Goldman Sachs has put a number on the AI buildout, and it is big enough to change how founders, investors and cloud buyers think about the next five years.
The latest AI story is not another chatbot demo or a clever video clip. It is the plumbing behind all of it. Goldman Sachs now expects Meta, Microsoft, Amazon and Alphabet to spend a combined $5.3 trillion on capital expenditure from fiscal 2025 through fiscal 2030, a figure that turns the AI race from a product battle into one of the largest private infrastructure cycles in modern technology.
That matters because the market had started to wonder whether AI spending was running out of momentum. Higher rates, nervous equity investors and questions about revenue have all made the boom look more fragile. But the money is still moving. It is moving into data centers, chips, power, cooling, networking and the model systems that sit on top of that stack.
As Reuters reported Wednesday, Alphabet increased the size of its equity offerings to $84.75 billion to help fund its AI ambitions, including infrastructure and computing capacity. That is not a small financing decision from a cash-rich company. It is a signal that even the strongest balance sheets are preparing for a longer, more expensive race than many investors expected.
It is tempting to read the $5.3 trillion forecast as a semiconductor story. Nvidia, AMD, Broadcom, TSMC, SK Hynix and Samsung will clearly remain central to it, because training and inference still need accelerators, memory and advanced packaging. But the bigger opportunity is spread across a wider supply chain.
AI factories need land, grid connections, fiber, backup power, cooling systems and specialized construction. They also need software that manages clusters, schedules workloads and squeezes more output from expensive hardware. For enterprise software builders, that is where the useful part of the story begins. If compute remains scarce, customers will pay for tools that reduce waste, route jobs intelligently or make smaller models perform like larger ones.
This is also why the spending cycle may not flow evenly to every vendor. Hyperscalers have the scale to negotiate directly with chipmakers, design custom silicon and absorb low near-term returns. Smaller infrastructure providers have less room for error. CoreWeave has become one of the most visible names in AI cloud, but its reliance on debt financing shows how hard it is to grow outside the balance sheets of the largest platforms.
That does not mean non-hyperscale players are finished. It means they need a sharper wedge. Specialized GPU clouds can still win where they offer speed, flexibility, better developer experience or access to capacity the big clouds cannot deliver quickly enough. But as consolidation accelerates, weak utilization and thin margins will be punished fast.
AI video is part of the same question
The user-facing angle is AI video. After OpenAI pulled back from Sora, some people read that as a sign that the market had cooled. A better reading is that video generation exposed the cost problem earlier than text did. High-quality video consumes enormous compute, and consumer pricing has not yet caught up with the infrastructure required to serve it at scale.
Google, Runway, Kling and other video players are still pushing forward, but the economics are becoming more disciplined. The winners will not simply be the models that produce the most impressive clips. They will be the companies that can connect video generation to real workflows: advertising, game development, enterprise training, product design and creative production.
So yes, AI video is back, but not in the loose hype-cycle sense. It is back as a workload that has to justify its place inside a massive compute budget. That is healthier for the market. A video tool that saves a studio time, helps a brand produce localized campaigns or lets a game team prototype faster has a clearer path to revenue than a novelty app burning premium GPUs for casual clips.
The Goldman forecast also changes how founders should think about timing. When four companies are preparing to spend trillions, the question is not whether infrastructure will exist. The question is who gets access to it, at what price and with what level of dependency on the major cloud platforms.
For investors, the practical takeaway is to look beyond the headline spend. The most durable value may sit in bottlenecks: power availability, memory supply, networking, inference efficiency and software that helps enterprises control AI costs. For founders, the lesson is simpler. Build where the spending creates a real problem for customers, not where it merely creates a larger press release.
The next phase of AI will be measured less by who has the flashiest model launch and more by who can turn expensive infrastructure into reliable economics. That is the market to watch now.
Also read: Tesla’s robotaxi reality is smaller than its AI ambition • AI data centers are turning water into an investment risk • Monterey Park has turned AI data centers into a local political risk
