Field report NE-36 / N4 · sheet /macro
Boardroom MacroFiled from the territory
AI Agents Push Internet Traffic Upstream, Pressuring Networks
U.S. household uploads grew 19.2% in the second quarter, three times the download rate, as agentic AI and physical systems push more traffic upstream rapidly.
- Grid
- NE-36 / N4
- Surveyed
- November 18, 2026
- Reading
- 5 min read
- Correspondent
- Elena Vasquez

Field notes
- OpenVault recorded 19.2% year-over-year U.S. household upload growth in the second quarter, versus 6.1% for downloads.
- Average monthly uploads reached 58.2 GB per subscriber, while the download-to-upload ratio fell to 11 times from 16.6 times in 2019.
- Ericsson tracked faster uplink growth at 43 of 55 providers in 2025 and projects traffic tripling or quintupling by 2031.
- Nokia forecasts AI will drive 44% of global network traffic by 2035, with more than 90% of physical-AI traffic starting upstream.
- AT&T targets 60 million fiber locations by 2030, about 1.5 times its footprint, after roughly 4 million additions this year and a $3 billion-plus Corning deal.
By Elena Vasquez
U.S. household upload traffic grew 19.2% year over year in the second quarter, more than three times the 6.1% growth in downloads, OpenVault data show, as agentic AI begins reversing the internet’s two-decade one-way design.
AT&T Chief Operating Officer Jeff McElfresh set out the operating case at Monday’s Fortune 500 Innovation Forum in Detroit: “The actual direction of the bandwidth of the internet is undergoing a fundamental shift ... we’re seeing agents tap the network in a 24 by seven, a very different approach. It’s not so episodic anymore.”
This changes the traffic model built around Netflix, YouTube and other streaming services, which sent content from centralized data centers to consumers’ homes for roughly two decades. Network builders designed last-mile systems around that downstream bias.
OpenVault, which tracks broadband usage for internet providers, said the average subscriber now uploads 58.2 GB a month. Households still download about 11 times as much as they upload, down from a 16.6-to-1 ratio in 2019, narrowing a gap that has defined network investment.
The money at risk extends to the hundreds of billions of dollars companies are spending on AI and the infrastructure carrying it. Many last-mile systems assumed customers would mostly watch and send relatively little data back, an assumption the new mix of continuous agent activity and machine communication is testing.
What is changing in traffic flows?
McElfresh, who has worked at AT&T for 31 years, said engineers once identified Netflix and YouTube easily from network-control-center feeds because human behavior produced predictable traffic patterns. The pandemic disrupted those patterns as work-from-home shifted traffic from cities to suburbs, strained many home networks and caused video calls to glitch.
AI agents run continuously rather than only when a person starts a video session. McElfresh cited an autonomous vehicle connected to AT&T’s wireless network that sends more data upstream to the cloud to train a model than it consumes, giving operators a concrete example of machine-to-machine demand.
Can networks handle sustained uploads?
Ericsson found that 43 of the 55 mobile service providers it tracked in 2025 saw uplink traffic grow faster than downlink traffic. Its scenarios project uplink traffic tripling or quintupling by 2031 if AI applications, smart glasses and agents gain adoption, while near-term growth comes from smartphone apps, user-generated video and cloud storage.
Mobile traffic will remain downlink-dominated in the near term, Ericsson said, but its warning to operators was direct: “Current networks are not dimensioned for sustained uplink demand.” The company’s 2031 scenarios make capacity planning the immediate issue for operators serving AI-heavy customers.
Nokia’s 2026 network traffic report goes further, forecasting that AI will drive 44% of global network traffic by 2035. For physical AI, including robots, cars and drones, Nokia expects more than 90% of traffic to begin in the uplink, challenging the asymmetric design of traditional access networks.
Why does fiber give AT&T an edge?
AT&T has a direct commercial interest in the shift. Fiber delivers equal downstream and upstream speeds, while cable broadband, its principal home-internet rival, has historically provided much slower uploads than downloads. McElfresh called fiber the “gold standard,” future-proof and able to respond rapidly to AI demand because it offers symmetrical capacity.
AT&T plans to reach 60 million fiber locations by 2030, about 1.5 times its present footprint. The carrier also signed a supply agreement worth more than $3 billion with Corning in September, securing fiber for a rollout that its technicians are carrying out in 10,000 neighborhoods on any given day.
That program will add roughly 4 million locations this year, taking the network toward the 60 million target. The investment links a household traffic trend to AT&T’s capital allocation: a larger symmetrical access footprint gives the carrier more capacity to serve upload-intensive applications than a slower-upload cable network.
OpenVault’s second-quarter figures support the commercial distinction. Fiber subscribers uploaded an average of 94.1 GB per month, against 58.2 GB on cable’s DOCSIS networks. The figures do not isolate AI traffic, but they show how much headroom fiber has relative to legacy cable.
OpenVault CEO Mark Trudeau said in October that gaming, Zoom and Teams calls account for much of the upstream increase, alongside AI content creation. Cable operators are responding with DOCSIS 4.0, which supports symmetrical speeds and could narrow fiber’s upload advantage.
What will customers notice?
Corning CEO Wendell Weeks, who shared the Detroit stage, said customers will experience the shift through a combination of available bandwidth and AI-enabled products, including robots and new device use cases. Corning makes the glass fiber AT&T installs, placing it in both access-network and data-center supply chains.
Weeks identified the data center as the next constraint. Operators will replace copper links with fiber and convert signals into light within one centimeter of the processor, reducing the distance over which electrical connections carry data as AI workloads intensify.
AT&T’s build is the near-term response: roughly 4 million locations this year, 60 million by 2030 and a $3 billion-plus Corning supply agreement. The trigger is already measurable in the 19.2% upload surge, but operators still face a warning: current networks are not dimensioned for sustained upstream demand.
via fortune.com/feed/ (Source)
More from Elena Vasquez
Show full bio
Staff writer covering companies, deals and strategy in corporate strategy and chief-executive leadership at CEO Frontier.
489 articles
Adjacent sheets
- Tate Reeves says data center backlash has lost sight of AI necessity
- Charter's $2 Billion Cost Cut Meets Subscriber Losses
- ChatGPT Hits One-Fifth of Google's Search Volume — Yet Barely Sends Traffic
- Data Centers Need Community Stakes and Regulatory Accountability
- Disney Plans 500 Daily Vertical Clips and a Free Disney+ Tier