Recently, Silicon Valley investment titan Peter Thiel punctured the biggest unspoken rule in the Valley during an interview—AI chips will eventually become as cheap as cabbage!

If you’re a Nvidia bull, this might leave you speechless.

If you’re an onlooker, you might suddenly see the light.

If you’re searching for the next AI investment direction, this statement could completely transform your perspective.

Nvidia’s profits essentially stem from “monopoly dividends.”

In the current AI landscape, 85% of the money is being pocketed by Nvidia.

Yet, the harsh but true reality is that Nvidia’s massive profits over the past few years haven’t been due to the magical nature of its GPUs themselves, but rather because—there were no alternatives.

The reason lies in CUDA’s ecosystem lock-in, its highly closed software stack, and its unparalleled training efficiency.

The result? Others had no choice but to buy.

This led to inflated GPU costs, with whole-machine prices skyrocketing severalfold, forcing AI companies to “pay tribute.”

In essence, Nvidia, reaping all the benefits, resembles the “Intel + Windows + Qualcomm” of the AI era.

But there’s a critical issue—this structure has never been stable.

Historical turning point: Alternatives are finally good enough

History repeatedly teaches us that when alternatives evolve from “terrible” to “good enough,” monopolies begin to crumble.

And now, AI chips are standing at this critical juncture.

First up is AMD. Its hardware performance has caught up with GPUs and even surpassed them in certain areas.

Many haven’t yet accepted this fact: in terms of pure hardware computing power, in some inference and training scenarios, AMD is no longer a follower but a parallel competitor.

Once the performance gap between its chips and Nvidia’s narrows to within 10%–20%, factors like price, supply, and customization capabilities will become decisive.

And these are precisely where Nvidia is weakest.

Besides AMD, there’s the comprehensive rise of ASICs.

ASICs have one fatal advantage: they’re designed for specific tasks, offering efficiency that crushes general-purpose GPUs.

They boast higher energy efficiency ratios, lower cost structures, and can be customized at scale.

In the past, the industry viewed ASICs as inflexible, but the reality is that the training paradigms for large models are converging, not diverging.

Once workloads stabilize, ASICs will, like mining rigs of yore, reduce general-purpose computing to a niche.

The giants have already voted with their feet

Now, facing Nvidia’s monopoly, the giants have responded.

Google’s Gemini 3 already runs entirely on TPUs.

This means TPUs have matured enough to support core models, with internal validation of their economic viability and stability.

After all, Google wouldn’t bet its most important models on TPUs if they weren’t up to par.

Anthropic’s extensive training and inference also run on Trainium.

As one of OpenAI’s strongest competitors, its models are extremely sensitive to computing power.

The reality is that Trainium has become its core computing source.

Will Nvidia collapse?

As the performance gap between GPUs and TPUs, ASICs, and Trainium gradually narrows, only one thing can happen next—profits will shift from the hardware layer to the application layer.

Just as when PCs ceased to be highly profitable and mobile SoCs became standard, the real money started flowing into applications and platforms. Hardware transformed from a “money printer” into a “tool.”

Of course, Nvidia won’t implode overnight but will quietly unravel.

When growth logic changes and sales volumes and profit margins can no longer sustain highs, capital markets will inevitably reassess its ceiling.

Nvidia’s biggest risk isn’t its competitors but AI itself.

Once AI moves from “compute scarcity” to “compute abundance,” Nvidia will no longer be the sole gateway to the era.

To put it bluntly, Peter Thiel isn’t singing the blues for AI; he’s reminding us that true opportunities never linger in the most crowded spaces.

While everyone fixates on compute power, chips, and GPUs, smart money is already shifting toward applications, ecosystems, and real commercial value.

AI’s first half belonged to Nvidia.

The second half? Not necessarily.

Nvidia’s “Circular Loan” Conundrum?

Coincidentally, just as Peter Thiel’s interview was released, foreign media published an exposé on “Nvidia’s circular loans.”

The article reveals that over the past few years, data centers and cloud service providers, seeking rapid expansion, have innovated a new financing model—”borrowing against GPUs.”

Companies use their high-end Nvidia GPUs as collateral to secure massive loans from banks and private equity firms, then reinvest the proceeds to buy more GPUs, build data centers, or lease them to major clients like OpenAI and Microsoft.

The reason? GPUs are high-value, highly liquid assets that can be treated as “capital” for loans.

A prime example of this model is CoreWeave, which pioneered GPU-backed financing.

By mid-2025, the debt market driven by this model had surpassed $20 billion.

CoreWeave alone accounted for over $12 billion in financing, with debt constituting the majority of its assets.

However, with exorbitant interest rates (reaching 10%) and rapid GPU depreciation, its assets could shrink rapidly.

Specifically, GPUs are rapidly depreciating assets; once older chips become obsolete, their collateral value plummets.

Massive inflows of private credit into the AI debt market carry extreme risks.

At this point, Nvidia becomes the “invisible lifeline.”

This is highly dangerous: if demand falls and AMD’s or Google’s self-developed chips rise, the entire ecosystem could shake.

This GPU debt resembles the “asset bubble + high leverage” combination that preceded past financial crises.

As one foreign journalist put it, “The tech industry has lent more debt than during the dot-com bubble of the 1990s.”

In short, Nvidia has now become the anchor of the global compute power and capital system.

If Nvidia experiences a systemic collapse, it would almost certainly confirm that global AI financialization has failed. A larger economic crisis could be on the horizon.