Open Source, Open Season on the AI Power Hype
In 1999, Peter Huber and Mark Mills warned in Forbes that “one billion PCs on the Web would represent electrical demand equal to the total power capacity of the U.S.” They were wrong. The internet didn’t fry the grid — it shrunk the need for malls, warehouses, and truck fleets. Power demand in the U.S. and Europe stayed flat for two decades. Today, the internet is under 2% of Western power use.
Fast-forward to 2025: the IEA says global data center electricity demand will more than double by 2030 to 945 TWh (Japan’s entire annual consumption) with AI-optimized facilities quadrupling their draw. Half of that spike? In the U.S.
Sounds scary. But are we just replaying the 1999 extrapolation mistake — assuming the dumbest, least efficient path, and betting against human ingenuity?
--------
The Open-Source Shortcut
China’s AI ecosystem already runs on shared weights and open datasets. Train once, update and fine-tune forever. Model A burns through a trillion tokens; Model B tweaks A with new information instead of re-running the whole furnace. Every downstream team rides the same base model, compounding efficiency gains and slashing aggregate GPU hours.
In the U.S., it’s the opposite: walled gardens. OpenAI, Anthropic, Google all pretrain from scratch. Same capabilities, duplicated compute, petawatt-hours up in smoke.
And here’s the kicker: Bill Gurley and Brad Gerstner just put it plainly on their latest podcast (see graph below) - Chinese open models are hitting 90% of the smarts for 10% of the cost. Mistral in Europe could be on that trajectory too, if it moves beyond open-weighs to full open-source.
Once that’s “good enough” for 95% of commercial workloads, proprietary turns into a luxury brand for edge cases — banks, pharma, defense. Everyone else? They’ll take the 90% model at a 90% discount.
----------
Shifting gravity?
If the world outside of the U.S. starts stacking on Chinese open weights instead of American closed ones, the center of gravity shifts:
• Standards, tooling, and guardrails flow from Beijing, not San Francisco.
• The base-layer innovation loop excludes U.S. players entirely.
And if Chinese labs keep dropping improved open models monthly, the gap closes faster than any frontier lab can widen it. That 10% IQ delta is a melting ice cube.
------------------
The Training Crash → Inference Boom
Training is the Olympic weightlifting of AI — massive energy burn, but rare. Inference is delivery driving — leaner, constant, everywhere. A single saved pretrain can fund billions of extra inferences.
If training collapses into fine-tuning and open reuse, demand shifts hard to inference. But inference is price elastic — it will run on the cheapest, most efficient stack. That’s not a wall of Nvidia GPUs; it’s Groq LPUs, Google TPUs, in-house ASICs, even edge chips.
This kills the “million-GPU megacluster” story Wall Street loves. Frontier training needs dense, liquid-cooled, high-bandwidth GPU pods. Inference? It scales out horizontally, runs cooler, and talks over commodity networking.
Capex pivots from exotic training temples to broad inference fleets. Nvidia’s monopoly cracks; ASIC vendors swarm in.
Result: total AI power demand still grows a lot, but the curve flattens, with infrastructure looking more like cloud + edge ASIC farms than sci-fi GPU cathedrals.
It could look something like this, starting from the assumed AI power demand of ~300 GWh in 2025. Closed-model training grows at 20% YoY vs inference at 10% YoY, so the curves don’t cross – at least not until we’ve solved the energy scarcity problem and graduated from a planetary to a stellar civilization.
---------------------
The Real Risk for Big Tech
Big Tech sees the open wave — they’re betting the “last 10%” gap will stay big enough to justify their bloated cost base.
If they’re wrong, we get a GPU overbuild bubble. Stranded assets. A hype cycle that ends in a whimper, not a bang.
The Value migrates up the stack, ie apps that run on cheap inference compute win. The build-out is steady, grid-friendly, and hardware-agnostic. The apps race could also be won by Big Tech, but the GPU cathedrals won’t be as important anymore.
But then what else to do with all the cash flowing from their core businesses?
------------------
If you’re a founder or investor, the takeaway is simple:
1. Don’t anchor your models to the walled-garden training treadmill — the economics are collapsing.
2. Bet on inference-native apps, efficient chips, and tools that make the most of open weights on the path to full open-source.
3. All solutions trying to get more from existing grids will do well, those betting on GPU cathedrals less so.
And remember: bubbles always look like the future until they pop.