In a debate over projected 2027 AI infrastructure spending—with the US eyeing around a trillion dollars versus China's roughly 123 billion—Hugging Face's CEO points to open-source practices as a way for teams to share compute resources and achieve frontier results far more efficiently.
How? Instead of closed-source frontier labs running the same training runs in secret and siloes, open science and open-source AI allows them to mutualize spending and compute making them an order of magnitude more efficient!
🇺🇸 The United States will spend $1 trillion on AI in 2027.
🇨🇳 China will spend $123 billion and still ship frontier models.
How?
Delangue's order-of-magnitude claim draws pushback as motivated advocacy, with no fresh empirical backing available for frontier-scale open-source training multipliers.
Analyst projections of hyperscaler outlays underscore why arguments about avoiding redundant closed efforts resonate, even as specifics on mutualization benefits stay unresolved.
Many users endorsed open-source AI mutualizing compute to cut frontier model costs, praising efficiency gains, strategic collaboration, innovation benefits, and user control over closed silos.
No Digg Deeper questions have been answered for this story yet.
@ClementDelangue Same logic holds past training runs too. Pool idle inference and fine-tuning capacity across a network instead of one lab cluster and you get a similar order-of-magnitude gain, just distributed.
@ClementDelangue this doesn't seem like a good explanation Clem. in fact it looks like motivated reasoning and talking your book which I don't usually expect from you
How? Instead of closed-source frontier labs running the same training runs in secret and siloes, open science and open-source AI allows them to mutualize spending and compute making them an order of magnitude more efficient!
@ClementDelangue Elon has also explained that it's engineering vs. science types of people.
@ClementDelangue $1T just to run the same experiments in secret.
@ClementDelangue I don't want to discount their ingenuity or the fact that they're piggybacking, have to be honest about both in this discussion right?
@ClementDelangue open source > closed silos, efficiency wins
@ClementDelangue if satya really believes in a vibrant ecosystem of customized models, then he should open source mai. that model isn't making them any money anyway, but open sourcing it could drive huge adoption and keep microsoft at heart of enterprises
@ClementDelangue What is preventing closed source labs from monetizing open source science and AI?
@ClementDelangue The spending stays within the walls of the US.
The knowledge also since people are free to move.
@ClementDelangue What I don't get is what financial gain does the Chinese companies have to open source their models?
I get your point but there’s nuance.
A lot of spend is on compute infrastructure which the US has the most share by a wide margin.
For actual labs, the experience is similar to pharma where the US spend 1000x others on drug discovery and others benefit.
Right now frontiers spending that on training and OS models get to distill and leverage.
@ClementDelangue one slice of that trillion: $100m+ to find 10,000 bugs in open-source code everyone depends on. the capex war isn't just about who trains the biggest model
@ClementDelangue Any projections on OpEx and ROI?
@ClementDelangue This should be the opening argument for the runners-up (UK, Europe, etc...)
You can’t afford the capital inefficiency of closed competition. Race toward leverage in an open ecosystem. If you lower the cost of intelligence faster than first movers can recoup capex, you can win.
@ClementDelangue easy pre training apps are needed
and training of adapters too output fellowers lisense (adapters can be gpl kind of if base is gpl or something like this)
@ClementDelangue secrecy gets expensive when everyone is solving the same bug
@ClementDelangue the open source multiplier is real, but the coordination cost is the real bottleneck nobody talks about
@ClementDelangue Or, hear me out, distillation?
@ClementDelangue Production Quality > Money
@ClementDelangue The mutualization logic is perfect though: Meta pays billions to train, you deploy free. That's definitely more efficient. For you.