By Joe Salesky, Head of AI Research

Based on interviews with 193,266 Americans conducted between March 2025 and February 2026

A massive and largely quiet transformation is underway in the American workforce. Recon Analytics tracks AI adoption through weekly pulse surveys of over 300,000 respondents annually. The data reveals that 42% of employed AI users now direct those tools toward work tasks, generating an estimated $420 billion in annual productivity value. Only half of these workers say their adoption came through any formal company program. The other half found their own way.

The Workforce Picture

33% of employed Americans now use AI for work tasks. Among knowledge workers in financial services, professional services, scientific and technical fields, communications, media, law, healthcare, education, and government, the rate is 37%. Among employed people who already use AI in any capacity, 42% use AI tools in their jobs.

Workers Are Leading, Companies Are Following

Work AI users who answered questions about their adoption pathway (n=20,934 since July 2025), roughly half report either using AI independently or working at a company with no formal AI initiative. 20% say they adopted AI through a deliberate, self-directed choice with no company involvement at all; another 38% report their company has no formal AI initiative or they are unsure whether one exists. The workforce is pulling AI adoption forward while many organizations are still debating governance frameworks.

For those who can identify initiative drivers, executive leadership accounts for 33% and IT departments for 47%. These figures overlap because respondents can select multiple drivers. The gap between the 47% who credit IT and the 50% operating outside any formal program is a governance challenge.

Who Pays for AI

29% of work AI users pay for a premium subscription to their primary tool. Employer or team funding covered 40% of paid subscriptions in August 2025. That figure has climbed to 50% in February 2026. Corporate procurement is catching up, while half of all paid work AI subscriptions still come out of workers’ own pockets.

Workers spending $20-30 per month of their own money on AI tools they believe make them more productive, without being asked and often without their employer’s knowledge, is a revealed preference. These workers have already conducted the cost-benefit analysis that many enterprise procurement processes are still debating.

Retention Is Strengthening, Not Weakening

For work AI users on paid subscriptions, cancellation intent has declined steadily from 26% in August 2025 to 20% in February 2026. The trajectory is consistent: Four in five paid AI subscribers plan to keep paying.

A 20% cancellation rate in a product category barely two years old compares favorably to mature SaaS categories that typically see 25-35% annual churn in their first 18 months. The downward trajectory suggests that workers who survive the first few months of paid use are settling into habitual usage patterns. Paid AI has crossed from experiment to infrastructure.

The Productivity Signal

Among 18,117 work AI users surveyed over the last 90 days, 97% report at least one tangible benefit and 94% report measurable time savings. The top benefit is increased productivity and efficiency at 45%, followed by quick generation of actionable information at 34%, enhanced decision-making at 28%, and automated repetitive tasks at 24%.

Workers who report hours saved (n=16,154) average 6.3 hours per week. Some of AI’s productivity value does not register as “hours saved”, a developer who writes cleaner code the first time, a marketing analyst who produces three campaign variants instead of one, a sales rep who personalizes 50 outreach emails instead of sending one generic blast: none of these show up as recovered hours. They show up as better output. Self-reported time savings captures one dimension of a multi-dimensional productivity gain.

Based on 161.3 million employed Americans (BLS, 2024 annual average), with 33% using AI and 42% of those directing it toward work, approximately 22 million workers are using AI on the job. At the self-reported mean of 6.3 hours per week and BLS total compensation of $48.05 per hour for civilian workers (Employer Costs for Employee Compensation, June 2025), the annualized productivity value is approximately $340 billion. Adjusted for the occupational skew of AI work users toward professional, technical, and financial roles that carry above-average compensation, the estimate reaches approximately $420 billion in annual productivity value.

What This Means

For enterprise IT leaders, internal AI adoption metrics almost certainly undercount actual usage. Dashboard-tracked deployments of Copilot, Gemini for Workspace, or other enterprise-licensed tools capture only the formal channel. Workers using personal ChatGPT Plus, Claude Pro, or Perplexity subscriptions on their own devices are invisible to those dashboards. The opportunity is to channel demonstrated demand into governed platforms.

For AI platform vendors, employer-funded subscriptions grew from 40% to 50% of paid work AI users in six months. That trajectory points toward corporate procurement absorbing what workers started on their own. The conversion trigger for free users considering paying is integration with existing productivity applications: Microsoft 365, Google Workspace, Salesforce. Prior reports have noted that Enterprise data infrastructure expansion will clearly grow both enterprise paid usage and business impact.

The comprehensive report providing deeper analysis, conclusions, and recommendations is available on ReconAnalytics.com.

---

Methodology: Recon Analytics AI Pulse Survey. Continuous weekly data collection, March 2025 to present. Total respondents: 193,266 (n=16,154 for hours-saved analysis, last 90 days). Monthly sample sizes range from 944 (March 2025, early panel) to 34,253 (October 2025, full panel). Quality controls reject unqualified responses at 4-10x the industry standard. Productivity value calculation: BLS Employer Costs for Employee Compensation (ECEC), June 2025 release, civilian workers total compensation $48.05/hour; BLS Current Population Survey, 2024 annual average, 161.3 million employed.

Contact: [email protected]

AI Revenue: Paid Adoption Requires Data, Not Better Models

February 17, 2026 | Joe Salesky, Analyst & Head of AI Research

---

Eight breakthrough model releases in seven months produced four percentage points of daily usage growth. 75% of Americans have tried AI, but only 25% use it daily. The bottleneck is not intelligence. It is trust and context. Privacy scores -30 cNPS across the entire AI ecosystem, the lowest attribute by 28 points, based on more than 120,000 respondents surveyed over 38 weeks. Users who connect AI to their private data convert to paid subscriptions at 3x the rate of those performing basic tasks. The model capability race has hit diminishing returns. The next phase belongs to whoever solves the data infrastructure problem for consumers and enterprise.

The 50-Point Gap Between Trial and Habit

Three out of four Americans have tried an AI tool. One in four uses it daily. The 50-point spread between trial (76%) and habitual use (26%) is not a marketing problem. It is a utility problem.

The middle of the funnel defines the conversion opportunity. 19% of respondents use AI weekly, suggesting it occupies a place in their routine but has not become essential. Another 10% use it monthly. Combined with the 21% who tried AI and walked away, these cohorts represent half the population: consumers who engaged with AI and did not find sufficient value to stay.

The demographic surprise: Millennials, not Gen Z, are the heaviest users. Daily usage peaks in the 30-44 age bracket at 36%, five points ahead of 18-to-29-year-olds at 31%. The explanation is context and compensation. Mid-career knowledge workers face more tasks that align with AI capabilities. The inflection arrives at 60, where daily usage drops to 11%. Seniors have not rejected AI. Nobody has shown them what it can do.

Privacy Is Both the Lock and the Key

45% of Americans remain outside the AI economy. Among those who have never tried AI, 44% cite lack of interest, 25% cite ethical concerns, and 24% cite distrust of AI answers. Cost ranks last in both the never-tried and tried-and-quit segments. Only 5-9% cite expense as a barrier. Free versions exist for every major platform. Price is not the obstacle.

Privacy is the top barrier and the top motivator. 21% of never-tried respondents say knowing their data would be safe and private would motivate them to try AI, the only response exceeding 20%. The same consumers who say they do not trust AI say privacy assurance would bring them in. This is not a contradiction. It indicates that trust is the gate through which new users must pass.

Among active users, the picture is worse. Privacy scores -30 cNPS, nearly 28 points below complete experience (-2). Every platform in the dataset reports negative privacy cNPS, from Perplexity at -18 to Meta AI at -43. Zero for twelve. Apple Intelligence scores -25, underperforming its privacy-centric brand positioning. Payment helps: paid users report -9 versus -34 for free users, a 25-point lift. But no platform achieves positive territory.

ChatGPT Dominates. Satisfaction Tells a Different Story.

ChatGPT leads primary platform selection at 50%. Google Gemini follows at 20%. Microsoft CoPilot (8%), Apple Intelligence (6%), and Meta AI (6%) form a competitive middle tier. The top four platforms represent 83% of primary selections.

Scale inversely correlates with satisfaction among free users. Perplexity, at less than 1% share, posts cNPS of +8. ChatGPT, at 48% of free users, registers -1. Google Gemini scores -6. The pattern is monotonic: larger user base, lower satisfaction. Smaller platforms attract self-selected enthusiasts who chose the tool deliberately. Larger platforms accumulate casual users through distribution advantages and brand awareness.

Paid users report higher satisfaction across every platform without exception. The aggregate cNPS gap: +18 for paid versus -7 for free, a 25-point differential. ChatGPT moves from -1 to +25. Payment is not a gamble. It is a reliable upgrade across the board. The challenge is getting users past the trust barrier to that first payment.

The 3x Conversion Divide

The use case hierarchy is steep. Web search leads adoption at 43%, followed by writing assistance at 33% and topical research at 26%. These information retrieval tasks require no private data, no account connections, no infrastructure beyond the prompt itself. Transformation use cases cluster at the bottom: data analysis at 20%, coding at 9%, automation at 9%.

The conversion gap tells the real story. Coding assistance converts at 41% to paid subscriptions. Automation follows at 37%. Data analysis converts at 31%. Information use cases trail: web search at 17%, topical research at 21%. The 2x gap between transformation and information use cases is not about features. Both tiers offer the same use cases. This is about user profile. Users who need transformation already have infrastructure that makes AI valuable.

Crossing work context with subscription status creates four segments with dramatically different satisfaction profiles. Work + Paid users report productivity cNPS of +23. Home + Free users report -20. The 43-point gap defines the challenge for consumer AI. A Home + Paid user (-1) is no more satisfied than a Work + Free user (-1). Work context provides more satisfaction improvement than payment alone. The difference between +23 and -1 is the combination of subscription plus infrastructure, not the subscription by itself.

The Second Inning Is About Data, Not Models

Data-dependent users, the 18% who employ AI for data analysis or automation, report productivity cNPS of +9 versus -13 for non-data users. They show 61% daily usage versus 28%. They convert to paid at 32% versus 9%, a 3x advantage. This minority demonstrates the engagement and monetization patterns every platform seeks. When users connect AI to their private data, they use it more, value it more, and pay for it more.

The entities with access to private data are not primarily AI companies. Apple holds photos, messages, health data, and financial transactions for over a billion users. Google holds Drive documents, Gmail archives, and location history. Microsoft holds OneDrive files and Outlook correspondence. These companies have the Organize and Unify layers that AI platforms lack. The strategic landscape creates natural partnership opportunities: AI platforms have models but need data access; device and cloud providers have data but need AI differentiation.

The shift is from “ask me anything” to “help me with my data.” The former is a party trick that saturates quickly. The latter is a utility that compounds with use. The stronger engine is ready. It needs better tires for traction.

---

Report Details

Genius Myopia: Why Smarter Models Aren’t Enough

How Trust and Context Unlock the Next Stage of Adoption | March – December 2025

The complete 23-page report includes detailed analysis of:

• Usage frequency distribution and demographic breakdowns across age and income segments
• Barrier analysis for the reluctant 45%, including never-tried and tried-and-quit segments
• Platform-by-platform market share, satisfaction, and subscription dynamics
• The O/U/T Framework: Organize, Unify, Transform as the value creation model for consumer AI
• Connectivity as infrastructure proxy: fiber converts at 2x the rate of cable for paid AI
• Privacy cNPS across all twelve platforms and the payment improvement effect
• The Second Inning Playbook for AI Platforms, Device OEMs, Cloud Providers, and Investors

📊 View Report Details & Purchase →

For licensing inquiries: [email protected]

 

AI Choice 2026: Why Licenses Don’t Equal Adoption

February 3, 2026 | Joe Salesky, Analyst & Head of AI Research

---

Despite Microsoft’s enterprise distribution advantages and Office 365 integration, Microsoft Copilot lost 7.3 percentage points of paid subscriber share in seven months while Google Gemini gained 2.9 points, based on more than 150,000 respondents. Distribution advantages do not lock in market position. Employees receiving enterprise AI tools evaluate options and select based on experience. The platform that delivers the most reliable results wins, regardless of vendor seat licenses.

The 39% Market Contraction

Copilot’s decline from 18.8% in July 2025 to 11.5% in January 2026 represents a 39% contraction in market position among U.S. paid AI subscribers. This occurred during a period when Microsoft actively invested in enterprise distribution and deepened Office 365 integration. The platform accesses the same OpenAI models as ChatGPT, so underlying capability is comparable. The divergence in user perception points to product experience and integration execution rather than model quality.

ChatGPT maintained dominant share near 55.2% with modest erosion from 60.7% in July. Gemini climbed from 12.8% to 15.7%, crossing Copilot in late November to claim the number-two position. The two platforms now separated by more than 4 percentage points, with Gemini’s trajectory continuing upward while Copilot stabilized in the 10-12% range.

Exhibit 1: Primary Platform Share Among U.S. Paid AI Subscribers

 

Source: Recon Analytics U.S. AI Survey, July 2025 – January 2026. U.S. paid subscribers only.

The Workplace Conversion Gap

When Copilot is the only AI platform an employer provides, 68% of workers adopt it as their primary tool. This demonstrates meaningful uptake when alternatives are absent. The competitive dynamics shift when employers offer multiple platforms.

Among workers with both Copilot and ChatGPT available, Copilot’s adoption falls to 18% while ChatGPT captures 76%. When all three major platforms are available, only 8% choose Copilot while 70% choose ChatGPT and 18% choose Gemini. The pattern is consistent: as worker choice expands, Copilot adoption collapses.

ChatGPT converts 83.1% of U.S. paid subscribers who have workplace access. Copilot converts 35.8%. Gemini converts 34.0%. The 47-point gap between ChatGPT and Copilot quantifies the adoption challenge. Microsoft’s Office 365 distribution creates exposure. That exposure does not automatically translate to preference when workers evaluate alternatives.

Exhibit 2: Workplace Conversion Rates by Platform

 

Source: Recon Analytics U.S. AI Survey, July 2025 – January 2026. Workers with paid AI subscriptions only.

Quality Perception Drives Share Movement

Gemini’s rise correlates with quality leadership. The platform posts the highest accuracy satisfaction scores among major competitors, 23 points above Copilot and 9 points above ChatGPT. Copilot’s decline correlates with the lowest accuracy perception in the market. Quality drives share movement, not deployment volume.

Copilot’s accuracy NPS remained persistently negative throughout the measurement period. July showed -3.5, September declined to -24.1, and January finished at -19.8. The gap is not closing. Users who tried Copilot and stopped using it cited distrust of answers at 44.2%, exceeding comparable figures for Gemini (42.8%) and ChatGPT (40.6%).

The correlation between accuracy perception and market share movement is direct. Platforms investing in quality gain share. Those relying on ecosystem integration without matching quality lose share.

The Enterprise Market Remains Contestable

Multi-platform enterprise deployment is common. Among U.S. paid subscribers with Copilot available at work, over half also have ChatGPT available. Workers evaluate options and select based on experience rather than defaulting to whatever platform their employer provisions.

The enterprise AI market is not winner-take-all. Every renewal cycle presents an opportunity for challengers to displace incumbents based on demonstrated superiority. Microsoft’s enterprise dominance in productivity software does not foreordain Copilot’s dominance in AI. Google’s Workspace position does not guarantee Gemini’s success. The platforms that execute on accuracy, integration, and use case demonstration will capture enterprise spend in 2026.

ChatGPT’s 83.1% workplace conversion rate and 55.2% market share reflect entrenched product-market fit. Competitors are not displacing ChatGPT. They are competing for second position. Gemini demonstrated that share can shift when product experience improves. Copilot demonstrated that share can decline when it does not.

---

 

On January 16, 2026, OpenAI announced plans to test advertisements in ChatGPT’s free tier and the new $8/month “Go” tier in the United States. The move was widely anticipated: advertising funded the scaling of Google Search and Facebook and has been expected as the monetization path for consumer AI services. With OpenAI reportedly losing over $11.5 billion in Q3 2025 and projecting infrastructure spending of $1.4 trillion over the next eight years, the decision reflects economic necessity rather than strategic pivot.

 

Commerce Secretary Lutnik signaled in an interview with Broadband Breakfast on March 5th, 2025 that the US government will rethink the BEAD program so that Americans “get the benefit of the bargain.” He elaborated that it could mean that homes get broadband through satellite instead of fiber. “We want the lowest cost broadband access to Americans,” he said.

Secretary Lutnik gets it. Connecting a location to broadband for $10k-$30k makes more sense than spending $60k to do the same thing. By approaching the issue from a technology-neutral perspective, we can connect a lot more people for a lot less money while improving connectivity and satisfaction with the connection.

From February 28th, 2024, to February 28th, 2025, Recon Analytics surveyed 160,848 people in the United States and asked them about their broadband experience with their current provider.

To determine satisfaction with their service, we asked consumers a standard net promoter question as developed by Bain, but modified to ask about specific components of the customer experience. Below you see a heatmap of Recon Analytics’ component NPS scores for connectivity. I omitted the customer interaction part of the heatmap for readability.

The 1,113 Starlink subscribers we interviewed over the last year were the most satisfied, followed by consumers of fixed wireless (FWA) with cable, and DSL customers being the least happy. Why are Starlink customers and FWA customers happier than cable and DLS?

While Starlink and FWA can be slower than speeds over cable, most consumers are not engineers or economists who make decisions based solely on technical merit or price. In the case of both Starlink and FWA, our data confirms that customers value the fact that the technology solution is easy to install and easy to get rid of if the consumer is not happy. In Starlink’s case, we have about three times as many former Starlink customers as current Starlink customers. This indicates that there are quite a few people who were unhappy with the solution and swapped out for a different one.

At the same time, the people for whom Starlink or FWA works are very happy with it, especially in comparison to what they had before. Interestingly, we consistently find in our polling data that the higher the cost for a service, the lower consumer satisfaction is, all other things held equal. The low satisfaction scores for cable, even though cable internet service is substantially cheaper than fiber, is a clear indication that cable needs to do a much better job in serving its customers. As confirmed in the quarterly reported financial data, customers leave services with low satisfaction and join services with high satisfaction. In the home internet case, customers choose a more expensive service because, in their experience, the cheaper service is not worth it.

We are also seeing in the data that the Starlink and FWA routers are state-of-the-art equipment. Investment in good routers leads to better scores for how well existing devices stay connected to the Wi-Fi network and how easy it is to connect to Wi-Fi. It also aids in every other connectivity and customer issues metric. This is demonstrated again in the table below.

We are also asking all of our home internet respondents how often they have subjectively experienced internet issues in the last 90 days. The data below is again from the 160,848 people we surveyed over the last year.

What is striking is how well Fiber, FWA, and Starlink are performing when it comes to reliability. When looking at reliability from a customer perspective, it is the interplay between the connection, which is determined by network technology; the router used or supplied; and the end user equipment. The end user equipment is the same for every customer – a mix of smartphones, laptops, desktops, and other connected devices. While Verizon FiOS and AT&T Fiber customers report fewer instances of their internet connection going down, the FWA and Starlink routers are able to mitigate a lot of the more difficult connection technology challenges.

When comparing the data presented in this research note with what was presented six months ago, the Starlink scores for connectivity increased as the company launched more satellites during that period. As Starlink continues to launch more satellites, its scores will change depending on what increases faster – the number of satellites or the number of customers.

Furthermore, it is interesting to see where Starlink customers came from. Eighty-five percent of Starlink respondents are from rural areas, consistent with Starlink’s reporting of where it sells. Almost 31% came from small ISPs. For more than 11% of respondents, Starlink is the first home internet provider that they have, followed by CenturyLink, Charter, Frontier and Comcast, who provide a lot of internet coverage in rural areas.

Recon Analytics data shows that a technology-neutral approach is the right way to go for allocating federal dollars to get affordable broadband out to as many Americans as will take it. There are many Americans, especially in very rural areas, who are very happy with Starlink’s service. Fixed wireless is also solving the broadband rubric for many customers in a satisfactory way. Fixed wireless is especially valuable in less densely populated areas, where ample spectrum and thereby speed and capacity is shared among fewer people resulting in higher speeds. Fiber, without a doubt, is the workhorse technology for more densely populated areas, where satellite and FWA do not have sufficient capacity given the current licensed full-power spectrum constraints to serve customers well.

What we need, and what Secretary Lutnik rightfully alluded to, is a technology-neutral approach where the Americans can choose how they want to be connected at the lowest price considering the circumstances.

Introduction

With the advent of 5G and, with it, a technology called network slicing, many parties, ranging from mobile network operators and enterprises to policymakers, are re-examining how to deploy customized networks that were previously unfeasible. The flexibility afforded by network slicing will allow wireless operators to more efficiently meet the needs of their enterprise customers, particularly those customers concerned about the potential costs or burdens of “do-it-yourself” or outsourced private networks. Alternatively enterprises may choose to deploy a private wireless network to meet their needs. This would require a specific plan to utilize spectrum they obtain or lease from another entity along with a private network managed by an operator, network supplier, or viable third party.

Network slicing allows a network operator using the same physical wireless network to provide virtual slices with different characteristics to serve different customer needs. It allows the operator to tailor the technical characteristics such as bandwidth, latency, and security for certain types of applications. This is especially significant for enterprise customers whose desire to enjoy the economic benefits associated with large-scale network operators; instead of defaulting to largely homogenized offerings, companies and operators now enjoy unprecedented flexibility to develop customized solutions at a much lower cost.

• Flexibility comes from the move towards virtualization and software-defined networking, where network infrastructure that was an integrated software and hardware component becomes disaggregated into a software program running on general purpose computers and servers.
• Cost savings come from the shared use of physical infrastructure when the new specialized tasks only require the installation of a new software application solving, the enterprise’s specific need.

The enhanced capabilities and cost savings allow enterprises to switch from a production process that has been hardwired to a flexible wireless approach, just as mobile phones have replaced landline phones as the preferred way to communicate. In particular, the development of network slicing allows providers and customers throughout the wireless ecosystem to continue to enjoy the efficiencies of flexible use licenses and larger geographic license sizes, while also benefitting from small-scale customization.

5G Network Slicing

Over the last five years, the concepts of software defined networking (SDN) and network function virtualization (NFV) have taken large strides in the telecom world. The change from application-specific hardware to networks that are built on a general-purpose computing foundation, implemented and orchestrated by software, is a watershed event. The typical wireless network of 2015 consisted of roughly 30,000 to 50,000 different pieces of hardware with integrated software. The hardware elements (also called SKU for Stock Keeping Units) included base stations, internet routers, messaging gateways, multimedia gateways, switches and many other components. For every function in the network, one new SKU was created. These SKUs are highly efficient at running the specific task for which they are designed as long as they are close to being fully utilized. The networks are also designed around the busiest time of the day for the specific function, plus a significant margin for growth so that customers can always use any function they might want to at any time of the day. Usage is highly variable both over time and across functions which are being used, leading to a significant amount of idle capacity and significant cost.

As MNOs have transformed their networks from hardware-centric to software-centric, wireless networks have become a lot more flexible. By being able to use each network function as an application, these developments have allowed MNOs to create network slices that provide service assurance by creating virtual wireless networks as part of the overall wireless network.

The key advantages of network slicing are:

1. Enhanced mobile broadband. It allows the operator to guarantee reliable, ultra-high speed data connections for applications such as live 4k video streams.
2. Very low latency. It makes applications such drone flying beyond visual range possible and among other things opens up the ability to inspect power lines and buildings.
3. Massive IoT. Thousands of IoT devices such as sensors can be installed per square mile allowing for a range of applications such extremely even temperature control on a shop floor where material variances are measured in mu requiring temperatures to be within one degree for product uniformity.

Importantly, these key advantages can be mixed and matched to suit different industry verticals, all over the same physical network.

Below in Exhibit 1, we have abstracted the network into three layers: The radio access network (RAN) layer, which is responsible for the wireless connectivity between the device and the network; the core network function layer (core) that controls how voice calls and data sessions are connected; and the network application layer on which the various services run. The network application layer consists of either a private or public cloud, centralized or at the edge of the network. This way, a specific amount of resources can be allocated to every network task. In the exhibit, the width of the bars is the amount of resources guaranteed to each slice. The MNO can thereby guarantee all customers who have bought a network slice a minimum performance through a service level agreement (SLA.)

For example, the MNO has set aside a specific amount of network resources for its retail or consumer customers (the MNO Network Slice). Since its customers are using a balanced number of applications, it needs a corresponding amount of core and RAN resources to serve all of its customers, both consumers and enterprise customers, well. The MNO is also hosting an MVNO which is smaller than the MNO and has therefore purchased fewer network resources that are also balanced. The same with a large industrial provider and an IoT provider who have created their own private networks by acquiring smaller slices of the network. On the right side of the exhibit are two examples that show the strength of the network slicing model. In the green case, a customer purchases a video-centric slice, either for internal use like a large close circuit video network or a company providing video services to its customers. Video is the RAN killer app, both in terms of popularity and in terms of RAN resources needed. For that reason, the company purchases a significant amount of RAN resources. At the same time, a single use needs less core resources and even fewer resources at the network application layer, may it be in the form of a private, public or mobile edge cloud, as it streams the video. A contrasting example would be a game-centric slice. Multi-player online games are not data intensive, using only a small fraction of the data bandwidth that a video stream uses, but are highly latency dependent. To ensure the required latency, gaming demands a service with a higher amount of core resources and an even higher amount of network application resources where the processor-intensive calculation takes place.

Exhibit 1:

Source: Recon Analytics, 2021

Network slicing allows MNO administrators to custom tailor which resources are needed to deliver the services that the customer wants and needs by guaranteeing their availability at the network level.

Building a wireless network involves a significant amount of investment regardless of whether the network serves one customer or one hundred million customers. Building a RAN with towers, antennas and radio heads, a core network and application network layer with servers is a significant undertaking, even for established MNOs. Network slicing allows non-telecom companies to have the benefits of a private, secure, custom-tailored network meeting their needs without having to build, manage, maintain and upgrade it themselves. It is the difference between building your own car and leasing a car to drive to work every day.

Telecom engineers were determined to discover better ways to run a network than with 50,000 individual components. Using data centers and even personal computers as an example, they posited that it would be more efficient, flexible and agile to virtualize network functions (NFV), turning the vast majority of SKUs into computer programs and using general purpose computers or servers to create what is known as software-defined networking (SDN). This solves various issues:

1. Agility: With an SDN it is much easier to launch new products and services. Instead of having to justify the cost and complexity of having to add another SKU to the network, mobile network operators (MNOs) can simply launch the service by starting the software program. Adding a new service becomes (almost) as easy as installing a new browser or word processor on a PC. If the service is successful, it takes up more data center resources, using excess capacity of a service that has fallen out of favor. If wildly successful, the MNO can simply add more servers. If the service fails to catch on, the application either uses minimal network resources or is deleted. Thus there is no need to pay for, install, and potentially remove hardware from the network.
2. Reduction in network complexity and cost: As mentioned before, an average mobile network has 30,000 to 50,000 different SKUs performing different network functions. The fully virtualized SDNs that are operational today have reduced that number to six.. Fewer network elements means less complicated and expensive maintenance and purchasing costs as well as fewer hardware incompatibilities that need to be overcome. With fewer network elements, fewer people are needed to operate the network, thereby lowering operating costs. Since the network runs on a large number of commoditized servers that are interchangeable, costs are lower due to larger purchase volumes of generally available commoditized hardware.
3. Speed and scale of innovation: By using standardized, general purpose server hardware, wireless networks are joining the larger and faster speed of innovation rather than continuing to rely on the speed of innovation in niche applications. Furthermore, by shifting innovation to the software layer, many more companies can develop new software than what would be possible in an integrated hardware and software model.
4. Security and vendor independence: By using standardized hardware to run the network, it becomes less likely that malware gets introduced to the network through compromised hardware as the removal of compromised hardware is expensive and time consuming. If any software is compromised, it can quickly and inexpensively be replaced by a competing product. Furthermore, by relying on software over standardized hardware, MNOs can much more easily switch from one vendor to another for technical or commercial reasons.
5. Ease of management: In an SDN, all virtualized network functions can be controlled and changed from one platform and one location. In a traditional wireless network, individual network elements have to be individually changed and monitored, often with different applications that do not work together, and the changes have to be made locally. Most importantly, when network software runs on standardized hardware, it is possible to allocate specific resources to predetermined tasks, something that is called network slicing.

Until now, enterprise customers whose operations were dependent on features like these were incented to build and operate (or outsource) their own customized networks, entailing significant capital and operational expenses since this approach would not incorporate the efficiencies passed through to customers by large-scale traditional mobile network operators. Network slicing now allows enterprise customers to share in those efficiencies of scale, while still gaining the advantages of network customization – the best of both worlds.

Network Slicing Use Cases

5G network slicing allows guaranteed network performance that was not possible in 4G. This allows mobile network operators to offer a whole new set of capabilities that were previously not possible.

Industrial use case

5G is ideal for factories as it removes the need to lay cables to run different machineries that cost in the hundreds of thousands of dollars depending on the size of the factory floor. Network slicing, with its guaranteed performance, allows factories to connect precision machinery that needs to react to changes in the production process with less than 10 millisecond reaction time, like metal works or chemical processes. It also allows the production process flow on the shop floor to be changed according to new and different tasks without incurring cost prohibitive expenses and time of rewiring the shop floor. In addition, the material flow on the shop floor can be automated and, with below 10 millisecond reaction time, centrally controlled. Work-in-progress material vehicles can shuttle material from work station to station. By shifting the work-in-progress material flow from manual labor to an automated process, the number of work-related accidents on the shop floor can be significantly reduced, improving employee health and decreasing work related healthcare costs.

Just like in the hospital use case, the industrial company can choose which elements and functions it wants to source from the mobile network operator and which parts it wants to own itself for maximum flexibility. For example, in an automobile manufacturing plant, the automobile manufacturer installs its own antenna network, uses its own industrial applications to run the robots and manufacturing street, but relies on the mobile network operator to integrate it and run it on the mobile operator core.

Drones beyond visual line of sight

One of the advantages of 5G and network slicing is the guaranteed low latency and high data through put. What has been an exclusive purview of the military, is coming to the civilian sector: Flying drones beyond the visual line of sight (BVLOS.) The FAA approved BVLOS flights for public safety in August 2020, with a commercial application approved in January 2021. Due to the reaction time limits by traditional technology the drone cannot be further away that 1,500 feet from the pilot and more than 50 feet above or within 400 feet horizontally of any obstacle. The 5G and network slicing operators can guarantee a stable connection and low latencies that will allow drones to fly wherever there is a network connection and much closer to buildings, cables, and anything else the FAA describes as obstacles. Powerlines, local and long-distances, as well as cell cites need to be visually inspected in regular intervals to ensure the structural integrity of the units is still warranted. Currently, a lot of these inspections are done either by car or on foot with binoculars or personnel has to climb up the structure to inspect it. Due to OSHA regulations, they have to be turn off to be inspected, which impacts customers. When the industry switches to BVLOS one drone pilot can inspect these structures from a central location leading to lower cost and improved inspections as the drone will deliver a 4k video stream that is then recorded. Successive videos of the same structures can be compared against previous recordings to evaluate how quickly the structure is withstanding environmental impacts and aging. This will allow for preventive maintenance reducing cost and uptime. For the public safety community, BVLOS provides the opportunity for firefighters to have a better view on forrest fires than what is possible with planes. Drones can fly closer and slower to the ground than planes or helicopters, without endangering a pilot, giving firefighters a better understanding of conditions before they fight the fire in person. Futhermore, use of augmented reality (AR) and virtual reality (VR) technology can aid flying the drones and better help with inspecting various pieces of equipment.

Augmented and virtual reality use case

Combining 5G with network slicing allows reliable and predictable performance of augmented and virtual reality applications. For example, television programs can use 5G to holographically capture interviews and events, transmit them to a TV studio, and project the live images, allowing people in the studio to interact with the holographic images. Examples of this include interviews from sports events where the athletes are projected straight into the studio to be interviewed. Network slicing is particularly useful for large, highly latency-sensitive data streams for multimedia applications, since it allows for the dedication of bandwidth and computational resources needed to ensure flawless delivery.

Secure banking use case

Mobile payment and banking applications that are currently operating on the common network can be enhanced by creating a network slice that is dedicated only to a bank or payment system’s dedicated network slice. This allows a complete separation of the payment and banking app from the commonly shared wireless network, allowing for greater security and flexibility.

International examples

The advent of network slicing increases the efficiency and attractiveness of larger geographic license sizes. Internationally, the most prevalent way spectrum is licensed is through nationwide licenses. Some large countries like Canada and Brazil follow the US model of splitting the country into a number of licenses. In particular, Canada with regional telecom providers has a similar system for the same reasons as the United States, namely to allow regional providers an opportunity to participate in the wireless market place.

With the advent of 5G, several European countries have set aside spectrum for Industry 4.0 companies in the 3.7 GHz to 3.8 GHz band. Despite its moniker, the licenses are available for all companies at costs that are in the thousands of dollars per year. One of the leading examples is Germany, with its significant manufacturing sector. The German government lays out four scenarios[1] of how German companies can get use a 5G campus network, ranging from a slice of a public network, a hybrid-shared RAN where the public RAN is supplemented with private small cells, a hybrid RAN with a private core, and a fully separate stand-alone network.

The premise of this approach is that allocating spectrum in smaller parcels to individual companies would be more efficient than allocating larger flexible use licenses to MNOs, but experience has not borne this out as the uptake has been limited and is projected to continue to be minor. The first companies that have made announcements on starting 5G private networks are either stand-alone networks or network slices, but almost all of them are being built together with an MNO, either as the provider of a slice or as the lead partner to the customer. For example, General Motors and Honeywell have a 5G indoor system deployed in the United States with Verizon[2]. In Germany, Lufthansa together with Vodafone has launched a stand-alone private 5G network[3] and OSRAM together with Deutsche Telekom has launched a hybrid slice network[4]. In the UK, Ford together with Vodafone[5] and Air France and Groupe ADP are planning a 5G network[6], and in China multiple companies are building or have already built 5G networks with China Mobile, China Telecom and China Unicom respectively[7].

Bitkom, the 2,700-member German industry association for the digital sector, conducted a survey of companies regarding their plans for 5G[8]. Half of all companies considered 5G an important factor for their company in the future. Thirty-six percent of companies are planning or discussing using a network slice from an MNO for their 5G needs and six percent are planning to use the 100 MHz license that the German government has set aside for private company networks. As one of the leading high-tech manufacturing countries in the world, the survey result showing that only six percent of German companies are interested in deploying a private 5G network, where the enterprise owns the radios, network core and applications, on dedicated spectrum is sobering. The advent of network slicing allows enterprises to avoid taking on those responsibilities while still enjoying the benefits of customization.

Opportunity cost of different license area sizes or licensing schemes

Exhibit 2: Industrial locations in Germany with more than 1,000 employees

Private networks, like the industrial networks envisioned in Germany, still require network design, someone who builds the network, operates it and maintains it. This is a challenge for even the largest industry providers. As a result, they are either looking to either large system integrators or mobile operators to provide these services.

An additional challenge is the relatively small license areas for these networks. The smaller the license area, the more challenging network design becomes. In the low- and mid-band, radio waves inevitably will travel further than the smallest, city-block-size census blocks in major metropolitan markets. This creates interference issues with neighboring antennas, especially if they are owned by a different licensee. The interference issue gets smaller with higher frequency spectrum as the signal is increasingly attenuated and has problems penetrating walls.

Another issue to consider is that if only a small number of companies are willing or eligible to use the spectrum, the unused spectrum lays fallow and cannot be used otherwise. This is particularly a problem for low and mid-band spectrum that would otherwise be deployed in wider areas.

Even in a country as large as Germany, with twice as many people than California on half the geography, larger businesses with more than 1,000 employees are highly concentrated as we can see in Exhibit 2[9]. In large swaths of Germany, 100 MHz of 5G spectrum that is reserved for companies will lie fallow as indicated by the large area of grey land and cannot be used for important tasks such as wireless broadband solutions to the home or businesses. This runs counter to the idea that MNOs in both urban and especially rural areas can solve connectivity issues driven by inadequate landline solutions. Network slicing by MNOs allows enterprise customers to avoid the inefficiencies of small license sizes, while still achieving the customization benefits of private networks,

Conclusion

5G and Network slicing are opening up new opportunities for consumers and businesses alike. It terms of added flexibility network sliceing is the best thing that happened to mobile networks since sliced bread. 5G, SND and network slicing eliminate the delineation between telecom and information technology and create a contiguous development space for software engineers. As software is eating the world, this includes wireless networks, just as wireless is eating the world as well. This development opens up new opportunities ranging from augmented and virtual reality to drones to more fully optimized factories.

The cost savings from replacing expensive cabling to reducing workplace accidents are significant as wireless connections replace wired connections. The increased flexibility of changing the workflow on a factory floor is substantial. The ability to fly drones beyond the line of sight will simplify many tasks requiring people on the ground to inspect structures. Considering that the core competency of regular companies is their specific focus and not that of running IT systems or even highly sophisticated 5G networks on a limited scale, the deployment scenarios we see in Germany and other European companies will be that of network slicing or a hybrid solution that MNOs will manage. While it is a possibility that we will see stand-alone networks, the trend in corporate support functions is towards outsourcing as it often provides a better solution at a lower cost.

Most companies have outsourced their entire IT segment to specialist companies that can provide a superior experience for a lower cost. Even email, a relatively simple application, has become a hosted solution provided by large IT companies. Network slicing allows for custom network applications previously deployed on private networks or not at all by essentially creating a network as a service approach, where connectivity is just one of the components of an enterprise solution. Network innovations that would take a separate investment in a private network come when using a network slice as the overall MNO network is being upgraded. By working with an MNO and using a network slice allows companies to benefit from technical innovation for free as the MNO will upgrade their network to remain competitive. As a stand-alone network the enterprise would have to bear the cost for it. Private networks are also an option, but each enterprise should consider the requirements to develop and manage such a network – from spectrum, to design, development, and ongoing operations.

Network slicing allows enterprises to have all the benefits of customized private networks without having to build, operate and keep current the network increasingly becoming essential to remain competitive.

This paper has been commissioned by CTIA – The Wireless Industry Association

---

[1] https://www.bmwi.de/Redaktion/EN/Publikationen/Digitale-Welt/guidelines-for-5g-campus-networks-orientation-for-small-and-medium-sized-businesses.pdf

[2] https://www.verizon.com/about/news/verizon-5g-ultra-wideband-general-motors-honeywell

[3] https://www.vodafone.com/business/news-and-insights/company-news/lufthansa-technik-and-vodafone-business-launch-5g-private-network

[4] https://www.telekom.com/en/company/details/5g-technology-in-campus-networks-556692

[5] https://newscentre.vodafone.co.uk/news/how-ford-and-vodafone-are-creating-the-5g-factory-of-the-future/

[6] https://www.air101.co.uk/2020/02/adp-group-company-hub-one-gets-10-year.html

[7] http://www.xinhuanet.com/english/2020-05/09/c_139043925.htm

[8] https://www.bitkom.org/Presse/Presseinformation/Grossteil-der-deutschen-Industrie-plant-mit-5G

[9] https://www.jpnaumann.de/fileadmin/user_upload/Publikationen/JPN_Publikation_Industriestandorte.pdf