The telecom industry says AI devices will overwhelm mobile uplinks and justify new spectrum and 6G investment. But beneath the forecasts, ratios and rhetoric, the evidence for an AI-driven radio supercycle remains surprisingly thin and largely unmeasured.
There are patterns in the radio industry where we manufacture an engineering crisis in service of marketing. AI devices send data upstream, networks were built for downstream, therefore the industry must invest in uplink capacity, new spectrum, and eventually 6G. Ericsson tells it with a ratio. The Wireless Infrastructure Association tells it to policymakers. AT&T tells investors on its Q2 call that agentic AI is reshaping traffic in volume, shape, symmetry, and criticality, and that its EchoStar and 600 MHz spectrum purchases exist to build a robust upstream for the agentic era.
These stories start with a true observation and then make a leap. Today, the true part is that AI traffic is more upstream-heavy than the traffic networks were tuned for. The leap is treating that observation as an investment requirement. Between the two sits a workload model that nobody, vendor nor operator, has published. The purpose of this piece is to show exactly what that model must contain and how little of it exists in public. Two different theses are currently traveling together under the AI banner and only one of them is in dispute here.
Telcos becoming inference infrastructure players is a real and separately-credible story. Verizon’s billion-dollar data center interconnect deal with Google, SK Telecom spinning up SK Hyper to run AI data centers, the general shift of inference from a handful of training campuses toward distributed metro sites – all of that monetizes fiber, facilities, and interconnect that operators already own. Cisco projects AI inference reaching a quarter of total network traffic by 2035, and the bulk of that runs between data centers over fiber. The thesis under examination is that AI device traffic will strain the cellular radio uplink, specifically, at a scale requiring a distinct capital cycle.
Watch for the conflation, because the wireline story keeps getting used as ambiance for the radio claim.
The four questions
“Will AI break the uplink” is really four separate questions.
– How many bytes does the workload generate?
– How many devices offer those bytes to the same sector at the same instant?
– Can a device at the cell edge or indoors achieve the rate the application needs at all?
– Does the application get its answer within the latency it needs to be useful?
A credible investment case has to publish numbers for all four, because a network can pass three and fail one. So far the industry has published a ratio, a shipment forecast, a modeled percentage, and a per-task multiplier. Let us see what the public evidence actually supports.
Question one: volume
Start with one heavy user of AI glasses. Assume a visual query, “what am I looking at,” uploads a compressed image of roughly 2 MB. Note that this is an illustrative assumption, not a measured standard. At 500 KB per query everything below shrinks to a quarter. With short video snippets it grows several fold. Give our user fifty queries a day, every day.
50 queries × 2 MB = 100 MB per day
100 MB × 30 days = 3 GB per month
For scale, the average smartphone already moves 22 GB a month, of which roughly 8 percent, about 1.8 GB, is uplink from ordinary use. So an intense AI user roughly doubles his own uplink. Nokia’s field measurements of AI apps on commercial Apple, Pixel, and Samsung devices found the same order of magnitude, an increase it described as tens of megabytes per day. Aggregate AI volume today is small. Ericsson’s Mobility Report measured generative AI at 0.06 percent of network traffic, majority downlink at that.
But volume is not the strong version of the case. The action is in the next three questions.
Question two: concurrency
Put 1,000 active glasses users on one sector, each making five queries in the busy hour, each query a 2 MB burst lasting about two seconds. Let’s look at the average offered load:
1,000 users × 5 queries × 2 MB = 10 GB in the hour
10 GB over 3,600 seconds ≈ 22 Mbps
Whether 22 Mbps is comfortable depends on the cell. Nokia’s public figures put typical 5G uplink at 10 to 15 Mbps on average and below 5 Mbps in many networks. On a well-provisioned mid-band sector, 22 Mbps of additional offered load is absorbable. On a constrained or heavily loaded one, it is material.
Averages assume the bursts spread out. Do they? If arrivals are independent, the expected number of users transmitting at any given instant is:
1,000 users × 5 bursts × 2 seconds ÷ 3,600 seconds ≈ 2.8 simultaneous
Under independence, this workload is trivial. Fifty users colliding in the same window would be a statistical freak.
But independence is the assumption the vendor side has been rejecting. Context is a common clock. Crowds look at the same event at the same moment. Tourists photograph the same facade. Workers start the same procedure together. Agentic systems add a second common clock, because the application decides when to transmit, and applications synchronize in ways people do not. The InterDigital and ABI report describes persistent contextual exchanges rather than randomly scattered questions, and AT&T’s statistic, that an AI agent generates up to 450 percent more traffic per task than a human doing the same work, points the same direction.
Fifty users offering 2 MB each in the same two-second window ask for 400 Mbps. A constrained sector does not drop the excess. The scheduler stretches it:
50 users × 16 Mb each = 800 Mb of work
800 Mb ÷ 50 Mbps of capacity = 16 seconds
The failure mode is latency inflation. A visual answer in two seconds is useful. The same answer in sixteen seconds is worthless. It means the panic side is right that correlated bursts can degrade the experience on ordinary cells. And it means the skeptics are right that nobody has measured the burst correlation factor.
Peak load = mean offered load × burst correlation factor
The public forecast supplies the first term through assumptions and leaves the second term blank. The 450 percent figure is traffic per task, it does not specify which link carries it, how much rides cellular rather than fiber and WiFi, or how the transmissions distribute across cells and seconds. Most agent chatter is API calls between servers, which is WAN traffic, not radio traffic. A per-task multiplier is directionally interesting and dimensionally unusable.
To see how wide the range is, vary the assumptions. A hundred users on an ordinary suburban cell making one busy-hour query each offer a mean load of about 0.4 Mbps. Two thousand users in a venue making five queries each offer about 44 Mbps. Adoption density and correlation decide which world a given cell lives in. Neither has been published for any real network.
Questions three and four: coverage and service quality
AI applications need consistent latency and adequate rate at the cell edge and indoors. Uplink has been the consistently weak direction, and best-effort uplink cannot promise this even on a lightly loaded network. Handsets transmit at a fraction of base station power. TDD frames allocate uplink a minority of airtime. A user behind two walls at the edge may not achieve 1 Mbps regardless of how empty the cell is. Ericsson’s material emphasizes exactly these dimensions. And this is the version AT&T is building against. Stankey’s stated logic for leaning into 600 MHz is that strong low-band reaches deep into buildings and delivers a consistent upstream, which is a coverage and consistency argument, not a tonnage argument.
A credible experience-assurance case must publish the application’s service target, the percentage of sessions currently failing it, where they fail, and which specific investment cures the failure. None of that is in the public record either. Cell-edge uplink has been weak since GSM. It justified uplink engineering in every generation without being attributed to any particular application. To convert it into an AI-specific capital requirement, you must show AI sessions failing at scale. The reframing from capacity to experience changes the vocabulary.
What has actually been measured
Against these four questions, the published record is thin and worth sorting by evidence type, because the debate now contains three different kinds of voice.
The strongest live-network test made public so far comes from Signals Research Group, which measured Meta Ray-Ban Display glasses and Samsung Galaxy phones running every supported AI and AR application on T-Mobile’s commercial 5G Standalone network. It found that image and video AI generate more uplink than text and that the overall impact was modest relative to other traffic. One study, one operator, current applications. It is sufficient to rebut the claim that today’s consumer AI inherently creates severe uplink loads. It is not sufficient to prove no operator has a localized problem.
The vendors’ own measurements confirm that present aggregate volume is small, even as the vendors interpret the burstiness and directional mix as reasons to begin upgrading. Nokia measured tens of additional megabytes per day and simultaneously flagged 25 Mbps two-second bursts as a planning concern. Both facts are real. The dispute is entirely about what they imply.
Examine what AT&T actually presented on its earnings call. The magnitude claims, 450 percent more traffic per agent task, 9x enterprise and 7x consumer traffic growth by 2035, were cited to industry research, the same Cisco-lineage forecasts the vendors use, not to measurements from AT&T’s own network. The trade press coverage itself noted the projections behind the strategy are unproven and are the industry’s own.
An operator repeating industry forecasts to investors is a forecast changing hands, not a measurement entering the record. The same reading applies to the most concrete uplink deployment anywhere, Huawei and China Unicom’s 100 MHz GigaUplink network in Beijing. It is real engineering serving real uplink demand, but the demand it serves today is China’s livestreaming economy, a massive, measured, pre-AI upstream workload. Huawei projecting a global AI uplink wave off the back of it is a deployment with a disputed attribution.
The Signals Research testing found that even when the glasses were not being used, the companion phone remained in RRC Connected substantially longer and generated notable control-plane signaling. Thelander declined to predict a repeat of the 3G signaling storms, but flagged connection behavior as potentially more consequential than payload volume. Signaling consumes real resources, random access opportunities, control channels, scheduler attention, and battery, and at scale it degrades capacity like any other load. Its first-line remedies, connection timers, batching, application discipline, and protocol work, are also real engineering with real cost. The only live-network study to surface a potentially consequential issue found it in connection behavior, while nearly all public investment arguments focus on payload capacity.
The measurement mess: 0.06, 0.2, and 4.2 are not the same object
Policymakers currently hear that AI is 4.2 percent of US wireless traffic, from the Wireless Infrastructure Association. Ericsson reports generative AI at 0.06 percent of measured network traffic. The GSMA’s 2026 analysis puts direct generative AI traffic around 0.2 percent and explicitly warns operators against overprovisioning on the strength of vendor forecasts.
These numbers cannot be reconciled from public information. They are different kinds of object. Ericsson’s figure is observed traffic classification on real networks. The GSMA’s is a direct-traffic estimate. WIA’s is a model output built from consumer surveys and platform activity, resting on a stated assumption that the average AI prompt generates 20 MB of mobile data traffic. That assumption is enormous for the text interactions that dominate actual AI use. WIA’s own report notes that the industrial and network-management components of its category were still too small to materially affect the result, so the breadth of the bucket is not the story.
The industry has no agreed method for measuring the thing it wants funded. A modeled 4.2, an estimated 0.2, and an observed 0.06 are not comparable facts, a is not yet a basis for spectrum policy.
The 1:8 ratio
Ericsson’s most quoted figure is that smart glasses generate eight bytes up for every byte down, against roughly ten to one the other way for conventional traffic. Uplink and downlink are asymmetric resources, and a compositional shift toward uplink is worth knowing about.
A device uploading 3 GB a month and one uploading 300 GB a month can both be 8:1. The ratio is identical, the network consequence differs a hundredfold. Ericsson’s own measurements reveal generative AI sessions run about 26 percent uplink against roughly 10 percent conventionally, and generative AI is simultaneously 0.06 percent of traffic. To turn a ratio into an investment case you need the absolute payloads, the busy-hour concurrency, the cellular share, and the location distribution.
It is not that vendors and operators research future demand. Of course they do, and nobody else has the data. It is that the public case jumps from device ratios, shipment forecasts, and per-task multipliers to network-level conclusions without publishing the assumptions that connect them. Show the bridge, and the argument can be evaluated. Until then it is an anecdote multiplied by a unit forecast.
The incentive map explains why the bridge stays unpublished. Vendors sell the gear the panic justifies. Lobbies sell the spectrum case to regulators. And operators sell their capex narrative to investors. AT&T committed billions to EchoStar spectrum and needs Wall Street to believe the purchase was foresight rather than opportunism, and “agentic AI demands robust upstream and our low-band wins” is that story in one sentence. None of this makes any party wrong. It means no participant in the public debate is a disinterested measurer, which is why the missing numbers matter more than the competing narratives.
The architecture
Prior predicted upload surges, from social sharing to videoconferencing, did not produce the forecast failures. Capacity investment, WiFi offload, codec improvement, and application adaptation all moved together, and a warning can be directionally right while investment prevents the predicted failure. What the record does support is forecasts that hold per-device behavior constant have consistently overestimated cellular load, because the one variable that always moves is where the computation happens.
Continuous capture, encoding, transmission, and heat dissipation collide inside a 69-gram frame with a watch-class battery. No mainstream consumer product today relies on continuous full-resolution camera upload as its normal operating mode, and the thermal and battery arithmetic explains why. The plausible future workload is intermittent, adaptive, contextual offload, sampled images, compressed clips, audio sessions. Its frequency and payload are unknown. This is an argument for measurement.
Most glasses ship without a cellular radio. The InterDigital and ABI forecast has only about 12 percent of 2030 shipments cellular-enabled. Tethered glasses still reach the mobile network through the phone whenever the phone is off WiFi, so the traffic is not deleted, it is mediated. A shipment forecast is not a forecast of cellular endpoints or cellular load, because phone connectivity, WiFi availability, and application behavior all intervene. Anyone multiplying glasses shipments into network demand is skipping those variables.
Every raw frame sent to the cloud costs the AI provider GPU money, so Apple, Google, Meta, and Qualcomm all profit from processing locally or preprocessing before transmission. Transcribe speech on-device and send text instead of audio, and the payload falls a hundredfold for that task. Recent research on intention-aware preprocessing for AI glasses demonstrated better than 50 percent bandwidth reduction across simulated scenarios while preserving task performance.
Granted, this is simulation, not fleet measurement, and the same work notes that heavy reasoning still offloads to edge or cloud, which supports the hybrid-network future as much as the skeptical one. And compression cuts both ways economically. Cheaper per-task network usage can mean many more tasks, so lower payload per query does not guarantee lower total traffic. The defensible conclusion is that semantic processing makes straight-line traffic forecasts unreliable in both directions, which is fatal to forecast-driven capex arguments without being a forecast of collapse.
The real battle
Networks are upgraded because service fails at the worst cell in the busiest moment, uplink is already the weak direction there, AI bursts are latency-sensitive and cannot be buffered like backups, so we build headroom before mass adoption rather than after. AT&T would add low-band spectrum that reaches indoors is how you build a consistent upstream, so the EchoStar deal is exactly this preparation.
My answer concedes most of it, because most of it recommends things operators should do anyway. Look at what AT&T actually bought. Low-band coverage spectrum and fiber-fed small cell densification improve the network for every traffic type, voice, video calls, fixed wireless, enterprise, and yes, whatever AI traffic materializes. It is ordinary, sensible modernization for the earnings call. The same is true of the vendor feature list, uplink carrier aggregation, transmit switching, scheduler intelligence. Much of it is justified by the existing traffic mix with no AI in the picture.
So the disagreement is not investment versus no investment. It is about scale, timing, targeting, and attribution. The evidence supports ordinary uplink modernization plus selective deployment where measurement shows need. It does not yet support attributing a nationwide investment cycle to AI, accelerating replacement on the strength of glasses forecasts and per-task multipliers, or the leap from “improve uplink where users congregate” to “the uplink imperative demands the next G.” That distance is the distance between an engineering budget and a supercycle, and the commercial weight of the industry, vendors and operators alike, is leaning on making it invisible.
What would change my mind
The thesis is falsifiable. It weakens materially if operators begin identifying AI applications as a meaningful contributor to busy-hour uplink utilization in congested cells, from their own network data rather than cited industry research. If AI traffic produces measurable uplink latency or loss increases after controlling for total load. If correlated AI activity causes repeated congestion outside specially engineered venues. If glasses payload and query frequency stay high despite on-device preprocessing. Or if operators change TDD configurations or densification plans specifically because measured AI load exceeded planning assumptions.
AT&T has, in effect, volunteered to run the test. It has the thesis, the spectrum, and the motive to prove the demand is real. The day its executives show busy-hour, cell-level agentic traffic measurements from their own network instead of the 450 percent industry statistic, this analysis updates. I found no public operator disclosure attributing material uplink congestion to AI applications, and the operator statements that exist rest on the same industry forecasts as the vendor decks. Vendor and analyst scenario models.
Where this lands
The industry has shown a direction of change. It has published ratios, shipment counts, modeled percentages, per-task multipliers, and application demonstrations. It has not published the cell-level workload distribution, the burst correlation, the cellular duty cycle, or the session failure rates that turn a directional observation into a capital requirement. Meanwhile the one live-network study on record found its most consequential result in connection behavior, a problem whose remedies begin with software and standards discipline. And the strong telco AI story of the moment, inference infrastructure built on fiber, facilities, and interconnect, does not need the uplink panic at all.
The right posture is to measure AI-specific busy-hour load, session latency, signaling behavior, and failure rates at the worst cells. Continue ordinary uplink modernization where it improves the whole traffic mix. And decline to label that spending an AI supercycle until measured AI workloads, rather than shipment forecasts, directional ratios, and per-task multipliers, account for the requirement.
