One of the things we're not very good at as a species is keeping two thoughts in our heads at once.
I recently attended a session at the Auckland Writers Festival, at which journalist and author Karen Hao spoke about how new AI data centres are drawing on old energy sources, such as coal and natural gas.
She's right, and the numbers bear it out, as you'll see in this International Energy Agency report: over half (56%) of the global electricity to supply data centres is powered by coal (30%) and natural gas (26%). The data centre markets are heavily concentrated in the United States and China; the biggest source of electricity for data centres in the United States is natural gas (over 40% share), whilst in China the data centre electricity supply is dominated by coal (nearly 70% share).
There were audible gasps from the audience last weekend as Karen Hao tore into what she terms the Empires of AI - also the title of her book, which I bought at the festival - not just for their dependency on fossil fuels to power their AI ambitions, but also for their extractive labour practices, their unrestrained access to our data and our culture, and their cool disregard for the mass unemployment that is already a side effect of AI growth.
If you have any serious regard for the state of our planet, you'd be forgiven for boycotting GenAI tools entirely.
And yet...
According to the same IEA report cited earlier, "renewables remain the fastest-growing source of electricity for data centres". Elsewhere, the report notes that "by 2035 low-emissions sources account for over half of the United States’ data centre electricity supply mix", and "By 2035 [renewables and nuclear] together make up nearly 60% of the data centre electricity supply in China."
We should also keep the bigger picture in mind here: the share of total electricity demand used by data centers in the US is only around 5%, and in Europe and China it's even less (at 1.9% and 1.1%, respectively); you can see how these numbers have been tracking since 2020 via Our World in Data.
So really there's nothing to worry about, right?
We're always under pressure to take a stand, and it can feel uncomfortable to believe two contradictory things at once: the growth of AI companies is powered by exploitation and extraction, yet the tools they've built are useful, and the harms relatively small. It's rarely either/or, but an ongoing balance of trade-offs and compromises.
I'm collaborating with a friend of mine to build something new, which we call Blue Dollars. We're still figuring out what kind of product we're building, but we know that AI will play a significant role in the business - in fact, it already does. And so, before even properly launching our product, we're already thinking about how we want to use AI to run our business and shape what we build.
Our first conviction is that frontier-lab AI is commercial infrastructure with shifting incentives, not a neutral tool. We wrote this before I attended Karen Hao's talk, but her uncompromising stance reinforced this conviction.
In practice, this means that we've chosen frontier models over running open-source models locally for the foreseeable future. The alternative would carry stronger privacy claims (the data remains in our control) but produce a meaningfully worse participant experience and an unsustainable engineering load for a two-person company.
That choice obligates us: we will remain transparent with participants about which AI processes their data, what we send, and what's retained where. The commercial incentives of frontier vendors aren't ours; what protects participants from the gap is the discipline we bring to our choices as a company, seeking participant consent at the point where new data is processed or new AI behaviour is introduced, and minimising the collection and processing of user data.
How are you thinking about the use of AI in your work? What are your convictions? Let me know, and take a look at The Pro-Human AI Declaration for some inspiration.