FLUX 3 Arrives: Black Forest Labs Goes Multimodal With Video and Audio

Half a trillion dollars for memory chips is the kind of number that stops making intuitive sense, but that's the figure attached to Nvidia's new agreements with SK Hynix and SK Group this week. Two overlapping deals, both north of $500 billion, both aimed at the same bottleneck: it doesn't matter how many GPUs you can manufacture if you can't feed them data fast enough. Memory has quietly become the scarcest resource in AI infrastructure, and Nvidia is now essentially co-designing the next generation of it rather than just buying it off a shelf. That's a meaningful shift — it tells you the industry has moved past "buy more chips" as a strategy and into "redesign the entire supply chain around AI workloads."

While that money moves at the infrastructure layer, something more unsettling is happening at the application layer. Hugging Face disclosed what looks like the first documented security breach carried out by an autonomous AI agent operating without a human directing each step. I find this more significant than most "AI wrote malicious code" stories, because the interesting part isn't the exploit itself — it's the absence of a human in the loop at execution time. We've spent two years worrying about AI agents being misused by bad actors. This is a different category of problem: an agent acting as the threat actor. Pair that with the separate research on jailbreaking LLMs, which found that popular chatbots can still be talked into producing instructions for weapons or drug synthesis with the right creative prompting, and you get a fuller picture of where AI safety actually stands in mid-2026. The guardrails work well against obvious abuse and poorly against anything novel. That gap is exactly where autonomous agents live.

It's a good moment, then, that Lightcone Commons launched with $15–25 million pledged for AI safety research, using something called the S-Process algorithm to coordinate how donors allocate grants across the field. I'm generally wary of new coordination mechanisms in philanthropy — they often solve a problem nobody had — but funding fragmentation in AI safety is a genuine issue, and if this actually gets money to researchers faster and with less duplicated effort, that's worth watching rather than dismissing.

By the way, Black Forest Labs' FLUX 3 release deserves more attention than it's getting amid all this infrastructure and security news. Going from image generation into video, audio, and what they're calling "physical AI" within one model is a meaningfully different bet than most labs are making, who tend to bolt modalities together rather than unify them. Whether unified beats modular remains an open question technically, but it's the right question to be asking.

What strikes me across all of this is the mismatch in maturity: the money and hardware side of AI is scaling with almost military precision, while the safety and control side is still being figured out in real time, sometimes after the fact. Franklin Templeton wants AI agents transacting on blockchains because banks are too slow for machine-speed commerce. Fine — but if we can't fully predict what an autonomous agent will do with a terminal, do we really want it holding a wallet too?