Category: 3. Business

Introduction

Diabetes represents one of the greatest public health challenges of the 21st century, contributing to morbidity, mortality, and economic loss worldwide. Currently, 589 million adults (aged 20-79 years) worldwide have diabetes (approximately 1 in 9 adults), and this number was projected to rise to 853 million by 2025 [,]. India accounts for a rapidly growing disease burden, particularly among its rural and underserved populations []. Escalating rates of type 2 diabetes and prediabetes are closely linked to shifts in lifestyle, dietary patterns, and limited access to preventive health care []. Rural communities are often disproportionately impacted due to limited health care access, low health literacy, and barriers to effective behavioral risk reduction [].

Mobile health (mHealth) interventions have emerged as promising tools for addressing such challenges by delivering targeted health education and behavior change messages directly through mobile device. A growing body of work also demonstrates that mHealth interventions, including SMS- and app-based educational programs, can improve diabetes-related health behaviors and clinical outcomes in low-resource settings []. Recent systematic reviews and trials show significant benefits for glycemic control, adherence, and lifestyle change in both urban and rural Indian populations [,]. At the same time, studies from 2024-2025 emphasize the emerging utility of artificial intelligence (AI)–powered platforms that dynamically tailor content using real-time user data, potentially setting a new standard in digital health communication []. Several contemporary randomized trials and reviews report that both traditional and AI-enabled mHealth modalities are feasible and beneficial, with limited but rapidly increasing direct comparisons in rural and underserved populations [].

Despite this progress, most previous interventions deploy generic, “one-size-fits-all” messages that do not consider individual behavioral patterns, motivational states, or changing needs over time []. This lack of personalization may lead to low engagement and suboptimal outcomes, and few published studies have directly contrasted static mHealth with AI-personalized approaches in rural Indian settings [].

Key uncertainties persist regarding the comparative effectiveness, user acceptability, and scalability of AI-driven personalized navigation versus simpler, static messaging strategies in real-world rural communities with varying digital access and literacy []. There is a pressing need to evaluate which digital strategies offer the maximum benefit for the least cost and the highest reach.

In this study, we rigorously compared, in a large rural cohort, an adaptive AI-driven mHealth intervention (dynamic) with a standard, static messaging system for diabetes prevention. To bridge the gap between the two types of interventions, this study leveraged reinforcement learning algorithms to optimize message delivery by sending customized messages based on an individual’s preferences (healthy food intake, unhealthy dietary habits, physical activity, diabetes symptom knowledge, and awareness of complications), aimed at promoting a healthy lifestyle based on the transtheoretical model of behavior change []. This individualized, data-driven communication framework represents an innovative strategy for enhancing engagement, contextual relevance, and behavioral adoption in diabetes prevention among rural populations. Our study will offer a model for future user-centered digital health interventions that can be customized for diverse behavioral domains beyond diabetes prevention. Therefore, this formative evaluation aimed to assess the effectiveness of an AI-enabled, personalized mHealth messaging intervention compared to traditional, nonpersonalized mHealth messaging in promoting engagement with diabetes risk reduction behaviors among adults in the rural district of Gulbarga, Karnataka, South India, while also examining how the intervention functions in real-world settings, what the participant engagement patterns are, and what the potential areas for improvement are [].

Methods

Study Design, Setting, and Population

The study used a quasi-experimental pre-post design with a control group and was conducted from January to November 2022, with recruitment and baseline data collection occurring between January and March 2022 and follow-up data collection completed from September to November 2022, among the rural population in the district of Gulbarga, Karnataka, South India. Gulbarga (also known as Kalaburagi), is situated in northeastern Karnataka, part of the Deccan Plateau, and is characterized by a semiarid climate and a predominantly agrarian economy. With a population of approximately 2.5 million (based on the 2011 Census, India), the district has a literacy rate of 64.8% []. Mobile phone usage among the rural population has experienced significant growth in recent years, with increasing adoption for communication, agricultural information access, and digital payments, although challenges in network connectivity persist in remote areas. The study population consisted of adults without diabetes aged 18-60 years.

Sample Size

Sample size calculations were based on a two-sided significance level of 95% and a statistical power of 80%, assuming a 10% difference in the primary outcome (physical activity) between the AI-enabled mHealth (intervention) group and the traditional mHealth (control) group. Based on these parameters, the calculated sample size was 415 participants for each group. To account for an anticipated 20% nonresponse rate, the target sample size was adjusted to 520 participants per group, resulting in a total target enrolment of 1040 participants.

Recruitment of Participants

Frontline workers (FLWs) facilitated the recruitment through community mobilization events in Gulbarga, aimed at raising awareness about the benefits of the AI-enabled, personalized mHealth messaging intervention used in this study. All participants received comprehensive information about the study’s purpose, objectives, potential risks, and benefits before they provided written consent for participation based on the eligibility criteria. Participants’ eligibility was determined based on self-reported or absence of physician-diagnosed diabetes (convenience sampling). Individuals with a known diagnosis of diabetes mellitus or that confirmed by medical records or current use of antidiabetic medication were excluded. This study followed a quasi-experimental design, without randomization. Eligible participants were divided into two groups, intervention and control, with 541 participants in each group.

Following recruitment, participants received an opt-in link, and FLWs guided them through the WhatsApp opt-in process. Upon successful opt-in, participants began receiving diabetes prevention messages starting the next day. The first follow-up was conducted 6 months after baseline data collection, using in-person interviews to fill out the survey at follow-up, facilitated by the FLWs. Blinding was not possible due to the nature of the intervention.

Intervention Design and Implementation

Arogya World’s mDiabetes program forms the foundation of the mHealth initiative evaluated in this study []. The mDiabetes program delivers diabetes prevention and control information directly to individuals’ mobile phones regardless of their risk status. Developed in 2011 in collaboration with the Rollins School of Public Health at Emory University, the standard program consists of 57 messages crafted using the transtheoretical model of behavior change. The system was built on a reinforcement learning framework using a Deep Q-Learning (DQN) algorithm, adapted for real-world mHealth deployment without prior training data, as described by Kinsey et al []. At baseline, participants completed a questionnaire assessing five behavioral domains: healthy food intake, unhealthy dietary habits, physical activity, diabetes symptom knowledge, and awareness of complications. Their responses were used in a warm-up phase to initialize state scores, ensuring that participants with lower baseline scores in a given domain received more relevant messages. Each week, the reinforcement learning agent delivered two customized health messages and two follow-up questions. Participant responses were used to update their individual state scores. A reward signal was generated when a participant’s score improved, and this was used to update the replay buffer.

The DQN was optimized weekly using minibatch gradient descent and a target network, with an ϵ-greedy policy applied to balance exploration and exploitation. Over time, this iterative process enabled the system to adaptively tailor content to participant needs (eg, used in the intervention arm: “You can reduce your risk of diabetes by walking briskly for 30 minutes daily; try walking to the temple or shops.”). These messages, available in 12 languages, were validated by Arogya World’s Behaviour Change Task Force comprising experts in diabetes, public health, and behavior change from both national and international spheres. Messages in the mDiabetes program are typically distributed as SMS texts, automated voice calls, or WhatsApp messages, sent twice weekly over a 6-month period. The program has reached approximately 2 million people to date, with this study using WhatsApp as the delivery platform.

A detailed description of building a customized messaging system for health intervention in underprivileged regions using reinforcement learning has been provided elsewhere () [].

The mDiabetes program includes an AI-based system to enhance traditional mHealth (diabetes) interventions by developing dynamic, customized text messages to improve adherence to diabetes prevention behaviors:

• AI-enabled mHealth: This intervention was different from the traditional mHealth program due to the addition of an AI system to develop a dynamic and customized text messaging intervention based on end-user feedback. Participants in the intervention group received two customized health-related messages on WhatsApp (containing information about diabetes complications and the impact of nutrition and physical activity on diabetes prevention), coupled with two questions probing their risk profile/behavior. The subsequent week’s messages for each participant in the intervention group were based on their responses to the two lifestyle-related questions from the previous week.
• Traditional mHealth: A total of 57 static mHealth messages were delivered twice a week via WhatsApp as per a standard scheduler for a period of 6 months, focusing on improving knowledge, attitudes, and practices related to diabetes prevention behaviors, including physical activity and dietary habits [].

‎

Data Collection

We assessed 2096 individuals for eligibility, of whom 1014 (48.4%) were excluded (n=598, 59%, did not meet the inclusion criteria and 416, 41%, declined participation). A total of 1082 (51.6%) participants were divided into two groups with an equal number of participants, an intervention group (n=541, 50%) and a control group (n=541, 50%). In the control group, 34 (6.3%) participants did not complete the WhatsApp opt-in process and therefore never received the intervention messages; these individuals did not contribute intervention or outcome data and were excluded from the analysis. Data for this study were collected from primary sources using structured questionnaires, direct interviews, and anthropometric measurements conducted by trained FLWs. Participant demographics, such as age, sex, education, and employment status, were recorded.

Physical activity was assessed through self-report questions on frequency and duration (≥30 minutes/day considered active), adapted from the World Health Organization guidelines on physical activity and sedentary behavior for adults [].

Dietary habits were measured by frequency of fruit and vegetable intake per day or per week and avoidance of high-fat foods. Secondary outcomes included knowledge of diabetes symptoms, complications, and preventive behaviors. These questionnaire domains underwent expert review by public health and behavioral science specialists and were pretested in a pilot sample for contextual relevance, clarity, and cultural appropriateness before field implementation. Anthropometric measurements, including height and weight, were obtained using calibrated instruments (a SECA 213 portable stadiometer and a SECA 803 digital scale, respectively). In addition, the BMI was calculated (kg/m²), and participants were categorized as underweight (<18.5 kg/m²), normal (18.5-24.9 kg/m²), and overweight/obese (≥25 kg/m²) []. All responses were self-reported except for anthropometric measurements. Awareness-related items (eg, “Are you aware of diabetes?”) and knowledge questions (eg, causes, complications) were scored dichotomously (yes/no). Lifestyle questions on diet and physical activity used frequency scales (eg, daily, 3-4 times/week, rarely) The same questionnaire was administered at baseline and endline to capture change over time (see ). Engagement and response data from WhatsApp-delivered messages were automatically logged by the mHealth platform.

Statistical Analysis

All statistical analyses were performed using StataMP 64 software (version 17.0). Descriptive statistics summarized the baseline characteristics and outcome variables for both groups. Categorical variables were reported as frequencies and percentages, and continuous variables were summarized using means (SDs).

Normality of continuous variables was assessed using the Shapiro-Wilk test. For normally distributed continuous variables (eg, BMI), independent-sample t tests were performed to compare group means. For categorical variables (eg, physical activity, dietary behaviors), chi-square tests were performed to examine group differences at both baseline and endline.

To evaluate the effect of the intervention, separate multivariable logistic regression models were constructed for each binary outcome variable (≥30 minutes of daily physical activity, daily fruit intake). Each model included the intervention group (AI-enabled vs traditional mHealth), the baseline value of the respective outcome, and additional covariates identified through univariate analysis (variables with P<.20 were considered for inclusion). Key demographic factors, such as age, sex, and employment status, were also retained, where appropriate. Both unadjusted odds ratios (ORs) and adjusted odds ratios (aORs) with corresponding 95% CIs were reported. For continuous outcomes, such as the BMI, ANCOVA was performed, with endline BMI as the dependent variable and baseline BMI included as a covariate to control for initial differences.

Where multiple comparisons were made across lifestyle behavior outcomes, Bonferroni correction was applied to control for type I error. The adjusted significance threshold was set at α/N, where α=0.05 and N is the number of comparisons. With eight outcomes, this yielded an adjusted threshold of P<.006.

Model diagnostics were conducted to assess multicollinearity and model fit (eg, pseudo R², Akaike Information Criterion). All statistical tests were two-sided, and P<.05 was considered statistically significant unless otherwise corrected via Bonferroni adjustment. Missing data were handled using complete-case analysis, and follow-up rates were reported to account for potential attrition bias.

Ethical Considerations

The study protocol was reviewed and approved by the Institutional Ethics Committee of Anusandhan Trust, Mumbai (reference number IEC26/2021) prior to implementation. The research was conducted in accordance with the ethical principles outlined in the Declaration of Helsinki and relevant national guidelines for biomedical and health research involving human participants. Informed consent was obtained from all participants before enrolment. Participants were informed about the study objectives, procedures, and potential risks and benefits and were assured that their participation was voluntary, with the right to withdraw at any time without affecting regular standard care. To ensure privacy and confidentiality, no personal identifiers were collected. The dataset used for analysis was anonymized and securely stored in password-protected files (server at Arogya World) and will be retained for 3 years, accessible only to the study investigators. This study was reported in accordance with the Transparent Reporting of Evaluations with Nonrandomized Designs (TREND) statement []. No monetary or material compensation was provided to participants for their participation. Additionally, no identifiable participant images or other personal visual materials were collected or included in the manuscript or appendices. In addition, regulatory guidelines of the Telecom Regulatory Authority of India (TRAI) were followed in sending text messages []. Our study was a quasi-experimental design and was carried out as a formative evaluation to understand how the intervention works in real settings, how participants engage with it, and how it could be improved, and there was no randomization procedure; hence, this study did not require mandatory clinical trials registration-India (CTRI) registration.

Results

Sociodemographic Characteristics of Participants

shows the participant selection flow diagram. Of the 1082 participants enrolled, 1048 (96.9%) completed the 6-month follow-up and were included in the analysis (n=541, 50%, in the AI-based mHealth [intervention] group and n=507, 48.1%, in the traditional mHealth [control] group), with 387 (36.9%) males and 661 (63.1%) females. The participants were largely within the age range of 26-50 years (n=723, 69%). Educational attainment varied across the sample, with nearly half of the participants (n=517, 49.3%) having completed some level of schooling and only 73 (6.9%) holding postgraduate or higher qualifications. Employment status showed that 76% (n=796) of the participants were employed, with a slightly higher proportion of employed individuals in the intervention group (n=426, 78.6%) compared to the control group (n=371, 73.1%). details the sociodemographic characteristics of both groups.

‎

Adherence to Different Components of Diabetes Prevention in Intervention and Control Groups

Physical Activity

The most notable intervention effects were observed in physical activity patterns, particularly in the intervention group. The percentage of participants engaging in regular physical activity increased from 66.7% (n=361) at baseline to 72.8% (n=394) at endline (P=.02) compared to the control group (n=333, 65.7%, at baseline to n=353, 69.6%, at endline; P=.18). Additionally, the duration of physical activity demonstrated substantial improvements in the intervention group, reporting a 15.4% increase in participants engaging in 30 minutes or more of daily exercise (P<.001) compared to a 14.8% increase in the control group (P<.001). details the adherence of both groups to different components of diabetes prevention.

Intake of Fruits and Vegetables

Both groups exhibited high baseline adherence to daily fruit and vegetable intake, with minimal changes observed postintervention. In the intervention group, 94.8% (n=513) of the participants continued to meet the recommended servings of fruits and 99.8% (n=540) for green vegetables, showing a slight improvement from baseline for vegetables (0.4%; P=.22). Similarly, the control group maintained comparable levels of dietary adherence, with only a negligible decrease in fruit intake (–1.7%; P<.001) and no change in vegetable intake ().

Behavioral Changes

Among participants receiving the AI-enabled mHealth intervention, the proportion who preferred walking short distances for daily chores rose from 96.7% (n=523) to 98.2% (n=531). This represented an increase of 1.5% (95% CI –0.4 to 3.4, P=.11). The use of stairs declined from 91.3% (n=494) to 87.6% (n=474), a change of –3.7% (95% CI –7.3 to 0.1, P=.48). In the control group, walking short distances fell from 97.8% (n=493) to 96.5% (n=489), a change of –1.3% (95% CI –2.9 to 1.4, P=.37), and stair use declined from 91.0% (n=461) to 88.1% (n=447), a change of –2.9% (95% CI –6.7 to 0.9, P=.07).

Participation in household chores increased significantly in both groups (). In the intervention group, the proportion of participants engaging in household chores rose from 93.5% (n=506) to 98% (530), an absolute gain of 4.5% (95% CI 2.1-6.9, P=.009). The control group showed a similar improvement (from n=474, 93.4%, to n=497, 98%), which also reached significance. Agricultural work participation increased from 75.7% (n=410) to 83.4% (n=451) in the intervention group (7.7%, 95% CI 2.8-12.4, P=.002) and from 73.5% (n=346) to 80.2% (n=407) in the control group (6.7%, 95% CI 1.5-11.9, P=.014).

Primary Outcomes

Physical Activity

shows the logistic regression results for achieving at least 30 minutes of daily physical activity at endline. After adjusting for baseline status and other covariates, there was no significant difference between the two groups. Participants in the intervention group had similar odds of meeting the 30-minute daily physical activity goal compared to those in the control group (aOR 1.0, 95% CI 0.7-1.3, P=.74). Baseline physical activity was a strong independent predictor of endline physical activity (aOR 2.1, 95% CI 1.5-3.1, P<.001). Older age was associated with greater odds of regular physical activity (aOR 3.8, 95% CI 1.6-9.3 for >50 years vs 18-25 years, P=.003), while being employed was associated with lower odds of daily physical activity (aOR 0.2, 95% CI 0.1-0.3, P<.001). These findings suggest that participant characteristics, rather than intervention type, are more influential in determining physical activity outcomes.

Daily Fruit Intake

As shown in , there was no significant difference in daily fruit consumption between the two groups at endline. After adjusting for baseline fruit intake and covariates, the odds of consuming fruit daily were modestly higher in the intervention group (aOR 1.4, 95% CI 0.8-2.3), though not statistically significant (P=.24). Baseline fruit intake was strongly predictive of endline fruit intake (P<.001). Age and employment were also associated with daily fruit intake: participants aged 26-35 years had higher odds of eating fruit daily than those aged 18-25 years (aOR 4.7, 95% CI 1.9-11.8, P=.001), while being employed was linked to lower odds of daily fruit intake (aOR 0.3, 95% CI 0.1-0.8, P=.02).

Body Mass Index

ANCOVA () showed no significant difference in the mean BMI between the two groups at endline. After adjusting for baseline BMI, the mean difference was essentially 0 (–0.0 kg/m², 95% CI –0.6 to 0.5, P=.95). Baseline BMI was a strong determinant of endline BMI (P<.001), but no intervention effect was detected.

Exploratory Behavioral Outcomes

Other incidental physical activity measures also showed no significant differences between the two groups. The aORs (95% CI) for the intervention versus the control group were as follows: stair use (aOR 0.9, 95% CI 0.7-1.4, P=.79), walking for chores (aOR 2.4, 95% CI 1.0-6.1, P=.06), helping with household chores (aOR 1.0, 95% CI 0.4-2.3, P=.94), and farm work (aOR 1.3, 95% CI 0.9-1.8, P=.19). See Tables S1-S4 in .

Message Delivery Rate, Responses to Feedback, and Engagement Over 6 months

Across both study arms, 85,000 WhatsApp messages were scheduled. About 82,600 (97.2%) of these were successfully delivered (95% CI 96.9-97.1), and approximately 77,000 (93.2%) of these were marked as read (95% CI 92.8-93.2) .At baseline, 65.1% (352) of intervention group participants reported making changes in response to messages; at endline, this proportion was 66.4% (n=359 participants). In comparison, 64.7% (n=328) of participants in the control group at both baseline and endline reported making changes in response to messages; see Table S5 in . In addition, engagement remained relatively stable over the 6‐month period ().

Discussion

Principal Findings

This study evaluated the effectiveness of an AI-enabled mHealth intervention (mDiabetes) versus traditional mHealth for promoting diabetes prevention behaviors in rural India. Our primary finding was that there were no significant differences between the AI-enabled and traditional mHealth groups for the primary outcomes of physical activity and dietary behaviors after 6 months of intervention. Both interventions demonstrated effectiveness in maintaining and promoting physical activity behaviors, with baseline activity being the strongest predictor (aOR 2.1, 95% CI 1.5-3.1, P<.001). This suggests that any form of consistent mHealth messaging may be beneficial for diabetes prevention regardless of AI customization. The finding that age>50 years is associated with higher odds of physical activity (aOR 3.8, 95% CI 1.6-9.3, P=.003) is particularly encouraging, as this demographic faces higher diabetes risk. Conversely, employment being associated with lower physical activity odds (aOR 0.2, 95% CI 0.1-0.3, P<.001) highlights the real-world barriers that working adults face in rural settings. Our finding are consistent with recent reviews that show that although mHealth interventions generally improve physical activity within groups, the incremental benefit of AI- or app-based enhancements over active comparators is often modest [,].

Our findings are in contrast with those of existing studies, which generally report declining physical activity with advancing age due to reduced mobility, chronic conditions, and competing health limitations []. However, some studies have reported that older adults may engage more consistently in routine or incidental activities, such as walking, household chores, or agricultural work, particularly in rural areas [].

For daily fruit intake, although the intervention effect was not significant, the age-specific patterns are noteworthy. Participants aged 26-35 years showed higher odds of daily fruit intake (aOR 4.7, 95% CI 1.9-11.8, P=.001), suggesting this age group may be more receptive to dietary behavior change. The negative association with employment (aOR 0.3, 95% CI 0.1-0.8, P=.022) again underscores socioeconomic barriers to healthy behaviors. Similar findings have been reported in other mHealth interventions in India, where improvements in fruit intake were observed but often without strong between-group differences. For instance, the SMART Eating trial in Chandigarh documented a significant rise in fruit consumption using IT-enabled strategies compared to standard education []. These findings suggest that AI-enabled approaches may promote healthier choices, such as fruit consumption [].

To the best of our knowledge, this study is the first in India to rigorously, directly compare, in a large rural cohort, an AI-enabled mHealth intervention (dynamic) with a traditional static mHealth intervention for diabetes prevention. Key novel features include its real-world rural setting with minimal exclusions, the innovative use of reinforcement learning AI to customize messages based on individual feedback, and its focus on adults without diabetes for prevention rather than diabetes management []. Additionally, the study achieved a high retention rate of 97%, demonstrating both feasibility and acceptability of mHealth in rural populations.

ANCOVA revealed no significant difference in the mean BMI between the two groups at endline after adjusting for the baseline BMI and other covariates (mean difference –0.0 kg/m², 95% CI –0.6 to 0.5, P=.95). Previous research suggests that even noncustomized mHealth interventions can yield positive outcomes, particularly in contexts where strong community support structures exist. Such interventions, regardless of personalization, have the potential to influence health behaviors; however, their effectiveness may be enhanced when integrated with complementary support mechanisms [,]. A recently published study [] showed that combined mHealth and community health education intervention improves diabetes awareness and healthy habits in rural areas, indicating potential for lasting outcomes and guiding future public health efforts in rural settings [].

The implications of these findings for public health practice are substantial. The demonstrated effectiveness of the AI-based mHealth intervention in increasing physical activity and maintaining healthy behaviors suggests that such tools could be crucial in diabetes prevention programs, especially in rural and underserved areas where health care resources are limited. Moreover, the ability of AI-driven interventions to provide customized guidance and real-time feedback makes them particularly suited for scalable, population-level health initiatives. These mHealth interventions can bridge significant gaps in health care delivery, particularly in resource-constrained settings []. However, the findings also suggest that a “one-size-fits-all” approach may not be sufficient. Integrating AI-driven mHealth interventions with existing health care systems, including community health workers and primary care providers, could enhance their effectiveness and sustainability.

Additionally, the existing literature also emphasizes the importance of physical activity and active lifestyles in managing health outcomes, particularly in mHealth interventions. Several studies have demonstrated that regular physical activity, such as engaging in 30 minutes or more of exercise daily, significantly reduces the risk of chronic conditions, including diabetes and cardiovascular diseases [,]. Moreover, nonexercise activities, such as household chores and farm work, which showed significant associations with better health outcomes in this study, are well recognized as beneficial contributors to physical and metabolic health. Evidence suggests that active engagement in household tasks and manual labor can improve cardiovascular health and reduce the risk of complications in chronic diseases like diabetes [,].

The significance of using stairs and walking for daily chores, particularly in the mHealth AI group, mirrors findings from other studies that promote incidental physical activity as a valuable component of overall health management []. Activities like stair climbing, which are simple to incorporate into daily routines, have been shown to improve cardiovascular function and aid in glucose regulation, both critical factors in diabetes prevention and management []. These findings underscore the importance of integrating physical activity, both structured and unstructured, into health interventions to enhance their effectiveness, particularly in AI-driven mHealth programs.

Strengths and Limitations

This study marks the first national attempt to use WhatsApp-based text messaging on mobile phones to support educational interventions aimed at preventing diabetes. The strengths of this study include a large sample size (N=1048), which offers adequate statistical power to detect intervention effects, and a high retention rate of 97% that minimizes selection bias and enhances the reliability of findings. Conducting the study in a real-world rural community setting further strengthened external validity. Additionally, rigorous statistical approaches were used, including appropriate adjustments for multiple comparisons, and a comprehensive set of outcome measures covered both behavioral changes and knowledge gains []. Moreover, in-person data collection by trained FLWs helped ensure data quality and reduced the potential for response bias, and the development of the intervention was guided by the transtheoretical model of behavior change, ensuring a solid theoretical basis.

However, the study has certain limitations. First, the reliance on self-reported data for physical activity and dietary habits (primary outcomes) is subject to recall and social desirability biases, potentially overestimating the true effects of the interventions. Second, the relatively short intervention period of 6 months limits the assessment of long-term sustainability of behavior changes. Third, the recruitment process involved an opt-in procedure, which could introduce selection bias, as participants who chose to participate may be more motivated to adopt healthy behaviors than the general population. Finally, biochemical markers were not objectively assessed to evaluate the clinical outcomes due to a lack of financial resources, which would have provided more detailed insights into the biological effects of the intervention. To address these limitations, further studies could incorporate objective measures of primary outcomes, such as accelerometers or pedometers for physical activity and validated dietary assessment tools, including food frequency questionnaires or 24-hour dietary recalls. Extending the intervention and follow-up periods would allow for evaluation of the sustainability of behavior changes over time. Recruitment strategies that ensure a more representative sample of the target population could help minimize selection bias. Furthermore, the inclusion of biochemical markers or other clinical endpoints would provide more robust evidence of the physiological and metabolic impacts of the intervention, enhancing the translational relevance of the findings. Additionally, reliance on WhatsApp messaging may have excluded individuals without smartphone access, limiting generalizability to economically disadvantaged populations. Collectively, these limitations likely bias the results toward the null hypothesis, suggesting that the true effects of the interventions may be underestimated rather than exaggerated.

Conclusion

This study revealed that engaging, well-designed static messages can be just as effective as complex AI-personalized approaches in diabetes prevention, challenging prevailing assumptions and pointing to cost-effective, scalable options for program managers and policymakers.

This study demonstrates that traditional mHealth interventions are as effective as AI-enabled approaches for promoting diabetes prevention behaviors in rural India. Although this finding challenges assumptions about the superior effectiveness of AI-powered health interventions, it provides valuable evidence for scalable, cost-effective diabetes prevention strategies. The high acceptability and retention rates of both AI-driven and traditional interventions suggest that consistent health messaging through accessible platforms like WhatsApp can effectively support diabetes prevention efforts in rural populations.

Rather than viewing the lack of AI superiority as a negative finding, this result should be interpreted as evidence for the democratization of effective health interventions. Simple, well-designed mHealth programs can achieve meaningful health behavior changes without requiring sophisticated technological infrastructure, making diabetes prevention more accessible to underserved rural populations.

Unlock the Editor’s Digest for free

Roula Khalaf, Editor of the FT, selects her favourite stories in this weekly newsletter.

When Dario Amodei left OpenAI in 2020, chief executive Sam Altman wished him well, anticipating that the project Amodei, his sister Daniela and other departees were planning would probably focus “less on product development and more on research”.

“We look forward to a collaborative relationship with them for years to come,” Altman wrote in a blog.

Instead, Amodei has become the sharpest thorn in Altman’s side. Anthropic, the artificial intelligence start-up he co-founded, will end the year with around $10bn in annualised revenue and is growing fast. It is in talks to raise funds at a valuation of over $300bn and is now laying the groundwork for a blockbuster initial public offering.

Amodei’s decision to leave OpenAI, and the work he has done since, is underpinned by two convictions, according to multiple people who know him. One, that he is more capable of building all-powerful AI than his former boss; and two, that the world would be safer if he does.

“He has a strong view on where he’s going. Dario understands you have to have a good business to pursue the mission,” says Matt Murphy, a partner at Silicon Valley investment firm Menlo Ventures, which led a funding round for Anthropic last year.

Investors in the company describe Amodei as someone whose evangelism for “safe AI” is wedded to keen commercial instincts. “Founders are either technical, good at product or sales. Dario is one of the few CEOs I’ve met in my life who does all three,” says Divesh Makan, founder of Iconiq Capital, which led Anthropic’s last funding round.

Amodei was born and raised in San Francisco by a mother who renovated libraries and a leathersmith father. The 42-year-old studied physics at Stanford and gained a PhD in biophysics before embarking on a career as an AI researcher. He joined Chinese internet giant Baidu in 2014 before a brief stint at Google Brain.

In 2016, he was among the earliest employees of OpenAI, founded by Altman, Elon Musk and nine others as a place to pursue AI research without the commercial pressures of a corporate tech parent. Amodei was instrumental in developing the large language models behind chatbot ChatGPT.

But after five years, Amodei left following disagreements with Altman over OpenAI’s direction and with concerns about AI’s potential for harm if appropriate guardrails were not put in place. In 2021, he co-founded Anthropic with his sister.

“What he thought was important was developing a company where these things could be deployed safely and transparently in the world,” says Ravi Mhatre, co-founder of VC firm Lightspeed Venture Partners, which invested over $1bn in Anthropic this year. “He felt he needed a clean slate.”

But this focus on safe AI development has earned him criticism in both Washington and Silicon Valley. David Sacks, Trump’s AI tsar, claimed in October that Anthropic was running a “sophisticated regulatory capture strategy based on fear-mongering”. Investor Marc Andreessen argues that extra AI regulation will impede US start-ups.

Critics cast Amodei as a “doomer” influenced by the effective altruism movement, which believes AI poses an existential threat to humanity. The Amodei siblings deny they are effective altruists or doomers. But the company’s early funding came from investors with ties to the movement, including Facebook co-founder Dustin Moskovitz and FTX co-founder Sam Bankman-Fried, who was later convicted of fraud.

There are also concerns from some in the AI sector that Anthropic’s rapid growth is now testing Amodei’s ability to balance the pursuit of “safe” AI with the needs of his shareholders.

“At the early stage of Anthropic, they very much said ‘we don’t want to fuel the AI race, we want to be just behind the frontier and do AI safety research.’ That’s clearly not the case now. Some people in the AI safety community are pretty unhappy with that,” said a person who works in AI safety.

Anthropic has received the backing of Google, Amazon, Microsoft and Nvidia, and earlier this year changed its stance on accepting funding from the Middle East. “Unfortunately, I think ‘no bad person should ever benefit from our success’ is a pretty difficult principle to run a business on,” Amodei told staff.

To articulate his strategic shifts and view of the world, Amodei likes to write lengthy public essays (the last one exceeded 13,000 words). During a five-hour podcast interview last year, he also took a brief diversion from discussing programming languages to hold forth on the meaning of life.

His earnest messages have been well received by the public. And his ebullience about his company’s mission is popular among employees, helping Anthropic retain top researchers in a competitive market. “He has cult leader status,” says the person in AI safety.

The company is now in the earliest stages of preparing for a public listing. There is strong demand from investors to own a slice of what Amodei has built.

“I can’t imagine the company without Dario. He is the person who spearheads the key technical challenges and motivates everyone,” says Lightspeed’s Mhatre. “What’s Apple without Steve Jobs or Microsoft without Bill Gates?”

george.hammond@ft.com

Additional reporting by Cristina Criddle and Tabby Kinder

Unlock the Editor’s Digest for free

Roula Khalaf, Editor of the FT, selects her favourite stories in this weekly newsletter.

US private capital group KKR is in talks to acquire Arctos Partners, one of the pioneers of the private equity industry’s push into professional sports, according to people briefed on the matter.

KKR has accumulated more than $700bn in assets, making it one of the largest players in private capital.

Its interest in Arctos comes as it looks to wealthy individual investors and ordinary retirement savers for future asset growth, making sports investments that attract interest from everyday investors an appealing new product offering.

The prospective deal would also signal KKR’s push into the booming market for second-hand private equity fund stakes, where Arctos’s executives have particular expertise.

Arctos owns stakes in some of the world’s most popular sports teams including European football giants Liverpool and Paris Saint-Germain.

One person with direct knowledge the discussions said KKR was in advanced talks to take a majority stake in Arctos. However, both people cautioned that the negotiations were ongoing and could yet collapse.

Arctos had also fielded interest from other large private capital groups and asset managers, the people said, and any deal would be likely to require sign off from the various professional leagues in which its portfolio teams compete.

Arctos was founded in 2019 by two executives who combined knowledge of sports, entertainment and private equity.

David O’Connor was previously an executive at the owner of the New York Rangers ice hockey team and the talent agency CAA. His co-founder Ian Charles was an early adviser to the now booming marketplace for second hand private equity fund stakes.

Together they have led private equity’s push into sports, buying minority stakes directly in popular sports teams.

In addition to its European football stakes, Arctos owns minority stakes in more than a dozen franchises.

Those include high profile US teams such as the National Basketball Association’s Golden State Warriors and Utah Jazz, the baseball World Series champion Los Angeles Dodgers, and two National Football League teams, the Los Angeles Chargers and Buffalo Bills.

It is also a minority investor in the Aston Martin Formula 1 team.

Arctos, which manages over $14bn in regulatory assets according to securities filings, has also built a business providing tailored debt and equity financing to the private capital industry.

Earlier this year, that unit of Arctos helped to finance the management buyout of private credit firm Hayfin from a Canadian pension fund.

KKR’s interest in Arctos comes as the New York based pioneer of big private equity takeovers has used its substantial cash reserves and valuable stock currency to expand into insurance and debt-based investments.

In 2023, KKR took full ownership of insurer Global Atlantic at a valuation of more than $7bn, its largest ever acquisition. While a price for Arctos could not be established, any deal would be among KKR’s biggest ever.

KKR and Arctos declined to comment.

Key insights:

• • • Blockchain’s transparency is a double-edged sword— While criminals use crypto for illicit activities, the permanent and public nature of the blockchain ledger creates an undeniable trail, making it a powerful tool for law enforcement to track and seize illicit funds.

    • The rise of crypto forensics— A growing industry of specialized firms and investigators is leveraging blockchain’s inherent design to unravel complex financial crimes, demonstrating that “lost” crypto funds can often be recovered.

    • An evolving battlefield— Despite the ongoing challenges posed by tools like mixers and privacy coins, blockchain technology is fundamentally shifting how financial crime is fought, turning the very system criminals exploit into the means of their capture.

---

Cryptocurrencies and other digital assets are used by criminals, which is great for catching them. Indeed, the biggest criticism of crypto since its inception has been its criminal use, which was estimated to be almost half of all activity by the end of 2017. In the past three months alone, asset seizures and forfeitures of more than $22 billion in crypto have been made by authorities in the United Kingdom, the United States, and their international partners.

These historic interceptions of illicit funds prove that the fundamental architecture of blockchain — the digital ledger that underpins most virtual transactions — makes it the perfect tool for catching criminals, validating the hypothesis of Satoshi Nakamoto, the presumed pseudonymous of the person or persons who developed bitcoin, that fraud could be prevented through intentional system design.

While criminals assumed they could optimize their illegal activities using crypto to obfuscate fund flows, the blockchain ledger’s immutability has created a niche for financial crime investigators seeking to unravel these cases. Companies like Chainalysis, Elliptic, and TRM Labs have become synonymous with these investigations, joined by a growing network of smaller firms that are democratizing crypto investigations, combating terrorist financing and online child abuse. Ultimately working to secure seized assets and prevent further harm. By all measures, the ecosystem is expanding rapidly.

Every crypto transaction creates a permanent trail that allows investigators to catch criminals even years after their crimes. This is how, a digital exchange hack in 2016 that resulted in the theft of 120,000 Bitcoin worth $72 million (at the time) and was chronicled in the Netflix documentary Biggest Heist Ever was wrapped up years later with the seizure of $4.5 billion in crypto and the arrest of the two alleged perpetrators in 2022. Law enforcement may not move as fast as crypto, but if the whale is big enough, they will catch it.

Indeed, the scale of cryptocurrency-enabled crime threatens Western economic stability. The FBI received 149,686 crypto-fraud complaints in 2024, totaling $9.3 billion in losses, likely significantly lower than the true figure. More than 100,000 people are trafficked and forced to operate scams from compounds in Cambodia and Myanmar. The Prince Holding Group, a transnational criminal organization headed by Chen Zhi, generated $30 million daily at its peak, approximately $10.95 billion annually.

Financial crime as economic warfare

These are just headlines. Further research in the Netherlands shows that only 11.8% of fraud victims actually report being victimized. While many dismiss fraud and blame victims, crypto-related fraud is becoming economic warfare systematically draining wealth from Western economies while enslaving hundreds of thousands in forced labor camps across the Global South. With potentially $80 billion lost annually to crypto fraud, the impact extends beyond the 1.14% of the US federal budget it represents. This illicit outflow causes loss of productive capital, tax base erosion, and reduced economic activity.

Yet the technology accused of enabling this new generation of fraud simultaneously provides the tools to detect and combat these criminal organizations more successfully than any financial crime fighting technology in history. The Chen Zhi case, easily the largest asset forfeiture in US history at around $15 billion, demonstrates this perfectly.

---

Every crypto transaction creates a permanent trail that allows investigators to catch criminals even years after their crimes.

---

This is why I’ve spent the last four years studying the crypto ATM industry. While most financial crime professionals saw a problematic service in a problematic industry, I saw a massive dataset of criminal activity that could predict other illicit activity beyond crypto ATMs. This dataset helped identify terrorist financiers, vendors of child sexual abuse material (CSAM), and countless scams and frauds. Layer data-rich sources like crypto ATMs with blockchain data, and a good investigator can achieve remarkable results.

Modern blockchain analytics leverage the features Nakamoto designed for trust and verification. Immutability makes evidence tampering impossible and investigations public; and verifiability allows investigators to validate every step of a criminal’s crypto trail. Consensus mechanisms create a distributed jury of millions, validating the evidence chain further. These features enabled authorities to map the Prince Holding Group’s entire criminal empire, revealing 76,000 fake social media accounts operated from facilities using 1,250 phones across 10 Cambodian compounds, and tie it to $15 billion in bitcoin.

The same technology facilitating billions of dollars in pig butchering scams annually enables law enforcement to catch the transnational criminals and recover funds. Traditional financial crimes disappear into offshore accounts and shell companies, often leaving investigators blind. However, as anyone in blockchain forensics knows, Locard’s Exchange Principle remains true: Every contact leaves a trace. Blockchain’s public ledger means every suspicious transaction leaves a permanent clue.

Nakamoto’s vision of “electronic transactions without relying on trust” inadvertently created a system for establishing criminal culpability. The blockchain’s public nature convinced criminals they could hide in plain sight, but Nakamoto saw that participants would be deterred from fraud by this transparency. The naive assumption that users had nothing to hide if doing nothing wrong quickly revealed plenty were doing wrong. Still, the system proved fit for purpose once tools were built to catch bad actors. Nakamoto’s white paper’s emphasis on preventing double-spending through public verification created a framework in which crime-spending leaves permanent evidence. All a good investigator needs is time.

The rise of crypto forensics

As crypto advances, tools like bridges, mixers, and privacy coins pose constant challenges for investigators, but claiming the money is gone when crypto is involved is simply false. As blockchain forensics advances, criminals face an uncomfortable truth: They’ve been conducting operations on a permanent, public, immutable ledger. Their only protection is time and cryptographic puzzles that an entire industry is working to unravel.

While some industry press reporting has been diligent in pointing out some of the challenges in the industry and some of what’s been missed, there are a lot more illicit fraud cases that never see the light of day because of what has been prevented by blockchain forensics. And while it may not be perfect, the fact that there is an industry working to build a safer financial system than what has gone before is commendable, and the accountability that public ledgers have enabled is energizing for those that must police it.

Unfortunately, the $15 billion Chen Zhi seizure isn’t the end but the beginning. With at least $64 billion stolen annually, these criminals have little incentive to stop. While some scam compounds have been dismantled, reports indicate they’re simply being relocated.

Nevertheless, blockchain is setting a new paradigm in financial crime, one in which the technology enabling crime will eventually become the weapon that defeats it.

---

As 2025 draws to a close, tech companies have continued to pour an unprecedented amount of capital into data centers, unveiling multi-billion-dollar investments to meet the surging demand for compute capacity. The United States is expected to continue to lead the way in new infrastructure development and deployment, though Western Europe and Asia are emerging as regions poised for significant growth in investment and build out.

By 2030, analysts are projecting that there will be over 2,000 new data centers constructed worldwide. Global data center infrastructure spending is projected to approach $7 trillion over the next five years, with the bulk of that capital flowing into servers and the chips that drive modern data center performance.  

Evolving Financing Models

While the financial figures continue to rise, considerations for investors across capital markets become increasingly complicated. Location, build-out timelines, and load capacity requirements are key factors shaping project risk. Given the significant capital expenditure involved, investors have had to adopt strategies that not only address the upfront spending required but also account for long-term economic and debt considerations.

To meet the scale of today’s buildouts, investors are increasingly relying on layered capital strategies that combine long-term financing with mechanisms that provide early cash flow. These include tenant prepayment structures, joint ventures with infrastructure funds and pension investors, and real estate–focused arrangements, such as sale-leaseback transactions.

At the same time, traditional project finance lenders are playing a growing role in the sector, underwriting large, syndicated loans supported by long-term leases, stable power-supply arrangements, and other risk-mitigation measures that can anchor billion-dollar developments.

Energy and Power Considerations

Despite the growing clarity around investment pathways, power reliability remains crucial with regard to the future of large-scale data center build-out efforts. With respect to data center power needs, two principal considerations emerge: ensuring reliable power generation and determining the vehicles by which that power is procured. Many hyperscalers have adopted ambitious clean energy goals and are prioritizing low-carbon alternatives like nuclear and renewables paired with energy storage systems.

Ensuring resilient transmission from those power sources to the data center can present significant challenges, particularly given the regulatory hurdles involved. Co-location has emerged as an option for data center developers, siting the facilities alongside existing power generation sources, along with Bring Your Own Generation (BYOG). Still, these arrangements also raise regulatory concerns, as the Federal Energy Regulatory Commission and state energy agencies have often scrutinized these efforts due to resource adequacy and reliability concerns.

Against this backdrop, nuclear power is quickly emerging as a central focus for data center operators seeking reliable, long-term, carbon-free power. Many hyperscalers are already forming partnerships with nuclear companies, exploring options that range from funding the development of new nuclear technology to the restart of previously retired plants.

Growing interest in new-build nuclear, both large-scale facilities and smaller, modular reactor designs, is positioning advanced nuclear generation as an increasingly viable component of future data center power strategies.

Server Connectivity

While reliable grid connectivity remains a necessity for any data center, so too is the high-capacity fiber connectivity that enables data centers to function at scale. Redundant, high-capacity fiber infrastructure is crucial to maintain a dependable connection between data centers themselves and with users and customers all over the world. Diversity in transmission connectivity can also boost overall reliability, with some data centers pairing multiple fiber lines with wireless or satellite point-to-point connections.

One of the primary ways global fiber networks are deployed is through submarine cables, which currently carry more than 98% of all international internet traffic and data. Hyperscalers have become the largest developers of these long-haul systems, leveraging direct ownership stakes and dedicated capacity arrangements to efficiently move data across their global facilities.

In contrast to the heavily regulated energy side, ownership and operation of data centers remain largely unregulated from a telecommunications perspective, at least in the United States. While licenses to operate data centers aren’t currently required by the Federal Communications Commission or most state telecom regulators, the network services provided by hyperscalers are often subject to various regulatory obligations. To secure connectivity, data center operators typically work with telecommunications carriers through established arrangements such as Indefeasible Rights of Use, capacity leases, and master services agreements that provide long-term access to fiber infrastructure.

As data center development accelerates, industry stakeholders will need to continue to navigate rising investment demands, evolving power strategies, and growing connectivity needs. Those able to balance these pressures with thoughtful planning and diversified infrastructure approaches will be best positioned to meet the next wave of global demand.

We invite you to subscribe to receive updates on our Data Center Bytes and join us during our Data Center Bytes webinar series.

Bruker (BRKR) just reported a classic mixed earnings result, topping quarterly profit estimates while simultaneously trimming its full year revenue and earnings outlook. This combination has clearly cooled investor enthusiasm.

See our latest analysis for Bruker.

The guidance cut comes after a sharp rebound in sentiment, with Bruker’s 30 day share price return of 19.8 percent and 90 day gain of 56.9 percent, in contrast with a weaker 1 year total shareholder return of negative 17.0 percent. This suggests that near term momentum is improving even as the longer term record remains underwhelming.

If this kind of volatility has you thinking about diversification, it may be worth exploring healthcare stocks as a way to uncover other healthcare names with different growth and risk profiles.

With earnings beating expectations but guidance moving lower, and the share price now hovering just below analyst targets after a strong rebound, is Bruker an underappreciated turnaround candidate, or is the market already pricing in any future recovery?

Compared with Bruker’s last close near 48 dollars, the most popular narrative pins fair value slightly higher, implying only marginal upside from here.

The analysts have a consensus price target of $46.727 for Bruker based on their expectations of its future earnings growth, profit margins and other risk factors. However, there is a degree of disagreement amongst analysts, with the most bullish reporting a price target of $65.0, and the most bearish reporting a price target of just $38.0.

Read the complete narrative.

Want to see what kind of revenue path, margin rebuild, and future earnings multiple are stitched together to justify this tight valuation gap? The answer might surprise you.

Result: Fair Value of $48.83 (ABOUT RIGHT)

Have a read of the narrative in full and understand what’s behind the forecasts.

However, persistent funding headwinds and execution risk around margin improvement could easily derail the constructive bookings story and reset expectations again.

Find out about the key risks to this Bruker narrative.

While the narrative fair value sits close to the market price, our DCF model paints a cooler picture, putting Bruker’s value nearer 36.72 dollars. This suggests the shares look overvalued at current levels. Is the market now leaning too heavily on the recovery story?

Look into how the SWS DCF model arrives at its fair value.

Simply Wall St performs a discounted cash flow (DCF) on every stock in the world every day (check out Bruker for example). We show the entire calculation in full. You can track the result in your watchlist or portfolio and be alerted when this changes, or use our stock screener to discover 910 undervalued stocks based on their cash flows. If you save a screener we even alert you when new companies match – so you never miss a potential opportunity.

If you see the story differently or want to dig into the numbers yourself, you can build a custom view in just minutes using Do it your way.

A great starting point for your Bruker research is our analysis highlighting 2 key rewards and 1 important warning sign that could impact your investment decision.

Before you move on, lock in a few fresh opportunities that match your style so you are not relying on just one turnaround story.

This article by Simply Wall St is general in nature. We provide commentary based on historical data and analyst forecasts only using an unbiased methodology and our articles are not intended to be financial advice. It does not constitute a recommendation to buy or sell any stock, and does not take account of your objectives, or your financial situation. We aim to bring you long-term focused analysis driven by fundamental data. Note that our analysis may not factor in the latest price-sensitive company announcements or qualitative material. Simply Wall St has no position in any stocks mentioned.

Companies discussed in this article include BRKR.

Have feedback on this article? Concerned about the content? Get in touch with us directly. Alternatively, email editorial-team@simplywallst.com

• Stablecoin could be used to collateralise government debt: Bilal.
• Pakistan seeks to be at forefront of digital innovation: crypto czar.
• “Why should we be at tail end when we have muscle and adoption?”

---

Pakistan is set to launch its own stablecoin as part of efforts to embrace digital financial innovation, Bilal Bin Saqib, Chairman of the Pakistan Virtual Assets Regulatory Authority (PVARA), announced on Thursday.

According to CoinDesk, a digital media outlet focusing on cryptocurrency, blockchain, and the digital asset economy, a stablecoin is a type of cryptocurrency whose value is pegged to another asset class, such as a fiat currency or gold, to stabilise its price.

Speaking at Binance Blockchain Week in Dubai, he said that the country is experimenting with both a stablecoin and a Central Bank Digital Currency (CBDC), but stressed that the stablecoin initiative will definitely move forward.

The crypto czar highlighted that the stablecoin could serve as a tool to collateralise government debt, adding that Pakistan aims to be at the forefront of global digital financial developments.

“Why should we be at the tail end of it when we have the muscle and the adoption?” he remarked, underlining the country’s ambition to lead rather than follow in this emerging financial sector.

Separately, the Pakistan Crypto Council said that Saqib spoke on a high-level panel discussing the future of virtual assets and emerging-market regulation.

“He emphasised that for countries like Pakistan, clear and innovation-friendly crypto regulation is a key driver of economic growth. Pakistan’s work on stablecoins, data frameworks, and banking the unbanked can become valuable case studies for the world,” the council wrote on X.

Pakistan’s crypto market is estimated to have more than 40 million users with an annual trading volume exceeding $300 billion, making it among the most active frontier markets for digital assets.

Earlier in May, Bilal unveiled the country’s first government-led Strategic Bitcoin Reserve, marking a historic pivot in the nation’s digital and financial outlook.

Addressing a global audience that included US Vice President JD Vance, Eric Trump, and Donald Trump Jr in Las Vegas, he positioned Pakistan as a forward-looking digital hub, powered by its tech-savvy youth and strengthened by a shift toward decentralised finance.

Bilal revealed that the reserve would not be used for speculation or trading but would serve as a sovereign holding — signalling long-term commitment to blockchain-based finance. The national Bitcoin wallet already holds assets under state custody.