If AI were a bubble, Big Tech wouldn’t be pouring trillions into compute — 2026 will be defined by who controls inference power.
There’s a lot of talk about the AI surge being a bubble that’s about to burst. But looking at the thinking behind recent compute infrastructure investments, rising AI demand and the rapid adoption of compute-hungry AI use cases paints a different picture of where AI is headed in 2026.
The past six months alone have yielded a flood of announcements from both Big Tech companies and public organizations about plans to ramp up investment in compute infrastructure. To some, this indicates an overheated market that will collapse catastrophically in the next 12 months. But a closer look at AI adoption rates and strategies makes a much stronger case: that of a vital investment crucial to meeting AI needs.
Does the AI boom presage an AI bust?
Companies and governments are building more data centers, upgrading chip production and improving their access to existing compute capabilities. One recent study estimates that on the heels of rising demand for AI, capital expenditures on data centers will exceed $1.1 trillion by 2029. In the meantime, infrastructure investments are mounting up.
The many recent announcements include:
- Meta Platforms expects to spend as much as $72 billion in 2026.
- Microsoft signed a $17.4 billion-deal for additional GPU capacity with the Nebius Group.
- OpenAI and Nvidia signed a $100 billion agreement, part of the $1 trillion in AI infrastructure deals OpenAI has signed.
- Alibaba will invest at least $52 billion in AI infrastructure over the next three years.
- Google will invest $40 billion in three new data centers through 2027.
- Amazon is investing up to $50 billion to expand AI and supercomputing capabilities.
For academic economists like Niall Ferguson and Servaas Storm, who argue that the AI market is a bubble that’s going to burst in dramatic fashion, this is just another sign of an overheated market. But the drive to expand computing power has been accompanied by increasing demand. Compute-hungry AI use cases such as GenAI video, agentic AI and low inference are rising, alongside detailed research underscoring compute needs.
Compute investments driven by demand
Recent investments are not just about enterprises trying to outdo each other with shiny AI toys. According to the tech companies themselves, fast expansion is required just to keep up with demand from AI customers. For example, Amazon is expanding its AI infrastructure to offer on-prem, fully managed AI capabilities for enterprise customers who need to comply with data sovereignty regulations and to provide better access for U.S. government users.
According to research by Morgan Stanley, global data center capacity will need to grow six-fold by 2035 just to meet the demands of cloud computing and AI, while McKinsey forecasts that global demand for data center capacity could almost triple as early as 2030, with about 70% of that demand coming from AI workloads.
Both rising AI adoption and the growing compute-hungry use cases drive the need for greater computing power. Governments are investing heavily in AI to ensure economic and military security and drive tech independence. At the same time, enterprises turn to solutions such as agentic AI and generative AI video to gain a competitive edge.
Agentic AI infrastructure is the new must-have
Deloitte picks agentic AI as one of the top trends driving AI compute needs in the next year, speculating that it could overtake SaaS tools. According to its research, up to 75% of companies may invest in agentic AI in 2026, driving demand for chips, data centers and AI infrastructure.
A mass of data backs up Deloitte’s estimates. Cisco forecasts that 56% of customer service and support interactions with tech vendors will be handled by agentic AI by the end of 2026, rising to 68% by 2028. The use of agentic AI has triggered an 8% drop in demand for software development skills.
Even IoT Analytics’ negative report about AI investments acknowledges that today’s projections could be insufficient if agentic AI delivers on its promise to automate workflows at scale.
AI generative video is in high demand
GenAI video is starting to appear everywhere. By May 2025, four of the top ten most popular YouTube channels had AI-generated material in every video, and it’s predicted that by 2027, more than 60% of all digital video will be AI-generated, at least in part. And that’s only the beginning. A recently announced partnership between AWS and Decart offers users a full-stack AI video generation platform for real-time AI-generated video and visual intelligence with implications across any number of use cases.
“Robots are already using models like this to envision infinite possibilities, infinite outcomes,” said Dean Leitersdorf, CEO of Decart, on stage at the recent re:Invent conference. “In the future, we’ll be able to train robots fully in generative simulations before they ever hit the ground. And when they hit the ground, they will use live visual intelligence to understand the world around them.”
These capabilities are powerful but also compute-power-hungry. Amazon’s investment in AI-generated video tools and the infrastructure to support them shows that the company sees a future worth investing in.
Inference is the new AI battlefield
Inference is when an AI model performs a task, whether that’s answering a question, generating an image, or triggering a security alert. People don’t want to wait for an AI response, so inference is crucial for any AI use case.
It’s especially important in agentic AI, where complex workflows mean that even short lags can build up into noticeable delays. No company wants to suffer the reputational blow of slow responses because of inference issues. But inference speed also requires more storage, which means more chips and data centers.
“The amount of inference compute needed is already 100x more” than it was initially for LLMs, “and that’s just the beginning,” said Nvidia CEO Jensen Huang in an earnings call in February. He’s backed up by Deloitte, which predicts that inference will make up two-thirds of AI compute by 2026.
Compute infrastructure is expanding for a reason
Everyone, including the Big Tech companies, is trying to keep up with AI demand. This is something they have signalled clearly in every announcement about increasing their compute capabilities. While some might see the rapid growth in AI infrastructure as cause for concern, a closer look reveals a real need behind the expansion. While the market might slow, there are a few signs that AI infrastructure is a bubble that will burst.
This article is published as part of the Foundry Expert Contributor Network.
Want to join?
