Is AI a Bubble? Nvidia and Data Center Risk

We have been hearing the phrase "AI bubble" almost every day for the last year, and every big company is investing heavily to win the game. That does not mean people were unaware of AI before now. The main reason everyone is focused on AI today is the speed of its recent evolution.

People have known about the term for well over a decade, but a paper published in 2012 — ImageNet classification using deep convolutional neural networks — changed the whole game. Before that, people focused mainly on computer vision and machine learning solutions that depended heavily on hand-crafted algorithms and techniques. People knew about neural networks but remained skeptical of them for two main reasons:

1. A neural network is not a fixed algorithm — it is an iterative process that adjusts itself to minimize error. Unless the error drops significantly with each pass, the system keeps feeding the previous errors back into itself without ever converging.
2. It requires massive hardware to run.

In that 2012 paper, however, the authors used GPU training and wrote custom CUDA code to train the neural network across multiple Nvidia GPUs. The ReLU activation function and dropout regularization made training many times faster and more efficient. From that point onward, data and hardware became the key drivers of deep learning, reshaping the field entirely. Deep learning solutions began emerging across tasks from classification to detection, spanning image, audio, and text.

---

The GenAI Race

Five years later, Google published a paper called "Attention Is All You Need" in 2017. It allowed neural networks to process an entire sentence at once and understand word context and relationships. Generative AI (GenAI), which uses large language models (LLMs), is also based on this architecture.

• OpenAI launched ChatGPT (GPT-3.5) in November 2022 as a free research preview and received a massive public response.
• Anthropic launched Claude as a closed developer API around March 14, 2023; its dedicated public chat interface followed in July 2023.
• Google launched its first GenAI model, Bard, in March 2023, which was rebranded as Gemini in February 2024.
• DeepSeek, a Chinese AI firm, launched its open-source model in November 2023.

---

Why These Companies Are Burning Money

All these companies are trying to grab market share first, even at the cost of losses. OpenAI has reported a record loss: it carries a large base of free-tier users that is costly to support, but the company accepts this cost to grow its market share. Its frontier AI models require a large number of GPUs and Nvidia chips for training and inference, and running them also requires data centers. On top of that, a significant share of OpenAI's API revenue is spent renting cloud infrastructure rather than kept as profit. However, OpenAI is already building its own data center, called Stargate, which should reduce its rental costs over time.

Anthropic is approaching profitability this quarter. That is because it offers less free-tier support and serves mainly enterprise and business clients — unlike OpenAI, which is prioritizing consumers and free-tier users to build market share first.

Google Gemini does not need outside funding, since it is backed by Alphabet.

DeepSeek, from China, offers a cost-efficient, open-source model with per-API pricing much cheaper than OpenAI's and Anthropic's. However, it has recently raised its prices to improve its margins.

---

What Is the Bubble Here?

The bubble is about the valuations these companies are holding — including AI labs (OpenAI, Anthropic, Gemini, and others), along with GPU providers like Nvidia, data centers and cloud providers, and electricity and power transmission companies.

Let's take Nvidia as an example. Since most AI companies rely on Nvidia GPUs, Nvidia profits whenever any LLM provider grows — OpenAI, Anthropic, Gemini, and many others — because it controls roughly 90% of the market. Some reports estimate Nvidia's revenue from OpenAI alone could exceed $600B by 2030. However, other reports suggest this revenue partly reflects circular financing rather than organic growth. For example, Nvidia might invest heavily in a data center for OpenAI, and in return OpenAI would be required to buy Nvidia chips — a purchase that is then booked as revenue.

Other companies like Google, Microsoft, and Meta are building their own custom chips, which could reduce Nvidia's market share over time. Separately, cheap Chinese AI models like DeepSeek and Kimi are competing heavily with OpenAI and Anthropic, which could erode those companies' market share as well. These models can also be built and trained with less computation — a real advantage at a time when high-end Nvidia AI chips face export restrictions in some countries.

Another issue is that, to sustain their valuations, these AI companies must grow 5 to 13 times faster than their current pace to justify the investment. The range is wide because of uncertainty around model optimization and future demand.

In the data center sector, Oracle, Microsoft, Amazon, and Alphabet are investing heavily — taking on debt and cutting jobs to fund it. Nvidia has also formed partnerships with BlackRock, Apollo, Goldman Sachs, and others to arrange $500B in AI infrastructure financing. All of this rests on the expectation of rising demand and much higher future profits. The bubble is not about whether AI is real — it is about expectations for future demand and profit margins. AI itself is not going away, but if growth slows more than expected, these companies' bets are at risk. Still, they are cash-flow-rich businesses, and to stay relevant, they have little choice but to keep investing in AI.

Meanwhile, businesses everywhere are adopting AI by paying for subscriptions to OpenAI, Anthropic, or others. This adds a new cost, and if a business cannot generate enough additional profit to offset it, the alternative is to cut costs by reducing headcount.

---

Companies Involved in the Global AI Ecosystem

---

Summary

AI's existence is not at risk — the main issue is future demand. If demand slows, it poses the biggest risk to companies building AI models. That slowdown would then hit companies building data centers — many of which have taken on heavy debt — and the pain would carry through to GPU supplier companies.

The repricing has already begun in enterprise IT: IBM crashed 25% in a single day after a frontier model disrupted its deals, and the same pressure runs through Indian IT results and the mid-cap IT names.

---

References

1. OpenAI
2. Anthropic
3. Google DeepMind — Gemini
4. DeepSeek
5. Nvidia

This article is for educational and informational purposes only and does not constitute investment advice. Please consult a registered investment adviser before making any investment decision.