneck for a typical consumer. The real constraint is the last mile. The connection from the network to your home.
The final few hundred meters of glass or copper that actually touches your life.
And this is where the architecture changes completely. If you're on cable internet, you're using DOCSIS — the latest version, DOCSIS 4.0, can theoretically do ten gigabits per second downstream. But that bandwidth is shared across a node serving anywhere from fifty to two hundred homes. If you're on fiber, you're probably on GPON — gigabit passive optical network — which gives you 2.5 gigabits downstream shared across thirty-two to sixty-four users. The newer XGS-PON bumps that to ten gigabits shared. Either way, you're sharing.
Like a party sub. Everyone gets some, but if the entire football team shows up at once, the slices get very thin.
That's exactly the contention ratio. ISPs oversubscribe — they sell more total bandwidth than they can actually deliver simultaneously — because they're betting that not everyone will use their full connection at the same time. The typical contention ratio in the US is somewhere between fifty-to-one and a hundred-to-one. So for every gigabit of actual capacity at the aggregation point, they've sold fifty to a hundred gigabits of service to customers.
Which sounds like fraud until you realize it's the same model as everything else in infrastructure.
It's exactly how electricity grids work. It's how water utilities work. We don't build pipes to every house capable of running every tap, every shower, and every sprinkler simultaneously. The system would be unaffordable. The entire engineering discipline of network dimensioning is built around the "busy hour" concept from old telephone networks — you build for the ninety-fifth percentile of demand, not the absolute theoretical peak.
Here's what I want to push on — the electricity analogy breaks down in an interesting way, doesn't it? Because if everyone on my block turns on every appliance at once, the worst case is a brownout or a transformer blowing. The grid doesn't gracefully degrade by giving everyone dimmer lights. It fails catastrophically or it doesn't. But the internet degrades continuously.
That's an excellent distinction, and it's one of the things that makes internet architecture genuinely elegant. TCP — the Transmission Control Protocol, the fundamental protocol of the internet — has congestion control built into its DNA. When packets start getting dropped because a link is saturated, the sending side interprets that as a signal to slow down. Every connection independently backs off, then gradually ramps up again. It's a distributed, cooperative congestion management system that requires no central coordination. The internet doesn't have a circuit breaker that trips — it has millions of tiny, polite negotiations happening constantly.
It's less like a power grid and more like a highway system where every car automatically slows down when traffic gets heavy, without anyone directing them.
And that elegance is why the internet has scaled as well as it has. But it's also why your individual experience of "throttling" is so hard to diagnose. You might be slowing down because TCP is doing its job, because your ISP is deliberately shaping traffic, because your neighbor's backup is saturating the node, or because a content provider's CDN is having issues. All of those look the same from your couch.
When my connection slows down at eight PM on a Tuesday, that's not a bug. That's the system working exactly as it was designed to work.
Working as designed, yes. Whether that's a good design is a separate question. The alternative exists — you can build a network with no contention at the last mile. It's called point-to-point Ethernet, or active Ethernet, where each home gets its own dedicated fiber strand all the way back to the central office. Full gigabit or ten gigabit, twenty-four seven, regardless of what your neighbors are doing.
This exists in the real world?
It costs about twenty to thirty dollars more per month than GPON service, and it's available from a handful of providers in select markets. But most ISPs don't offer it because most consumers won't pay the premium. The cost to deploy fiber-to-the-home in the US runs between twelve hundred and eighteen hundred dollars per home passed, according to the FCC's 2024 Broadband Deployment Report. If you want dedicated fiber with no contention, that number goes up, and the monthly price has to reflect it.
The limit isn't physical — it's economic. We could build a network with no peak-time slowdowns. We just don't want to pay for it.
That's the core insight. And this is where the question about throttling gets interesting. When your connection slows down at peak times, it's almost never because a backbone fiber somewhere is saturated. It's because the aggregation point in your neighborhood — the CMTS if you're on cable, the OLT if you're on fiber — has hit its capacity. A few dozen of your neighbors all decided to stream 4K video at the same moment, and the shared pipe can't keep up.
The ISP's traffic management software starts making decisions about whose packets get through first.
And they're not always transparent about how those decisions get made. Some ISPs use protocol-level shaping — they'll identify streaming video and throttle it specifically, while letting web browsing and email through at full speed. Others use user-level shaping — if you've transferred a lot of data in the current billing cycle, you might get deprioritized during peak hours regardless of what you're doing now. The net neutrality debates of the last decade were fundamentally about whether ISPs should be allowed to make these distinctions at all.
The counterargument being that if you can't manage traffic during congestion, everyone's experience degrades equally, including the person trying to make a VoIP call while their neighbor torrents the entire Criterion Collection.
That's the genuine engineering tension. Let me give you a concrete example. In 2018, firefighters in Santa Clara County had their wireless data throttled during a wildfire response because they'd exceeded their plan's data cap. Verizon's traffic management system didn't know or care that these were emergency responders — it just saw heavy users and applied the policy. After public outcry, Verizon removed throttling for first responders during emergencies, but the incident illustrates exactly the problem. The network's automated systems can't always distinguish between "this person is torrenting" and "this person is coordinating an evacuation.
That's terrifying. And it makes the net neutrality debate feel a lot less abstract.
It's not abstract at all. The question the prompt raises — "is throttling a byproduct of some other factor?" — the answer is yes, and that factor is the economic decision to oversubscribe. ISPs could eliminate peak-time throttling tomorrow by either reducing contention ratios or building dedicated last-mile connections. Either option would require significant capital expenditure — digging up streets, pulling new fiber, upgrading aggregation hardware — and the cost would flow through to consumer bills.
Let me put some numbers on this. What does the average household actually consume?
According to Sandvine's data, the average US household consumed about five hundred eighty-six gigabytes per month in 2025, up from three hundred forty-four gigabytes in 2020. That's a seventy percent increase in five years. And the trajectory isn't flattening — 4K streaming, cloud gaming, and AI inference moving to the edge are all pushing per-household demand higher.
Yet the network hasn't collapsed.
Because ISPs keep upgrading. They light new fiber pairs — that costs about thirty thousand dollars per kilometer to deploy, by the way. They move to higher-order modulation — going from 16-QAM to 64-QAM to 256-QAM, which packs more bits into each symbol but requires better signal-to-noise ratios. They're experimenting with space-division multiplexing — multi-core fiber where each core carries its own set of wavelengths, effectively multiplying capacity by the number of cores.
Multi-core fiber. So instead of one strand of glass, you have several cores inside the same cladding, each acting as an independent channel. How many cores are we talking about?
Lab demonstrations have gone as high as nineteen cores in a single fiber. But the practical challenge is enormous. When you pack multiple cores into the same cladding, you get cross-talk between them — light leaks from one core to its neighbors. The closer the cores, the more cross-talk. So there's a fundamental trade-off between density and signal quality. It's the same kind of engineering tension as the Kerr effect — you're fighting physics for every additional bit of capacity.
Then there's hollow-core fiber, which you've mentioned before. That sounds like science fiction.
Hollow-core fiber is exciting. NTT and a company called Lumenisity demonstrated in 2024 a hollow-core fiber that carries light through an air-filled center channel rather than through solid glass. Light travels at ninety-nine point seven percent of the speed of light in a vacuum — compared to about sixty-seven percent in conventional glass. That lower latency matters for high-frequency trading and certain scientific applications, but the bigger deal is that hollow-core fiber has much lower nonlinear effects, which means you can push more power and more wavelengths through it before hitting the Kerr effect ceiling.
Wait — why does the speed of light in the medium matter for capacity? I thought we were talking about how much data, not how fast it gets there.
The latency improvement is a separate benefit. The capacity benefit comes from the reduced nonlinear effects. In conventional fiber, the Kerr effect limits how much optical power you can launch into the fiber before distortion becomes unacceptable. Hollow-core fiber essentially eliminates that constraint because the light is traveling through air, not glass. You can pump far more power into the channel, which means you can support more wavelengths at higher modulation orders. The capacity ceiling gets pushed way up.
The "capacity crunch" is real but we keep inventing our way around it.
The question is whether the pace of innovation keeps up with demand growth. And this is where I want to circle back to something the prompt hints at — the idea that the internet has no start and no end, that there isn't really such a thing as "internet supply." That's philosophically appealing but technically wrong.
The internet is a network of networks, yes. But every one of those networks has physical infrastructure — fiber optic cables, routers, switches, amplifiers — and every piece of that infrastructure has finite capacity. The "supply" of internet in a given city is the sum total of all the lit fiber capacity connecting that city to the rest of the world, plus the internal metro fiber rings, plus the last-mile connections to every building. That total has a number. It's a very large number, but it's a number.
Someone, somewhere, is responsible for making sure that number stays ahead of demand.
Backbone operators like Lumen and Zayo and the big content networks — Google, Meta, Amazon, Microsoft — who build their own private fiber routes. Last-mile ISPs. They're all making independent decisions about when to upgrade, and those decisions don't always align perfectly. You can have a situation where the backbone has plenty of headroom but a particular neighborhood's aggregation node is saturated, and from the user's perspective, the internet is "slow.
It's a coordination problem as much as a capacity problem.
And the coordination is largely market-driven, which means it responds to incentives, not to some central plan. If a neighborhood consistently saturates its node, the ISP eventually upgrades it — because churning customers cost more than new hardware. But "eventually" can mean months or years of degraded peak-time service.
Which brings us to the international comparison. I've heard you mention South Korea before.
South Korea is the poster child for doing this differently. Their government mandates a contention ratio of thirty-to-one, compared to the US average of a hundred-to-one. That means for the same advertised speed, a Korean household is far less likely to experience peak-time slowdowns. The trade-off is that Korean ISPs had to invest more in infrastructure per subscriber, and those costs are reflected in the market structure — it's a denser population, easier to wire, and there's more government subsidy.
The US could do this. We've chosen not to.
We've chosen a market-driven approach that prioritizes lower monthly bills over guaranteed peak performance. And to be fair, for the vast majority of users, the vast majority of the time, it works fine. The average American household's actual experienced throughput is more than adequate for multiple simultaneous 4K streams. The problems show up at the margins — the neighborhood with older infrastructure, the apartment building where the building's internal wiring hasn't been upgraded since the Bush administration, the rural area where the only option is DSL running over copper that was buried in 1978.
If someone's listening to this and they consistently experience slowdowns at peak times, what should they actually do?
First, figure out what you're actually working with. Tools like the FCC's broadband map or the SamKnows measurement platform can show you the difference between your advertised speed and your actual throughput during peak hours. If the degradation is more than about twenty percent of what you're paying for, the FCC has had enforcement teeth for this since 2025 — you can file a complaint and the ISP is required to respond.
That's a specific number.
It's the threshold the FCC settled on for "materially different from advertised performance." If you're paying for a gig and consistently getting seven hundred megabits or less during peak hours, that's a legitimate complaint.
Check what kind of connection you actually have. If you're on GPON fiber, you're sharing with your neighbors. If you're on point-to-point Ethernet, you're not. Most ISPs don't advertise this distinction prominently, but it's usually in the fine print of the service agreement. Look for terms like "contention ratio" or "oversubscription rate." Anything above fifty-to-one means you will feel peak-time effects.
If you're on cable?
Cable is inherently shared at the node level, and there's not much you can do about it short of switching to fiber. But you can check how many homes are on your node — some ISPs will tell you if you ask — and if you're on an overloaded node, you can sometimes get moved to a less congested one. The squeaky wheel approach.
The third thing — and this is the one most people don't want to hear — is pay more.
If you need guaranteed throughput, business-grade connections exist. They cost more, they come with service level agreements that specify minimum performance, and they're typically not oversubscribed at the same ratios as consumer connections. For most people this is overkill, but if you're working from home and your livelihood depends on stable video conferencing, it might be worth the premium.
Let me ask you something that's been lurking behind this whole conversation. The oversubscription model works because usage is bursty — people stream for a bit, then browse, then walk away. But what happens when AI-generated content becomes dominant? If everyone's running real-time AI video generation or persistent augmented reality overlays, doesn't that shift the traffic profile from bursty to constant high-throughput?
That's the question that keeps network architects up at night. The entire economic model of consumer internet is built on the assumption that the busy hour is an hour — that demand peaks and then recedes. If AI inference at the edge means every device is constantly pushing and pulling high-bandwidth data, the oversubscription model breaks. You'd need to either massively overbuild — which means much higher prices — or accept that "up to one gig" becomes "up to one gig, sometimes, when the AI agents aren't busy.
It's not just AI, right? The whole trajectory of consumer applications has been toward more persistent connectivity. We went from checking email a few times a day to push notifications to constant background syncing. AI agents that are always running, always processing, always pulling context from the cloud — that's the logical endpoint of that trajectory.
That's the nightmare scenario for network dimensioning. The busy hour model assumes a certain statistical multiplexing gain — the ratio of peak aggregate demand to average aggregate demand. For traditional internet traffic, that ratio is about two-to-one. The network needs to be roughly twice as capable as the average load to handle peaks. But if traffic becomes constant, the multiplexing gain drops toward one-to-one. You lose the statistical efficiency that makes oversubscription viable. Every subscriber would need something much closer to dedicated capacity.
Which brings us back to the economic question. The network we have is optimized for the traffic patterns we've had. If the patterns change, the economics change.
And the power draw question. You mentioned this before we started recording — routers and switches are becoming the bottleneck, not fiber.
Right, because all those bits need to be processed, not just transported.
Cisco's 2025 Silicon One chips can handle 25.6 terabits per second per chip. That's extraordinary. But they draw about a hundred watts per terabit per second of throughput. When you're moving petabits through a data center or a major internet exchange point, the power consumption becomes the limiting factor before the fiber capacity does. We're approaching a point where the cost of electricity to run the switching fabric rivals the cost of the fiber itself.
The internet doesn't run out — it just gets more expensive to keep up.
That's the real ceiling. Not Shannon's theorem, not the Kerr effect, not the contention ratio in your neighborhood. It's the collective willingness to pay for the next generation of infrastructure. Every upgrade — hollow-core fiber, multi-core fiber, higher-order modulation, faster switches — requires capital. Someone has to spend the money. And the question of who pays — consumers through higher bills, governments through subsidies, or content providers through their own private networks — is a political question dressed up as an engineering one.
Which feels like the right place to land. The prompt asked whether there's a genuine bandwidth limit, and the answer is yes — there are genuine physical limits, genuine economic limits, and genuine architectural limits. But none of them are a "sold out" sign. They're thresholds where someone has to decide whether to spend money.
In the meantime, when your Netflix buffers during the big game, you can take comfort in knowing it's not because the internet is full. It's because your neighbors are all watching the same thing, and your ISP gambled that they wouldn't be.
The internet — powered by physics, constrained by accounting.
That should be on a T-shirt.
Now: Hilbert's daily fun fact.
Hilbert: The platypus hunts with its eyes, ears, and nostrils closed, relying entirely on electrolocation — it detects the tiny electric fields generated by its prey's muscle contractions. This makes it the acoustic equivalent of a passive sonar system that listens not for sound waves but for bioelectric signatures, a sensory modality unique among mammals and first documented by European naturalists in the high medieval period of the fourteenth century, though the platypus itself had been perfecting the technique in the rivers of what is now Eritrea's geological cousin, the Australian continent, for millions of years prior.
...right.
This has been My Weird Prompts. Thanks to our producer Hilbert Flumingtop. If you want to dig deeper into the numbers behind your own connection, head to myweirdprompts.com for links to the tools and reports we mentioned. And if you found this useful, leave us a review wherever you get your podcasts — it helps.
Until next time.
