Hook
Crypto Briefing dropped a bombshell: "US hyperscalers to invest over $750B in AI infrastructure this year."
That number is not just ambitious. It is structurally impossible.
Microsoft spent $80B on AI capex in fiscal 2025. Amazon, Google, and Meta combined another ~$170B. Total AI infrastructure spend across the four hyperscalers? Roughly $250B. The $750B claim is three times that—a mathematical hallucination pumped by a crypto-native outlet that should know better.
Context
The article appeared on Crypto Briefing, a platform built to cover token launches and DeFi exploits. Its pivot to AI infrastructure should already raise flags. The piece cites no primary source, no earnings call transcript, no analyst report. It just floats a number so large it can only be a typo, a deliberate lie, or an aggregation of multi-year projections masquerading as an annual figure.
But the market is hungry. AI euphoria is at peak froth. A number like $750B feeds the narrative that we are witnessing a once-in-a-generation capex cycle. Every crypto native who has survived an ICO bubble or a DeFi summer knows this pattern: an authoritative-sounding headline, no verifiable data, and a mob ready to bet on the story.
Core
Let me stress-test this number the way I stress-tested Compound Finance’s liquidation thresholds in 2020. I’ll start with physics.
A single NVIDIA B200 GPU costs ~$30,000 at retail. Each GPU consumes ~700W under load, plus cooling overhead. A standard hyperscale data center has a capacity of 150MW. To house $750B worth of infrastructure, assuming 70% of that spend goes to GPU servers (a generous ratio for pure compute), you get $525B in GPUs. That buys 17.5 million B200 units.
Seventeen million high-end GPUs. Each requiring power, networking, and physical space. The top limit of NVIDIA’s production capacity for B200 in 2025 is around 2 million units. You would need nearly a decade of NVIDIA’s total output to fulfill that single-year order. The supply chain breaks before the check clears.
Now power. 17.5 million GPUs at 700W each give a raw compute power draw of 12.25GW. Add networking, memory, and cooling, and you are looking at 18GW of continuous electrical load. That is the equivalent of 18 nuclear power plants dedicated solely to AI. The U.S. currently struggles to add 10GW of new capacity per year. The article assumes a tripling of national grid additions for one sector alone.
This is not investment. This is fantasy.
I have seen this before. In 2017, I reverse-engineered the TON whitepaper and found 60% of tokens allocated to insiders. The community didn’t want to look. In 2021, I traced wash trading on OpenSea—15 wallets driving $2M of fake floor volume. The market didn’t care. The pattern is the same: a headline crafted to justify a price move, not to convey truth.
Crypto Briefing’s $750B claim is the AI version of a fake TVL. It is narrative inflation designed to attract clicks and—if you follow the money—to pump AI-related tokens or stocks. The authors likely aggregated multiple analyst forecasts for 2025–2030 and called it a single year. Or they confused billions with millions. Either way, the ledger lies; the code tells.
Contrarian Angle
But the bulls have a point. The trend is real. All four hyperscalers are investing aggressively. Microsoft’s $80B is real. Amazon’s $75B is real. The macro is supportive. AI workloads are growing 2x annually. The infrastructure has to be built.
The contrarian truth: the narrative, not the number, is what matters. Even an exaggerated story can drive capital flows. If enough investors believe the $750B figure, they will buy NVIDIA, Vertiv, or even crypto assets tied to AI compute (like Render or Akash). Short-term price action follows perception, not reality.
I learned this during the Terra collapse analysis. The death spiral was inevitable after I recreated the mechanism in a sandbox. But the market kept buying Luna until the last block. The narrative sustained itself longer than the fundamentals. The same is happening here. The $750B myth will circulate for weeks before someone fact-checks it.
Takeaway
Demand primary sources. Every capex figure that matters is published in an SEC 10-K or an earnings call transcript. Spend five minutes searching. If you can’t find the original, treat the number as noise.
Friction reveals the true structure. The bottleneck is not money—it is physical: GPUs, power plants, cooling towers. Those constraints are not negotiable. Gravity doesn’t care about your narrative.
Watch the quarterly earnings calls. When Microsoft reports its AI revenue growth rate and compares it to its capex, you will see the spread widen or narrow. That spread is the real signal. The $750B headline is just the echo of a bubble waiting to pop.
Silence is the first red flag. When an outlet refuses to cite a source, they are trying to sell you a story, not inform you. History is just data waiting to be read—if you bother to read it.