The Bandwidth Wall: Why I Think Machine Traffic Could Strain the Internet

The idea that the world could run out of internet sounds crazy when your phone streams video without a hiccup. My argument is that growing machine traffic could strain capacity, not that the global internet is already full.

Financial television spends plenty of time on AI chips, data-center electricity and power equipment. My concern is that bandwidth could become another constraint over the next few years as agents and robots scale.

Built for people who sleep

The internet has largely served human attention and working hours. Software can keep making requests without waiting for someone to read each response. That contrast is my starting point.

Musk spoke with Ron Baron for the 33rd Annual Baron Investment Conference on November 6, 2026, an event in New York. In the public recording’s transcript, he said his understanding was that agentic traffic already exceeded human traffic, if he had read Cloudflare correctly. He also described Cloudflare’s five-year forecast as agentic AI traffic reaching 1,000 times human traffic, and thought that underestimated it.

The corrected transcript of Cloudflare’s August 6, 2026 earnings call identifies the speaker as CEO Matthew Prince, not CFO Thomas Seifert as I said on tape. Prince said his team reported in May 2026 that non-human traffic had passed human traffic. With a caveat that his previous forecasts had been wrong, he said continued trends could produce a non-human-to-human ratio as high as 1,000 to one in five years. He said, "humans will be a rounding error on the Internet". This is Cloudflare’s account and conditional forecast, not a measurement of global bytes or proof that AI agents alone have passed humans.

Four layers of pipe

How much bandwidth does Earth have? The internet feels weightless. It isn't. It's physical plumbing made of glass fiber and ocean cable, and there's no single number for all of it. I think of it as a road system delivering packages to your house.

Access is the driveway. That's the fiber into your garage, a cable modem, a cell tower down the block, or a Starlink dish on the roof. Metro and regional is the city highways, fiber loops that gather homes and offices and route them toward data hubs. Backbone and long haul is the interstates and ocean bridges, meaning buried land fiber and submarine cables. Spectrum and satellite throughput is the flight paths overhead. Starlink uses radio between space and the ground and lasers between satellites. Those radio links share spectrum; weather and geometry can limit usable capacity.

Lanes versus cars

Capacity is how many lanes are available; traffic is how many cars travel on them. But TeleGeography’s term used bandwidth means capacity deployed by several classes of network operator, not measured traffic. Peak and average utilization are separate measures again.

TeleGeography’s September 21, 2026 report puts mid-2026 international internet bandwidth at 2,259 Tbps, about 2.26 Pbps. One petabit per second is 1,000 terabits per second. It reports 23% growth in 2026 and a 23% compound annual growth rate from 2022 to 2026. This is deployed international internet-backbone capacity, not all installed fiber, all private networks or a ceiling on the world’s internet.

A separate TeleGeography report published May 12, 2025 says used international bandwidth surpassed 6.4 Pbps in 2024. Its broader scope includes content-provider, research, enterprise and internet-backbone networks. Content and cloud networks accounted for almost three-quarters of that demand in 2024. These are different dates and scopes. The larger figure is not traffic overflowing the smaller capacity figure.

One prompt, ten sub-agents

A chatbot answer can be a small exchange of text. An agent doing repeated browsing, downloads or video work could move much more data. The amount depends on the job. It is not a fixed requirement for every agent.

Imagine asking an agent to research 50 suppliers and build a prototype. In this hypothetical workflow, it starts 10 sub-agents to browse pages, call APIs, download code, run tests and revise errors. They exchange files without waiting for a human to read each screen. This is an illustration, not a measured workload.

Musk wrote on August 9, 2026: "AI agentic Internet traffic will obviously VASTLY exceed human usage. Not a close call at all." He added, "Cloudflare’s forecast is accurate." On October 24, he wrote, "Agentic demand is growing faster than I’ve ever seen anything grow". These are his forecasts and assessment.

SpaceX’s official investor page confirms an October 20, 2026 Q3 earnings webcast. A third-party transcript of that call attributes to Musk a forecast of 1,000 times current bandwidth demand in three to five years, because agents could use more bandwidth per agent and outnumber people. This is a reported company forecast, not a measured outcome. My example of one worker running 50 agents around the clock is illustrative. Local GPU memory transfers are not automatically traffic on the public internet.

Different metrics, different forecasts

At a constant 23% annual growth rate, a quantity doubles in about 3.35 years. That arithmetic does not make the measures comparable. Cloudflare’s future non-human-to-human traffic ratio is not a 1,000-fold increase in bytes from today. Musk’s global bandwidth-demand forecast is a different claim again. Neither establishes a 1,000-fold requirement for international public links. Request size, caching, private networks and where traffic stays all matter. TeleGeography’s September report says AI has not yet generated a massive surge in public IP-backbone traffic. My concern is that machine workloads could require much more capacity, not that these figures prove a dated global shortage.

You can't plug in a moving car

A moving car or untethered robot cannot depend on a fixed fiber tether. It needs wireless for remote data exchange. That does not mean every driving or walking decision needs a continuous cloud connection.

Musk forecast on August 8, 2026 that AI and robotics would greatly increase bandwidth demand, writing, "Their need for data transfer is orders of magnitude more than humans!" Tesla describes fleet-data training and over-the-air updates. Selected clip uploads and model downloads illustrate how fleets could exchange data, not what every vehicle or robot currently transmits. Bandwidth needs depend on what is transmitted and when.

The third-party earnings transcript attributes to SpaceX president Gwynne Shotwell the expectation that connected cars, robots and superintelligence would expand communications demand and give Starlink a significant share of internet traffic. Musk wrote on August 9, "All cars will have Starlink in the future." He added, "It’s the only way to get super high bandwidth to billions of vehicles." That is his prediction, not a confirmed vehicle deployment plan or proof that cellular networks cannot serve these workloads.

At the conference, Musk said a robot could initially cost about the same as a car to build and less over time. Baron floated a $30,000 manufacturing cost, and Musk agreed. This is not a quoted retail price. Musk said the robots would probably be rented. He also projected that a humanoid working 168 hours a week could deliver roughly five times the weekly output of a person working about 40 hours, with assumptions about speed and breaks. That is not a demonstrated Optimus benchmark. I think millions of robots could be deployed if cost and capability targets are met. Their networking requirements would depend on the work. A rural farm without fiber could benefit from wireless, but that need not mean Starlink or continuous uploads of every camera feed.

The 90% claim

Musk’s August 9, 2026 post said, "And the only system that can support the insanely fast bandwidth growth needed by AI is Starlink". In a separate August 10 post, he wrote, "It’s possible that Starlink may end up doing >90% of IP traffic, even if competitors 10X their bandwidth". IP refers to Internet Protocol. These are his speculative market and technology claims, not an established future traffic share.

New fiber routes can require permits, construction and cable ships, but operators can also light additional capacity on existing systems. TeleGeography says such equipment work can take months; its 23% historical growth rate is not a physical ceiling. Satellite launches add fleet capacity over time, but do not instantly multiply usable capacity in every market. In the reported earnings transcript, Shotwell says users would need a few hundred V3 satellites to feel a consistent benefit.

SpaceX’s filed Q3 results report 13.6 million consumer subscribers at September 30, 2026, and $5.248 billion in connectivity revenue, about $5.25 billion, up 69% year over year. In his September 15 post, Musk used an approximately 11,000-satellite constellation as the baseline for a projected eventual more-than-100-fold bandwidth increase from V3 deployment. That is a dated company baseline, not a fresh count of active satellites. The third-party Q3 transcript reports his roughly three-year, 100-fold network-capacity forecast. In an August 9 post, he said he expected the total constellation, including direct-to-cell, to exceed 100,000 satellites. These are forecasts, not achieved capacity or approved final deployment counts.

At the conference, Musk described current laser links as around 100 Gbps and forecast next-generation links at 10 Tbps. That is a 100-fold per-link increase, not a demonstrated 10Tbps service to a user on the ground. He also forecast V4 at ten times V3’s capability. In his August 9 post, he said future satellites may exceed 50 times a V2’s throughput. Per-link, per-satellite and constellation-wide figures are different measures, and these future specifications remain his projections.

Satellites by the freight load

SpaceX’s 2025 progress report assigns deployment of next-generation Starlink constellations to Starship. Its filed Q3 results confirm 26 V3 satellites on Flight 14 and approximately ten times the downlink capacity added by a Falcon 9 Starlink launch. The third-party earnings transcript reports Musk describing 26 Tbps of added capacity and a ship that was "about maybe half loaded". That is company-reported capacity, not measured traffic. Doubling a comparable payload would roughly double its added capacity; that is my extrapolation, not a verified full-load specification.

Where fiber still wins

None of this kills fiber, whatever social posts say about satellites beating undersea cables. Four physical limits keep both in play.

Spectrum is the first constraint. Laser links between satellites do not remove the capacity limits of shared Earth-space radio links. Dense groups of users can compete for capacity, and rain or obstructions can impair the radio path. The result depends on frequency, geometry and network design.

Gateways are the second. Traffic destined for terrestrial services reaches ground gateways and internet points of presence, sometimes after traveling through the inter-satellite laser mesh. Starlink’s progress report describes routing around outages to unaffected ground stations.

Software is the third. In an August 9 reply, Cloudflare’s CTO said bandwidth was only half the issue: congestion control often assumes a stable path and round-trip time, while LEO does not provide either. He proposed that Cloudflare and Starlink work on it together. Musk replied on August 10 that he had sent the post to Starlink. This exchange does not establish a partnership or a specific packet-loss rate.

Density is the fourth. Fiber can carry large volumes between fixed locations without consuming shared radio spectrum. Rain does not attenuate light inside a buried cable as it does a radio path, though terrestrial infrastructure can still be damaged. Adding strands or upgrading terminal equipment also takes money and infrastructure; capacity is not unlimited.

Four gears, one loop

I see Musk’s rockets, satellites, AI data centers and robotics as parts of a connected strategy: four gears of one machine. That is my interpretation, not proof that every product is already integrated.

The brain is AI data centers, on the ground and potentially in orbit. Musk and SpaceX have proposed solar-powered orbital compute; that remains a development plan, not an existing large-scale computing fleet. It would need communications links to exchange data with Earth. The nervous system in my analogy is Starlink, the delivery vessel is Starship, and the hands are cars and robots.

In my scenario, cars and robots exchange selected sensor data and model updates with data centers. Better models could enable more tasks, which could increase bandwidth demand and help fund launches and satellites. That is a possible business loop, not evidence that every Tesla or Optimus currently sends its data through Starlink or that more data automatically makes a model better.

The conference forecasts belong to their speakers. Baron described a possible $1 trillion in annual Starlink revenue, $700 billion to $800 billion in EBITDA and $14 trillion to $15 trillion in value in ten years. He described all of SpaceX as then worth less than $2 trillion. Those are his estimates, not an independently established valuation. Musk forecast a December 2026 annualized revenue run rate above $100 billion for SpaceX and a best guess above $500 billion in December 2027. The event was November 6, 2026, so his "December of next year" means 2027. He said it was not a promise and gave a roughly 2% chance to $1 trillion. The Q3 transcript defines ARR here as revenue annualized from December’s pace, not revenue already earned over a year. These SpaceX-wide forecasts are not Starlink-only figures or guaranteed contracted recurring revenue.

The human internet was the warm-up

My bet is that software agents and physical robots could make machine-to-machine connectivity far more important. The human internet was the warm-up in that argument. I think companies that build useful capacity could benefit. That is my thesis, not a finding that the world has already run out of bandwidth or that one supplier must carry it all.