AI is everywhere. When you use a search engine, shop online or interact with a company’s customer support, AI is often embedded in the process. At the same time, active consumer use of generative AI services like ChatGPT, Gemini, Claude or Perplexity is rising. The recent launch of Meta’s Muse brought this to the forefront, with high early adoption proving the appeal (and market value) of consumer-focused AI services.
While AI is computationally intensive, the good news for mobile operators is that its current network requirements are fairly modest – especially when focusing on the ways mobile consumers interact with generative AI services. While increasing numbers of consumers interact with AI with voice and video (more on that below), much interaction, if not most, is still text-based. Our research suggests a text prompt needs just 0.01Mbps download and 0.2Mbps upload. This is just a fraction of what any modern mobile connection can deliver. And yet averaged across the 140 countries we tested, just short of three-quarters of mobile network tests were good enough for text AI. The leading cause of failure here is time to first byte – analogous to time to first token. This is how long it takes for a connection to be established. A slow time to first byte could make AI conversations feel slower and immersive.
In this report we unpack how ready mobile networks around the world are for the different ways consumers interact with AI services, the different network limitations affecting experience, and the broader implications for equitable global AI adoption.
Consumers' use of AI in their personal lives is growing
Across the world AI adoption is growing, with 18.8% of the world’s working-age population using AI as of Q2 2026. While much research focuses on AI in the workplace, personal AI use is rising. On ChatGPT, the volume of non-work messages grew faster than the volume of work-related messages between June 2024 and June 2025. On Gemini, 86% of interactions are now outside of work.
Mobile devices are central to this adoption. Device manufacturers have built AI functionality directly into handsets, meaning users can access generative AI services without having to directly open an app or browse to a webpage. As of January 2026, Apple partners with Google’s Gemini to augment its Siri assistant. Meanwhile, Samsung’s Galaxy S26 allows users to select which AI service they would like to use for their AI assistant, as well as interact with AI directly by long-pressing the power button. Some requests can be handled on the handset, but those that don’t get routed to a generative AI service.
Meanwhile, AI services like Perplexity have explored working with mobile operators, offering year-long AI trials to an operator’s subscribers. This pattern repeats across Japan (Softbank), India (Airtel), Canada (Bell), Germany (Telekom), Singapore (Singtel) and Australia (Optus) among others. Softbank’s announcement of the trial explicitly references the goal to “bring the benefits of AI to all walks of life as soon as possible”, emphasising the role of mobile operators (and their networks) in connecting AI to consumers.
Where (and how) consumers use AI shapes their usage, and changes AI’s network requirements
Mobile interaction with AI also tends to be more conversational than on a computer. On Copilot “users reach for their phones [to interact with AI] for the topics that are more personal or pique their interest beyond work”. Users’ top mobile interactions overwhelmingly centered on searching for information or advice on various topics. In contrast, on computers, top interactions included more intensive and work-focused behaviors: editing, creating and learning.
These personal uses take a very conversational, personal tone: for example, conversations on Anthropic’s Claude outside of work center around emotional and physical wellbeing, finances, and general companionship. This type of conversational, advice-centered interaction demands immediacy. While consumers may be more understanding of delays working on a large file, these back and forth conversations quickly expose when a network causes delays.
Consumers largely interact with AI via text today, but this is beginning to shift. In recent months, increasing numbers of consumers now use voice and live video to interact with AI services. As ofAugust 2026, 63% of Gemini users now use voice at some point in their interactions with Gemini. One in five Gemini Live interactions also involve live camera feeds or screen sharing.
Understanding the requirements of consumer AI interactions
Text AI services have an incredibly small impact on networks. A text prompt can be as little as a few kB of data. As such, our thresholds for it are very low – just 0.01Mbps down, and 0.2Mbps upload. Similarly, for voice, 0.2Mbps is enough to transceive high quality audio in both directions.
Video is more challenging, however video calling is a well established use case, requiring just 2Mbps of upload speed. This exceeds Microsoft Teams’ suggestions for HD video in a one-to-one call (1.5Mbps). Generative AI services then typically respond over voice or text.
Mobile networks are capable of reliable text interactions in the majority of countries around the world. In 117 of the 140 countries we tested, users can use text AI services more than half of the time. However this falls to 83 of 140 measured countries who are text AI ready at least 75% of the time. Furthermore, as AI interactions get more challenging this number drops. For voice, 45% of countries meet this 75% threshold, while for video this drops to 39%.
AI readiness diverges sharply by country income - driven by network inconsistency
Digital inclusion started as simply a question of access. Now as more users connect, the disparity in connection adequacy is clear. Looking at AI Readiness we can see the stark gaps at different country income tiers. AI text readiness in low/lower middle income countries is about 35pp lower than we see in high income countries. On top of this, the decline in score between text interaction AI Readiness and video interaction AI Readiness is even greater. 11pp separates text and video AI interactions in high income countries, compared to 20pp in low income countries.
Moving from text to voice interactions with AI adds conditions around network responsiveness and stability (latency, jitter and packet loss). Moving from voice conversations to video then layers on higher upload requirements.
For all low/lower middle income countries we measured, we see a notable decline (greater than 5pp) in AI readiness just moving from text to voice. This indicates just how much of a hurdle network responsiveness and stability continues to be in these countries. This is echoed when we dig deeper into the exact KPIs tests fail against. An average of 24% of completed tests in low/lower middle income countries fail the latency conditions for voice and video interactions, compared to 1% in high income countries, and 24% fail packet discard (compared to 6%).
Upload’s role in AI is clearest among high-income countries
High income countries face challenges with network responsiveness and inconsistency too, albeit to a lesser degree. 20 out of 52 high income countries see a notable decline in voice AI readiness compared to text, and 22 see a decline from the upload challenges of video conversations compared to voice. This would seem to echo industry concerns about the demands of an “uplink-first world”. However, it’s important to keep an eye on scale. An average of just 6% of tests per country fail to meet our video AI upload requirements – slightly less than the number that fail for packet loss. Over three-quarters (78%) of tests in high income countries still exceed all of our video AI Readiness thresholds.
Some AI applications – like image and document editing – do further increase upload requirements. For our image editing AI Readiness score, we use an upload throughput threshold of 3Mbps (enough to transfer a typical photo within a few seconds). Here, 10% of tests in high income countries fail to meet this slightly-higher threshold. However overall image editing AI Readiness is actually marginally higher than Video AI Readiness (80% vs 78%), because it does not need the same degree of responsiveness.
Future use cases likely will increase upload needs, but the thresholds here so far remain moderate. Ericsson predicts smartglasses users will need up to 5Mbps to share video snippets with AI, while autonomous vehicles could need up to 10Mbps to upload telemetry data. For some perspective, among high income countries, average upload speeds per country range from 6.6Mbps to 28.7Mbps. 90% of high income countries have an average upload speed that exceeds that 10Mbps threshold. The challenge is moving the boundary from this being the average experience, to the baseline.
Building for a more equitable AI adoption future
The countries with the strongest networks offer their consumers the greatest flexibility to benefit from AI – whether that’s from existing use cases or as-of-yet unknown potential applications down the line.
Yet even now, as AI’s network requirements remain so marginal, the current disparity between countries risks deepening digital divides. Network instability compounds with device availabilityand the strains local AI datacenters can place on local water and power resources. The net impact is that lower income countries stand to once again miss out.
This is reflected in the variation in consumer AI adoption seen at different country income thresholds. The countries with worse mobile AI readiness, even just for text, are also the ones where AI adoption is constrained.
This matters because using AI, even in consumers’ personal lives, frees up time. Google research suggests that in the US alone, just 30 minutes of personal time saved per week equates to unpaid productivity gains that “could be worth approximately $100 billion”. Lower income countries currently miss out on this opportunity. Not only this, but AI familiarity is becoming increasingly important in the workplace. Restricting opportunities to access and develop skills around AI may limit the economic benefits AI could bring.
Solutions to address the challenges of AI mobile network readiness
While a common challenge of network consistency unites all countries, no one solution overcomes all challenges. Ensuring a country’s mobile networks can stand up to current – and future – AI demands will require the combined efforts of many stakeholders.
Governments and NGOs: work alongside local operators to prioritize access to modern network infrastructure in developing markets
In markets where mobile users spend less time connected to 4G or 5G, mobile AI readiness is notably worse. Based on our 2026 Q2 Global Network Excellence Index, there are 60 countries where users spend less than 90% of their time on a 4G or better connection. For governments and non-governmental organizations (NGOs) focused on closing the digital divide, these represent prime targets for additional private investment to support local operators with connecting users. The GSMA’s 2026 State of Mobile Internet Connectivity report highlights the impact of device access inequality. Investment in devices needs to also come with investment in backhaul and the core network, as well as broadening networks’ reach. The rise of AI makes this need more acute than ever.
Regulators: broaden focus from coverage to day-to-day consistency
Consumer day-to-day AI use does not radically reshape network requirements, nor does it need high peak speeds or ultra-low latency. However, it does reinforce the need to monitor both basic access and adequacy at scale, rather than peak capabilities or latest-generation access and adoption. Setting expectations with operators (either as guidelines or using other regulatory levers), around both latency and the percentage of tests meeting low thresholds of upload, and prioritizing widespread adequate access is needed for current and near-term future consumer AI needs.
Operators in mature mobile markets: highly-targeted, data-led interventions offer best results
Discussion of networks for AI often talks about a revolution in global network traffic, requiring radical shifts in network planning. However, even as AI traffic is rising, it is currently negligible. Cisco’s numbers suggest that it could become a meaningful component of overall network traffic by 2035, however that assumes sustained growth at these current rates – a big assumption. This likely means wide-scale network improvements are unnecessary, especially in the near future. Of course, in a small handful of rural locations, operators may benefit from improving last-mile connectivity. The goal is not uniform 5G access. Given the limited requirements from AI services, even some of the near-term non-terrestrial networks may be good enough for AI.
Elsewhere, treating AI as a last-mile capacity challenge will deliver limited returns. Instead, operators should look at targeted interventions. If congestion affects AI traffic, dedicated network slices for AI can present a revenue opportunity (if consumers or AI services are willing to pay for it), as well as provide more reliable service for a low-impact set of data traffic.
AI service providers: Work with operators on moving AI closer to the consumer
Where last-mile network does not represent a limitation, the biggest opportunity for improving AI performance and reliability for consumers is moving AI closer to the user.
This means investing in new edge locations, such as local sovereign data centers, as well as reinforcing fiber backhaul and core network infrastructure, focusing on reducing latency – and improving time to first token for users. As consumers become more impatient, working in partnership ensures that neither telcos nor AI companies become the perceived roadblock to a good AI experience.
None of these interventions require operators to radically rethink network resources, or even spend heavily on the latest 5G Advanced (or 6G) technology. Consumer AI readiness is simply an extension of existing expectations of reliable, ubiquitous connectivity. But the price of not doing them, or getting them wrong, isn't a slower AI experience. Instead, it's a widening gap in who gets to benefit from AI at all.
Regional leaders
Text chatVoice conversationalVideo conversational
Sub-Saharan AfricaSouth AfricaSouth AfricaSouth Africa
Europe and Central AsiaEstoniaEstoniaEstonia
Middle East and North AfricaTunisiaTunisiaOman
Latin America and CaribbeanPuerto RicoArgentina, UruguayUruguay
East Asia and PacificSouth KoreaSouth KoreaSouth Korea
South AsiaBangladeshIndiaIndia, Maldives
Methodology note
Opensignal’s testing is designed to measure the real-world experience of users, from their devices, in the places they actually want to use networks. We test against a range of common internet endpoints including Content Delivery Networks (CDNs) such as Google, Akamai and Amazon. This is identical to the way connections are made each day to typical websites and content and provides a truly representative end-to-end view of network experience. At present, no standardized global convention defines the precise network requirements for consumer AI applications. Because generative AI models and interaction modalities evolve continuously, network performance demands will inevitably shift. Rather than relying on static industry assumptions, our framework combines empirical testing across real-world endpoints with established benchmarks for general internet responsiveness, voice transmission, and interactive video quality to create these baseline thresholds.
While this report focuses on primary consumer interaction modalities – text, voice, and interactive video – Opensignal’s Mobile AI Readiness framework also evaluates domain-specific workloads, including coding and media/document generation.
©2026 Opensignal, Limited - All rights reserved.
Opensignal Limited retains ownership of this insight including all intellectual property rights, data, content, graphs & analysis. Reports and insights produced by Opensignal Limited may not be quoted, reproduced, distributed, published for any commercial purpose (including use in advertisements or other promotional content) without prior written consent. Journalists are encouraged to quote information included in Opensignal reports and insights provided they include clear source attribution. For more information, contact press@opensignal.com.
