Since joining Fastly, a few people have asked me, “Why is an internet company thinking about sustainability?” It’s a fair question. When most people think of the internet, they think of something omnipresent - always there, always accessible. This wasn’t the case twenty-five years ago, but today, we expect connectivity wherever we go. Most of our devices are connected to the internet 24/7, but to most, it feels unseen and intangible.
However, at Fastly, we are acutely aware of the internet’s physicality because we provide the physical infrastructure. The internet's physicality manifests in various forms: servers processing countless requests every second, sitting in data centres, requiring power and cooling, connected by cables spanning vast distances.
When you recognise the Internet as something physical, it becomes immediately obvious that you need to think about sustainability: supplying all that physical infrastructure and providing it with energy has consequences. The good news is that even small changes can have a significant impact when applied at scale. Here’s how.
Download the PDF version of the infographic.
Renewable Energy in Data centres
Data centres require massive amounts of electricity to power and cool servers: the IEA estimated global data centre energy consumption in 2022 was 1-1.3% of total energy demand. When data centres incorporate energy from renewable sources, energy-related greenhouse gas (GHG) emissions are significantly reduced.
This is why Fastly prioritises working with colocation providers that are leading the charge in renewable energy adoption. Our largest colocation data centre supplier, Equinix, last year covered its global data centre network with 96% renewable energy. Fastly’s operations in Equinix data centres have been covered by at least 95% renewable energy since 2021. Working with data centre operators that incorporate renewable energy, such as Equinix, reduces the energy-related GHG emissions of running our platform, while still maintaining the high-performance service our customers rely on us for.
Reducing Power Consumption and Data Transmission
Reducing the total amount of energy consumption is also a critical sustainability consideration. One major focus area we can directly influence is to reduce the energy consumed by our IT equipment.
Fastly works hard to improve the energy efficiency of our equipment by:
considering energy consumption in our hardware procurement exercises,
optimising the configuration of both our software and the hardware it runs on to minimise energy consumption
More broadly, PUE (Power Usage Effectiveness) is a metric that measures the efficiency of a data centre's overall energy consumption. It's calculated by dividing the total energy used by the data centre facility (including cooling, lighting, and other infrastructure) by the energy used specifically by IT equipment (servers, storage, networking). Understanding and monitoring PUE across a distributed network is an important way to assess and optimise overall power consumption.
It’s also important to reduce the amount of data being transmitted. Every gigabit per second of data requires physical infrastructure, electricity, and expensive, resource-intensive cables. If we can optimise how data moves across the internet we’ll consume less of these resources, benefitting our customers, our company, and the environment.
Data transmissions can be reduced by:
Optimizing image sizes so they load faster and consume less bandwidth
Reducing unnecessary data transfers by caching frequently accessed content
Delivering data from servers closer to users to cut down on energy-intensive long-distance transmissions
Focussing on the full technology lifecycle
We’ve identified sustainability opportunities during the use-phase of technology, but there‘s a full lifecycle to consider, from the materials and components used in hardware through to what happens at end of life.
One of the best ways to reduce impact across the lifecycle is to use the least amount of hardware to achieve the same outcome, such as:
using larger, more powerful cache servers in fewer locations,
leveraging switched-based architecture that eliminates the need for large chassis routers, and
extending hardware lifespans beyond the typical four to five years.
When hardware eventually reaches its end of life, we can still make sure it’s not the end of the road. Older equipment can be repurposed to use cases with lower computing needs, components can be harvested for spares and remanufacturing and e-waste can be dealt with responsibly by ensuring valuable source materials such as precious metals are recycled. The future is circular.
Making AI Technology More Sustainable
Emerging technologies pose new challenges in the quest for sustainability. Cryptocurrencies have already shown their potential for enormous energy consumption. Now, artificial intelligence (AI) and large language models (LLMs) are raising similar concerns.
These technologies are computationally intensive and require significant energy resources. To put this into perspective, one study suggests that the AI industry could consume as much energy as a country the size of the Netherlands by 2027. No wonder we've heard discussions about restarting Three Mile Island and other nuclear reactors just to power AI. This issue will only grow in the future as generative AI use cases continue to proliferate, and that is why we need to get ahead of it. As a sector, we need to be very conscious of how to deal with the sustainability implications of growth.
Key Questions for AI Sustainability:
How can we mitigate the environmental impact of AI queries?
How can we reduce the number of queries being processed?
How can we cache queries more effectively? (as we’re doing with our AI accelerator)
How can we protect against malicious queries to prevent unnecessary work?
How can we make queries more efficient?
How can we process queries closer to users to reduce data transmission?
There’s no single solution to these questions, but a collective effort with small interventions could add up to a significant impact.
Caching AI Queries More Efficiently
The challenge with large language models is that no two queries are identical. They aren’t structured in a way that computers can easily process, making them difficult to cache. As a result, every query requires significant computational work, making it extremely expensive.
This is reminiscent of the early web era when database-driven websites were sluggish and costly. Over time, we optimised databases, making them faster and less resource-intensive. We also developed strategies around caching and protection. At a large scale, caching is a form of protection—it reduces unnecessary work, blocks malicious queries, and minimises costs and environmental impact.
At Fastly, we've spent years working on the intersection of caching and protection. We have moved our security capabilities to the edge with a flexible, signal-based approach rather than rigid rules. This has allowed us to handle a wider range of threats.
As AI adoption grew, companies turned to us for help. In response, we developed a semantic cache—a system that breaks down queries into their core meaning rather than focusing on the exact wording. While the underlying mathematics are complex, the result is simple: we can cache more AI queries, significantly reducing the number of computations required.
Initially, we built this for OpenAI, but we’ve since expanded it to support multiple LLMs, each with its unique semantics. This has become one of our most popular products, helping companies cut costs while reducing environmental impact. Our goal is to continue developing innovations that make the internet more sustainable, one optimization at a time.
Want to build a faster, greener internet? Check out these 5 expert tips on how to make your website more sustainable.
