7 Reasons 'DeepSeek Moments' Pour Gasoline on the AI Progress Fire
Using less is good, using more is better
1. DeepSeek Was a Signpost, Not a Summit
When DeepSeek unveiled its open-weight frontier model in January 2025, markets whipsawed and pundits shouted “peak AI.” Yet the release was just another dot on the same exponential cost-performance curve that’s driven every major leap since ENIAC.cset.georgetown.edu
2. Scaling Laws Still Pay Rent
DeepMind’s Chinchilla paper showed that shuffling the mix of parameters and data can squeeze far more capability from the same FLOPs.arxiv.org DeepSeek rode that wave, and so will the next dozen contenders. Efficiency gains don’t cap demand; they slash the effective price of computation and invite us to use more of it.
3. Jevons Paradox, Silicon Edition
Enter Jevons: make a resource cheaper and society consumes even more. The same 19th-century logic that wrecked coal forecasts now wrecks “AI will plateau” takes.news.northeastern.edu Every token that costs less to train or serve simply fuels bigger models and entirely new workloads.
4. Watches vs. Moonshots—75 Years of Escalation
A modern smartphone outguns the Apollo Guidance Computer by roughly five orders of magnitude.independent.co.uk If raw silicon were the limiting factor, we’d have hung up our GPU spurs long ago. Instead, each hardware breakthrough widens the ambition envelope—from guidance systems to GPT-class reasoning, from Pong to photorealistic open worlds.
5. Pocket AI Won’t Kill the Data Center
Sure, your 2027 phone will host a 10-billion-parameter personal assistant offline. Fantastic. Meanwhile:
Drug discovery teams now spin up generative chem-bio models that demand racks of H100s and Blackwells.intuitionlabs.ai
Enterprise agents will crawl entire corporate memory graphs, running month-long simulations that dwarf anything a handset can stomach. We are not going to shut down the data centers.
6. The Build-Out Numbers Are Going Vertical
$7 trillion: McKinsey’s estimate for global AI-compute capex this decade.mckinsey.com
+165 % power draw: Goldman Sachs’s forecast for data-center electricity demand by 2030.goldmansachs.com
10 GW: new capacity breaking ground this year alone, per JLL’s 2025 outlook.jll.com
If demand were about to taper, hyperscalers and real-estate lenders missed the memo.
7. Bigger Keeps Being Better
History’s lesson is blunt: larger compute pools unlock qualitatively new things—weather models, protein folding, autonomous supply chains. The AI era is just pushing that pattern faster. Every “DeepSeek moment” will look seismic on day one and routine by day 180, clearing runway for the next jump.
Bottom line: Betting that compute demand will flatten is effectively shorting human ambition. Good luck with that.
