The AI race is no longer just about models.
It is about power grids, debt markets, GPUs, land, cooling systems, utility mergers, and trillion-dollar forecasting assumptions that may decide who survives the next decade.
Over the last few weeks, three developments quietly revealed how extreme the infrastructure buildout has become:
- NextEra agreed to acquire Dominion Energy for $67 billion
- Anthropic committed more than $40 billion in compute spending with xAI
- CoreWeave reported massive growth while carrying $25 billion in debt
Different companies. Different industries. Same underlying belief:
AI demand is going to be so large that the world needs to rebuild its entire infrastructure stack around it.
And the numbers now involved are unlike anything the tech industry has ever attempted before.
The New Arms Race Is Physical
For years, software scaled cheaply.
A startup could add millions of users without building factories, power plants, or railroads. AI changes that equation completely.
Training and running large AI models requires:
- Massive GPU clusters
- Dedicated energy infrastructure
- Cooling systems
- Fiber connectivity
- Land near power access
- Long-term chip supply agreements
This is no longer a software cycle.
It increasingly looks like an industrial cycle.
The hyperscalers alone are expected to spend roughly $725 billion in capex in 2026, with most of that tied directly to AI infrastructure.
That includes companies like:
- Amazon
- Microsoft
- Alphabet
- Meta
- Oracle
Some of these firms are now spending nearly half their annual revenue on infrastructure.
Historically, those kinds of capex-to-revenue ratios belonged to utilities and telecom operators with guaranteed cash flows.
The difference is that utilities operate on regulated returns.
AI companies operate on projections.
And projections can break.
The Deal That Explains Everything
The most revealing part of the entire AI buildout may actually be Anthropic’s compute agreement with xAI.
According to disclosures tied to SpaceX’s filing, Anthropic committed to paying $1.25 billion per month for access to xAI’s Colossus infrastructure cluster in Memphis.
That works out to over $15 billion annually.
The cluster reportedly includes:
- More than 220,000 NVIDIA GPUs
- Roughly 300 megawatts of power capacity
- Multi-year infrastructure commitments running through 2029
What makes this fascinating is not just the size of the deal.
It is why the deal exists.
Anthropic CEO Dario Amodei openly admitted earlier this year that the company planned its business around 2x to 3x growth.
Instead, demand exploded roughly 80x.
That kind of forecasting miss creates chaos.
If you underbuild, customers hit outages and competitors gain market share.
If you overbuild, you are stuck paying for infrastructure that may not generate enough revenue to justify the investment.
That is the central tension inside the AI economy right now.
Everyone is building aggressively because the cost of being too small may be fatal.
But being too large could also become fatal.
The Entire Supply Chain Is Leveraged to the Same Assumption
The most important part of this cycle is how interconnected the system has become.
Every layer is effectively betting on the growth of the layer above it.
Utilities are betting on data centers
NextEra’s acquisition of Dominion is fundamentally a power demand trade.
Northern Virginia remains the largest data center market in the world, and utilities increasingly view AI infrastructure as long-duration electricity demand similar to industrial manufacturing.
The combined company is positioning itself around more than 130 gigawatts of large-load opportunities.
That is enormous.
Data centers are betting on hyperscalers
Companies like CoreWeave are spending aggressively because they believe Microsoft, OpenAI, Anthropic, and others will continue consuming compute at exponential rates.
The problem is that this business model requires enormous leverage.
CoreWeave reportedly carries around $25 billion in debt, much of it at double-digit interest rates.
That creates very little room for forecasting mistakes.
If utilization slips or contracts get delayed, the financial pressure compounds quickly.
Hyperscalers are betting on AI demand
Amazon alone plans to spend around $200 billion in capex next year.
Meta, Microsoft, and Alphabet are all operating at spending levels previously considered unimaginable for software businesses.
They are making the calculation that missing the AI transition would be more dangerous than overspending during the buildout phase.
AI model companies are betting on enterprise adoption
At the top of the chain sits the actual monetization question.
Will enterprises spend enough on AI products to justify all of this infrastructure?
That remains the biggest unknown.
The demand today is real.
The question is whether the current pace of growth becomes durable enough to support the capital structure being built underneath it.
The Forecasting Problem Nobody Can Solve
One of the most important insights came from Goldman Sachs’ recent infrastructure analysis.
The firm estimated that cumulative AI infrastructure spending could reach $7.6 trillion between 2026 and 2031.
But buried inside the report was a more important point:
Small assumption changes massively alter the economics.
For example:
- If GPU lifespans shrink from five years to three years
- If utilization rates fall
- If power efficiency assumptions change
- If AI demand grows slower than expected
…the entire profitability structure shifts.
A single variable adjustment can move industry costs by hundreds of billions of dollars.
That makes forecasting incredibly difficult because nobody has historical precedent for this scale of AI infrastructure demand.
This is not cloud computing 2.0.
This is an entirely new industrial layer being built in real time.
Why The Market Still Keeps Spending
Despite all the risks, the spending has not slowed.
In fact, it continues accelerating.
Why?
Because so far, the pessimists have mostly been wrong.
Cloud demand remains strong.
AI workloads continue increasing.
Enterprise adoption is growing.
GPU shortages still exist in many markets.
And hyperscalers continue reporting strong revenue growth tied directly to AI products.
From their perspective, underinvesting may be the bigger risk.
If AI truly becomes the next computing platform, companies that fail to build enough infrastructure early could permanently lose strategic positioning.
That fear is driving the cycle.
Not necessarily irrational optimism.
But fear of irrelevance.
What Happens If The Math Is Wrong?
This is ultimately the question the market is trying to answer.
Not whether AI matters.
That debate is largely over.
The real question is whether infrastructure spending has overshot the actual monetization curve.
Because if demand slows even slightly:
- Debt-heavy infrastructure companies get squeezed
- GPU pricing weakens
- Power demand assumptions soften
- Long-term contracts get repriced
- Capex cycles compress
And since every layer depends on the one above it, pressure can travel quickly through the chain.
But if demand continues compounding at current rates, then today’s spending could eventually look conservative in hindsight.
That is what makes this moment so unusual.
The industry is effectively building infrastructure for a future that nobody can confidently model yet.
The Bigger Takeaway
The AI boom is no longer just a technology story.
It is becoming a capital allocation story.
A power infrastructure story.
A debt market story.
A geopolitical story.
And maybe most importantly, a forecasting story.
Because the difference between being early and being wrong may come down to just a few years.
Right now, the entire ecosystem is making trillion-dollar decisions based on assumptions about demand curves that have never existed before.
The demand appears real.
The uncertainty is whether anyone can size it correctly.
And that may end up defining the winners and losers of the AI era more than the models themselves.