The number is too clean.
Seven point five trillion. Over five years. AI infrastructure.
Goldman Sachs dropped this projection without blinking. No qualifying footnote. No contingency table. Just a headline meant to reshape the tech industry.
But I've spent the last eight years auditing smart contracts that promised the moon and delivered a rekt. I've traced integer overflows through vesting schedules that would have drained millions. I've watched NFT standards fail on interoperability because someone assumed ERC-721 and ERC-1155 would play nice.
Numbers that clean are always hiding something.
Let's open the hood.
Context
The prediction: $7.5 trillion in capital expenditures on AI infrastructure—chips, data centers, networking, cooling, power—between 2026 and 2030. That's $1.5 trillion per year. For reference, the entire global semiconductor market today is roughly $600 billion. The cloud computing market is about $600 billion in annual revenue.
Goldman is saying we'll spend two and a half times the current chip market on new AI hardware alone, every year, for five years.
The assumption is that AI model scaling continues unabated. That GPT-5, GPT-6, and whatever follows will need exponentially more compute. That inference at scale will dwarf training. That autonomous agents, robotics, and AI-driven workflows will generate enough economic value to justify this spend.
The unspoken assumption is that none of the fundamental physical and economic constraints will bite back.
Core Analysis
Let's break this down the way I break down a yield contract before I deploy it.
The Chip Math
Current flagship: NVIDIA B200. 20 petaflops of training compute. Street price around $30,000 per GPU.
If 50% of that $7.5 trillion goes to chips—$3.75 trillion—that buys 125 million B200s. Even assuming a 50% cluster utilization (Model FLOP Utilization, or MFU), that's effective compute of 1.25 zettaflops per second. That's roughly 10,000x the training cluster used for GPT-4.
But chip capacity isn't elastic. TSMC's CoWoS advanced packaging is already backordered two years. HBM memory production is constrained. The lead time for a new fab is four years minimum. Goldman's model essentially assumes we can scale fab capacity 10x in five years. That's possible. But it's never been done before.
The Energy Wall
A B200 draws over 700 watts under load. Multiply by 125 million chips and you get 87.5 gigawatts of continuous draw. Add data center overhead—cooling, networking, lighting—and you're looking at 150–200 gigawatts of new power demand.
The entire US grid today has about 1,200 gigawatts of generation capacity. So we're talking about building the equivalent of 15% of the current US grid for AI alone. New nuclear plants take 15 years. Renewables can't deliver baseline load. Natural gas has methane leakage issues.
Goldman's number assumes energy is a solved problem.
The Cooling Reality
Every AI chip above 300 watts forces liquid cooling. The industry standard for air cooling maxes out around 250 watts per square foot. Data center operators are scrambling to install direct-to-chip liquid cooling and immersion tanks. Companies like Vertiv and CoolIT see demand doubling every 18 months.
But retrofitting existing data centers is like patching a Solidity contract after the exploit. You can do it, but you'll never fully remove the entropy.
The ROI Contradiction
Here's where it gets uncomfortable.
Assume the entire $7.5 trillion is capitalized and depreciated over five years. That's $1.5 trillion in annual depreciation. Add operating expenses—power, staff, network—at roughly 20% of CapEx, and you need annual revenue of $1.8 trillion just to break even.
The entire cloud market today generates about $600 billion in revenue. To make AI infrastructure pay, AI services would have to consume three times the entire cloud market within five years.
That's not impossible. But it requires adoption rates that far exceed any previous technology wave. The internet took 15 years to reach that scale. Mobile took 10. AI would need to do it in 5.
The Historical Precedent
The internet bubble of 1996–2000 saw $1.5 trillion in fiber optic infrastructure investments. By 2002, 90% of that fiber was dark. Dark fiber doesn't depreciate quickly. AI chips do—three to five years, and they're obsolete.
If AI revenue doesn't materialize, those $30,000 B200s become landfill. That's a risk that Goldman's spreadsheet doesn't seem to price.
Contrarian Angle
Here's the part nobody wants to talk about: this prediction serves a narrative.
Goldman sells products. They underwrite bonds for data center REITs. They advise on M&A for chip companies. They have a long book on AI.
Crypto Briefing, which published this news, serves an audience that buys into large-scale transformative narratives. The same audience that FOMO'd into AI tokens last cycle.
The narrative hides a structural vulnerability: overcapacity.
We saw this in crypto. During the 2021 bull run, we built out massive GPU mining farms. By 2022, they were uneconomical. AI chip investment faces the same risk—only at 1,000x the scale.
The gas isn't free. It's the friction of poor architecture—architectural assumptions that ignore the hard constraints of physics and economics.
The Geopolitical Blind Spot
Goldman's $7.5 trillion assumes global supply chains remain open. But export controls on advanced chips to China are tightening. China will build its own AI infrastructure, but at 30-50% lower efficiency due to inferior lithography. Two separate AI stacks are forming, each with its own underutilization risk.
If the US buildout assumes selling chips to China, that revenue won't come.
If the Chinese buildout assumes access to Western tooling, that tooling won't come.
The Security Dimension
Vulnerabilities aren't just in code; they're in assumptions.
A $7.5 trillion AI infrastructure means every layer of the stack becomes a target. The chip supply chain. The data center physical security. The power grid. The cooling systems. The orchestration software.
We've already seen attacks on electrical grids. A coordinated attack on 100 large AI data centers could disrupt power to entire regions.
And then there's the AI agent layer. I recently identified a prompt-injection vulnerability in an oracle feed that would have let a malicious agent drain a zk-rollup's bridge. The damage was simulated at $2 million. Scale that to a $7.5 trillion ecosystem, and the attack surface becomes planetary.
What Nobody Modeled
AI efficiency gains. If model efficiency improves 10x (through better architectures, quantization, distillation), then the compute required for a given intelligence level drops 10x. That would crater the demand for new hardware.
Jevons paradox applies: cheaper compute increases usage. But the elasticity isn't infinite. There are only so many useful AI queries the economy can absorb.
The combination of efficiency improvements and adoption saturation could leave half that 7.5 trillion stranded.
Takeaway
Goldman's prediction is a scenario, not a forecast. It's the upper bound of a bull case that assumes everything breaks right: no energy bottlenecks, no chip supply constraints, no geopolitical fractures, no efficiency breakthroughs, no regulatory speedbumps.
I've seen this playbook before. In 2017, every ICO projected hockey-stick user growth. Most didn't survive the bear.
Code that doesn't earn its keep gets forked or abandoned. Infrastructure that doesn't generate returns gets written off.
The real question isn't whether we'll spend $7.5 trillion. It's whether we'll spend it wisely—or repeat the same mistakes we made in crypto, on a global scale.
Optimization isn't about squeezing gas; it's about respecting the user's money. The same applies to AI infrastructure. If you can't fix the economic model, you don't understand the technology.
And right now, Goldman's model looks like it skipped the audit.