Lumian Gen AI Newsletter Issue #58

Welcome to the 58th edition of the Lumian Gen AI Newsletter!

DeepSeek—the once-obscure Chinese AI startup—shocked the world last week when it claimed it trained an advanced AI model, DeepSeek-V3, for under $6 million. At first, it sounded like wishful marketing. Then, over the course of a single Monday, Nvidia’s stock plunged 17%—erasing $589 billion in market value, the largest one-day wipeout of any company in history. Suddenly, people began asking whether these new models truly need the kinds of billion-dollar data centers we’ve been told are essential for advanced AI.

The irony is that this massive wave of doubt arrived just as the American tech sector was buzzing about Stargate, a rumored $500 billion investment partnership among OpenAI, SoftBank, and MGX. The premise behind that deal is straightforward: with enough servers, you can dominate AI. But DeepSeek’s success—20 to 50 times cheaper inference than certain U.S. counterparts—suggests software tricks and more efficient approaches might turn that logic on its head.

The tension is straightforward. On the one hand, SoftBank and other mega-funders want to deploy enormous sums into AI because they believe it could be the next trillion-dollar industry—think self-driving, augmented reality, or anything else that once promised to reshape the world. On the other hand, DeepSeek is showing that maybe you don’t need a gargantuan budget if you have the right software tricks or enough open-source savvy. That possibility has rattled hardware suppliers like Nvidia, which thrive on the narrative that AI requires limitless GPU spending.

This all might feel like déjà vu if you’ve followed tech cycles before. A starry-eyed newcomer arrives with a “breakthrough,” triggers an investment frenzy, and then the big incumbents—Google, Meta, Microsoft—quietly keep pouring billions into R&D. Meanwhile, SoftBank joins the party with an outlandish check, and everyone wonders if it’s time to bet big or head for the exits. The difference now is how quickly these memes and markets move. Thanks to viral hype, a single lab’s announcement can evaporate tens of billions in market cap almost overnight.

But does that mean we should worry? Well, not in the sense that DeepSeek’s cheaper training pipeline is about to topple the entire AI infrastructure market. Most likely, we’ll see a repricing or a reevaluation of which parts of AI need massive capital. Training a big model from scratch can be cheaper if you’re clever with data, open-source tools, or specialized optimizations. Yet scaling that model to serve millions of daily users still demands an epic amount of compute. The real power players—Google and Meta—already own giant data centers and can amortize those costs over an entire ecosystem of products.

In that light, Stargate’s rumored $500 billion might not be as absurd as it sounds—at least for the corporate backers who see it as a longer-term bet on AI’s infrastructure future. Even if smaller labs demonstrate new efficiencies, they don’t eliminate the underlying resource demand. They might reduce training budgets here and there, but as soon as DeepSeek or any other newcomer needs to handle a billion queries a day, it’s going to face the same massive hardware bills everyone else does.

That’s where the idea of tech as deflationary comes in. Software is cheap to copy—once a model’s built, you can spin up replicas with relative ease. And open-source communities tend to reverse-engineer breakthroughs quickly, making it even cheaper. From a macro perspective, that means you can’t just pour infinite cash into AI and expect guaranteed profits. Sooner or later, every new technique propagates, and returns on giant capital investments might slip. But from a practical standpoint, “deflationary” doesn’t mean “no money is needed.” It just means the capital requirements evolve and may focus more on inference (serving model outputs to users) or specialized hardware, rather than pure R&D blowouts.

If you’re betting on AI in 2025, you still can’t go too wrong with the incumbents (not investment advice). Google, Meta, Microsoft, and Amazon are the safest plays because they’ve got user distribution, data-center heft, and strong AI talent. If you’re looking for the next upstart, be aware that most smaller labs end up partnering with or selling to a bigger platform. Even OpenAI, the darling underdog, leans heavily on Microsoft’s cloud muscle.

In the end, DeepSeek is a great reminder that AI hype moves lightning-fast, but big budgets and distribution matter more. Yes, the news cycle can cause Nvidia to lose tens of billions in market cap overnight and then gain it back if the next headline says we actually need more GPUs than ever. No, it doesn’t mean AI is a bubble about to burst. It just means we’re in a phase where technology and massive funding dance around each other, and occasionally the music gets so loud, the floor shakes.

That’s the heart of this story: with AI, you have big believers writing enormous checks, lean newcomers proving it can be done cheaper, and giant incumbents unfazed because they own the infrastructure everyone else eventually needs. This cycle repeats: a meme arises, valuations spike, a new model appears, valuations wobble, but the fundamental advantage still lives with whichever player can distribute the product at scale. If you like making bets on hype, there’s plenty of that. But if you’re looking at long-term viability, watch the big platforms—and keep a side-eye on SoftBank’s next dramatic announcement.

Happy reading! 📚🤖🎵

In this week’s issue:

• News Flash: Stargate, Meta’s $65B, Perplexity bids TikTok

• AI Frontier: AI Social Media Growth and Automation Tools you can use today

• Fundraising: The biggest deals in AI

• Nerd Out: Technical and Business Content for Everyone

🚀 AI in Practice

Cutting-Edge AI Social Media Growth and Automation Tools You Can Use Today

• FeedHive – AI-powered social media scheduling & engagement

• Tweet Hunter – AI-driven Twitter/X growth and automation