Technology · Artificial Intelligence and Economics
The AI Bubble: When Intelligence Becomes Cheap, What Happens to Its Value?
DeepSeek challenged the economics of artificial intelligence. As models become more efficient and accessible, the value of AI may shift from foundation models and infrastructure towards applications, integration and the ability to solve real problems.
On January 27, 2025, NVIDIA lost approximately $593 billion in market value in a single trading session. It was one of the largest single-day losses in the history of the stock market, and the interesting part was that nothing particularly catastrophic had happened to the company. Its technology had not suddenly stopped working, its customers had not disappeared, and the demand for artificial intelligence had not collapsed. Instead, investors had encountered a possibility that challenged one of the most important assumptions behind the AI industry: perhaps building powerful artificial intelligence did not require as much computing power and money as everyone had expected.
The development came from a relatively young Chinese company called DeepSeek, which demonstrated that competitive AI models could be developed using a considerably more efficient approach to computation. The immediate market reaction was dramatic because NVIDIA had become one of the central beneficiaries of an investment cycle built around the assumption that increasingly powerful AI would require increasingly large quantities of specialized hardware. If the same level of intelligence could be achieved with fewer computational resources, the economics of that entire investment cycle deserved another look.
But this is not simply a story about one Chinese company challenging an American technology giant. It raises a much more interesting question about the relationship between technological progress and economic value. Throughout history, the technologies that transformed civilization did not always create lasting wealth for the companies that first invested in them. Sometimes the technology became more valuable precisely because its cost declined, making it accessible to people and businesses that had previously been excluded.
The Economics Behind the AI Race
The current AI industry has developed around an assumption that is technically reasonable but economically complicated: larger models, more computing resources and better hardware can produce increasingly capable systems. Training a modern neural network requires enormous quantities of mathematical operations, and performing those operations efficiently requires specialized processors, high-speed memory, networking infrastructure and substantial amounts of electricity.
Companies such as Microsoft, Google, Amazon and Meta have therefore been investing heavily in data centres and AI infrastructure. NVIDIA has benefited enormously from this demand because its GPUs and software ecosystem are particularly suitable for the parallel computation required by modern machine learning. What began as a competition to build better software has gradually become an industrial competition involving semiconductor manufacturing, energy supply, cooling systems and physical infrastructure.
There is a financial assumption behind these investments. Companies are spending money today because they expect artificial intelligence to become sufficiently important in the future that the infrastructure being constructed now will generate substantial returns. If AI becomes part of software development, scientific research, financial services, manufacturing, communication and everyday business operations, the demand for computing could become enormous.
However, the existence of a large future market does not automatically mean that every company investing in it will become profitable. A company can correctly predict that demand for a technology will increase and still make a poor investment if it overestimates the price customers will pay or underestimates how quickly competitors will reduce its costs.
This is where the DeepSeek development becomes particularly relevant.
DeepSeek Changed the Question
DeepSeek's achievement was not that it discovered an entirely new principle of artificial intelligence. Its importance was that it demonstrated how much could be achieved through architectural and computational efficiency.
DeepSeek-V3, for example, uses a Mixture-of-Experts architecture. Instead of activating every parameter of the model for every operation, the system activates only selected groups of parameters depending on the input. Although the model contains 671 billion total parameters, approximately 37 billion are activated for each token. This allows the model to retain a very large overall capacity without requiring every part of the network to participate in every computation.
The researchers also introduced improvements in attention mechanisms, load balancing and multi-token prediction. These are engineering decisions, but their economic implications are significant because they demonstrate that computational efficiency is not determined only by the number of processors available. The architecture of a model and the algorithms used to operate it can change the amount of hardware required to achieve a particular result.
DeepSeek reported that the official training process for V3 required approximately 2.788 million NVIDIA H800 GPU-hours, estimating the associated GPU rental cost at around $5.576 million using an assumed rate of two dollars per hour. This number received enormous attention, although it is important to understand what it represents. It does not account for the complete historical cost of research, previous experiments, infrastructure, personnel and other development activities.
Even with that qualification, the engineering achievement remains important. It challenged the idea that progress in AI would necessarily be dominated by organizations capable of spending the most money on computing infrastructure.
There is a familiar pattern here in the history of technology. A faster processor can improve computing performance, but a better algorithm can sometimes produce an even greater improvement without changing the hardware at all. The history of computing is full of examples where mathematical and engineering improvements allowed existing resources to perform tasks that previously appeared impractical.
DeepSeek brought this question directly into the economics of artificial intelligence: if intelligence can be produced more efficiently, how much should we really be paying for it?
Open Models and the Changing Economics of Access
Another important part of the story is the availability of model weights. Traditionally, when someone uses a proprietary AI service, they are purchasing access to a capability controlled by another company. The model runs on infrastructure operated by the provider, and the customer generally interacts with it through an application or API. The customer receives the result but does not necessarily have control over the underlying model.
Open-weight models change this arrangement. Developers can obtain the model parameters and deploy them on infrastructure they control, modify or adapt them where permitted, and integrate them into their own applications. This does not mean that every model is fully open-source, because releasing model weights is not the same as publishing the complete training data, source code and development process. It also does not mean that a large model can suddenly run efficiently on an ordinary laptop. Hardware requirements remain an important limitation.
Nevertheless, the economic implications are considerable. A company that previously had to pay an external provider for every AI request may now have the option of running a model on its own infrastructure. A university can experiment with language models without building one from the beginning. A startup can develop a specialized application around an existing model rather than spending enormous resources on foundation-model training.
This changes the competitive structure of the industry. When a technology is controlled by a small number of providers, those providers have considerable influence over pricing and access. When capable alternatives become widely available, customers gain more choices, and the economic value of simply providing access to a model can decline.
The same process has happened in other areas of computing. The availability of open-source operating systems, databases, programming languages and development frameworks made it possible for smaller organizations to build products that previously required access to expensive proprietary technology. The value gradually shifted from owning the underlying tools to building useful services and systems around them.
Artificial intelligence may be entering a similar stage.
When Efficiency Increases Demand
There is, however, an important reason why cheaper AI does not necessarily mean that the demand for GPUs will decline.
When a technology becomes more efficient and less expensive, people often find new ways to use it. An activity that was previously too costly to perform becomes economically practical, and the total market can expand. This is related to the Jevons paradox, an economic observation that improvements in resource efficiency can sometimes increase total consumption rather than reduce it.
Consider computing itself. The cost of processing information has declined dramatically over several decades, but humanity has not responded by using fewer computers. Instead, computing has become part of almost every industry. Smartphones, cloud services, online banking, scientific simulations and countless software applications have created demand that would have been difficult to imagine when computers were expensive and relatively inaccessible.
AI could follow a similar path. If the cost of generating an intelligent response declines, companies may begin using models for tasks that previously did not justify the expense. Software agents could continuously analyse documents, monitor networks, coordinate business operations and assist engineers. A system that once used AI occasionally might eventually make hundreds or thousands of model requests as part of a single workflow.
The price of each operation may decline while the total number of operations increases.
This creates an interesting situation for the semiconductor industry. More efficient models could reduce the computational requirements of individual tasks, but the expansion of AI into new applications could increase the total demand for computation. Both outcomes are possible at the same time, and their relative importance will depend on how quickly efficiency improves compared with how quickly new applications emerge.
Therefore, the DeepSeek development should not be interpreted as proof that the demand for NVIDIA's hardware will disappear. It is evidence that the relationship between AI capability, computing requirements and commercial pricing is more complicated than a simple race to purchase more GPUs.
The Difference Between a Technology Revolution and an Investment Bubble
The discussion about an AI bubble often becomes confused because two separate questions are treated as though they are the same. The first is whether artificial intelligence is a genuinely transformative technology. The second is whether the financial expectations surrounding AI companies and infrastructure investments are realistic.
There is no reason to assume that the answer to both questions must be identical.
The internet provides a useful historical comparison. During the late 1990s, investors poured enormous amounts of money into companies built around the expectation that the internet would transform commerce and communication. Many businesses received extraordinary valuations despite having limited revenue or unclear business models. When the dot-com bubble collapsed, numerous companies disappeared, and investors suffered substantial losses.
Yet the internet itself did not become less important. Much of the physical infrastructure built during that period remained available. Fibre-optic cables, networking equipment and data centres continued to provide the foundation for the expansion of internet services. Some of the infrastructure had initially been built ahead of demand, but it became useful as the digital economy developed.
The companies that failed did not invalidate the technology they had invested in.
This distinction is particularly important when discussing AI. It is possible that some companies are overestimating future demand, that some infrastructure investments will not generate sufficient returns, and that the market is assigning valuations based on expectations that may take much longer to materialize. None of these possibilities means that artificial intelligence is a temporary phenomenon.
The more useful question is whether the financial returns expected from today's infrastructure will remain possible as AI becomes cheaper and more widely available.
If competing companies can offer similar capabilities at lower prices, the businesses that have invested the most money may face considerable pressure. Infrastructure that was profitable under one pricing structure may become less attractive under another.
But if cheaper intelligence creates enough new demand, the additional usage could compensate for declining prices. At present, there is no simple way to know which effect will dominate.
The Opportunity for Countries Outside the Technology Centres
For countries like Bangladesh, the changing economics of artificial intelligence could be more important than the stock-market competition between NVIDIA, Microsoft and other major technology companies.
The ability to access capable models without developing them from scratch creates opportunities for universities, startups, software companies and established institutions. A local software company could build a Bangla-language assistant around an existing model. A bank could investigate private AI deployment for internal document processing. An engineering team could connect models to industrial monitoring systems, telecommunications infrastructure or business automation.
The important point is that a country does not necessarily need to manufacture the processors or train the world's largest foundation model to participate in the AI economy.
It needs people who understand the technology well enough to apply it to real problems.
There are still practical limitations. Running large models requires appropriate hardware, memory, electricity and technical expertise. A model that performs well in English may not perform equally well in Bangla. Local deployment also requires proper security, data management, evaluation and maintenance. Simply downloading an open model does not create a useful AI product.
However, as the cost of capable models declines, the barrier to experimentation becomes lower. This is particularly relevant for developing countries, where access to expensive proprietary services can become a constraint on smaller organizations.
Perhaps the more realistic opportunity for Bangladesh is not to compete in the construction of enormous AI infrastructure, but to build useful systems around increasingly accessible intelligence.
What Happens After Intelligence Becomes Cheap?
The most interesting possibility is that the AI industry may eventually become a very different business from the one investors are currently imagining.
Today, much of the competition is focused on producing increasingly capable foundation models. Companies invest in training, computing infrastructure and research because they believe that better models will create a lasting commercial advantage. But if capable models become widely available, the model itself may no longer be the most valuable part of the system.
The value could move towards proprietary information, specialized applications, integration, reliability and the ability to solve specific problems. A company that understands how a hospital operates, how a telecommunications network behaves or how a manufacturing process can be improved may create more commercial value by applying an existing model than by attempting to build a foundation model of its own.
This is not an entirely new economic pattern. The availability of inexpensive computing did not make every computer manufacturer successful. The expansion of the internet did not make every website profitable. Technologies create opportunities, but the businesses that benefit from those opportunities are not necessarily the ones that originally developed the technology.
The same could happen with AI.
The companies investing billions of dollars in computing infrastructure may help create the technological foundation for a future in which intelligence becomes significantly cheaper. Some of those companies may capture enormous value, while others may discover that their investments were based on expectations that no longer reflect the economics of the market.
The technology can succeed even when individual business models fail.
And perhaps that is the central lesson of the DeepSeek moment. The AI industry is not simply competing to build the most powerful machines. It is also discovering the economic consequences of making intelligence increasingly accessible.
We are still at an early stage of this transition. We do not know how quickly computational efficiency will improve, how much new demand cheaper models will generate, or which parts of the AI industry will retain the greatest commercial value. Predicting an inevitable collapse or an uninterrupted technological boom would be equally premature.
But one thing is becoming increasingly clear: as intelligence becomes easier to produce, the ability to access it may become less valuable than the ability to use it effectively.
The next major opportunity may not belong to the company that builds the largest model. It may belong to the people who understand what can be built once powerful models are no longer prohibitively expensive.