Sent to 16 readers
Race to AGI Daily Digest - Friday, December 26, 2025
Share:
TLDR
Tech giants are pushing $120bn of AI data center costs into off-book SPVs, signaling that AI compute now demands its own financing structures.
China has formally registered over 700 generative AI models under its CAC regime as part of a broader ‘AI+’ industrial push.
Italy’s new IPCEI program for AI and chips shows Europe treating compute like strategic infrastructure, not just another tech sector.
A viral Salesforce post about regret over AI-driven layoffs underlines how corporate AI bets are still colliding with real-world execution.
The Full Story
Following Monday’s Mango-vs-Sora model war, Tuesday’s world-model bet from AMI Labs, Wednesday’s Grok deployment at the Pentagon, and yesterday’s copyright clash, today zooms out to the question underneath all of it: who actually owns the compute?
Start with the balance sheets. Meta, xAI, Oracle, CoreWeave, and BlackRock are shifting around $120bn of AI data center costs into special-purpose vehicles. That’s not a tweak; it’s a new financing playbook for AI infrastructure, where investors hold the steel and silicon while the labs rent capacity. How the $120bn SPVs are structured ->
On the state side, China just crossed 700 registered generative models under its CAC filing regime, wrapped in an “AI+” push that bakes models into every sector. It’s the regulatory backdrop for the “China’s AGI Champions Head for Public Markets” storyline we’ve been tracking—Beijing is not just tolerating models, it’s cataloguing them. China’s 700-model registry snapshot ->
Europe wants in too. Italy’s launch of an EU IPCEI for AI and chips shows Brussels and Rome treating compute more like energy or rail, with dedicated public money for fabs and clusters. That plugs straight into the “AI Governance Moves from Talk to Law” arc: rules plus subsidies, not just ethics papers.
Then there’s the rhetoric. xAI is now openly vowing to surpass the compute of Microsoft, OpenAI, Google, Anthropic, Amazon, and Meta within five years. That’s the “AI Enters the Defense Stack” story, but in peacetime language: whoever owns the most flops defines the frontier. xAI’s five-year compute pledge ->
Meanwhile, the social feed reminds us the human side is messy. A viral thread about Salesforce regretting firing thousands and backfilling with AI shows how uneven the short-term productivity story still is. Salesforce layoff debate on HN ->
So by Friday, the through-line is clear: models and lawsuits grabbed the early headlines, but the real game is shifting to who can quietly bankroll and govern the next trillion GPU-hours.
Get This Delivered Daily
Join thousands of AI professionals who start their day with Race to AGI.