7.2.1 Data Generating Infrastructures: Floats and Sensors

Data does not exist in the state of nature – in contrast to natural resources to which data is sometimes analogized.Footnote ²² It is generated – in this case, by Argo sensors and the floats they ride on. Zooming in on how and what data is produced by Argo floats reveals the immense importance of materialities of data generation. Argo’s network of autonomous floats drift with the ocean currents and cycle between the surface and depths of 2,000 meters (“Core Argo,” the continuation of the original version initiated in the 1990s) or up to 6,000 meters (“Deep Argo”).Footnote ²³ The technology behind the floats was inherited by Argo from the World Ocean Circulation Experiment (WOCE) – the first global hydrographic survey that deployed autonomous floats. Only a small group of manufacturers has the technological sophistication to cater for the relatively low demand for floats – some of them design and equip the floats in dialogue with scientific groups who plan to operate them.Footnote ²⁴ Performance assessments, design ideas, and new needs or requirements circulate through industry channels and through the Argo governance structures. Argo’s data generating infrastructure is dependent on float manufacturers and their supply chain infrastructures, and also on laboratories for development and testing of floats and instruments, shipping and logistics infrastructures for transporting floats to vessels of launch, and the launching vessels themselves.

Access to vessels that travel at the requisite time needs to match the availability of funding, floats, and the scientific team to specific geographic areas. Where several floats are launched at intervals, the ship needs to be at sea for a sufficient period of time to enable spatial and temporal distribution of the floats.Footnote ²⁵ Institutions that have access to funds, or that can partner with other research programs, can commission research vessels.Footnote ²⁶ In other instances, commercial and tourist cruises,Footnote ²⁷ and even military vessels are used for deployment. Lack of reliable launch vessels creates not only geographic but also geopolitical dependencies. Russia’s invasion of Ukraine unexpectedly complicated deployment efforts when militaries around the world were for a time put on heightened alert, with some navies no longer willing to have their ships used for Argo float deployment to avoid inadvertent disclosure of their fleets’ locations by backtracking public Argo data. Earlier, the COVID-19 pandemic had temporarily set back deployment efforts with fewer ships going out to sea and quarantine rules preventing researchers from boarding vessels.

With lifespans ranging from five to seven years (and even less for Deep Argo floats) and no opportunity for post-deployment service, every component of the float, its sensors, and communication capabilities through the satellite system needs to be tested pre-launch, and the release of the float needs to be monitored as well. This requires trained human labor and time.Footnote ²⁸ The ultimate pre-deployment assessment can take between one and one-and-a-half hours, the last direct human-machine interaction in the series of numerous tests, calibrations, and validations that floats and sensors have undergone at the facilities of their respective manufacturers as well as at the research centers responsible for their deployment.Footnote ²⁹

Once released, the float descends to drifting depth (usually 1,000 meters), where it drifts for roughly ten days before descending to profiling depth of 2,000 meters (or up to 6,000 meters for Deep Argo) before resurfacing. During their ascent, the floats take a series of measurements via sensors, and, once on the surface, transmit the data via satellites, while receiving instructions for their next mission. For the majority of the Argo fleet, this surface interval is between fifteen minutes and one hour, after which the float sinks to a drift depth for about nine days. It then descends anew, repeating the ten-day cycle. See Figure 7.1.

Figure 7.1

Argo float deployment, data generation, and transmission. (C) Argo Program. Reprinted with permission from the Argo Program.

Figure 7.1 Long description

The illustration shows a boat and floating sensors on the ocean’s surface. Sensors are shown under the ocean up to a depth of 6000 metres, measuring salinity and temperature and drifting for a 10-day period. The illustration also shows a satellite above the ocean’s surface transmitting the data collected from the sensors. Three fish and a whale are also included in the depiction of the ocean.

The floats – made by different manufacturers and carrying different sensors – produce what is sometimes termed “raw data” about certain physical characteristics of the ocean, such as temperature and pressure.Footnote ³⁰ Given the relatively circumscribed market for floats and sensors, researchers and manufacturers often collaborate in the development, testing, calibration, and customization of relevant instruments. Frictions can arise when instruments fail or produce flawed data, especially about whether the cash-strapped lab or the manufacturer will carry the loss or cost of replacement. A sensor malfunction on an already deployed float can be very costly, as floats are usually not recoverable for repairs or upgrades.Footnote ³¹

Although all floats are equipped with conductivity-temperature-depth (CTD) sensors, only some floats generate profiles of biogeochemical properties (BGC-Argo). In addition to CTD sensors, BGC-Argo floats can integrate sensors to measure chlorophyll fluorescence, particle backscatter, dissolved oxygen, nitrate, pH, and irradiance. The variability is partly a function of different scientific interests and needs and partly a function of launching parties’ capacity and availability of resources. Historically, except for oxygen, a single sensor was used for the other BGC variables. However, increasingly, for most of these variables there are different types of sensors available, some differing on measurement principles. Introduction of new sensors requires ensuring interoperability with float models and cross-sensor calibration to ensure standardized data output, which depends on the willingness of sensor manufacturers to share the necessary metadata among themselves as well as with float manufacturers and Argo scientists. Evaluation of new sensors may also require review of baseline standards: should the first integrated sensor be used as a reference point against which any new entrant is evaluated? Introducing new sensors also requires additional testing, but trials of sensors on floats is a slow and expensive process. Any minor malfunction (e.g., a failure in a trivial connector cable) can lead to loss of data, and since floats are difficult to recover, on-float testing is not an efficient proposition. In-lab testing requires specialized lab settings that enable sensor testing under controlled circumstances, but these are not very widely available.

7.2.2 Data Transmission, Location, and Control Infrastructures: Satellite Systems

To transmit and receive data, Argo relies on satellite infrastructures. The less time a float spends on the sea surface for transmission, the less vulnerable it is to external elements. The Iridium satellite system, which consists of sixty-six satellites at a height of approximately 781 km, can transmit more data within a shorter period of time than other satellite systems dating from the period prior to Starlink-like mega-constellations, because Iridium allows “data calls” to be relayed from one satellite to another until they reach the satellite that registers the float. Iridium also allows for two-way communication, making it possible to send instructions to the float for troubleshooting or for changing the float’s mission (e.g., a float can be redirected to observe a developing hurricane). In this way, Iridium is both a data transmission and control infrastructure.Footnote ³²

Iridium, however, lacks the ability to geo-locate the floats. Hence, each float with an Iridium modem also includes a Global Positioning System (GPS) receiver to record the location of the float. A controller board on the float interfaces with the GPS receiver, the Iridium modem, the CTD sensors, and any other additionally installed sensors. The GPS receiver and the Iridium modem share an antenna that is connected to the radiofrequency circuit board, which serves as a platform for the GPS receiver, the Iridium modem, and any other communications devices.Footnote ³³ Every additional instrument mounted on the float comes with a cost that is not purely financial: increased weight of the float, higher energy consumption, and shorter battery lifespan are tradeoffs and compromises that are often made through discussion among members of Argo. These shape not only what data is being generated but also who can generate it given capacity constraints.Footnote ³⁴

Once on the surface, floats using the Iridium network perform a series of tasks in sequential order, with the floats creating their own temporality.Footnote ³⁵ First, if a float is equipped with an oxygen sensor, it will collect in-air samples of dissolved oxygen. Second, it will acquire a position fix using GPS. This typically takes several minutes as the float establishes the connection with the GPS satellites and subsequently determines the position. Third, the float connects with an Iridium satellite and the Iridium gateway, logs on to the launching group’s data server, and queries the server for any changes that might be necessary in the mission configuration for the next profile. Lastly, the float transmits the profile (i.e., data collected from the water column during the float cycle), surface O₂ samples, GPS, and engineering data (i.e., float diagnostic data), packaged into a series of highly compressed packets which are formatted according to Argo Data Management Team standards, via Iridium (for ways in which standards impact data production, see discussion infra, Section 7.3.1).Footnote ³⁶ Due to interruptions in the transmission process, it is often necessary for the data to be transmitted several times before all data have been successfully uploaded.Footnote ³⁷ On average, floats using Iridium spend twenty to thirty minutes on the sea surface for each cycle.Footnote ³⁸

In the early days of Argo, almost all floats transmitted their data via the Système Argos – a location and data transmission system operated by Collecte Localisation Satellites based in Toulouse, France and Maryland, United States. Despite the similar name – both inspired by Greek mythology – Argo and Argos developed separately and at different times, although both were motivated by the desire to enhance opportunities for ocean observing. Argos satellites were initially developed as one-way, low-bandwidth satellites, with an effective data throughput of no more than 1 bit per second. In contrast to Iridium, Argos satellites have both data transmission and location data capabilities. However, Argos’ low data transmission rates required floats to spend between six and eighteen hours at the sea surface to ensure error-free data reception in all weather conditions. A later model (Argos-3) offers bidirectional transmission and a higher throughput, but the high-data-rate mode suffered from electromagnetic noise around Europe.Footnote ³⁹ Some of the floats continue to use Argos satellites, and other satellite transmission systems have also been used on profiling floats, including transmission systems such as BeiDou in AsiaFootnote ⁴⁰ and ORBCOMM in North America.Footnote ⁴¹

7.2.3 Data Processing, Storage, and Dissemination Infrastructures: Data Assembly Centers

Data is always stored somewhere (though not necessarily in one place) and needs to be processed and disseminated to render it accessible and useful for scientific knowledge production. Argo float data is transmitted via satellites to eleven national Data Assembly Centers (DACs) for processing, with each float being allocated to a specific DAC. Processed data from the national DACs is subsequently transmitted to two Argo Global Data Assembly Centers (GDACs) – one located at the Coriolis Data Center in France and the other at the US Navy’s Fleet Numerical Meteorology and Oceanography Center, where data is checked for format and content consistency.Footnote ⁴² Once per day, data holdings at the two Argo GDACs are being synchronized. The US National Centers for Environmental Information (NCEI) operates and manages the Global Argo Data Repository (GADR), which provides long-term archive services to store and preserve data and implements reanalysis updates and corrections provided by the GDACs.Footnote ⁴³

The Argo Data Management Team (DMT) is responsible for coordinating data management and ensuring access to Argo data. The Argo DMT is composed of at least one representative of each institute involved in Argo data management activities.Footnote ⁴⁴ DMT also defines Argo’s data management infrastructure, data formats, and quality control procedures.Footnote ⁴⁵ There are numerous publicly available and regularly updated manuals, handbooks, and “cookbooks”Footnote ⁴⁶ addressing all aspects of data processing for each parameter. The Argo data management teams aim to ensure not only that its data is openly available but that its potential users understand what data they receive. Argo data is made available from GDACs in the Common Data Format, a file format for storing multidimensional scientific data (variables such as temperature, humidity, pressure, wind speed, and direction).Footnote ⁴⁷ All Argo data is publicly available without charge. Anyone can access Argo data via direct file transfer or via a World Wide Web interface. The digital data in this case rapidly becomes part of the public domain, not an accumulating private resource from which other well-funded scientists are excluded. It is not a form of club good in which non-contributors cannot get access.

7.3.1 Infrastructural Data Inequality: Who Decides What Data Is (Not) Being Produced?

Who decides what ocean data is (or is not) being produced? Although the Argo array currently reflects contributions from thirty countries, significant forms of “data inequality” persist.Footnote ⁵⁴ These inequalities reflect, at least in part, colonial legacies that led to the uneven distribution of oceanographic knowledge, as former colonial powers and dominant trading nations made significant historical investment in oceanography and associated technologies of ocean data generation, recording, and mapping.Footnote ⁵⁵ In 2022, 80 percent of BGC-Argo float deployments were done by research labs from only three countries (US, France, and Canada), with more than 50 percent of the deployments by the United States.Footnote ⁵⁶ Coverage for different parameters is uneven, with only a limited number of BGC-Argo floats carrying the complete suite of sensors (with 40 percent carrying only oxygen as the only extra sensor in addition to CTD). Overall sampling of the oceans is also unevenly distributed in Argo, with the Atlantic Ocean oversampled while the Indian, Pacific, and Southern Oceans remain undersampled.Footnote ⁵⁷ Only nine countries participate in the Deep Argo array, with the United States, France, and Japan being leading national contributors. Some of the undersampled regions – such as the Southern Ocean – for many years were not the priority for main float contributors. As Dean Roemmich – one of Argo founders – noted in 2001, “[t[hat is one of the hardest implementation items for the Argo science team. In order to achieve a global array we will need to set some fraction of floats aside for a region that is not the top priority of any single government.”Footnote ⁵⁸

Any decision on which data to produce must take into account technological and funding constraints and the interests of different communities. Expanding data production must be sensitive to the demands imposed on data managers, because additional sensors increase data complexity, which can require new standards, guidelines, and manuals.Footnote ⁵⁹ Originally, Argo floats carried only CTD sensors. However, understanding the role of oceans in regulating climate change and, conversely, the impact of climate change on ocean ecosystems, requires observations of the ocean biogeochemistry. The Intergovernmental Panel on Climate Change’s Fifth Assessment Report listed temperature, salinity, acidification (pH), oxygen, nutrients (nitrate), and light as major drivers of ocean health. BGC-Argo introduced six new variables (oxygen, nitrate, pH, chlorophyll fluorescence, optical backscatter, and solar irradiance). The choice of these variables (and not others) was based not only on scientific research needs but also on availability of operational sensors.Footnote ⁶⁰ Proponents of BGC-Argo acknowledged that specific scientific goals, regional issues, funding sources, and even sensor configurations may motivate different missions, requiring tradeoffs and compromises. For example, current radiometric sensors on BGC-Argo have four color bands available, which is not sufficient to characterize the spectral variability of the underwater light field and thus cannot capture the diversity of phytoplankton. Hyperspectral radiometry, which captures light over many wavelengths and enables recognition of tens or hundreds of colors, can enable better discrimination among different phytoplankton types, potentially improve knowledge of oceanic carbon stocks, pathways, and fluxes.Footnote ⁶¹ Enabling BGC-Argo with hyperspectral capability could also produce knowledge useful for the management of living resources and ecosystem services (e.g., habitat suitability, fish stocks, and recruitment) and for biohazard surveillance (i.e., harmful algal blooms).Footnote ⁶² Hyperspectral radiometric sensors, however, are more expensive, require more energy consumption, and impose additional labor costs for the Argo DMT. The working group on radiometry, created within the AST, recommended that those programs that have capacity move to hyperspectral radiometry but it could not mandate a wider adoption of the sensors.

Argo floats generate only some “ocean data” directly from the sensors (e.g., measuring temperature and electrical conductivity of the seawater sample); other data is derived from computations. For example, all Argo floats are equipped with CTD sensors. The “C” in CTD is conductivity, which measures how well a solution conducts electricity. Since dissolved salts and other inorganic chemicals conduct electrical current, conductivity increases as salinity increases. The measure of conductivity thus relates to the measures of salinity. However, salinity is also dependent on pressure and temperature. Whereas “T” measures temperature directly, “D” (depth) actually measures pressure. Using temperature and pressure, salinity can be derived via mathematical equations.Footnote ⁶³ The density of water is calculated from temperature and salinity, and because pressure increases with increasing depth, depth can be derived in turn. Even with direct measurements, there are accuracy errors.Footnote ⁶⁴ Sensor drift can introduce larger errors, requiring corrections by “experts,” which may entail comparing older floats with newly deployed instruments and with ship-based data.

Decisions of how to allocate limited resources – whether to invest in increase of spatial coverage, maintenance of Core Argo, enhanced longevity of floats, enabling more (up to six) BGC sensors on all BGC floats, expanding the reach and capabilities of Deep Argo, improving sensors to enable better data on specific parameters, or adding resources to the DMT – are discussed at annual meetings of the AST as well as at periodic meetings of various working groups formed to focus on specific issues or topics. Limitations and inequalities of funding are a recurring theme in these meetings. The phenomenon of underfunded infrastructures corresponds with broader insights from infrastructure studies that observe and sometimes lament the invisibility of infrastructures that are being “taken for granted” until their break down or fail.Footnote ⁶⁵ Data infrastructures may be particularly susceptible to invisibility (as compared e.g., with physical infrastructures) because their failures and malfunctions are not immediately evident, even to those working directly with data processing and quality control. Errors may be revealed only when data is being validated by comparing it to other data sources (e.g., Argo data is often validated by data from the Global Ocean Hydrographic Investigations Program (GO-SHIP)) or when data from numerous floats is processed in the aggregate (e.g., to reveal a sensor drift). The value-added of the data infrastructure is also not immediately visible. Argo’s importance lies in having a reliable time-series data over substantial period of time, which enables comparative, historical, and predictive analysis of oceanic temperatures, for example. Yet, on a daily basis, Argo participants have to educate audiences, institutions, policymakers, and funders about the program’s importance and highlight new, potential, or previously unforeseen uses of Argo data in continuous efforts to sustain the program. Without sustained efforts at making Argo data infrastructure visible, Argo risks receding not only from public imagination but also from the gaze of those who supply information and knowledge that directly shapes earth systems governance. As a transnational infrastructure, Argo faces additional challenges that not only require international coordination but would also benefit from transnational cooperation in form of technology-transfers and capacity-building. Governmentally funded research, however, often remains stubbornly territorialized and hence misaligned with the transnational infrastructures necessary for planetary-scale science.Footnote ⁶⁶

7.3.2 Infrastructural Space: Where Is Data (Not) Being Generated?

Argo seeks to generate ocean data on a planetary scale. Naturally, Argo’s floats cannot generate data everywhere as data generation depends on float location. In this way, data generation via floats differs from data generation via satellites. While the latter achieves planetary scale more easily by orbiting around the planet, the former manages to reach deeper into the ocean. In its originally proposed design, Argo was to comprise around 3,300 floats, each profiling the ocean between the surface and a depth of 2,000 meter around 25 times per year over an estimated lifetime of 3–4 years.Footnote ⁶⁷ The OneArgo program aims to expand the spatial coverage of Core Argo while also increasing the number of BGC and Deep Argo floats. Given the significant financial and human resources involved in procuring, testing, and deploying the floats and processing subsequent data, tradeoffs and tensions arise along the way, shaping both the type of data that is produced and the choices about data’s target users and uses. By one estimate, given current life expectancies of floats, more than 800 floats per year would need to be deployed to develop OneArgo while maintaining the Core Argo array.Footnote ⁶⁸ Funding remains a core challenge for maintaining and expanding the Argo program’s scale. However, spatial distribution is also impacted by natural phenomena, technical limitations, business models, and legal constraints.

Some areas of the ocean require more intense concentration of floats due to natural phenomena. For example, mesoscale variability in the Western Boundary Current regions require enhanced sampling to reduce noise in tracking the large-scale temperature and salinity fields.Footnote ⁶⁹ One alternative to increasing float density is to increase the frequency of sampling. Another option is to rely on data provided by other instruments, such as glidersFootnote ⁷⁰ or nonhuman species equipped with sensors,Footnote ⁷¹ to supplement Argo data. Both choices, however, impose costs associated with data generation, processing, and dissemination (e.g., additional processing to ensure standardization and interoperability of measurements).

Longevity of floats also impacts spatial coverage. Many floats are not recoverable (they can sink, get lost, be collected upon washing on shore and not returned to the owner, etc.) and thus cannot be repaired and redeployed. Areas with lower launch frequencies typically have older floats. Sufficient sampling would require increased float density to replace aging floats as well as to compensate for the original under-sampling. Increasing longevity of floats (and of their components), however, can be difficult both technologically and due to misalignment of incentives. Given the limited demand and supply of floats and sensors, manufacturers do not necessarily have incentives to prolong the life of the float and may need to balance demands for float longevity with demands from wealthier purchasers to enhance float capabilities or to customize floats.

Dependencies on a single producer or a limited group of producers aid in consistency of data and in implementation of uniform procedures for data processing across programs. However, such dependencies also create single points of failure: past incidents of faulty pressure and pH negatively impacted Argo coverage.Footnote ⁷² Greater diversification of sensor manufacturers would alleviate this problem but would also cause challenges relating to interoperability and data standardization.

Lastly, coastal areas remain undersampled due to uncertainties about whether Argo should be interpreted as “marine scientific research” under the UNCLOS.Footnote ⁷³ Under Article 248 of UNCLOS, states or international organizations that intend to undertake marine scientific research within the exclusive economic zone (EEZ) or on the continental shelf of a coastal state are required to obtain that state’s consent, with typically a requirement of at least six-month advance notice describing the research project. Some states have agreed to deployment of Argo floats in their EEZs under the conditions of free and unrestricted data exchange and the transparent implementation through OceanOPS monitoring (a joint World Meteorological Organization (WMO)–Intergovernmental Oceanic Commission (IOC)/UNESCO Support Center for oceanography and marine meteorology). Other coastal states reportedly do not consider Argo to be “marine scientific research” and hence allow float deployment without consent.Footnote ⁷⁴ Yet other states maintain the position that every float deployment needs to comply with Article 248. Thus, as a practical matter, deployment of Argo floats within EEZ is arranged on a bilateral basis. However, because there is no articulated procedure for “clearing” marine scientific research projects, compliance with Article 248 is cumbersome and deployment of floats in EEZs is often avoided.Footnote ⁷⁵ This leads to significant under-sampling, given that EEZs represent nearly 30 percent of the world’s ocean surface.

7.3.3 Infrastructural Time: When Is Data (Not) Being Generated and Transmitted?

Argo generates a set of synoptically sampled time-series data that enables ongoing analysis of ocean parameters, over-time comparisons not only over the course of Argo tenure but also retroactively,Footnote ⁷⁶ and forward-looking forecasts, models, and predictions. The timeliness of data generation and transmission are important infrastructural decisions that affect different scientific communities in different ways.

Fluorescence observations obtained as a measure of phytoplankton mass, for example, are best done at night. In contrast, assessment of biomassFootnote ⁷⁷ using radiometry and comparison of BGC-Argo sensor results to satellite remote sensing motivates measurements around noon.Footnote ⁷⁸ When manufacturers set the default for the floats to come up at noon, they put an onus on operators to alter the setting as needed. Lack of standardized guidance on the timing of sampling can create challenges for data comparability. The Argo Steering Committee and the Argo DMT deliberated on whether to adopt a sampling frequency of 10.2 days (245 hours) to ensure profiles are not taken at the same time each day. This compromise ensures temporal compatibility of measurements across the array but potentially comes at the cost of not optimizing data of particular interest for specific research communities (e.g., those interested in phytoplankton might want all data to be sampled at night).

Measurement of oxygen also poses challenges on temporal scales. Oxygen sensors can be used to measure both changes in global ocean oxygen inventories and local biological activity. Changes in the whole-ocean oxygen inventories, however, are expected on decadal timescales, necessitating sustained array observations and significant commitment to sensor stability assessments. In contrast, net community production measurements (i.e., gross oxygen production minus oxygen consumption by all organisms), especially if they are to be scaled to satellite remote sensing of biomass, require high temporal and spatial frequency observations. This is because oxygen levels shift throughout the day due to changes in photosynthesis, which rises through the day and falls to zero at night.Footnote ⁷⁹ Frequent sampling increases energy consumption, which may decrease the longevity of the float.

At the time of its founding, the Argo program aimed to meet the requirements of operational oceanographers (i.e., weather monitoring and forecasting) on the one hand and the scientific community engaged in climate monitoring and modeling on the other hand. These two communities and research agendas demand different kinds of data on different time scales. Operational oceanographers need access to Argo data as soon as possible, even if the data is not of the highest quality. In contrast, for climate monitoring and modeling communities, the quality of data is of paramount importance. To meet the objectives of the two communities, Argo’s designers decided to make vertical profile and trajectory data available in “real-time” and in “delayed-mode.”Footnote ⁸⁰ “Real-time” data is not immediate or continuous but is provided within twenty-four hours of satellite transmission and is subjected to a set of automatic quality tests, which assign “quality flags” to grossly “bad” data. Flagged data is preliminarily adjusted and both the original and the adjusted data are subsequently inserted into the Global Telecommunications System (GTS) used by operational meteorological agencies. In contrast, delayed-mode data is typically made available within a year (or more) after transmission. The quality control for delayed-mode data involves a combination of human expert review (e.g., data is visually examined by oceanographic experts for additional (re-)flagging) and complex statistical processes.Footnote ⁸¹ In these ways, Argo creates different kinds of data depending on when data is being generated, when data is being processed, and when it is being made available. These different kinds of data correspond to the needs of diverse scientific communities, or: Argo’s infrastructural publics.

7.3.4 Infrastructural Publics: Scientific Communities and Beyond

Infrastructures generate publics.Footnote ⁸² Argo’s foundation and trajectory as a data infrastructure that is mostly, but not exclusively, used by various scientific communities informs many of the contests about what data is being produced, when, where, how, and for whom.

Argo operates as a self-defined community that formed around the creation and operation of the Argo infrastructure. Argo’s trajectory has been shaped by dedicated scientists, many of whom have been deeply involved since the program’s inception. Their priorities are animated by the sensibilities of scientific practices and the desire to ensure that users of the data understand what the data represents and what its limitations are, with outermost attention placed on the quality and consistency of data.Footnote ⁸³ The proposal for the array, put forth in 1998 by a US oceanographer from the Scripps Institution of Oceanography, Dean Roemmich, found resonance with the climate research interests of the Climate Variability and Predictability (CLIVAR) ProgramFootnote ⁸⁴ and the operational estimation objectives of the Global Ocean Data Assimilation Experiment (GODAE).Footnote ⁸⁵ Inheriting the legacies and technologies of prior ocean observation infrastructures, Argo founders aimed to develop a program that would respond to the needs of meteorologists and climate modeling communities. The incorporation of Argo into GOOS further entrenched these twin goals,Footnote ⁸⁶ and the incorporation of Argo into the priorities of the Ocean Decade envisioned further uses and users of Argo data, including via aggregation with data from other sources and infrastructures.

Fulfilling the needs of different communities, however, presents challenges. Scientific communities deploy different methodologies,Footnote ⁸⁷ pursue divergent political priorities,Footnote ⁸⁸ and encounter variegated funding and organizational problems.Footnote ⁸⁹ To date, neither GOOS nor the IOC provide funding for Argo (or any other ocean observation network), but instead aim to standardize and systematize what knowledge about the ocean is being produced transnationally and for what purpose.Footnote ⁹⁰ Securing funding for Argo hence remains a recurring challenge for the Argo community.

In a quest to ensure its sustainability and in response to growing demand for Argo data from diverse scientific communities, Argo’s data infrastructure is constitutive of and responsive to diverse infrastructural publics.Footnote ⁹¹ Scientific communities interested in operational ocean science and related modeling, for example, are affected when Argo’s data is being updated, including retroactively. To better connect with these operational and modeling communities through regular information exchange and communication, the AST created a regular virtual forum for Argo and ForeSea/OceanPredict.Footnote ⁹²

The rarity of having consistent time-series of ocean observations has made Argo infrastructural not only to its intended users (i.e., operational oceanographers, climate modelers, etc.) but also to those who see Argo as a foundation upon which additional types of data can be generated. In this way, Argo’s spatial and temporal data generation can generate new publics, or re-make existing publics. For example, Argo’s success in creating observations of ocean’s interior at “adequate temporal and spatial scales” prompted scientists to note that there were no similar advances in observations of large-scale biogeochemical and biological states of the ocean.Footnote ⁹³ The only program mapping out the distribution of biogeochemical tracers in the ocean had been the global CO₂ survey undertaken jointly by WOCE, the Joint Global Ocean Flux Study (JGOFS), and a few other national groupings – that program took a decade to complete, and was largely a one-time snapshot, providing limited information about evolution over time. The Repeat Hydrography program suffered from similar temporal challenges, with about a decade between repeats of hydrographic sections.Footnote ⁹⁴ Although such temporal sampling was adequate for determining the long-term increase of the oceanic carbon content in response to the increase in atmospheric CO₂ levels, it was incompatible with the seasonal to sub-decadal timescale of variability in many biogeochemical parameters within the transition layer between warmer mixed water at the surface and cooler deep water.Footnote ⁹⁵ In 2007 it was proposed that oxygen sensors be added onto Argo floats. While this proposal was being considered, contemporaneous developments in sensor technologies for oxygen, chlorophyll, particles, and nitrate had enabled the deployment of sensors on long-endurance missions on autonomous platforms.Footnote ⁹⁶ This, in turn spurred the creation of BGC-Argo – a program that would see floats carry not only oxygen sensors but other biogeochemical sensors as well, bringing into the collaborative fold of this Argo program various BGC modeling communities, along with the physical ocean data assimilation community.Footnote ⁹⁷

The introduction of BGC sensors and the expansion of the originally intended Argo users, however, have introduced new challenges in determining what constitutes “bad” data and based on which criteria “flagging” of such data should occur. For example, bio-optical sensors provide proxy information on the size structure of the phytoplankton assemblage, which are responsible for about half of the biological uptake of CO₂ on Earth through photosynthesis.Footnote ⁹⁸ The light from bio-optical sensors, however, often attracts larger mid-water organisms, creating “spikes” in data that would be typically flagged as “bad” or “probably bad” data since it is not the type of data expectedFootnote ⁹⁹ from the sensors. However, this data could be useful to scientists interested in tracking migrating organisms that live in the mesopelagic zone of the ocean (e.g., zooplankton, fish, squids, and jellyfish).Footnote ¹⁰⁰ The eventual compromise arrived at during the 2023 meeting of the AST was to retain the automated flagging of “bad” or “probably bad” but provide an index file of profiles or data segments that are believed to contain signals of mesopelagic organisms.

Building synergies with other “communities” continues to be a strategy for the Argo program as it struggles to secure consistent and reliable funding. At the AST meeting in March 2023, discussions focused on potential collaboration with other “communities” that could benefit from Argo data and that have access to funding,Footnote ¹⁰¹ such as “ocean color” communities that use satellite-based remote-sensing infrastructures to monitor health of the oceans and entities and collectives focused on marine carbon dioxide removal (CDR). Diversified uses of Argo data can enrich outcomes of other knowledge infrastructures (e.g. providing the “vertical dimension” for validating satellite data)Footnote ¹⁰² while also making Argo more relevant and sustainable by creating dependencies with a broader array of users (e.g., Argo’s oxygen monitors may become indispensable to CDR projects).

At the same time, the expanded user base can lead to changes in the type of instrumentation and data, creating interoperability issues and requiring additional resources for the DMT, thereby posing governance challenges. Additionally, as data infrastructures expand or are being repurposed, contests over their normative orientation may emerge. Consider, for example, discussions at the AST meeting in March 2023 on potential collaboration with CDR projects. Should BGC-Argo act as a neutral “referee,” independently evaluating the effects of ocean manipulation associated with CDR on oxygen levels and ecosystems? Such a role would expand uses of Argo data but would not give Argo access to funding associated with CDR initiatives. Or should Argo collaborate with those involved in CDR to open up funding opportunities for Argo? Such an endeavor would implicate the Argo community in an activity that might not normatively align with its members’ priorities and expectations, and might lead to unforeseen externalities.Footnote ¹⁰³

Argo’s infrastructural publics also interact with legal publics (states) and their representatives (governments),Footnote ¹⁰⁴ as they fund, support, coordinate, prioritize, and otherwise contribute to the array. More attenuated but nonetheless affected publics are the individuals and communities that experience extreme weather conditions and the impacts of climate change and for whom weather and climate forecasts (or absence thereof) can have existential implications.

Given the “openness” (in the sense of public accessibility without charge) of Argo data not all of the data infrastructure’s publics may be immediately evident and some may be unintended. The tenor of Argo is one of scientific knowledge production – chief impacts touted by Argo members and documented by states in national reports are the number of scientific publications using Argo data. Yet data produced through BGC sensors is relevant for fisheries management,Footnote ¹⁰⁵ and variability of temperature and salinity with depth contributes to understanding the structure of the sea for the purpose of submarine warfare.Footnote ¹⁰⁶

Argo also engages nonhuman publics – the species and organisms of various sizes inhabiting different layers of the ocean and unwittingly “collaborating” (in a Latourian sense)Footnote ¹⁰⁷ with the floats and their instrumentation to generate data about themselves (e.g., as in when migrating organisms are attracted to the light sensors to “reveal” their presence to observing scientists), subverting Argo’s mission by “polluting” the data or damaging the floats, or complementing and validating Argo data (e.g., where sensor-carrying animals are used to verify or supplant Argo data). The interests of the nonhuman publics are not really recognized or represented in the Argo data infrastructure, however, except insofar as their datafied representation becomes relevant to the expanded base of users of the Argo data infrastructure (e.g., those concerned with biodiversity conservation or species preservation).

8.2.1 Application Functionality Through Data-Driven Supply Chains

With a typical supply chain model, a client application (often accessed through a lower-powered device) makes requests to a server through an Application Programming Interface (API). The server – usually in a data centre – processes the request and responds with the result, which the client displays, uses for functionality, or inputs to another service for further processing. One actor (a customer) can control the client, while another (a provider) performs the computation. Data flow links actors’ systems and enables the functionalities they produce. This allows expensive computing resources to be centralised, with costs spread across a provider’s customer base and customers paying based on usage or subscription (or some combination thereof). Some providers’ data centres work at hyperscale, where computing resources rapidly grow and shrink in response to demand.

Through this, hardware resources and software capabilities are modularised into plug-and-play services, with vendors coordinating their integration into a coherent application. This application may itself be a cloud service working as a component in other applications, with the vendor accordingly being a provider to their own customers. As supply chains comprise multiple applications and services with their own supply chains, coordinated by different actors, no single party has overall control. Moreover, each instance of a supply chain may differ in composition and structure: data’s processing determines the direction of its subsequent flow through the chain (going to different actors depending on the result) and therefore the actors involved.Footnote ¹⁹ However, companies and organisations typically have ongoing legal, technical and economic relationships across chains. As such, while chains may differ with each instance, they broadly follow patterns reflecting these more persistent relationships, taking a particular pattern – or combination of patterns – each time. Supply chains thus involve services and applications dynamically strung together and linked by data flows – each potentially with their own supply chain – which may support many types of service, application or functionality and serve multiple customers and purposes simultaneously.

8.2.2 Data-Driven Supply Chains and Data Production

Platformisation – as a process of platforms replacing markets – involves capture as described by Agre,Footnote ²⁰ whereby human activities are (re-)configured to make them legible and trackable by and with computers.Footnote ²¹ As Poell et al argue, platform relations are ‘organised through the systematic collection, algorithmic processing, monetisation, and circulation of data’.Footnote ²² This produces social and economic interactions represented, according to Cohen, by ‘structured flows of data suitable for continuous collection and analysis at the platform level’.Footnote ²³ Agre characterisesFootnote ²⁴ capture as a complex process by which activities are broken down into their component parts, reconceptualised to permit datafication and re-constructed accordingly to enable computerised control – allowing (as Cohen puts it) the platform to serve its own priorities.Footnote ²⁵

A key feature of informational capitalism is that data and other intangible informational resources are reconceptualised as property and commodified.Footnote ²⁶ Indeed, the production and processing of data generates significant revenue for platform companies, while datasets are extensively traded between a vast array of data brokers and widely leveraged for profit. The datafication found in capture therefore operates as the ‘principal logic of commodification’ in informational capitalism,Footnote ²⁷ whereby social and economic relations reconfigured around platforms are represented in data, commodified and put to work in platform business models. While, as Cohen argues, platforms represent points of ‘friction’ in networksFootnote ²⁸ – in that social and economic relations must pass through them – capture thus permits frictionless value extraction by platform providers.

This involves complex patterns of data production and flow – not just around ‘a platform’, but between a variety of platforms and services controlled by multiple actors who may together produce an application, its functionalities and its effects, and through which significant data may be extracted. As such, supply chains are important infrastructures for data production, both for AI production and developmentFootnote ²⁹ and for tracking, profiling and targeting. Data produced through supply chains feeds into social sorting systems for making decisions about people’s lives in the private sector; and in eligibility assessments for welfare, housing or other public services. This is all enabled by supply chains stretching across societies: producing, transporting and processing data; linking actors and their systems; and giving effect to decisions, outcomes and functionalities.

In digital advertising, supply chains are extensive.Footnote ³⁰ Trackers produce dataFootnote ³¹ for profiling, bidding and targeting systems,Footnote ³² which is widely processed and traded.Footnote ³³ Flows of data through these chains link audiences, advertisers, and many ad brokers, ad exchanges and data brokers. Advertising chains are often comingled with others – a mobile app might integrate trackers from multiple companies,Footnote ³⁴ or use component services which themselves integrate trackers, while displaying targeted advertising for revenue purposes. The technologies which underpin application functionality therefore often also produce data through tracking systems and spirit it away to targeting processes.

8.2.3 Data-Driven Supply Chains and Informational Capitalism

Supply chains thus have a dual role: working to produce application functionality, and to produce and extract data and enable targeting. Reorganising social and economic life around platforms and supply chains therefore means reorganising around infrastructures for data’s production, aggregation and use. This data – relating both to the markets reconstituted around supply chains and to the interactions of individuals and groups – is subject to ‘active appropriation strategies’Footnote ³⁵ (as Cohen says) by providers seeking competitive advantage. Data-driven supply chains are therefore how various platforms are constructed, how certain platforms are strung together to construct others, and how much of social and economic life has been capturedFootnote ³⁶ by the logics and imperatives of profit extraction under informational capitalism.Footnote ³⁷

As such, supply chains are key infrastructures in informational economies. With the reorganisation of markets and social relations around platforms, and of platforms around supply chains, data and its flow plays an important structural role in addition to its propertisation and commodification.Footnote ³⁸ Data’s flow links actors to produce applications’ functionalities and enables their platformised relations. Data’s processing determines which actors take part in each instance. Given data’s structural and economic importance, asymmetries in the ability to produce and productively use data affect the position and power of various actors. Those with greater capacity to construct platforms and provide services by doing so can position themselves advantageously, extract significant revenue and build significant power. As such, the relations constituted by and through supply chains are significant sites for governance in contemporary societies, as I now explore.

8.3.1 Operation and Control of Technology

Distribution of modularised computing resources through services and their integration into applications involves legal, technical and economic relationships through which control over various aspects of technologies is assigned between multiple actors. In integrating services into applications, developers delegate control over certain aspects of their application to providers. The corollary is that providers delegate limited and conditional control of certain aspects of their systems to customers or end-users,Footnote ⁴¹ while retaining control of other aspects of that technology (delegated to them from the customer’s perspective). This would be impossible without data’s flow linking geographically dislocated actors and enabling control to be split between and exercised separately by them.

A key feature of data-driven supply chains is therefore that different people simultaneously operate or control various aspects of the same systems, allowing them to provide some service, undertake some computation, supply some functionality or produce some application. This is enabled by modularisation of computing resources and underpins their distribution through APIs. ‘Operate’ here means running computing infrastructures an actor owns, under their retained control or another party’s delegated control. ‘Control’ means deciding questions around systems’ design and development, who they are used by, in what way, for what purposes and so on. Some such questions relate to a technology’s essential aspects, going to the system’s core function.Footnote ⁴² These are generally uniform across instances, as variation would affect the service’s core features, with providers retaining control for technical and commercial reasons. Other aspects are incidental,Footnote ⁴³ not relating to core functions and could vary between instances (particularly as supply chains unfold dynamically across various patterns). These may be delegated or retained.

There is therefore an exchange of control of certain aspects of technologies between actors and a potentially complex distribution of responsibilityFootnote ⁴⁴ for them (in that multiple actors are factually responsible in different ways for different practical aspects of their production, deployment and use). Application vendors may sub-delegate some of the control delegated to them by providers while retaining aspects of that control themselves (the degree of delegated control often diminishing with each step in a chain). For example, the food delivery app vendor may allow restaurants to personalise menus to end-users with a recommender service from the vendor’s cloud provider, thereby sub-delegating some control over that underlying technology to restaurants. Yet they may retain some (delegated) control over other aspects of the recommender service, allowing it to be integrated with other platform features (particularly if those aspects of the recommender service go to the platform’s core intermediary function and are therefore essential). Operators may have their own providers, contributing to development or operation of their technologies, with exchanges of control up and down the chain.

Exchanges of control necessarily vary: different things are essential – and control over them typically retained – to different services and applications. Providers may also retain control over incidental aspects relating to data production and policy enforcement (for example) to minimise their potential legal and commercial risks and maximise their benefit from the platform. Moreover, different kinds of service or application – with different core functions – necessitate different retention and delegation. In Software as a Service (for example), where customers access whole user-facing applications remotely, delegated control is more limited. In Infrastructure as a Service, however, where customers can install and run software they choose on lower-level computing resources, delegated control is necessarily greater and retained control correspondingly less. In AI services, providers retain significant control over producing and operating their system while delegating much of its deployment and use to customers.Footnote ⁴⁵

Digital technologies thus often involve complex patterns of operation and (sub-)delegation of control which produce supply chains. For example, the storage provider Dropbox used AWS’s infrastructural services to host customer data.Footnote ⁴⁶ Dropbox necessarily delegated control over some incidental aspects of their application to AWS, who operated the underlying technology. AWS delegated to Dropbox control of other aspects of that technology (incidental to an infrastructure service): what kind of information is stored, on whose behalf, and when and how it is retrieved. Dropbox, in turn, sub-delegated some control to their customers (such as the specific information stored). However, other aspects of the control AWS delegated to Dropbox form essential aspects of a storage service, going to its core function, and Dropbox therefore retained them (such as around which customer’s data is stored, with what access controls, and mechanisms for storage and retrieval). Through this, both AWS and Dropbox retained control over different aspects of the same technologies which were essential to their respective services, while delegating control over other aspects of those and other technologies to each other and to the end-users.

8.3.2 Exchange of Control and Interdependence

Interdependence is thus a key dynamic in data-driven supply chains. Through retention and (sub-)delegation of control, each actor does something which triggers, enables, or underpins others’ activities. As such, each actor depends on othersFootnote ⁴⁷ in ways which reflect and shape patterns of operation and control: giving inputs, undertaking computation, applying outputs, providing services which support and enable their activities, and much more.

With the hypothetical food delivery app, for example, each actor does things which others depend on – through the supply chain – enabled by data flows between them. Cloud providers operate technical infrastructure underpinning the platform, controlling retained aspects of their service but otherwise acting under control delegated to the vendor, who pays service fees. The vendor controls essential aspects of their platform to intermediate between restaurants, customers and riders (such as for assigning deliveries to riders), as well as over incidental aspects which shape their relations, while customers and restaurants pay fees. Riders control some aspects of the platform relating to them – accepting or declining jobs and so on – and deliver food from restaurants to customers, in return for fees from vendors. Other supporting providers contribute to various aspects of the platform (around payments, security, content delivery and others). This involves a constellation of actors, where data’s production in client devices, processing in servers and flow through supply chain components underpins exchanges of control which link them dynamically into a fluctuating web of interdependent relations and enables their various activities and interactions.

Yet providers’ role in supply chains – and therefore their business model and market position – depends on others as much as anyone else’s. Discussions around digital technologies sometimes overlook this fundamental interdependence, framing people, organisations or societies as dependent on providers. Yet without others doing things which use their services, providers would play no role in a supply chain; indeed, without others’ activities, there would be no supply chain to be part of at all. Moreover, supply chains’ dual purposes – also producing data for development processesFootnote ⁴⁸ or advertising-based revenue streamsFootnote ⁴⁹ – means production and ongoing development of the technologies through which providers partake in chains depends in part on others in them, as does their income. Choices which enable providers’ activities are of course conditioned both by their capture of markets and by dynamics which tend towards supply chain consolidation around certain players (and which providers proactively pursue). But the fact remains that providers’ position is constituted by and sustains only through mutually dependent relations.

8.3.3 Structuring Patterns of Interdependence

Supply chain interdependencies are, however, asymmetric in ways which reflect asymmetries in patterns of operation and control. Patterns of operation are often asymmetric both within and across chains, shaped by differences in technical and financial resourcing and by commercial, legal, and political economic influences. Services and applications rely on data production and processing, for example, yet operating computing infrastructures to produce and productively use data may require significant financial and technical resources – consequently, global markets for infrastructural computing services are particularly (and increasingly) concentrated around three companies.Footnote ⁵⁰ Law is important in constituting both capitalismFootnote ⁵¹ (generally) and informational capitalismFootnote ⁵² (specifically) – enabling platformisation,Footnote ⁵³ supporting surveillance business modelsFootnote ⁵⁴ and shaping technological development.Footnote ⁵⁵ Intellectual property laws protect closed-source, proprietary systems.Footnote ⁵⁶ Intermediary liability protections reduce legal risks of operating services at scaleFootnote ⁵⁷ (which might otherwise be insurmountable) by limiting claims for customers’ or users’ activities.

Providers may adopt strategies to reinforce and amplify asymmetries in operation – hoarding patents and other intellectual property; lobbying to create or exacerbate regulatory barriers to entry; or (relatedly) lobbying around licensing, labour or data protection law (for example). Providers typically pursue vertical and horizontal integration, seeking to maximise economies of scale and reduce the extent to which their own platforms and services rely on infrastructure operated by others (some providers, such as Google, Amazon and Microsoft, have achieved high levels of integration). Moreover, providers may incentivise customers to use several of their services rather than those of competitors, centralising applications around their technologies and amplifying asymmetries in operationFootnote ⁵⁸ (as a provider can operate not just some technologies underlying an application, but potentially many).

Patterns of control are typically asymmetric in ways shaped by patterns of operation. Indeed, providers can leverage asymmetries in operation towards asymmetric patterns of control: more vertically integrated providers – operating much of the infrastructure for their services – can delegate to fewer of their own providers (for example). Yet patterns of operation do not determine exchanges of control. Rather, control within and across chains is also shaped by legal and political economic context, by technical necessity (delegating control to individual customers of aspects of services which are uniform across instances would not necessarily work), and by providers’ commercial interests in retaining control over both essential and incidental aspects. Indeed, providers may delegate the minimum control needed to allow customers to use their service, retaining as much as possible for themselves. While legal, technical, economic relationships between actors in captured markets are configured in many ways around providers’ interests, patterns of control thus arise from interactions of all sorts of influences, shaping relations and exchanges between providers, vendors, suppliers, customers and others.

Law, for instance, structures patterns of control in a general sense, shaping and underpinning the relations through which providers’ market position arises. For example, delivery riders’ legal classification as workers, employees or self-employed affects their individual and collective protections and rights against the vendor and the degree of control they may have over certain aspects of the platformFootnote ⁵⁹ (and may prompt reform to address social and legal implications of new ways of working).Footnote ⁶⁰ Restrictions on industrial action and collective bargaining affect riders’ abilities to act together to protect their rights and interests by limiting the vendor’s control over aspects of the platform which shape their work and interactions with customers and restaurants.Footnote ⁶¹ Data protection and privacy laws shape the processes of capture and of data’s production, processing and flow (with the commodification and economic value of data – a key reason why differential abilities to produce and productively use data matter in structuring supply chains – itself being legally constructed).Footnote ⁶² These data protection and privacy laws typically rely on the principle of individual control by data subjects,Footnote ⁶³ limiting opportunities to challenge systemic problems or represent collective interests.Footnote ⁶⁴ Regulatory compliance obligations may present barriers to new market entrants, and lax competition enforcement combined with providers’ financial resources often allows them to buy out potential competitors.

Law thus affords conditions for patterns of operation and control and shapes interdependencies across societies. Asymmetries in these produce asymmetries in supply chain interdependencies, which are often heavily unequal both within and across chains. Within chains, individual customers (or users) generally depend much more on a provider’s services – potentially underpinning essential aspects of the application’s core function – than the provider depends on their custom. Across chains, due to market concentration, many customers may depend a great deal on a single provider, who seemingly barely depends on the others individually.Footnote ⁶⁵ Individual delivery riders may depend on the food app vendor for much of their income, whereas each rider plays a relatively minor role in the vendor’s service (which may have many potential riders). Yet often-overlooked collective dimensions matter: if riders unionise or otherwise act collectively, the vendor’s dependence on riders as a whole, rather than individually, quickly becomes apparent.

8.3.4 Regulatory Power

Asymmetries in supply chain interdependencies arising from asymmetric patterns of operation and control produce, shape and intensify power dynamics within and across chains which are also asymmetric (often greatly). These asymmetries potentially allow providers to make access to technologies which are essential to applications in many markets and sectors conditional on acceptance of certain requirements.Footnote ⁶⁶ Indeed, Cohen argues that platforms’ legal and technical protocols perform a ‘double function’: ‘affording access but also points of contact for exercises of technological and political authority’.Footnote ⁶⁷ Through these points of contact, providers aim to extend their power up and down chains in attempts to maintain their position, minimise legal and other risks and responsibilities and maximise their commercial benefit.

For example, the vendor’s control of aspects of the food delivery platform which are essential to its core function of intermediating between restaurants, customers and riders – itself conditioned by law and by the vendor’s relationships with providers who operate or control underlying technologies – produces key asymmetries in their collective interdependence which favour the vendor. This allows the vendor to insist on retaining incidental aspects of the platform which amplify asymmetries in interdependence with and between platform users and perform some governance function – algorithmically shaping interactions between restaurants, customers and riders, producing and extracting data for targeting and enforcing policies – thereby permitting them to structure those relations to maximise their commercial benefit from the platform’s core function of intermediation. A cloud provider may insist on retaining control over incidental aspects of services to restrict purposes or contexts of use to reduce legal or reputational risk, or which limit integration with competitors’ services.

Sizeable asymmetries in interdependence afford certain providers regulatory power in their relations with others:Footnote ⁶⁸ the capacity to durably affect or constrain their behaviour. This regulatory power is not held by one actor exclusively, but is multipolar: cloud providers may have some regulatory power in their relations with the food app vendor, for example, while the vendor has some regulatory power in their relations with restaurants and riders (itself conditioned by the cloud provider’s regulatory power in their relations with the vendor). Yet the power dynamics between the app vendor and a delivery rider, for example, result also from the relations between riders themselves and whether they act individually or collectively. Power dynamics in data-driven supply chains therefore arise relationally but multilaterally through the interactions of the many actors therein (conditioned by the legal, social, technological and market factors which shape those relations). Providers may exploit their regulatory power in efforts to build, maintain, and amplify further legal and technological asymmetries both within and across chains, shaping the activities of individual customers and users and power dynamics in captured markets to their advantage. Although regulatory power is multipolar, asymmetries in supply chains thus produce asymmetries in that power which offer opportunities for extensive private ordering by providers.

The services model – affording the ability to dictate terms and alter, withdraw or cancel services at will – itself offers providers opportunities to retain some control over deployment of their technologies in customers’ applications in ways which would have been impossible with traditional software licensing via physical media.Footnote ⁶⁹ A primary mechanism for this is standard form, take-it-or-leave-it contractual Terms of Service.Footnote ⁷⁰ These establish legal relationships which underpin technical and economic relationships and through which customers and users exchange control in particular ways (in the context of commercial, technical, legal and other influences setting a general structure to them). At the same time, however, providers also impose terms for exchange of control which seek to minimise legal costs, risks, responsibilities and accountabilities by passing them to others.Footnote ⁷¹ As such, they are influenced by (for instance) intermediary liability, intellectual property, data protection and employment laws, but they also make normative claims about those laws’ meanings in the context of their relations with customers and users and – as a result – shape their effects to providers’ advantage.

Employment status for delivery riders (for example) depends on broader legal context, but the food app vendor can construct their Terms of Service (and their practices and the platform’s operation) towards one classification or another. In other cases, providers’ Terms may make intellectual property law claims to restrict access to their proprietary technologies and foreclose efforts to better understand how those technologies work, while claiming rights to use others’ intellectual property to train models and improve services.Footnote ⁷² Providers may similarly seek to legitimate their practices in data protection law, claiming questionable legal bases for data processing and rights to use data for development, service improvement, and tracking and targetingFootnote ⁷³ (obtaining data being an important aspect of platforms’ competitive strategiesFootnote ⁷⁴ and producing data being a key function of supply chains). Providers may claim to be processors in data protection law,Footnote ⁷⁵ seeking to place primary compliance and regulatory responsibilities for their data practices onto customers (aided by data protection law’s binary distinction between controllers and processors).Footnote ⁷⁶ Terms may also disclaim liability for things not covered by intermediary liability protections, even where problems originate in technologies providers operate or where they retain control.Footnote ⁷⁷ Moreover, Terms may contain choice of law or arbitration clauses to resolve disputes over meaning in favourable jurisdictions or away from formal legal processes.

In effect, through their Terms of Service, providers use various areas of law as both swords and shields in the relations through which their position and power arises, structuring their legal relationships with – and the activities of – other supply chain actors. Providers seek maximum retained control over their technologies (depending on the type of service) as far down chains as possible – and maximum benefit from that control – while externalising costs or responsibilities which might otherwise come with it (often attaching in these agreements to the more limited control delegated to others). Through this, enabled by simultaneous control by separate actors of different aspects of the same systems, legal rights and benefits around operation and control (on the one hand) are decoupled from legal risks, responsibilities and obligations (on the other) and arranged in ways which maintain and amplify technological and legal asymmetries to providers’ benefit.

The regulatory effects of softwareFootnote ⁷⁸ are also leveraged by providers in the design of platforms and services to prevent or affordFootnote ⁷⁹ different activities by users in line with platform policies. In other cases, providers may retain control over incidental aspects of technologies allowing them to enforce policies and restrictions to limit customers’ delegated control through APIs, the design of services, or other means. The recommender service provider which underpins personalisation in the food delivery app (for example) may delegate control over the system such that the vendor can integrate it into the platform, while (for policy purposes) retaining control allowing them to limit the kinds of content restaurants input for that personalisation. A content delivery network provider may delegate control to allow customers to use that service to underpin applications of many kinds, while also retaining rights to refuse content they find objectionable or commercially difficult. If the content network provider is several steps up or down a chain from the point where content is ‘submitted’, those rights will necessarily be extended through those steps (i.e. the provider extends retained control up the chain to govern the activities of actors they do not have a direct relationship with). A facial recognition service provider may use automated checks to restrict why and where customers can deploy that service. Such mechanisms may extend an application’s supply chain – the recommender service provider may use another provider’s content moderation service to check inputs for compliance, for instance (involving another exchange of control between the recommender service provider and the moderation service provider).

Asymmetries in supply chain interdependencies thus provide conditions by which certain actors shape others’ activities to their benefit. Through these strategies, providers’ regulatory power affects what customers can do for their customers and end-users in turn and therefore exchanges of control within and across chains more generally. Asymmetric patterns of ownership and control, constructed by and combined with asymmetries in legal rights and obligations, thus contribute to structuring supply chains in ways that allow providers to reinforce and amplify those asymmetries and further consolidate capture of markets and social relations. These dynamics unfold in contexts where various influences shape and amplify asymmetries in interdependence and balances of power dynamically in ways which were not inevitable and are not unchangeable. Providers’ position and regulatory power in data-driven supply chains is therefore not a technological phenomenon, but a socio-legal one which can – in principle – be conditioned, contested, and dismantled.

8.4.1 Asymmetries in Operation and Control

Asymmetric interdependencies – reflecting asymmetries in operation and control – provide underlying conditions for providers’ regulatory power in supply chains; they should therefore be core considerations for governance of supply chains. Providers’ control over essential aspects affords asymmetric interdependencies, allowing providers to insist on retaining control over incidental aspects, shape others’ activities to their benefit, and impose conditions of access which reinforce and amplify asymmetries. As such, a key regulatory goal should be arranging patterns of control to address asymmetries in interdependence: which aspects of what kinds of service are retained by and delegated to whom; who controls essential aspects relating to core functions and who controls others. Where platforms algorithmically intermediate markets, for instance, providers typically retain significant control for technical and commercial reasons. This might instead be distributed as appropriate between the actors who actually constitute those markets to mitigate or eliminate the provider’s regulatory power and balance interdependencies between them. Where conflicts between different interests arise – for example, where incidental aspects of a platform affect both riders and restaurants and could therefore be controlled by either – control should be assigned according to the goal of balancing interdependencies. What is appropriate differs between markets and sectors, and between services with different core functions, so (re-)distributing control requires a specific understanding of services and of the social and economic relations around them.

Importantly, individualised control has limitations, not least because of any remaining asymmetries between providers and individual customers or end-users. Moreover, features of services which are uniform across instances – such as essential aspects of systems going to their core function – may not be individually variable for technical reasons. Legal and regulatory governance mechanisms may therefore seek collective control of some or all such aspects. For the food delivery app, for example, this could be mutualised, cooperative, or other collective control over certain essential aspects of the platform’s intermediation function by particular groups with concentrated interestsFootnote ⁹⁸ – such as delivery riders – or of incidental aspects which perform some governance function. Collective control over essential aspects could rebalance core interdependencies, which between providers and customers as a whole are perhaps not so asymmetric. This can be combined with other strategies aimed at incidental aspects of services: for instance, delivery riders as a whole might collectively control the food app’s essential aspects even while particular riders and restaurants have more individualised control over incidental aspects relating to their activities through the platform. More generally, legal restrictions which prevent certain actors from acting collectively,Footnote ⁹⁹ or frameworks such as data protection which focus on individual control and limit opportunities for collective intervention or redress,Footnote ¹⁰⁰ should be revisited with a view to balancing interdependencies in this way.

Collective control over essential aspects of infrastructural services is difficult given the diverse and diffuseFootnote ¹⁰¹ interests of customers across markets and sectors accessing the same general-purpose technologies. Moreover, there may be multiple ‘steps’ up and down chains from the point of infrastructural services, which makes distributing control to customers, end-users, or others difficult. However, addressing asymmetries in interdependence fundamentally involves reducing the relative importance of any single actor’s retained control. As such, meaningful representation or protection for groups of shared interest – such as data subjects’ shared interest in the protection of their personal data,Footnote ¹⁰² or content creator’s shared interest in protecting their intellectual property from appropriation into AI production – to collectively control or contest retained aspects of providers’ services may help. Law may also seek to restrict infrastructural providers’ ability to leverage control of essential aspects of services into retaining control of incidental aspects which amplify asymmetries in interdependence and afford regulatory power. This could involve reducing providers’ discretion in insisting on particular exchanges of control with customers, by identifying the essential aspects of different kinds of service (based in the core function they perform), regulating the conditions for providers’ retention of those aspects, and passing control over others to customers where possible. Again, this requires specific understandings of particular kinds of service, which will fulfil different core functions.

Mitigating asymmetries in operation (or their effects) at infrastructural level may be possible through interventions to reduce the relative importance of operating particular technologies and create space for alternatives, such as requiring open sourcing of proprietary systemsFootnote ¹⁰³ or interoperabilityFootnote ¹⁰⁴ between platforms.Footnote ¹⁰⁵ Enforced separations between infrastructural and user-facing services (challenging vertical integration), or requiring providers to divest certain services, limit reuse of data in development across services,Footnote ¹⁰⁶ or limit reuse of models or systems across services (challenging horizontal integration) may also help. Some of these interventions have been proposed or sought in certain jurisdictions.

8.4.2 Asymmetries in Legal Rights and Obligations

Asymmetric patterns of operation and control do not fully explain governance in supply chains, however. Various other influences and supply chain dynamics shape and amplify asymmetries in interdependence, including laws which afford providers’ position in patterns of control. Moreover, providers’ regulatory power works in part through certain legal arrangements with others – their Terms of Service strategically interpret and make normative claims about relevant laws and legal frameworks to their advantage. Addressing regulatory power therefore requires addressing asymmetries in legal rights and obligations which influence supply chain interdependencies and exchanges of control in complex ways, reinforcing and amplifying technological and other asymmetries.

As described above, providers’ Terms work with various frameworks to decouple legal rights and benefits around operation and control of technologies from legal risks and responsibilities, arranging them to instead maximise providers’ commercial benefit and externalise liabilities and accountabilities to the limited control delegated to customers. A key goal for legal and regulatory frameworks in structuring interdependencies between actors is therefore to reliably align legal risk and responsibility with operation and controlFootnote ¹⁰⁷ – particularly for infrastructural services where re-distributing control is difficult. This means assigning roles, responsibilities, rights, obligations and accountabilities to reflect factual distributions of control over both essential and incidental aspects of technologies in accurate, sufficiently nuanced, and clearly specified waysFootnote ¹⁰⁸ – to ensure that providers who benefit from services used across societies and economies bear the appropriate costs, that legal responsibilities around production and operation go to those who are factually responsible for them, and that customers’ and end-users’ rights and interests further down chains are properly protected. Addressing providers’ ability to unilaterally determine Terms of Service violations at their discretion, remove customers’ access at will, and limit avenues for challenge or redress may also help reduce the relative importance of providers’ retained control.

9.4.1 Environmental Harms of Data Infrastructures

Data-driven infrastructures and data value chains contribute to environmental harms in direct ways due to the massive resource extraction, water and energy consumption, and contribution to global waste that they entail.Footnote ⁶⁹ Data-driven environmental monitoring infrastructures rely on sensors, satellites, data servers, data centers, cables, chips, semiconductors, batteries, and other material equipment that is required for them to run. These in turn require extraction of lithium and rare earth minerals to build, which form part of the global data supply chain underpinning data-driven infrastructures.Footnote ⁷⁰ Moreover, the running of data centers contributes to significant water and energy consumption. These infrastructures require massive amounts of energy consumption to power the equipment. This electricity generates heat which then requires water to cool the equipment, straining the water resources of local communities that are already facing drought conditions.Footnote ⁷¹ Significant amounts of water are also used in mineral extraction processes, which can also result in toxic chemical leaks that affect local water supplies and ecologies.Footnote ⁷² As drought conditions are one of the impacts of climate change, data-driven infrastructures are further exacerbating this problem as they require water for mineral extraction and cooling of equipment in places already facing water shortages.

Digital devices and infrastructures also contribute to e-waste (or electronic waste), one of the fastest growing forms of global waste, as well as to space debris in the context of climate monitoring satellites. The toxic e-waste that is produced is then transported and dumped in the Global South, including Mongolia, Ghana, and Nigeria.Footnote ⁷³ These dynamics can be placed in a broader pattern of trade in global waste markets.Footnote ⁷⁴

These impacts are increasing as more data is collected, and as more intensive data-driven infrastructures, algorithms, and machine learning models are used to analyze and make use of that data. As will be further discussed, this dynamic is driven by a growth-oriented production model and political economy in which tech corporations such as Google, Microsoft, and Amazon play a significant role.Footnote ⁷⁵ Further, environmental data infrastructures have raised concerns about their potential uses as extractive technologies that can increase “the efficiency of resource extraction.”Footnote ⁷⁶ Oil and gas companies, for example, benefit from environmental monitoring, such as ocean observation.Footnote ⁷⁷ These infrastructures can be used not only for environmental monitoring, but also for more efficient resource extraction that would further exacerbate environmental degradation.

9.4.2 Political Economy

The UNEP and other international organizations use environmental data to develop policy recommendations, monitoring compliance with international treaties and climate targets, including those in the UNFCCC Paris Agreement, make impact assessments, coordinate emergency responses, and support research into climate change. These data-driven initiatives are facilitated by corporate partnerships for cloud computing, data visualization, and mapping. These partnerships are necessitated by private corporations’ considerable access to and ownership of cloud infrastructure and algorithmic capabilities. Private corporations’ capabilities for algorithmic data analytics based on their access to large-scale datasets to train their machine learning models shape the public-private relations in this context since public institutions and international organizations often lack those technical capabilities on their own. The reliance on private technology corporations in this context, however, risks replicating the same economic growth, production, and extractive models that have contributed to ecological crisis in the first place.Footnote ⁷⁸

UNEP partners with companies like Google and Microsoft for use of their data infrastructures, including geospatial data analytics, mapping, visualization, and cloud computing capabilities.Footnote ⁷⁹ For example, Google Earth Engine, powered by Google’s cloud infrastructure, provides a “planetary-scale platform for Earth science data & analysis.”Footnote ⁸⁰ As further described by Google:

The datasets used for this platform are sourced from, among others, UNEP’s World Conservation Monitoring Centre (UNEP-WCMC), along with Landsat, the European Space Agency (ESA), and the US National Aeronautics and Space Administration (NASA).Footnote ⁸² Google Earth Engine is free for use by non-commercial users and for a fee for commercial users for everything from managing their supply chains to managing climate-related business risks.Footnote ⁸³ The data and the use of algorithms that compile, assemble, and render those datasets visible are commodified in the process.

As Jennifer Cobbe argues, the infrastructural platform services of Google and Microsoft illustrate some of the interdependencies afforded by data-driven technologies and how certain players leverage the “structuring power of technology” to create asymmetrical interdependencies between platforms and users.Footnote ⁸⁴ The relationship between UNEP’s environmental data and monitoring initiative and these platforms is complex; there are interdependencies within the ecosystem of actors and technologies formed around data-driven infrastructures. There are also interdependencies between platforms and contemporary forms of global environmental governance that incorporate data-driven infrastructures.

The uses of environmental data by UNEP and other organizations that are shifting to more open-source, accessible forms of data are also changing the business practices and data-driven business models of platforms providing digital infrastructure for environmental data analytics, visualization, and cloud computing. The increasing use of environmental observation and monitoring data that is free and openly accessible has “shifted the value proposition for many companies from data towards algorithms, applications and tailored interfaces based on [Earth observation (EO)] data and related services.”Footnote ⁸⁵ Earth observation data is thus helping shape environmental data economies that promote the profitability and economic interests of private technology corporations.

The European Commission has made efforts toward developing an environmental data economy as part of its plan to develop EU industry competitiveness. One of the claimed benefits of environmental observation data obtained from the EU Copernicus satellite program is the opportunity to support economic growth in the EU.Footnote ⁸⁶ While the vast majority of the environmental observation data obtained from the EU Copernicus satellite program is free and open data accessible to the public, commercializing uses of this data is central to initiatives to maintain economic competitiveness for the EU.Footnote ⁸⁷ For example, in 2016, the European Commission launched the EU Copernicus Accelerator to increase the use of EO data obtained from the EU Copernicus satellite program by incentivizing startups to use the data for commercial applications.Footnote ⁸⁸ This example too points to a developing environmental data economy built around commercializing and commodifying open environmental data.

These initiatives leave questions on the political economy of data-driven infrastructures, and the role of law in shaping that political economy, in the background. These questions are important to consider when there is infrastructural reliance on private technology corporations. The use of data-driven infrastructures in global environmental governance will further empower and entrench already powerful technology companies and has created profitable opportunities for private companies focusing on environmental data. This is illustrated by the increased use of private satellites for EO data, which is visualized and commodified using automated software, algorithms, and machine learning analytics to process satellite images.Footnote ⁸⁹ For example, one company, Planet, has over 200 satellites in orbit for EO, outnumbering NASA and the ESA combined. The company sells products based on this data for commercial and humanitarian use, allowing customers to derive value from their datasets.Footnote ⁹⁰ This raises issues of inequality in relation to the production of and access to environmental datasets, who stands to benefit from their production and commodification, and who might be vulnerable to their harms, including both their associated environmental harms and the embedding of racial inequalities in environmental data.

While public institutions behind open and free environmental monitoring and observation data offer benefits in terms of accessibility by the public and are preferable to commercialized environmental data which offer less accessibility, these public institutions are often reliant upon partnerships with private technology companies and platforms for data analytics, storage, and visualization services. One reason this is problematic is that private corporations are less responsive to public concerns and interests and are driven by motives of capital accumulation that are at odds with public concerns over climate change. Moreover, as Isin and Ruppert note, “[t]he data that sensory power produces is transparent (‘open data’) but how such data is transformed into analytics or intelligence remains opaque as are the infrastructures, code, algorithms and machine learning practices that are part of the relations that make up sensory assemblages.”Footnote ⁹¹ Part of this opacity can be traced to the need to gain competitive advantages in a capitalist economy.Footnote ⁹²

In many ways, climate change has become a legitimizing force for big data and the “practice of corporate collection and targeting of user data” for profit.Footnote ⁹³ Platforms, data-driven corporate actors, and their technological capabilities have become essential infrastructures for responding to climate change with data-driven environmental governance. By incorporating the uses of private technology corporations’ software and hardware infrastructure to inform environmental governance, the processing of environmental datasets to make them actionable for policymakers entangles commodified environmental data and algorithms with political decisions on how to best govern the climate.Footnote ⁹⁴

Digital technologies cannot be understood without also considering “capitalist understandings of value-market relations.”Footnote ⁹⁵ The use of data-driven infrastructures in environmental governance reflects a “technological solutionism”Footnote ⁹⁶ that is inadequate to address the broader structural and political economic problems associated with ecological crisis, including the role of data-driven infrastructures in perpetuating the same growth-oriented and extractive political economy of capitalism that underlies environmental degradation in the first place.