Don’t be too scared of the AI bears. They are wondering aloud if the big boom in AI investment already came and went, if a lot of market excitement and spending on massive AI training systems powered by multitudes of high-performance GPUs has played itself out, and if expectations for the AI era should be radically scaled back.
But if you take a closer look at the plans of the major hyperscalers, AI investment is alive and well. Meta, Amazon, Microsoft, and Google have all recently doubled down on investing in AI technology. Their collective commitment for 2025 totals well over $300 billion, according to a recent story in the Financial Times. Microsoft CEO Satya Nadella said Microsoft could spend $80 billion alone on AI this year. Meta Founder and CEO Mark Zuckerberg said on Facebook, “We’re planning to invest $60-65B in capex this year while also growing our AI teams significantly, and we have the capital to continue investing in the years ahead.”
This is not the sound of an AI boom going bust, but there has been a growing unease around how much money is being spent on enabling AI applications. After at least two years of technology giants saying they were seeing clear demand for more computing power to help train massive AI models, 2025 has begun with those same companies being called on the carpet daily by business media for building up so much AI hype.
Why has there been such a sudden shift from hope to concern? The answer can be found partly in the rapid rise of a new AI application from China. But to fully understand what is really happening, and what it means for AI investment and technology programs in the coming years, we must acknowledge that the AI era is shifting into a new phase of its evolution.
DeepSeeking the Truth
By now, the world knows all about DeepSeek, the Chinese AI company touting how it used inference engines and statistical reasoning to train large language models much more efficiently and with less cost than other firms have trained their models.
Specifically, DeepSeek claimed its techniques resulted in it requiring far fewer GPUs (as few as 2,048 GPUs), as well as less powerful GPUs (Nvidia H800s) than the hundreds of thousands of premium-performance GPUs (think Nvidia H100s) that some hyperscale companies have required to train their models. In terms of cost savings, while OpenAI spent billions of dollars on training ChatGPT, DeepSeek reportedly spent as little as $6.5 million to train its R1 model.
It should be noted that many experts have doubted DeepSeek’s spending claims, but the damage was done, as news of its different methods drove a deep plunge in the stock values of the hyperscalers and the companies whose GPUs they have spent billions on to train their AI models.
However, a couple of important points were lost amid the chaos. One was an understanding that DeepSeek did not “invent” a new way to work with AI. The second is that much of the AI ecosystem has been well aware of an imminent shift in how AI investment dollars need to be spent, and how AI itself will be put to work in the coming years.
Regarding DeepSeek’s methods, the notion of using AI inference engines and statistical reasoning is nothing new. The use of statistical reasoning is one aspect of the broader concept of inference model reasoning, which involves AI being able to draw inferences based on pattern recognition. This is essentially similar to the human capability to learn different ways of approaching a problem and compare them to find the best possible solution. Inference-based model reasoning can be used today and is not exclusive to a Chinese startup.
Meanwhile, the AI ecosystem for some time already has been anticipating a fundamental change in how we work with AI and the computing resources required. The initial years of the AI era have been all about the big job of training large AI models on very large data sets, all of which required a lot of processing, complex calculations, weight adjustments, and memory reliance. After AI models have been trained, things change. AI is able to use inference to apply everything it has learned to new data sets, tasks, and problems. Inference, as a less computationally intense process than training, does not require as many GPUs or other computing resources.
The ultimate truth about DeepSeek is that while its methods did not shock most of us in the AI ecosystem as much as it did casually interested stock market investors, it did highlight one of the ways in which inference will be core to the next phase of AI’s evolution.
AI: The Next Generation
The promise and potential of AI has not changed. The ongoing massive AI investments by the major hyperscalers show the faith they have in the future value they can unlock from AI, as well as the ways in which AI can change how virtually every industry works, and how virtually all people go about their everyday lives.
What has changed for those hyperscalers is how those dollars are likely to be spent. In the initial years of the AI era, most of the investment was necessarily on training. If you think about AI as a child, with a mind still in development, we have been spending a lot of money to send it to the best schools and universities. Now, that child is an educated adult–and it needs to get a job to support itself. In real world terms, we have invested a lot in training AI, and now we need to see the return on that investment by using AI to generate new revenue.
To achieve this return on investment, AI needs to become more efficient and less costly to help companies maximize its market appeal and its utility for as many applications as possible. The most lucrative new services will be the autonomous ones that don’t require human monitoring and management.
For many companies, that means leveraging resource-efficient AI computing techniques, such as inference model reasoning, to quickly and cost-effectively enable autonomous machine-to-machine communications. For example, in the wireless industry, AI can be used to autonomously analyze real-time data on spectrum utilization on a mobile network to optimize channel usage and mitigate interference between users, which ultimately allows a mobile operator to support more dynamic spectrum sharing across its network. This type of more efficient, autonomous AI-powered machine-to-machine communication will define AI’s next generation.
As has been the case with every other major computing era, AI computing continues to evolve. If the history of computing has taught us anything, it is that new technology always requires a lot of upfront investment, but costs will come down and efficiency will go up as we start to leverage improved techniques and better practices to create more beneficial and affordable products and services to appeal to the largest possible markets. Innovation always finds a way.
The AI sector may have recently appeared to suffer a setback if you listen to the AI bears, but the dollars the hyperscalers plan to spend this year and the increasing use of inference-based techniques tell a different story: AI computing is indeed changing, but AI’s promise is fully intact.
