We are Building the Wrong AI Infrastructure and Paying Trillions for It
If you are someone that likes to see lots of zeros all in a row, you should pay attention to what’s going on in the data center buildout for artificial intelligence (AI). Check out these price tags:
OpenAI signed a $300,000,000,000 deal with Oracle for compute capacity in 2025. The deal is to create 4.5 gigawatts of compute capacity that Open AI can use for its model training work. For context, 4.5 gigawatts is 4,500,000,000 watts.
The numbers get higher and higher when you look across the sector, so high that some consulting companies have estimated that AI companies need to turn $2,000,000,000,000 in revenue in less than five years.
Billions of dollars spent. Billions of gigawatts required. Trillions in revenue needed to justify the investments.
This must leave you scratching your head. What is it that these companies are buying for all these trillions of dollars in chips, data centers, and energy plants? What are we building toward?
The question about what we are building toward is an uncomfortable one in the AI industry and leads to answers from CEOs that cite competition with China. The flaw in this argument is that we are making a simplistic assumption: That vaguely “building bigger” will somehow “win” this race with China. Nothing in the world is this simple, and we should not base the expenditure of trillions of dollars around it.
Yes, the AI industry needs infrastructure to run on, which is what the major AI companies are building now. But are we building the right infrastructure? There are still fundamental problems with how AI models perform in a variety of situations and edge cases leading to financial, physical, and reputational harm for organizations and individuals. These problems are unsolved and the industry’s answer is to just keep building, which must lead us to ask whether the issues we have with AI today are doomed to get worse as trillions of dollars more are poured into better and better infrastructure.
The Wall Street Journal recently looked at these issues and asked whether AI might be heading for a bubble. There are reasons to consider this possibility, but what jumps out is the infrastructure question. Are we building the infrastructure we need or are we in an ambiguous “just build bigger” race against an unknown competitor?
A LOT of Zeros
The OpenAI-Oracle deal is just one of several deals, but first a word on compute capacity for data centers. Data centers are large server farms housed in massive warehouses and those centers require a lot of electricity. In July of 2025, the US Department of Energy published a report on the need for over 50 gigawatt hours of additional power in the US grid or it risks a 100 fold increase in blackouts by 2030. The report goes on to say that those 50-gigawatt hours are required to “win the AI arms race.” Leaving the geopolitical commentary aside for a moment, the point is that these big warehouses full of server stacks need A LOT of power and there are consequences if we don’t provide it. As a result, it is possible to measure the compute capacity of a data center not in its storage capacity but in the electricity required to run it.
With this in mind, OpenAI estimates it costs about $50 billion for one gigawatt of compute when you factor in building the data center and the huge number of Nvidia chips required to create that compute capacity. If 1 gigawatt is $50 billion, then the 7-gigawatt center that OpenAI is hoping to create would cost $350,000,000,000 (zeros written out for effect). Remember that OpenAI generated $3.7 billion in revenue in 2024, way, WAY short of covering this kind of cost. If OpenAI paid Oracle ALL its 2024 revenue every year, it would need 94 years to pay off the $350B bill for a single data center.
Some experts estimate that the compute required to advance AI models to the true next level (unlike the incremental improvements of GPT5) would take a data center that has 100 gigawatts of compute, which would run $5 trillion.
OpenAI isn’t the only one spending. Amazon plans to spend $100 billion in Capex in 2026 while Microsoft is looking at $80 billion to build out its AI infrastructure. All in, estimates are that the big tech companies will spend $400 billion just on AI build. In July of 2025, Google CEO Sundar Pichai said, “The risk of under-investing is dramatically greater than the risk of over-investing for us here.” This quote has a distinct FOMO vibe to it. A sort of, “Everyone else is doing it, so we are going to do it too” mindset.
According to the Wall Street Journal, the projected spend for AI infrastructure in the next year is more than what was spend on the interstate highway system over a 40-year period.
If you are running a business, you can’t just spend money. You need to make money. According to the consulting firm Bain, the AI industry would need $2 trillion in new revenue to keep up with the infrastructure scaling trend by 2030. That’s pretty ambitious when one of the biggest companies in the world brought in just $3.7 billion in 2024.
Again, a lot of zeros. A lot of expense, a lot of compute, and a lot of ambition. But what it is we are building toward is unclear and the answers on how we solve the problems AI already has are in short supply. Is this the right infrastructure or should we be building something else?
Safety Infrastructure
I want to stick with the comparison with the interstate highway system. The highway system is one of the clearest examples of infrastructure that most Americans can connect with because so many use it and have used it. The highway system facilitated safe travel across the United States by car. Without it, we would effectively travel the way the pioneers on the Oregon Trail did. But the money spent on the interstate highway system over those 40 years was not just on paved roads. The builders of the system knew they had to include safety measures becauase driving in a car is one of the more dangerous things that humans do. So, the roads were accompanied by guardrails. We also built runaway truck ramps for long downhills. We put up signs that warn of freezing bridges and sharp turns. We built restaurants, gas stations, and rest stops along the way. These additions are a cost above the roads, which are strictly necessary to move a car from one destination to another. However, the road was not sufficient. We needed to build safety measures to help grow the transportation infrastructure envisioned. It wasn’t just the road.
Likewise, the infrastructure into which AI companies are pouring hundreds of billions of dollars needs to include safety measures. We need the guardrails, signs, and runaway truck ramps because the people using it need these protections. Right now, the AI companies are only focused on building more and building bigger, they are not focused on building safer or smarter.
Safety infrastructure means robust and continuous testing of AI systems. This is not as simple as a sign or a ramp, but it means continuous testing of models. The testing must be continuous because the state of the model changes consistently due to training and learning. A test at a single point in time is not useful for model users; it is a rubber stamp that a system implementer can use to say the model WAS safe or in compliance. But truly understanding how safe, compliant, and/or accurate your model is requires consistent testing. That testing must take place at the interaction level, not in the lab.
AI is tested in the lab, but that’s not where AI is deployed. AI is deployed to humans and humans introduce complexity. A small change in language or an accidental misspelling could cause an AI model to break a guardrail with potentially harmful consequences. Runaway truck ramps on highways are rarely used, but they are there because the edge case of a runaway truck could be devastating. In the same way, AI edge cases have the potential to cause real harm. Models are designed with guardrails, but those guardrails are broken frequently. Model builders mitigate risks they know about, but human create a level of complexity that can’t be designed for. It can be tested for.
Ask Questions
If the biggest technology companies in the world are pouring hundreds of billions of dollars into data centers, we should ask to what end. If those data centers require so much power that the Department of Energy is warning of a 100-fold increase in blackouts, we should ask what that power is going to. If there are issues with our current AI in terms of safety and accuracy, we should ask what scaling that AI will cause.
If you accept the premise that the US is in an “AI arms race” you should also question whether the path to “winning” that arms race (whatever winning means) is equal to building more and more. Underlying assumptions have gotten investors and policy makers alike into trouble throughout history. The underlying assumption of building bigger data centers is that AI will always require more computers and more GPUs. This sounds a lot like the assumption that the housing market in the early 2000s would always appreciate it or that if you had a company with .com at the end of the 1990s, you would get a lot of money. Both were bubbles because the assumptions were wrong. Right now, the assumptions are costing companies hundreds of billions of dollars and threatening the electrical grid. Are those assumptions correct? Or should we be questioning them?
Infrastructure without safety measures is destined for a failure at scale. Even if the companies are right and these trillions of dollars in data center compute are required, we still aren’t building the safety infrastructure we need. All this AI, present and future, will impact humans, and we need to make sure those impacts will not be harmful. We need to test continuously and test at the interaction level. We also need to see safety testing as a fundamental part of the infrastructure we are building just like the highway system.
All those zeros need to serve us. Not an abstract future AI.