The internet is a fossil fuel industry.
Ben Tarnoff, 21.05.2020
Ben Tarnoff, co-founder of Logic Magazine, explores the devastating impact of cloud computing on the climate — and makes the case for a radical transformation of the internet as we know it.
As of writing, roughly half of the world’s population is living under lockdown.
Not everyone can remain indoors, of course: millions of working-class people put their lives at risk every day to be the nurses, grocery store clerks, and other essential workers on whom everyone else’s survival depends. But globally, a substantial share of humanity is staying home.
One consequence is a sharp increase in internet usage. Trapped inside, people are spending more time online. The New York Times reports that in January, as China locked down Hubei province — home to Wuhan, the original epicenter of Covid-19 — mobile broadband speeds dropped by more than half because of network congestion. Internet speeds have suffered similar drops across Europe and the United States, as stay-at-home orders have led to spikes in traffic. In Italy, which has one of the highest coronavirus death tolls in the world, home internet use has increased 90 percent.
The internet is already deeply integrated into the daily rhythms of life in much of the world. Under the pressures of the pandemic, however, it has become something more: the place where, for many, life is mostly lived. It is where one spends time with family and friends, goes to class, attends concerts and religious services, buys meals and groceries. It is a source of sustenance, culture, and social interaction; for those who can work from home, it is also a source of income. Quarantine is an ancient practice. Connected quarantine is a paradox produced by a networked age.
Anything that can help people endure long periods of isolation is useful for containing the virus. In this respect, the internet is a blessing — if an unevenly distributed one. Indeed, the pandemic is highlighting the inequalities both within and across countries when it comes to connectivity, and underlining why internet access should be considered a basic human right.
But the new reality of connected quarantine also brings certain risks. The first is social: the greater reliance on online services will place more power in the hands of telecoms and platforms. Our undemocratic digital sphere will only become more so, as the firms that own the physical and virtual infrastructures of the internet come to mediate, and to mold, even more of our existence. The second danger is ecological. The internet already makes very large demands of the earth’s natural systems. As usage increases, those demands will grow.
In our efforts to mitigate the current crisis, then, we may end up making other crises worse. A world in which the internet as it is currently organized becomes more central to our lives will be one in which tech companies exercise more influence over our lives. It may also be one in which life of all kinds becomes harder to sustain, as the environmental impact of a precipitously growing internet accelerates the ongoing collapse of the biosphere — above all, by making the planet hotter.
Machines Heat the Planet
To understand how the internet makes the planet hotter, it helps to begin with a simplified model of what the internet is. The internet is, more or less, a collection of machines that talk to one another. These machines can be big or small — servers or smartphones, say. Every year they become more ubiquitous; in a couple of years, there will be thirty billion of them.
These machines heat the planet in three ways. First, they are made from metals and minerals that are extracted and refined with large inputs of energy, and this energy is generated from burning fossil fuels. Second, their assembly and manufacture is similarly energy-intensive, and thus similarly carbon-intensive. Finally, after the machines are made, there is the matter of keeping them running, which also consumes energy and emits carbon.
Given the breadth and complexity of this picture, it would take a considerable amount of time to map the entire carbon footprint of the internet precisely. So let’s zero in on a single slice: the cloud. If the internet is a collection of machines that talk to one another, the cloud is the subset of machines that do most of the talking. More concretely, the cloud is millions of climate-controlled buildings — ”data centers” — filled with servers. These servers supply the storage and perform the computation for the software running on the internet — the software behind Zoom seders, Twitch concerts, Instacart deliveries, drone strikes, financial trades, and countless other algorithmically organized activities.
The amount of energy consumed by these activities is immense, and much of it comes from coal and natural gas. Data centers currently require 200 terawatt hours per year, roughly the same amount as South Africa. Anders Andrae, a researcher at Huawei, predicts that number will grow 4 to 5 times by 2030. This would put the cloud on par with Japan, the fourth-biggest energy consumer on the planet. Andrae made these predictions before the pandemic, however. All indications suggest that the crisis will supercharge the growth of the cloud, as people spend more time online. This means we could be looking at a cloud even bigger than Japan by 2030 — perhaps even the size of India, the world’s third-biggest energy consumer.
Machine Learning is a Fossil Fuel Industry
What can be done to avert the climate damage of such a development? One approach is to make the cloud run on renewable energy. This doesn’t entirely decarbonize data centers, given the carbon costs associated with the construction of the servers inside of them, but it does reduce their impact. Greenpeace has been waging a campaign along these lines for years, with some success. The use of renewables by data centers has grown, although progress is uneven: according to a recent Greenpeace report, Chinese data centers are still primarily powered by coal. It also remains difficult to accurately gauge how much progress has been made, since corporate commitments to lower carbon emissions are often little more than greenwashing PR. “Greening” one’s data centers can mean any number of things, given a general lack of transparency and reporting standards. A company might buy some carbon offsets, put out a celebratory press release, and call it a day.
Another approach is to increase the energy efficiency of data centers. This is an easier sell for companies, because they have a strong financial incentive to lower their electricity costs: powering and cooling data centers can be extraordinarily expensive. In recent years, they have come up with a number of ways to improve efficiency. The emergence of “hyperscale” data centers, first developed by Facebook, has been especially important. These are vast, automated, streamlined facilities that represent the rationalization of the cloud: they are the digital equivalent of the Fordist assembly line, displacing the more artisanal arrangements of an earlier era. Their economies of scale and obsessive optimizations make them highly energy-efficient, which has in turn moderated the cloud’s power consumption in recent years.
This trend won’t last forever, however. The hyperscalers will max out their efficiency, while the cloud will continue to grow. Even the more conscientious companies will have trouble procuring enough renewables to keep pace with demand. This is why we may also have to contemplate another possibility: not just greening the cloud, or making it more efficient, but constraining its growth.
To consider how we might do that, let’s first consider why the cloud is growing so fast. One of the most important factors is the rise of machine learning (ML). ML is the field behind the current “AI boom.” A powerful tool for pattern recognition, ML can be put to many purposes, from analyzing faces to predicting consumer preferences. To recognize a pattern, though, an ML system must first “learn” the pattern. The way that ML learns patterns is by training on large quantities of data, which is a computationally demanding process. Streaming Netflix doesn’t place much strain on the servers inside a data center; training the ML model that Netflix uses for its recommendation engine probably does.
Because ML hogs processing power, it also carries a large carbon footprint. In a paper that made waves in the ML community, a team at the University of Massachusetts, Amherst found that training a model for natural-language processing — the field that helps “virtual assistants” like Alexa understand what you’re saying — can emit as much as 626,155 pounds of carbon dioxide. That’s about the same amount produced by flying roundtrip between New York and Beijing 125 times.
Training models isn’t the only way that ML contributes to climate change. It has also stimulated a hunger for data that is probably the single biggest driver of the digitization of everything. Corporations and governments now have an incentive to acquire as much data as possible, because that data, with the help of ML, might yield valuable patterns. It might tell them who to fire, who to arrest, when to perform maintenance on a machine, or how to promote a new product. It might even help them build new kinds of services, like facial recognition software or customer-service chatbots. One of the best ways to make more data is to put small connected computers everywhere—in homes and stores and offices and factories and hospitals and cars. Aside from the energy required to manufacture and maintain those devices, the data they produce will live in the carbon-intensive cloud.
The good news is that awareness of ML’s climate impacts is growing, as is the interest among practitioners and activists in mitigating them. Towards that end, one group of researchers is calling for new reporting standards under the banner of “Green AI.” They propose adding a carbon “price tag” to each ML model, which would reflect the costs of building, training, and running it, and which could drive the development of more efficient models.
This is important work, but it needs a qualitative dimension as well as a quantitative one. We shouldn’t just be asking how much carbon an ML application produces. We should also be asking what those applications do.
Do they enable people to lead freer and more self-determined lives? Do they cultivate community and solidarity? Do they encourage more equitable and more cooperative forms of living? Or do they extend corporate and state surveillance and control? Do they give advertisers, employers, and security agencies new ways to monitor and manipulate us? Do they strengthen capitalist class power, and intensify racism, sexism, and other oppressions?
Resistance with Transformation
A good place to start when we contemplate curbing the growth of the cloud, then, is asking whether the activities that are driving its growth contribute to the creation of a democratic society. This question will acquire new urgency in the pandemic, as our societies become more enmeshed in the internet. It is a question that cannot be resolved on a technical basis, however — it is not an optimization problem, like trying to maximize energy efficiency in a data center. That’s because it involves choices about values, and choices about values are necessarily political. Therefore, we need political mechanisms for making these choices collectively.
Politics is necessarily a conflictual affair, and there will be plenty of conflicts that arise in the course of trying to both decarbonize and democratize the internet. For one, there are obvious tensions between the moral imperative of improving and expanding access and the ecological imperative of keeping the associated energy inputs within a sustainable range. But there will also be many cases where restricting and even eliminating certain uses of the internet will serve both social and environmental ends simultaneously.
Consider the fight against facial recognition software that has erupted across the world, from protesters in Hong Kong using lasers to disrupt police cameras to organizers in the United States pushing for municipal bans. Such software is incompatible with basic democratic values; it also helps heat the planet by relying on computationally intensive ML models. Its abolition would thus serve both the people and the planet.
But we need more than abolition. We also need to envision and construct an alternative. A substantive project to decarbonize and democratize the internet must combine resistance with transformation; namely, it must transform how the internet is owned and organized. So long as the internet is held by private firms and run for profit, it will destabilize natural systems and preclude the possibility of democratic control. The supreme law of capitalism is accumulation for accumulation’s sake. Under such a regime, the earth is a set of resources to be extracted, not a set of systems to be repaired, stewarded, and protected. Moreover, there is little room for people to freely choose the course of their lives, because everyone’s choices — even those of capitalists — are constrained by the imperative of infinite accumulation.
Dissolving this law, and formulating a new one, will of course involve a much broader array of struggles than those aimed at building a better internet. But the internet, as its size and significance grows through the pandemic, may very well become a central point of struggle. In the past, the internet has been a difficult issue to inspire mass mobilization around; its current highly privatized form, in fact, is partly due to the absence of popular pressure. The new life patterns of connected quarantine might reverse this trend, as online services become, for many, both a window to the world and a substitute for it, a lifeline and a habitat. Perhaps then the internet will be a place worth struggling to transform, as well as a tool for those struggling to transform everything else.