This claim about fossil fuels sounds compelling—but it’s misleading. Let’s break it down 🧵
Yes, fossil fuels were ~77% of global energy in 1995 and ~76% today.
But that stat hides what actually changed.
Hydro hasn’t really moved since 1995 and wind/solar came from nowhere.
The key issue: global energy demand has exploded.
So even if fossil fuels stayed a similar percentage, the total energy pie got MUCH bigger.
That means renewables didn’t “fail”—they grew massively, just alongside rising demand.
In fact, renewables are the fastest-growing energy sources in history.
•Solar costs ↓ ~90% since 2010
•Wind costs ↓ ~60–70%
•Deployment has scaled to almost 100% of all TWh growth last year (with new Nuclear)
That’s not “barely a dent.” That’s exponential growth.
The stat also relies on “primary energy,” which is a flawed comparison.
Fossil fuels waste a lot of energy as heat. Solar/wind/geothermal/hydro don’t.
So depending on how you measure, fossil fuels can look artificially dominant.
Look at electricity (where renewables actually compete):
Renewables now generate ~30%+ of global electricity—and rising fast.
Many regions already hit 50%+ at times.
That’s a real shift.
The “renewables are just add-ons” argument is outdated.
The US electric grid hasn’t grown meaningfully in 20 years. Now data centers want to add the equivalent of three Californias to Texas alone by 2030. You can’t just plug that into existing utility ratemaking math. A new paradigm is needed. 🧵
@TKavulla describes the core problem: utility pricing recovers average embedded costs. But every new data center megawatt today costs more than average to serve — materials up 152% since 2019, capital costs rising fast. Under current rules, legacy ratepayers automatically subsidize new data center load. That’s the trap.
The Trump admin’s Ratepayer Protection Pledge (March 2026) names the right goal: AI companies pay for their own infrastructure, full stop. But a pledge isn’t a mechanism. Three mechanisms actually work: open season auctions for grid access, customer-prepaid capital, and Bring Your Own Generation madness.
76 countries are now in “emergency measures” for the oil crisis, up from 55 just six weeks ago. That’s not a statistic. That’s the fastest spreading energy emergency in history. And the number is still climbing. 🧵
What does “emergency measures” actually mean? It’s a huge range. Lithuania is cutting train fares. Australia made public transport free. Myanmar and South Korea are telling people they can only drive on certain days. That’s the mild end.
The middle: Philippines declared a full state of national emergency. Indonesia, Japan, South Korea and India are spending billions on fuel subsidies which should be spent on solar instead. Bangladesh asked businesses to turn off unnecessary lights. Pakistan is rationing fuel.
Bain & Standard Chartered’s 2026 SE Asia Green Economy Report: data centres, EVs & green industrial parks will drive 100+ TWh of new power demand in 3-4 years, requiring $200B+ in investment. $540B in announced green spending is on a credible path to deployment. 🧵
@JavierBlas on Odd Lots today: everyone expected oil at $200+ with 60+ days of Hormuz closed. We’re not there because bypass pipelines, SPR drawdowns, and inventory burns have cushioned the shock. But the biggest surprise? ~5% demand destruction that nobody saw coming.
@JavierBlas Where did that 5% oil demand destruction come from? Not just EVs — it’s price-driven behavioural change, Asia absorbing the sharpest hit, and crucially: countries with existing clean energy infrastructure simply felt less pain. The energy transition was quietly doing work.
America cannot lead the AI race without building the power and computing infrastructure AI requires. On that I agree completely. But the piece gets the diagnosis wrong — and a wrong diagnosis produces the wrong cure. 🧵
The Gallup poll it cites shows a majority of Americans oppose data centers near their home, including 63% of Republicans. The real enemies of AI are electric utilities that use data centers to jack up rates and GW+ developers that fail to engage communities.
The electricity price comparison also misses the structural problem. Investor owned utilities don’t make money when you use what we have already paid for more efficiently. They only make money when they invest new money even if they don’t need to.
.@EpochAIResearch has quietly assembled the most rigorous data set on AI infrastructure in existence. Here's what it tells us about how much compute we need by 2030, how many giant campuses are actually required, and where the real distributed inference opportunity lies. 🧵
The baseline: AI compute stock is growing at 3.4× per year, doubling every 7 months. Training compute for frontier models grows at 5× per year. US AI data center capacity will exceed 50 GW by 2030 — approaching 5% of total US generation capacity.
This is not scary for the grid.
But not all of that 50 GW needs to be concentrated. Final frontier training runs — the kind that require extreme GPU synchrony — represent only ~10% of total R&D compute spend. The other 90% is experiments, fine-tuning, inference, and synthetic data. Distributable. epoch.ai/gradient-updat…
Start with the obvious: data centers need firm power, not really 24/7 as they run at 50% capcity factors.
Solar/Wind are fuel, batteries are really providing the capacity.
So the assumption is gas & nuclear carry the load.
Big Tech energy portfolios tell a more nuanced story.
Amazon & Microsoft: 40+ GW wind/solar each
Google & Meta: ~15 GW wind/solar each
All four signed nuclear deals as well
The nuclear numbers are smaller
In short they are sticking to their commitments to clean energy and buying NG units for "capacity" for speed to power.