Which Cloud Computing Service Is the Most Environmentally Friendly?

Alex Poulin
Sustainable Finance
7 min readSep 16, 2020

--

What’s the difference between AWS, Microsoft Azure and Google Cloud in environmental terms? With an investor’s perspective, if your portfolio has tech companies with a SaaS segment looking towards environmental sustainability in mind, it is important to deconstruct which cloud provider offers better environmental performance on two fronts: energy efficiency and Co2 emissions. Depending on the cloud provider, they can at times be mutually inclusive or exclusive. With the onset of AI with companies flocking to the cloud for big data, investors will want to know at least which of the three are the most energy efficient.

Which Cloud provider is the Most Efficient?

Although the cloud computing by definition reduces the usage of necessary servers by mimicking the hardware in software, one must still look to the efficiency of this hardware regardless. The leading metric in determining if servers in data centers are efficient is the Power Usage Effectiveness (PUE) ratio. It calculates the total power for the IT infrastructure divided by the total power used in data centers. A ratio of 1 means for every unit of power entering into your data center, it goes directly into powering the IT hardware. Usually, energy has to be diverted towards cooling servers so it is noteworthy to see which ratios are lowest in terms of energy use. The more efficient a data center is, the more the consumption of energy is related directly to the creation of value by the company.

https://unsplash.com/s/photos/google-cloud Courtesy of@morningbrew

Which Cloud provider is the Most Efficient?

Although the cloud computing by definition reduces the usage of necessary servers by mimicking the hardware in software, one must still look to the efficiency of this hardware regardless. The leading metric in determining if servers in data centers are efficient is the Power Usage Effectiveness (PUE) ratio. It calculates the total power for the IT infrastructure divided by the total power used in data centers. A ratio of 1 means for every unit of power entering into your data center, it goes directly into powering the IT hardware. Usually, energy has to be diverted towards cooling servers so it is noteworthy to see which ratios are lowest in terms of energy use. The more efficient a data center is, the more the consumption of energy is related directly to the creation of value by the company.

Here’s the bottom line when comparing the PUE of Google, Amazon and Microsoft:

- Google posts their ratios over the last few years at around 1.11 and now in 2020 at 1.10.¹ This is the best performer among the three cloud providers.

- Amazon is vague and does not publish the exact PUE. It has as a footnote on one of its sites citing PUE at under 1.2. That was in 2014.² It is unclear of its past or current overall efficiency.

- Microsoft, through its sustainability reports, does not publish its PUE other than citing that it is improving its ratio score.³

- If a portion of data centers are outsourced for each, then its PUE becomes industry standard at 1.67.⁴

The transparency Google displays in its PUE puts in in stark contrast to its two other cloud providers. Google’s performance can be attributed to a myriad of factors. It builds most of data centers with a careful design for efficiency and leverages the use of technology. For one, Google’s Tensor Processing Units accelerate the training of machine learning models with the lowest bound estimates at 10x greater efficiency compared to GPUs.⁵ Google’s brain child Deepmind (pun intended) published an article back in 2016 demonstrating that using AI can reduce energy consumption by 40% and reduce cost. In that article, they claimed that overall, there was a 15% reduction in PUE. However, in Google’s sustainability report — and if we assume AI use in the energy controls in data centers were conducted in 2016 — there is only an overall 0.8% reduction in PUE and not 15% as they claimed when comparing 2015 to 2016 figures.⁶ It’s performance is still noteworthy but caution should be drawn in using AI to achieve energy efficiency because using itself is very energy expensive. Companies may become more efficient with AI, but that allows them to consume more energy. Depending on the energy source, using AI becomes self-defeating.

How Green is Your Cloud Provider for AI?

In terms of Co2 emissions, both Microsoft and Google offset their emissions with renewable offset credits. As a result, they are carbon neutral. A significant footnote must be added in regards to these offsets. The fact remains that they still produce Co2 emissions while creating strong incentives for renewable production. It all depends how one looks at this situation. Depending on the level of impact desired through investing in clean cloud providers, Google or Microsoft can be deemed as 100% clean. Alternatively, their sustainability reports still show their Co2 emissions and one could peer through these reports to determine the least Co2 intensive. This situation is analogous to the European Cap-and-Trade system: as selling credits to more intensive polluters helps offset Co2 emissions of some companies, overall levels do not necessarily decrease proportionally. A study conducted on the EU Cap-and-Trade found that the emissions across the EU only decreased by 3.8% despite the emissions mitigating mechanism.⁷ The impact desired hinges on a micro or macro perspective. Even then, in the micro perspective when observing emissions from one company, if a carbon tax is enacted by governments, both Microsoft and Google would be impacted by the tax given that they directly or indirectly emit GHG emissions. Climate risks with offsets do not disappear.

Why offsets are least effective in terms of true impact is that when buying energy from renewable sources, these producers send their renewable energy through the power grid. The clean energy gets mixed up with the electricity generated from fossil fuels — electrons from both sources do not and cannot get differentiated in the grid. Hence, if we are to find the true value, we must take the Co2 emission coefficient numbers that account for the mix energy sources.

If you plan to invest in one of these three cloud provider leaders for their energy and CO2 efficiency in terms of artificial intelligence, then below I provide a quick calculation on which is the most efficient and least co2 intensive based on this previous article:

# data of electricity and PUE keep in mind for 2017 with tpu and GPU efficiency
pue_average = 1.58
am_pue = 1.19
goo_pue = 1.11
re_emi_coef = 0
emit_coef = 0.954
gpu_eff = 1
tpu_eff = 0.1
# nb of hours to train the model
nb_grid_hyperparameter_searches = 123
nb_of_jobs = 4789
avr_job_length_hours = 52
total_hours_from_paper = 239942
average_watts = 250
total_time_of_RD_deploy = nb_of_jobs * avr_job_length_hours
print(str(total_time_of_RD_deploy) + " " + "hours")
# Co2 intensity of training models when energy is sourced from the grid and taking TPU for google which is 10x efficient
def co2_emissions_from_grid(var1, var2):
energy_consumption_micro = (pue_average*(var1 * var2))/1000
co2_emit_micro = energy_consumption_micro * emit_coef
energy_consumption_am = (am_pue*(var1 * var2))/1000
co2_emit_am = energy_consumption_am * emit_coef
energy_consumption_goo = ((goo_pue*(var1 * var2 * 0.1))/1000)
co2_emit_goo = energy_consumption_goo * emit_coef
energy_from_grid = (pue_average*(var1 * var2))/1000
co2_emit_all_grid = energy_from_grid * emit_coef
return co2_emit_micro, co2_emit_am, co2_emit_goo, co2_emit_all_grid

Here are the results to build a machine learning model from scratch:

  • Microsoft = 93841 tons of Co2
  • Amazon = 70677 tons of Co2
  • Google = 6592 tons of Co2 (accounting for TPU efficiency, if not the number is 65926)
  • For all when energy from the grid = 93841 tons of Co2

From the calculations, developing a machine learning from scratch is best with Google given its TPU’s outperform GPU’s used by the other providers Microsoft and Amazon by a few thousand tons of Co2 emissions. If taken from the grid, there’s little difference except for the enhanced technology Google Cloud can leverage with TPUs. However, returning to the notion of offsets, Amazon’s is further ahead in terms of in-house renewable energy production compared to Google and Microsoft as 40% of Amazon is powered by renewables.⁸

Conclusion: What to do as an Investor Focused on SaaS ESG for the Cloud

The limits of this article are the partial observation of the energy and Co2 emissions for machine learning on these cloud platforms. There are other components to factor into this equation when observing the Co2 emissions from these SaaS providers investors must take note of. It depends if through the perspective one adopts that renewable energy purchase agreements offsetting purchases of fossil fuel electricity does make a business 100% renewable. In terms of energy efficiency and Co2 emissions from a levelled playing field by observing only energy drawn from the grid in the US, we can see Google is unrivalled in both. Then again, Amazon with 40% of renewables and on track to be 100% sourced by 2030, compared to Google’s and Microsoft’s considerably lower renewable energy in-house production share, Amazon could be on track to be the leader. It all depends on the time horizon or impact and risk perspective with respect to offsets an investor chooses. What is certain: these nuances may not be exhibited in the ESG data service providers feed investors. Do not take things at face value.

References

[1] https://www.google.com/about/datacenters/efficiency/#servers “Efficiency — Data Centers.” Google. Accessed September 16, 2020. https://www.google.com/about/datacenters/efficiency/#servers.

[2] “AWS & Sustainability” Amazon. Accessed September 14, 2020 https://aws.amazon.com/about-aws/sustainability/

[3] “Global Infrastructure: Microsoft Azure.” Global Infrastructure | Microsoft Azure. Accessed September 14, 2020. https://azure.microsoft.com/en-ca/global-infrastructure/.https://azure.microsoft.com/en-ca/global-infrastructure/

[4] George, Aaron St. “Uptime Institute 2019 Annual Data Center Survey Results.” Uptime Institute LLC. Accessed August 24, 2020. https://datacenter.com/wp-content/uploads/2019/06/data-center-survey-2019.pdf

[5] Dean, Jeff. “Machine Learning for Systems and Systems for Machine Learning Presentation. Accessed September 9, 2020. https://research.google/teams/brain/

[6] “Google 2019 Environmental Web Report.” Google Sustainability. Accessed September 10, 2020. https://sustainability.google/reports/environmental-report-2019/.

[7] Bayer, Patrick, and Michaël Aklin. “The European Union Emissions Trading System Reduced CO2 Emissions despite Low Prices.” PNAS. April 21, 2020. Accessed September 14, 2020. https://www.pnas.org/content/117/16/8804.

[8] Hoium, Travis. “Amazon Is Now Taking Renewable Energy Seriously.” The Motley Fool. October 05, 2019. Accessed September 15, 2020. https://www.fool.com/investing/2019/10/05/amazon-is-finally-taking-renewable-energy-seriousl.aspx.

--

--

Alex Poulin
Sustainable Finance

Aspiring polymath. Driven by questions and ideas to reduce existential risks.