Let's start with a number that barely makes sense
In November 2022, NVIDIA reported quarterly datacenter revenues of $3.8 billion. Not bad for a chipmaker. Fast forward exactly three years, and that number had grown to $51 billion per quarter. Same company. Same basic product category. Thirteen times larger.
That kind of growth does not happen because of hype alone. Something real changed. And understanding what changed is the key to understanding why the world's biggest companies are now committing to a cumulative $6.4 trillion in AI infrastructure spending over the next five years. PitchBook describes this as the single largest deployment of investment capital in human history.
The moment everything shifted: DeepSeek's surprise
For most of 2022, 2023, and 2024, AI was impressive but unreliable. Models hallucinated, misunderstood instructions, and struggled with anything requiring genuine reasoning. Useful for drafting emails, but not quite ready to run a business.
Then in January 2025, a Chinese AI lab called DeepSeek released a model called R1. What made R1 special was not just its performance but its architecture. It used something called Mixture of Experts, which means the model does not activate its entire brain for every question. It routes each query to the relevant specialist. The result was dramatically better performance at dramatically lower cost.
At the time of R1's release, OpenAI's GPT-4.5 scored 10.3% on a key reasoning benchmark at a cost of $2.10 per task. R1 scored 15.3% at just $0.08. Better performance and 26 times cheaper.
This was the moment the industry stopped debating whether AI could be genuinely useful and started figuring out how fast it could scale. If expert-level intelligence was within reach and the economics were falling rapidly, then the only question left was how much compute you could get your hands on.
So where is all that money actually going?
When analysts talk about AI infrastructure spending, they are not talking about apps or software subscriptions. They are talking about the physical stuff: chips, datacenters, power systems, and cooling equipment. It is closer to building a factory than launching a software product.
Here is how the $6.4 trillion roughly breaks down:
And 80 to 85% of this spending is coming from just five companies: Microsoft, Amazon Web Services, Google, Meta, and Oracle. These are not startups swinging for the fences. They are the most cash-generative businesses ever created, and they are spending at a record 77% of their operating cash flow on capital expenditure.
The hidden bottleneck: it is not the chips
Here is something surprising. The biggest constraint on AI growth right now is not semiconductors. It is electricity.
Modern AI clusters require staggering amounts of power. A traditional datacenter rack consumes maybe 10 to 15 kilowatts. An AI-optimised rack today needs 50 to 100 kilowatts. And the US and European power grids have interconnection queues stretching three to five years, meaning even if you wanted to build a new datacenter tomorrow, you might not be able to connect it to the grid until 2029.
Power and cooling equipment is forecast to grow at a 34% annual rate through 2030, faster than any other segment in the entire advanced computing stack.
This is why legacy datacenters are effectively obsolete for AI workloads. Blowing cold air over server racks cannot keep up with the heat output of modern GPU clusters. The industry is rapidly shifting to liquid cooling: running coolant directly over chips, or in some cases submerging servers entirely in dielectric fluid. The physics leave no other option.
Why this is not the dot-com bubble
It is a fair question. Trillion-dollar infrastructure bets made on a technology that is not yet fully proven. Does that not sound familiar?
The critical difference is where the money is coming from. During the dot-com era, spending was funded by speculative public markets and venture capital betting on future demand that had not yet materialised. Today's AI investment is funded almost entirely from the operating cash flows of profitable businesses responding to actual customer demand.
In 2025, the Big Five hyperscalers generated $576.9 billion in operating cash flow and spent $444.5 billion on capex. That is aggressive, but it is not reckless. These are rational businesses with real customers paying for real services, and the spending will continue as long as AI keeps delivering productivity gains.
The four things that could go wrong
Power and cooling
Grid constraints hit a ceiling
If interconnection queues do not shorten, the number of GPUs that can be deployed is physically capped regardless of how much capital is available.
AI fidelity
Models plateau before they are truly useful
If AI cannot achieve reliable expert-level reasoning, the business case for autonomous agents and physical AI collapses, reducing demand sharply.
Supply chain
Taiwan becomes a flashpoint
TSMC manufactures around 90% of advanced AI chips. Any disruption to Taiwan Strait stability could halt the global AI hardware supply almost overnight.
Concentration
One hyperscaler pulls back
80 to 85% of spending flows from just five companies. If any one of them faces shareholder pressure or a revenue shortfall, the ripple effects hit the entire semiconductor value chain
Where we actually are right now
The best way to understand the current moment is through the ARC-AGI-2 benchmark, a test designed to measure genuine reasoning ability rather than pattern matching. The average human scores 60%. A human expert scores 100%.
As of late 2025, Google's Gemini 3 scored 32% and OpenAI's GPT-5.1 scored 18%, at costs of $0.81 and $1.17 per task respectively. We are not at the finish line, but the trajectory is clear and the finish line is in sight.
The 60% threshold is the critical milestone to watch. Crossing it would validate the idea that AI can reliably replace entry and intermediate human knowledge work, and justify every dollar of infrastructure being built today.
A brief timeline of how we got here
November 2022
OpenAI launches ChatGPT-3.5. NVIDIA ships the H100 GPU. The era begins.
May 2023
NVIDIA hits a $1 trillion market cap after guiding revenue 50% above expectations. The market wakes up to what is coming.
January 2025
DeepSeek R1 proves Mixture of Experts architecture works at scale. The big bang moment for hyperscaler IT spending.
Q3 2025
Oracle reports a $455 billion AI backlog, up 360% quarter on quarter. TSMC begins work on the 2nm chip node.
Q4 2025
Gemini 3 Pro scores 32% on ARC-AGI-2. AMD forecasts a $1 trillion AI chip market by 2030.
The bottom line
The $6.4 trillion number is not a forecast pulled from thin air. It follows a straightforward logic chain. AI gets smarter as you give it more compute. The world's most profitable companies have the cash flow to buy that compute. And the productivity gains they are seeing justify continuing to spend. Until one of those three things breaks, the infrastructure buildout continues.
What is being built right now — the datacenters, the chip fabs, the power infrastructure — is the physical substrate of the next era of computing. Whether or not AI delivers on its most ambitious promises, that infrastructure will exist. The companies, investors, and countries that secured their place in it early will have a very significant head start.
The key metric to watch over the next two years: hyperscaler capex as a percentage of operating cash flow. If it stays above 60% and revenue growth holds, the bull case is intact. If either breaks, expect a rapid reassessment.
