More AI, More fun— is a new climate crisis coming?
Global leaders just concluded the United Nations Climate Conference (COP26) last November. If there’s one thing you can be sure of, it’s that there will be a lot more COPs happening before any noteworthy action is seen on the ground.
The unfortunate part is that the powers that be had hardly taken any serious action on the climate front until recently, when hurricanes, storms, and freak weather in the past few years raised environmental alarms across the globe.
Numerous well-documented factors worldwide have contributed to global warming. The majority of pollution is a result of burning fossil fuels, which also cause industrial and vehicular pollution. In addition to this, pollution from animals and our everyday white goods — like refrigerators and air conditioners — also exists.
To be accurate, we have built up a comfort economy premised on disrupting the planet’s fragile ecosystem!
Multiple COPs and several previous protocols aimed to shift us from fossil fuels to more environmentally friendly alternatives.
We believed that it would set us on a much-needed correction course.
We believed that disaster would unite us to address the most pressing global problems.
Just as it appeared that the world had agreed in principle after decades of wrangling with big oil, there was some kind of consensus: then, without any warning, a new combustable element gets added to an already polluted atmosphere — silicon.
The Real AI Revolution
Although AI has been with us for quite some time now (recall autocomplete from Google and facial detection from Facebook appeared more than 10 years ago), it was Generative AI (GenAI) that put this power in the hands of the masses.
It was ChatGPT, the poster boy of GenAI, that brought the magic of AI to ordinary people via a simple prompting interface.
However, with great power also comes greater environmental responsibility.
The GenAI revolution was sped up by the availability of GPUs (Graphical Processing Units). GPUs are special-purpose processors that process instructions needed for AI models quickly. They do it at scale and at a cost, both monetarily and environmentally.
Here’s why this matters.
Large language models (LLMs) like ChatGPT are trained using GPUs in large-scale server farms, which number in the thousands.
GPUs require significantly more energy than CPUs. The figure below displays the energy consumption of some of the popular GPT AI models in a study conducted by Cornell University.
Now, lets for instance examine the top 7 GenAI firms in the list below. (Note that this is clearly not an exhaustive list)
You can take away a few points from the table above.
- Public cloud providers like Micosoft and Google probably own or have a stake in the biggest models around today, which may not always be the case.
- Huge giants like Meta and Baidu have their own data centres to train their own foundational models, such as Llama and Ernie.
- Smaller players such as Anthropic and Stability AI cannot afford their own data centres and therefore train their models on public clouds such as Google or AWS.
As you can see, the higher the demand for AI, the more firms would want to jump into the game. This would result in an insatiable demand for chips, an enormous demand for the supporting infrastructure, and an unending need for unlimited power.
Infrastructure — where will it come from
The cost of training huge models with more than 100 billion parameters is costly and time-consuming. Which is why smaller AI vendors turn to the top three cloud providers to train their models (see table above).
Cloud providers like Amazon (AWS), Microsoft Azure, and Google have data centres that span the entire globe. Their reach is breathtakingly enormous, as is their political clout in so many places.
These cloud providers have, after all, invested almost a decade or more in establishing the needed infrastructure (power, servers, networking, etc.) to run their centres efficiently.
In fact, one of the three cloud providers, Google, prides itself on being the most environmentally friendly among the other cloud providers on the planet. Wind turbines partly power most of their data centres.
As easy as it may sound, the top three cloud providers cannot handle the upcoming onslaught of demand for more infrastructure at planet scale.
Infrastructure — the big ramp-up
It has already started happening.
You just need to follow the news on AI, and not a day goes by without companies spending more on the cloud or on their own infrastructure for AI.
New firms from countries where millions of netizens reside are now challenging the semiconductor fab manufacturing dominance once held by the likes of TSMC and Samsung from Taiwan and Korea.
Both India and China, for instance, are now pumping millions, if not billions, into building fabs, completely aware that it may take a few years for them to attain the economies of scale that firms in Taiwan and Korea have.
We should remember that most of the countries where fabs are being set up are signatories to the Paris Climate Agreement, which means they have an obligation to ensure that firms setting up shop abide by these rules.
In fact, a study done by Mckinsey shows that most of the semiconductor giants have already been working on such measures, to reduce their carbon footprint, (which is large for fabs)
Chips are not the only infrastructure bottleneck in play!
Last week the Cloud Leader AWS announced that it was buying a data centre Talen Energy that is powered by nuclear energy.
That Amazon chose to do this, also probably sends out a signal that it is serious about meeting its renewal energy commitments by 2025.
It remains to be seen, however, whether nuclear energy becomes the mainstay of cloud providers going forward.
As private companies now stampede for the AI gold rush, governments must also hold firms accountable for environmental sustainability in the early stages. Once an industry has matured, reversals are normally difficult, if not impossible. (Conversion to electric or hybrid cars, for example.)
What good is a world without diesel or petrol cars polluting the skies, but smoky data centres and polluting fabs doing the damage?
