Barriers for startups tackling data center energy consumption
How to measure the problem, the uncertainty around the current situation, and the limited number of growth buyers
Recently I have spoken to several people asking about startups in data center energy, the area of focus of my MSc research at Imperial College London. The most common question is: “Where are the interesting opportunities?”, followed by “What interesting startups have you seen tackling this?”
If you have an idea for a new software product, it is a rule that another startup already exists doing the same thing, and they have probably raised significant VC funding! The on-prem to SaaS transition has happened and everything you can think of has multiple businesses already competing. Standing out requires more thoughtful innovation, perhaps by vertical specialisation, better UX, or reimagining a “done” product with a strong opinion.
Innovation in data centers, and energy consumption of IT, is different. There are few companies tackling this and the startups I have seen to-date have not been impressive. Why is this?
How much energy do data centers use?
As I have written in the past:
Estimates of annual data center electricity usage vary from 200 terawatt hours (TWh) (Jones, 2018) to 500 TWh (Bashroush & Lawrence, 2020). The lower of these figures would suggest that data centers consume 1% of global electricity (Jones, 2018), but this could be significantly higher. One study suggests that global data center energy usage was 270 TWh in 2012 ( Van Heddeghem et al, 2014). Another study estimates that 104 TWh will be used by European Union data centers in 2020, which makes a global total of 200 TWh unlikely (Avgerinou, Bertoldi & Castellazzi L, 2017).
- How much energy do data centers use? Mytton, 2020
This highlights the big issue of transparency. Not only do we not know how much energy data centers consume, there are competing projections for whether data center energy consumption is going to fall (Masanet et al., 2020), plateau (Shehabi et al, 2018) or grow (Andrae and Edler, 2015).
There are several complex factors which mean past trends may not play out the same way in the future. The industry moves rapidly and trends such as low power ARM chips are offset by the huge number of IoT devices coming online. Further, new types of components such as ML specialised chips have unknown energy profiles which need analysis (García-Martín et al., 2019) with some estimates showing that training complex models can generate more emissions than 300 flights from San Francisco to New York (Strubell, Ganesh & McCallum, 2019).
Understanding the trajectory is crucial to whether we consider data center energy consumption a problem or not. If data center energy consumption is falling then it may not be a top priority. However, if it is true that 30% of energy demand in Ireland will be from data centers in 2028 (EirGrid, 2019), and 15% in Denmark by 2030 (Danish Energy Agency, 2019), then this is a serious issue that makes sense for startups to tackle.
Who is the buyer?
- Who is the buyer of electricity and where does that appear on the P&L?
- Who is responsible for that cost?
If you are buying physical servers and racking them in a data center, you are the buyer of electricity. You care about how much power servers consume in aggregate, and at peak loads.
If you are renting dedicated servers as a product from a hosting vendor you are paying a blended price of hardware, networking and support. That includes electricity, but it is not broken out. The hosting vendor is the buyer and they provision power for peak capacity. If the server has good power proportionality and it is under-utilised then that will show up as as saving for the hosting vendor, but the customer sees no benefit.
If you are buying cloud resources in the form of VMs (or other direct compute, such as dedicated instances or containers), you are another level up the stack. The cloud vendor is procuring the hardware which consumes electricity. The cloud vendor customer has no visibility of the underlying electricity consumption (Mytton, 2020) and any efficiencies in the infrastructure accrue to the cloud vendor.
At the top of the infrastructure stack, if you are buying products like managed databases or queues, you do not even see the underlying compute resources being consumed. Energy efficiencies accrue to the hosting vendor.
Finally, if you are end-user of a SaaS product e.g. Gmail, you might be buying nothing (as a free user), or are paying on a per seat basis e.g. for Microsoft 365. There is no correlation between price and resource consumption. Again, any efficiencies accrue to the hosting vendor who sells resources to the SaaS vendor (which may be the same company).
Note there is one place where energy efficiencies are noticed: if you own/rack your own hardware, either as a hosting vendor, cloud vendor, or a business that runs its own infrastructure.
Where is the growth?
All trends point towards public cloud as the source of growth for infrastructure spend. The colo market is not dead:
The datacenter colocation market has now grown to $40B per year. Seems less and less likely that “on-premise” enterprise software will completely disappear as some had predicted.
- Chetan Puttagunta via Twitter (2020)
But $40B/yr is small compared to the $288bn in cloud computing market 2019 (Forrester, 2019).
There are only 3 cloud computing companies that matter: Amazon, Google and Microsoft. Nobody else has the Capex to be a relevant player. They are all working on their own innovations in energy consumption, which makes the sales process even harder because you have to compete with private, internal projects.
Challenging, but not impossible
Startups exist to find unusual opportunities and overcome structural challenges. Data centers are just one of the two components which make up the majority of the energy footprint of IT (Jones, 2018). Networking is the second. And given the expected growth of data volumes over the coming decade (Cisco, 2020), potentially a bigger opportunity than data centers e.g. 5G is initially expected to consume x3 the energy of 4G networks (Koziol, 2019), although the industry says this will fall to become x20 more efficient by 2030 (Orange, 2020).
There is scope to start something in this space and tackle the problem of data center energy consumption. However, there are significant challenges: how to measure the problem, the uncertainty around the current situation, and the limited number of growth buyers, mean that this is a difficult area for startups to tackle.
Originally published at https://davidmytton.blog on September 3, 2020.