August is ending and we have a juicy newsletter this week. Last week I mentioned the news about Amazon’s gigantic gas plans. This deserved a dedicated analysis, so that’s what we have today.
If you missed it: Cleanview recently reported that Amazon will build the largest gas plant in US history, at 7.65 GW of capacity, powering the company’s first off-grid data center.
The largest gas plant in the world
7.65 GW is a lot. Looking at GEM data, this is larger than any single gas power station currently operating in the world (or second largest, depending how you count).
The largest single power station now is Sabiya in Kuwait, at 6.8 GW running and another 1.1 GW in construction. The Jebel Ali complex in Dubai scores 8.7 GW in GEM data across several power stations (9.5 GW in Dubai government’s data), if you treat it as one power plant, it’s larger.
If we look at announced / planned / expanding power stations, the Amazon one would be number 3-5 globally, after Jebel Ali, Sabiya, Ghazlan in Saudi Arabia (8.8 GW target capacity) and Taichung in Taiwan (8080 MW target total, but it’s half gas half coal). What distinguishes Amazon’s plant is that it’s not powering a country or a megalopolis, but a single data center.
And one other reflection - it’s not an oil and gas company building it. It’s a publicly traded US tech company. With a sustainability strategy and a 2040 net-zero goal, supposedly already “matching 100% of the electricity with renewable energy”. With that gas plant, it’s going to be a lot of matching…
A single data center burning more gas than a country
The capacity of Amazon’s facility is one thing.
But most gas power plants don’t run 24/7. Their utilizations (or capacity factors) vary, the US fleetwide average for CCGTs was 58% in 2025, 14% for OCGTs. In an off-grid data center, the plant would have to run close to 100%.
The permit describes 5 GW of nominal output for the 7.65 GW nameplate capacity, I guess 2.65 GW would be reserves. It’s 35 turbines of around 200 MW, if they’re OCGTs with 40% efficiency (likely in the beginning), the plant would consume 11 bcm of gas a year (at 100% load). For CCGTs with 60% efficiency this drops to 7 bcm. A disclaimer - there’s also some solar and battery storage onsite, but that’s like 3% of the power generation.
Let’s compare the gas consumption to European countries.
In the OCGT scenario, 20 out of 27 EU countries consume LESS gas in a year than the Amazon plant will. And that’s in the whole economy, not just in the power sector. Among them are Czechia, Hungary, Austria, Greece with roughly 10 million populations, or Romania with 19 million. In the CCGT scenario, the Amazon plant is still comparable to Czechia on gas consumption…
Running it in Europe would cost EUR 4-7 billion a year
Now the really juicy bit - costs. I know in theory data centers don’t care about power costs, because they generate so much revenue from compute and the capex is so much higher than the opex.
But still, if the plant sat in Europe, at the current TTF gas price of around 60 EUR/MWh, Amazon’s annual gas bill would land between EUR 4 and 7 billion.
Yes, I know it’s Texas and Waha gas is 5 times cheaper.
But here’s an interesting statement - an Amazon spokesperson said the site will be “powered by new on-site generation that won’t raise electricity costs for Texas families”. If you consume all the stranded Permian gas, Waha gas prices will go up. Which will translate to higher electricity costs in the area anyways…
AI reducing emissions is looking increasingly tricky
Finally - emissions. Nobody seems to care anymore, but some folks keep claiming AI will reduce emissions - e.g. by optimizing grids or better forecasting. Sure, those are positives, and AI can have many benefits.
But AI will not reduce emissions if we run it on gas.
A recent Nature paper shows that quite clearly - at current pathways, AI adoption adds more fossil fuel emissions than clean power related reductions. And that’s before we account for giga-projects like the one we’re discussing today.
The point of this newsletter is not to complain about data centers. But this is clearly not the way to do things.
I wrote about the xAI Colossus project a while ago - and the gas route that led to a lawsuit.
The backlash against data centers in the US is growing every day and Europe should learn from those mistakes. And, as I keep showing, it doesn’t have to mean we quit the AI race altogether - there are clean power + storage + flex + distributed AI options that can be deployed faster than buying 8 GW of gas turbines...
P.S. I’ve decided to leave Ember. Will share more on the decision next week and a bit of what’s next :)
i don’t buy there anyway.
Oh
What will you do now?