AI Meltdown

Unless you have been under a rock this week or at a dark room retreat, you have been pelted with headlines screaming about the end of US AI dominance and, accordingly, the end of the world as we know it. The news has certainly had a significant effect on the stock market, with Nvidia (maker of the chips needed to run large language models) down almost 17% since the release.

The TLDR is that a Chinese company released, as an open-source project, an AI model that rivals OpenAI’s ChatGPT-4 in some scenarios and did so with a mere $5.6M investment—pennies compared to OpenAI’s billions.

A Sputnik Moment

This is what we refer to as a disruptive technological advance—where someone comes out of left field with an exponentially more efficient way of doing something that the established technology leaders, including the Magnificent Seven, are focused on. To top it all off, the release came directly on the heels of the announcement by the Trump administration of its $500B (that is half a trillion) AI infrastructure investment called "Stargate." If you remember the Cold War and the space race, you’ll see why people are calling this a ‘Sputnik moment' Like the Russians coming out of the blue to announce they had beaten America by being the first to put a man into orbit, the Chinese have just taken the wind out of the United States' assumed technological dominance by releasing a model for free that rivals our best at a fraction of the cost.

The Elephant in the room

I am just skimming the surface here on the details, as I want the context understood before we begin talking about what I believe is the elephant in the room as it relates to AI and the direction this event may take us. That elephant has two parts—the front and the back. DeepSeek is the front, even though it appeared last.

Deeper into DeepSeek

To understand this, we have to look at exactly how DeepSeek has been able to achieve this disruption. To date, the major investment requirement to develop large language models has been in the training. It takes a huge amount of computing power to do this in data centers filled with high-end AI-specific Nvidia GPUs. These chips are forbidden for export to China—they can only get what are essentially consumer-grade GPUs, the kind everyone has access to. So, they took inferior hardware and performed some very innovative optimizations, essentially getting similar results with decimals less of precision as part of the early training. And then—and here is the kicker—they completed the training via the use of other AI models to fine-tune the model according to human preferences. This kind of inference is commonly known as distillation.

Think of it like this: You build a very sophisticated probe (model) from scratch that is launched (trained) into a closed box (data set) to learn everything there is to know about the box—its contents, dimensions, makeup, limitations—everything. If you ask the probe about the box (you are from the box), it uses your pattern of inquiry to match your question to things it knows about the box that match the pattern. That is basically how ChatGPT works.

What DeepSeek did is they went to RadioShack, used off-the-shelf hardware to assemble a cheaper but cleverly constructed probe, and then had it ask the original probe questions (training) until it knew everything about the box that the first probe knew. This makes training much cheaper, allows the whole thing to work on inferior hardware, and makes the answers feel more humanly relevant.

Crossroads

Of course, I am WAY oversimplifying things here, but this is the gist of it. While it has yet to be fully proven out (can the optimizations laid out in the open-source model be replicated? Can we prove DeepSeek distilled its training using the OpenAI API?), the implications for further development of AI are huge.

Essentially, we are now at a crossroads from an R&D perspective. Does the industry continue to invest in expensive training and development of LLMs on more real-world data, or do companies now take the shortcut approach and train their models on the backs of the work (and investment) that has already been done?

The Other End of the Elephant

The thing that everyone keeps ignoring about AI is that the model is only as good as the data it is trained on. To this end, there are all these posts and headlines about how open and unrestricted DeepSeek is, and the question that immediately arises is—if that is so, then how? Since we know that DeepSeek was trained on other LLMs and we know that this training includes bias and narrative shaping, it makes absolutely no sense that DeepSeek would be less biased or narrative-driven.

In my brief experience with DeepSeek, it isn't. Ask it about mRNA gene therapy (vaccines), and you get the same whitewashed platitudes of ChatGPT. Ask it about political repression by the PRC or genocidal atrocities or anything else politically incorrect according to the current prevailing narrative, and you get the same kind of responses. Which makes sense, as AI is only as good as the data it is trained on. Which leads us to —Stargate.

Stargate

For those who may not have been paying attention when this announcement was made by the President: "The Stargate Project is a new company that intends to invest $500 billion over the next four years building new AI infrastructure for OpenAI in the United States." In fact, $100 billion (disputed by Musk) is being "deployed immediately" by the company, with four 500k sq ft data centers already under construction in Texas. That is a lot of GPUs with a huge energy footprint!

Stargate will be operated by OpenAI, financed by SoftBank, and has OpenAI, Microsoft, Nvidia, Oracle, and ARM as its technology partners. That announcement from Microsoft a few weeks back about the creation of a brand-new AI division is making more and more sense now...

Final Thoughts

So, where does this leave us? The US is doubling down on AI dominance through Stargate, a massive private-public initiative with unfathomable computational power and, more crucially, access to vast amounts of data. But this isn’t just any data—it’s government data.

Think about what that means. If OpenAI already built its models by scraping the internet with impunity, what happens when the next generation of AI is trained directly on classified intelligence, government records, sensitive internal communications and private citizens data? AI isn’t just about generating better chatbot responses; it’s about shaping reality itself—what’s considered truth, what narratives are prioritized, and ultimately, what information is accessible at all.

Meanwhile, China has just demonstrated that a scrappy, resourceful approach can achieve near-parity at a fraction of the cost. The real battle isn’t just about hardware, investment, or even innovation—it’s about control. Control over the models, the data they are trained on, and the perspectives they reinforce.

Both DeepSeek and Stargate signal a new phase in the AI arms race, one where access to data and the ability to shape its interpretation may prove more decisive than sheer processing power. And here we are, caught in the middle, watching as trillion-dollar interests and geopolitical forces collide. The question isn’t just who wins the AI race—it’s what kind of future that winner will create, and whether we’ll have any say in it at all.