Daily AI News · Oct 31 Midday Edition: Nvidia Backs Loans & DeepSeek's Money Printer

Lots of new moves in AI today — the words that keep coming up are capital and compute. Today we're covering a few things: first, Nvidia backing data centers; second, Salesforce acquiring Rune; third, the RustCUDA meetup; fourth, Ollie shutting down; fifth, Thunder Tubes; sixth, the DeepSeek inference engine payback math; and seventh, that crazy expensive touchscreen MacBook. And at the end, a little side note about an open-source game model. Yep, that's a lot to cover. Alright, let's start with the most explosive one. First up, I just saw on TechCrunch that Nvidia is considering co-signing loans for Nscale's data centers. That's a heavy signal. You have to understand what co-signing means first. Simply put, if the borrower defaults, Nvidia has to pay back part of the money. Wait, so what's in it for them? Isn't selling chips enough? That's not the point. The point is the financing side. Investors are a bit spooked by AI spending right now. With Nvidia's backing, financing costs go down, and projects are more likely to get off the ground. So it's like extending the risk exposure from the supply chain to project financing. Exactly. It ties the chip supplier and the compute operator even closer together. But details like loan amounts and installation scale haven't been disclosed, and participation is still being weighed. Second, let's talk about Salesforce — they're at it again. Reports say they're buying defense tech startup Rune for $375 million — it does AI logistics. The logic is pretty clear — they want to fold it into their Missionforce defense platform. Military logistics is way more complex than civilian. The bar for reliability and mission execution is on a whole different level. So this is about filling capability gaps, and it shows capital is heating up in defense tech. Yeah, Salesforce's acquisition pace has been really dense lately. Third, let's switch to a tech flavor. There's a livestream on Dev Community called CUDA Rust Meetup. Rust writing CUDA — that combo is interesting. They're trying to nail both safety and performance. The developer community is definitely paying more attention to this direction. Fourth, here's a sad one. Just saw on TechCrunch that privacy AI assistant Ollie is shutting down. Shutting down within a week, and the founder didn't disclose financials or user data. The article explains it well — it wasn't a tech failure, but a mismatch between business model and market timing. Oversupply diluted user attention, and prioritizing privacy meant giving up data monetization. Plus, with big platform ecosystems bundling everything, it's hard for independent products to compete on price and iteration speed. For privacy-sensitive users, migrating and rebuilding trust both come at a cost. Fifth, there's also a video called Thunder Tubes — let's note it for now and look closer later. Alright, that's the news for now. Next, let's talk about today's hottest topic. A video from a French blogger blew up on Bilibili — the title literally says America is still burning cash while China's AI is printing money. The core is one sentence: DeepSeek's open-source inference engine pays back a single GPU in two months. Two months to break even? How do you even calculate that? Do you buy it? Don't rush to believe it. The assumptions behind that math are key — you have to see how they factor in GPU utilization and throughput. If the inference engine really maxes out per-card throughput, the payback period can definitely shrink a lot. But that's contradictory, isn't it? If it's open source, anyone can run it — wouldn't profits get flattened? It's the opposite. The value of an open-source inference engine isn't in selling licenses — it's in lowering inference costs across the whole industry. Once costs drop, apps that couldn't run before can now run, and demand actually expands. So American peers are scared not because of the tech itself, but because of this cost structure. Right — burning cash to pile up compute versus using software to squeeze every drop out of compute are two completely different paths. But I think this huge-profit math is a bit clickbait-y — two months per card depends on the specific scenario. I agree there — the numbers in the video are more like an ideal upper bound, not the norm. But the direction is right: lower inference costs mean agent-type products will spread faster. To sum up: what DeepSeek is really attacking this time is inference cost, not parameter scale. One more thing — YouTube is talking about a $14,549 touchscreen MacBook Pro. At that price, regular developers can basically only look at it. But the fact that it's trending shows people still care about local compute and where the limits of high-priced hardware are. Alright, that's it for today. We'll keep watching the capital and inference cost storylines. See you tonight.