The Subprime Secrets of Generative AI: America’s Next Great Bubble?

America might be the home of the free, but it is also the land of the big short and the long con—a place where ambition masquerades as innovation, and selling a dream is sometimes more lucrative than building one. The economy functions through cycles of excessive confidence, rampant speculation on the commodification of hype, and the relentless drive to create something from nothing. When it works, you get Apple and Tesla. When it doesn't, you get Theranos and FTX—grand visions fuelled by a cocktail of storytelling and creative accounting.

This trick isn't new. The ghosts of Enron, the dot-com bubble, the great 2008 CDO heist, and the $65 billion Ponzi scheme where Bernie made off with investor funds still haunt Wall Street. So, are today's financial and tech elites simply running the same old playbook—packaging cheap tech, selling it as gold, and hoping no one looks too closely?

Now, in 2025, is generative AI shaping up to be Silicon Valley's most successful illusion? One that has captivated investors, policymakers, and the public alike? The fundamental building blocks of AI—open-source software, datasets, algorithms, and compute—seem as abundant and relatively cheap as lead. If that's the case, have American AI VCs discovered Newton's Philosopher's Stone, spinning base metals into billion-dollar, gold-plated valuations?

The Moat Illusion

For years, Silicon Valley has spun a grand narrative: AI is so powerful and revolutionary that only trillion-dollar corporations could possibly build it. They claimed to possess the “secret sauce” - that critical mass of money, intellect, experience, and data no outsider could match. Emboldened by this self-proclaimed exclusivity, these tech oligarchs believed they had engineered an unassailable moat, as wide as it is deep, around their AI citadel.

Then, a teneral AI butterfly flapped its wings in China, and a storm rattled the gilded gates of Wall Street with DeepSeek AI 's $5.6 million bargain bucket deployment of a GPT-4-level model using older Nvidia H800 GPU chips that were introduced almost two years ago. That’s almost a lifetime within tech circles. Shortly after its release on Jan. 20, the DeepSeek-R1 AI assistant — powered by V3 — became the top download within Apple’s Top Free App category and, in doing so, splintered the veneer of Silicon Valley exclusivity. Suddenly, you had to wonder: Is their proprietary AI truly an exclusive club, or just another commodity rebranded and sold at a premium? Perhaps that imposing moat was never there at all.

Has DeepSeek exposed the San Andreas Fault of Silicon Valley's proprietary AI?

They say history doesn’t repeat itself, but it sure knows how to keep a beat, and DeepSeek’s meteoric rise might be the opening snare roll of yet another U.S. bubble about to burst. Its rapid ascent has laid bare an uncomfortable truth: generative AI isn’t the rare, mystical marvel Silicon Valley has hyped. Instead, it thrives on open-source research, standardised techniques, and readily available hardware.

Tomorrow, Jan 29th, and unusually on the very first day of the Lunar New Year, when most Chinese people are off celebrating with family, Alibaba plans to unveil Qwen 2.5, an AI model touted to outperform DeepSeek-V3. The timing alone speaks volumes about the pressure DeepSeek’s success has heaped on both overseas rivals and domestic peers, pushing them to churn out comparable models at breakneck speed. Of course, with such rapid turnaround, it’s hard to maintain that building these systems is extraordinarily costly or time-consuming. Much of it is simply fine-tuning a remarkable computational feat. But is it evidence that computers are on the cusp of sentient thought? Not on your Nelly.

In other words, this is no top-shelf elixir: think less Jimmie’s gourmet roast and more Jules’ freeze-dried Taster’s Choice (full disclosure: I buy pods).

The Playbook Never Changes

To paraphrase Yogi Berra, is this déjà vu all over again?

The mere existence of mortgages didn’t cause the 2008 financial meltdown; rather, it was the result of American investment bankers bundling risky, subprime loans into so-called “high-grade” collateralised debt obligations (CDOs). These products were cloaked in the synthetic complexity of triple-A ratings, which suggested they were nearly as safe as U.S. Treasury bonds. Financial institutions raked in enormous fees for creating, repackaging, and selling them—only to short the very same CDOs, profiting from their eventual downfall. The moral conflict was stark: these banks publicly touted the safety of CDOs while privately wagering bets against their failure.

Is AI following the same arc? If mortgage-backed securities were the illusion of wealth, could ClosedAI models be the illusion of exclusive artificial intelligence capabilities, packaged to seem lucrative, destined for a reckoning?

Every pillar that once propped up US AI stock values seems to be crumbling. Cutting-edge research is freely available, compute costs are falling, and open-source models are rapidly closing the gap on their trillion-dollar counterparts. The fundamental question is whether closed-source AI can maintain market dominance (and valuations) when open-source models are catching up so quickly and right now, its looking increasingly fragile.

A few inconvenient facts:

• AI research is public. Transformer models, diffusion techniques, and reinforcement learning are all documented in open-access papers.
• Performance Gap is Closing – A year ago, GPT-4 was miles ahead. Now, models like Mistral, DeepSeek, and LLaMA 2/3 are getting close, and open-source fine-tuning is accelerating that.
• The hardware gap is closing. Open-source models are optimised to run on older hardware, meaning businesses don't need massive GPU clusters to get good AI performance, and sovereign AI initiatives are gaining ground globally.

Open-source AI is exploding. Models like Llama, Mistral, and Stability AI prove that AI can be replicated at a fraction of the cost.

What Happens When Open-Source Catches Up?

• Proprietary AI Loses its Premium Appeal - If an open model matches GPT-4 in performance and can be self-hosted, proprietary AI’s main differentiator disappears.
• Enterprise Adoption Shifts - Right now, businesses use proprietary AI because it’s relatively reliable and easy. But as open models improve, businesses will start owning their AI stacks, cutting companies like OpenAI out.
• Regulatory Pressure and Costs - Proprietary AI’s closed nature means it takes on more risk (e.g., AI liability, data privacy regulations), whereas open-source AI spreads that across the community.

So, Is proprietary AI Actually in Trouble?

Not yet—but the long-term trajectory is concerning for them. If LLaMA 3, DeepSeek, and Mistralkeep improving, ClosedAI will need something radically better to justify its market values. Otherwise, it risks becoming another walled-garden tech giant fighting an open-source wave—like Microsoft in the 90s or Google vs. open Android forks.

For now, ClosedAI is still ahead, but the clock is ticking. If, once again, something that is increasingly ubiquitous is being sold as cutting edge and exlusive, the only real question is: how long can the illusion hold?

The Unreliable Narrators of American Business

If modern American capitalism excels at one thing, it's making reality optional.

Goldman Sachs proved that even if you can't polish a turd, you can undoubtedly roll it glitter and sell it as a triple, a rated collateralised debt obligation (CDO), to gullible customers while quietly betting against the very same securities. After all, why settle for getting paid once when you can double-dip? Meanwhile, Elizabeth Holmes peddled a "revolutionary" blood-testing device that operated about as well as an ashtray on a motorbike. Sam Bankman-Fried proved that "customer funds" can make for a rather convenient personal piggy bank. Hats off to all involved.

And now? US AI startups, backed by venture capitalists, breathlessy claim their models are light-years ahead of the competition—while quietly licensing open-source alternatives. If the highest levels of business operate with creative storytelling, should it shock anyone that the tech industry is no different?

America has long celebrated the audacity of its business class, but perhaps the ultimate unreliable narrator now sits in the White House. In Donald Trump, the country has a convicted felon as President—a man whose financial statements were once described as unreliable by his auditors, Mazars. Within his first week back in office, he fired, suspended, or otherwise sidelined multiple Inspectors General—federal officials tasked with protecting the public and maintaining confidence in American markets.

If financial markets rely on trust, what happens when that trust erodes at the highest levels? When does deception become standard operating procedure? Could it be that Silicon Valley, Wall Street, and Washington are all selling different variations of the same illusion: a run-of-the-mill product wrapped in layers of branding, marketing, and regulatory theatre, designed to look like something more valuable than it is?

At this point, ClosedAI's most significant innovation could not be in technology— it could be in financial storytelling.

The Inevitable AI Reckoning

Like every speculative bubble before, AI's moment of truth will come if reality fails to match the sentient thought hype. And when that moment arrives, key questions will demand answers:

• Are ClosedAI models really that complex and costly? Or can anyone train and replicate them cheaply?
• Was the moat always artificial? Has branding and PR merely disguised the lack of true technological exclusivity?
• Is open-source AI the great equaliser? Are barriers to entry disappearing year by year (quarter by quarter or month)?
• How do Silicon Valley's ClosedAI giants justify their multi-billion-dollar valuations if an LLM with GPT-4-level benchmarks can be built for $5.6 million?
• If selling a model amounts to merely offering an AI platform, ecosystem, and brand, then aren't we inevitably heading toward a cost race to the bottom.

Billionaire investor Ray Dalio had drawn comparisons between the current AI investment boom and the dot-com bubble of the late 1990s when companies were valued not on profitability but on burn rate and aggressive customer acquisition, no matter the cost. He warns that soaring valuations and rising interest rates could trigger a market correction. Much like the tech euphoria of the late 90s, the frenzy surrounding AI rests on an illusion of infinite growth, where investor FOMO fuels ever-rising valuations. But as history shows, bubbles don't burst when investors stop believing in a technology's potential; they burst when they realise they've been paying fantasy prices. When the dust settles, who will be left standing? I suspect the real winners won't be those hoarding proprietary models but those who figure out practical applications and business models beyond the AI-as-a-hype cycle.

The Final Act

Silicon Valley seems convinced it can financially engineer reality indefinitely—pouring billions into high-burn startups with no viable revenue model, all on the premise that if you squint hard enough, you might make out a unicorn shimmering on the horizon. But do the laws of business, like the laws of physics, bend forever?

DeepSeek's $5.6 million lurch forward may have been the first crack in the fa?ade. The response has been predictable: OpenAI, the maker of ChatGPT, has suggested that rivals—including those in China—are making rapid AI advances using the same open-source foundations it built upon. It's the business-world equivalent of "They copied what we copied!"—a grievance that would be easier to take seriously if the most prominent players in ClosedAI weren't themselves built on the work of others.

More than just an engineering feat, DeepSeek has shattered a fundamental Silicon Valley illusion—that its AI must become as ubiquitous as the dollar and its tech oligarchs as Internationally dominant as the US Federal Reserve.

The tech elite sought to monopolise the web by enclosing it within corporate walled. gardensThe very same technologies that promised to democratize business are now threatening to make it more monopolistic.and now they are attempting the same with AI. But this time, it looks like the genie escaped the bottle before they could buy it and cork it. The illusion of generative AI's exclusivity may be fraying—and when the reckoning comes, AI will likely still be valuable, just not at the fantastical prices Silicon Valley has cajoled investors into believing.

But what stings them the most is that Liang Wenfeng pulled it off for the price of a Malibu condo—displaying a level of audacity that, under different circumstances, would be celebrated as classic American chutzpah.