Is AI in a Bubble? Breaking Down the Bear Thesis

FUD has permeated the market and “gone viral,” especially around the AI sector. The bear thesis for AI has gained strength. But is it fundamentally sound?

Sam Altman (CEO of OpenAI):

“Are we in a phase where investors as a whole are over-excited about AI? My opinion is yes.”

Jeff Bezos (Founder of Amazon):

At a tech event, he described the current AI surge as a “good kind of industrial bubble”:

“This is kind of an industrial bubble as opposed to financial bubbles … the ones that are industrial are not nearly as bad, they can even be good.”

Ed Zitron (Tech writer and critic):

Declares the 2020s AI boom a bubble and predicts it will pop.

Michael Cembalest (Analyst):

Describes today’s AI boom as a “vendor-financing circle” — capital raised by one set of companies is recycled through others, raising concerns of financial engineering rather than innovation.

A Bank of America Global Research survey found that 54% of fund managers believe AI stocks are in a bubble, versus 38% who do not.

Clearly, influential voices see bubble dynamics forming.

Is AI a Bubble?

Let’s break down the causes of fear, skepticism, and the foundation of this bear thesis.

I. What is a Bubble?

A bubble occurs when an asset, or an entire asset class, appreciates far beyond what is justified by its underlying economics. For businesses, this happens when expectations of future cash flows diverge widely from what the company can realistically deliver within a reasonable timeframe.

So, are some AI businesses in bubble territory, for sure they are. Some, just as always, won’t produce the necessary economics to justify their valuations or even their existence. Is this an issue?, absolutely not, this is the way that entrepreneurship and business reality works, and has worked for a while.

The real question isn’t whether some companies will fail, some will, but whether the entire AI sector is structurally overvalued, and whether the specific companies you own are driven by hype or by durable economics.

II. The Three Elements Behind the “AI Bubble” Narrative

1. Adoption & economic impact

2. Financing of the buildout (vendor financing and circular capital flows)

3. Economic feasibility (profitability/scalability of AI data centers, especially NeoClouds and GaaS)

1) Adoption & Economic Impact

The Dot-Com Analogy (and Why It Breaks Down)

Many arguing that “AI is a bubble” reason by analogy to the 2000 dot-com boom, saying “history rhymes.” But the analogy breaks once you examine what’s actually happening.

During the dot-com era, infrastructure came first and demand lagged by nearly a decade. Massive over-investment in fiber created “dark fiber,” and plenty of dot-com businesses never turned a profit. Adoption was slower than expected, use cases immature, and real economic gains modest.

The key difference in the AI revolution is actual demand, driven by faster adoption and clear economic benefit. AI is being deployed at an unprecedented pace, and in many cases it delivers immediate, measurable returns.

I don’t want to dismiss concerns about the technology itself; some are valid. There are legitimate questions about the ultimate capacity and scalability of AI models. If AI proves less transformational than expected, or if models hit a scaling wall and fail to reach the required capabilities, the industry could face real headwinds.

Adding to the debate, Andrej Karpathy recently discussed these issues in a podcast, questioning current model architectures, their economic impact, and the assumptions around progress toward AGI.

I won’t touch on every point here, but after watching the interview twice, I don’t think Karpathy is pessimistic about AI itself. He is skeptical of the idea that AGI will arrive tomorrow, replace humans, and instantly drive 20% GDP growth.

He highlights model limitations, the breakthroughs still needed for AI to become more broadly useful, and the fact that humans do not want or need AI to replace us; we want to work with it.

Karpathy’s view is that AGI may be a decade away, not a year.

Many took his comments as “the AI-bubble-bursting podcast,” much like the overreaction to DeepSeek’s efficiency breakthrough. In reality, his perspective is bullish: it reflects a field that is maturing while still driving rapid innovation. And it doesn’t change the fact that AI adoption keeps accelerating, with a growing share of the global economy already running on it.

Adoption and Economic Benefit Are Already Here

Adoption is broad, measurable, and economically meaningful. The largest AI spenders, including Meta, Microsoft, Amazon, and Google, are integrating AI into the core of their business models. Their combined CAPEX over the past twelve months exceeds about 291 billion dollars, against roughly 359 billion in combined net income. At that scale, there are no structural red flags.

We can also look at current use cases and see tangible, often immediate, economic rewards:

Consumer and enterprise software: search, ranking, recommendations; ad targeting; copilots improving output and reducing labor hours.

Commerce and logistics: Amazon’s recommendations, robotics, and forecasting; retailers’ dynamic pricing and shrink reduction; route optimization cutting fuel use.

Finance and insurance: fraud detection saving billions; JPMorgan’s COiN platform saving hundreds of thousands of legal hours; instant underwriting and claims.

Healthcare and life sciences: radiology-level models; accelerated drug discovery; documentation copilots freeing clinician time.

Energy and industrials: predictive maintenance; digital twins; grid forecasting for renewables.

Media and entertainment: generative tools reshaping creative workflows; procedural content in gaming.

Public sector and defense: cybersecurity, satellite analysis, decision support.

So yes, there are legitimate questions about the limits of current AI models, but there should be no doubt that today’s systems already generate real, tangible economic value for the companies that know how to deploy them effectively.

2) The “Circular Money Loop”

Some critics argue that Nvidia and other chipmakers are fueling a “circular money loop.” The claim is that Nvidia invests in AI startups such as OpenAI, xAI, and Anthropic, and those startups then use that capital to buy Nvidia GPUs, inflating valuations and repeating the cycle.

The key question is whether this is a closed loop or an open system. The main rebuttal is that real cash flows enter from outside. It is similar to an automaker providing financing to a dealership: if there are no customers, the money just circulates inside the loop. But once customers buy cars, external cash enters the loop with genuine demand. The same logic applies here. AI startups are not buying GPUs in a vacuum; they serve clients, run workloads, and generate revenue that flows back into the system.

Cross-investment between hardware, software, and cloud providers has long been part of platform industries. Apple and TSMC collaborate to secure advanced process nodes. Microsoft’s partnership with OpenAI guarantees Azure utilization. AWS co-funded startups that anchor workloads on its cloud. Even in traditional manufacturing, automakers and equipment makers often support downstream partners to accelerate distribution.

Nvidia’s ecosystem strategy follows the same pattern. By funding companies that build products and services around its GPUs, it strengthens the broader software stack, including CUDA and TensorRT, while seeding workloads and data-center capacity for future hardware. This approach echoes Intel’s past work on compiler optimization and AWS’s early support of AI startups that later became core customers. The goal is straightforward: accelerate real demand that would have emerged naturally.

Markets, as George Soros noted, are reflexive. Feedback loops exist, but not all are harmful. Healthy reflexivity amplifies fundamentals: better GPUs lead to better models, better products, more users, higher revenue, and more compute demand. Unhealthy reflexivity fuels speculation, where inflated valuations fund unprofitable ventures with little real usage.

So far, Nvidia’s ecosystem looks much closer to the healthy version. Its investments seem to amplify real adoption and infrastructure build-out rather than paper demand. Still, investors should recognize that reflexive growth can reverse if fundamentals weaken.

In short, what some call a “circular money loop” looks more like a strategic feedback system: capital reinforcing ecosystem growth. Nvidia’s role in AI’s expansion reflects a platform leader investing to strengthen and extend its technological base, not the financial engineering seen in earlier speculative cycles.

3) Economic Feasibility: Profitability and Scalability of NeoClouds and GaaS Models

Concerns about the economic feasibility of GPU-as-a-Service (GaaS) and NeoCloud data centers have grown louder, driven by a mix of misunderstanding of GPU economics, technology, flawed assumptions, and, in some cases, biased analysis.

The Critics and Their Claims

One of the most vocal critics has been legendary short-seller Jim Chanos, who argues that GaaS operators can’t earn sufficient returns on capital. When someone with his record raises questions, it’s worth listening, but his framework appears misapplied. He uses total CAPEX to calculate ROIC, rather than the portion of capital already producing cash flows. That’s like judging a hotel’s returns before it finishes construction.

Chanos also criticizes depreciation schedules, assuming GPU hardware will be obsolete within two or three years. That assumption ignores both historical precedent and the interplay between software and hardware improvements. I’ll return to that point shortly, but suffice to say: the evidence doesn’t support such an accelerated write-down.

Other critiques have come from The Information and The Financial Times, which questioned the profitability of GaaS models. The Information reported that Oracle lost $100 million on its AI cloud segment last quarter. That was likely a timing effect, a temporary mismatch between new capacity going live and revenue recognition as customers ramp up workloads. Oracle’s own commentary in its latest AI World conference effectively put those concerns to rest.

Meanwhile, the Financial Times argued that GPUs are “too expensive to buy and too cheap to rent.” Their math was flawed: they compared the price of a full 8-GPU node to the rental rate of a single GPU, an 8x error. Such oversights feed the perception of unsustainable economics where none exists.

The Obsolescence Misconception

Another persistent narrative is that GPUs will become obsolete every two or three years, as each new generation dramatically outperforms the last. Some suggest that future chips, like the Rubin CPX, will render current generations, such as GB300, uneconomic overnight.

If that were true, the entire NeoCloud sector would be at risk. But history points to a more measured reality: a regression to the mean.

Server hardware has historically remained productive for 7–9 years. Microsoft Azure’s own retirement schedule shows this clearly:

The K80, P100, and P40 GPUs (launched between 2014–2016) will retire in 2023, implying an 8–9 year life.

The V100 (launched in 2017) will retire in 2025, around 7.5 years later.

The A100(launched in 2020) and it remains economically viable.

That aligns closely with long-term IT depreciation norms. GPUs don’t vanish when the next model launches, they simply move down the value chain to workloads that don’t require the latest performance tier.

The Three Goals and the Three Forces Driving AI Economics

At its core, the AI industry is pursuing three fundamental goals:

A. Build smarter models.

B. Find and scale real-world use cases.

C. Drive down the cost per token.

These goals, in turn, activate three interacting forces that define the long-term economics of GPU infrastructure: demand growth, software optimization, and hardware innovation.

A. Demand Growth - The Expanding Workload Spectrum

The push to discover and scale new use cases continually broadens demand for compute. Every successful application (search, copilots, generative media, robotics) creates new workloads and revenue streams. As more industries integrate AI into operations, aggregate compute demand grows faster than supply expansion.

This persistent imbalance ensures that nearly all available hardware remains valuable. New GPUs power the frontier, massive model training, high-speed inference, while older chips cascade down to less intensive workloads where they still generate strong returns. Demand diversity stabilizes the system.

B. Software Optimization - Extending Hardware Lifetimes

The industry’s drive to reduce cost per token channels directly into software efficiency. Every layer of the stack (compilers, kernels, memory scheduling, quantization, fine-tuning algorithms) improves throughput without requiring new hardware.

Most of the gains in tokens-per-second efficiency today come from software, often by orders of magnitude compared to hardware alone. Continuous tuning by hardware vendors, hyperscalers, open-source contributors and AI labs, keep legacy GPUs productive far beyond initial expectations.

In effect, software innovation acts as economic gravity: it pulls down the effective cost of compute and extends the profitable lifespan of installed hardware.

C. Hardware Innovation - The Step Function of Capability

New architectures (like Nvidia’s Blackwell and Rubin) deliver large efficiency jumps but require long deployment and integration cycles. Power, cooling, and software adaptation slow immediate adoption. As a result, new hardware doesn’t instantly obsolete the old, it gradually shifts workloads upward along a capability curve.

Frontier training moves to the latest generation, while inference and specialized workloads continue profitably on prior ones. This interplay between innovation and inertia keeps the overall system balanced and capital-efficient.

The Interlocking System

These three forces and goals reinforce one another in a self-stabilizing loop:

The search for smarter models and new use cases expand demand.

The need for profitability drives software optimization.

Hardware innovation provides the next step in capability, which feeds back into new use cases and more demand.

Together, they create a layered equilibrium: demand outpaces supply, software stretches the useful life of capital, and hardware advances reset the frontier. The result is a sector that compounds through iteration rather than bubbles through speculation.

Fears about the economic viability of GaaS stem more from misunderstanding than reality. Demand is real and broad-based. Depreciation cycles are stable and supported by precedent. And cost-per-token reductions are coming primarily from software, not just hardware refreshes.

The result is an industry where nearly all compute remains economically useful, newer hardware drives the frontier, and older hardware serves the long tail. Far from fragile, the economics of GaaS look increasingly sustainable.

III. Conclusion

The question isn’t whether AI contains pockets of speculation, it does. The real question is whether the foundation beneath it is real, durable, and economically sound. Across all three dimensions: adoption, financing, and feasibility. The evidence points toward resilience.

Adoption and economic impact are already visible across every major sector. AI isn’t waiting for customers; it’s embedded in products, workflows, and infrastructure.

Financing of the buildout reflects strategic ecosystem development, not financial circularity. Nvidia, Microsoft, Amazon, and others are seeding complementary demand through partnerships and investments. Capital here is not being recycled in a vacuum; it’s reinforcing real economic activity driven by customers, workloads, and adoption.

GPU-as-a-Service operators and NeoClouds are showing early signs of sustainable unit economics. Depreciation cycles are likely to revert toward historical norms, while cost-per-token efficiency continues to improve, driven primarily by software innovation. The ecosystem’s self-balancing loop, rising demand, smarter software, and steady hardware progress, ensures that nearly every watt of compute finds productive use.

AI as an industry is not in a bubble. That doesn’t mean a bubble can’t form. It’s important to stay vigilant and know what to look for. It’s equally important to keep in mind the economic realities not only of the industry but of any individual company we choose to invest in.

Thanks for reading.