Goldman Sachs Projects $5.3 Trillion AI Data Center Spending by 2030

https://www.goldmansachs.com/insights/articles/private-markets-expected-to-have-growing-role-in-data-center-financing

@rohanpaul_ai The $5.3T isn't the scary part. It's that much of it is GPUs depreciating on a 3-year clock, funded by debt on a 10-year clock — financing short-lived assets with long-dated money rarely ends quietly.

@rohanpaul_ai Lol, GDP would have to be clocking 10-20% growth in perpetuity to even have a chance of justifying that kind of spend with a generous ROI.

Yet another GS "please select us for your IPO" marketing pitch disguised as market research.

@rohanpaul_ai The headline is $6B.

The market implication is that compliance-based sanctions relief is becoming a key tool of economic diplomacy.

@rohanpaul_ai The biggest AI constraint may no longer be GPUs.

It's whether enough capital, power, and infrastructure can be mobilized to support the buildout.

@rohanpaul_ai to be or not to be a bitch of $NVDA, crazy that bottleneck of supply chain will break hyperscalers in next 2-3 years.

AI's next margin fight is watts per answer.

If most data-center energy is going into chips, the scarce resource is not just GPUs. It is useful compute per watt.

That changes the scoreboard. Better models matter, but so do cooling, utilization, power contracts, and inference efficiency.

The winners will not just buy more chips. They will turn each megawatt into more paid work, with margins that survive real usage.

Very interesting assessment from Goldman Sachs on AI spending:

"Hyperscalers will need financing from across markets, structures, and currencies as they potentially bump up against saturation consideration in liquid credit markets and issuers concentration constraints."

That means the spending has already outrun the capital willing to fund it on normal terms. Pushing the debt into private credit, SPVs and asset-backed vehicles scatters the risk to where it's harder to see. When you have to invent new ways to finance something, the easy money is already gone.

@rohanpaul_ai 👇

@bbxjasper @rohanpaul_ai Yep. People comparing it to dotcom and railroads bubbles are wrong, because after those bubbled popped the infrastructure stayed long term. With data centers that won't be true

@rohanpaul_ai Taps starts playing in the background 🎶

@rohanpaul_ai $5.3T over 5 years is about $3B a day—wild to think about. The real challenge isn’t building the infrastructure, but whether inference demand grows fast enough to make all that silicon worth it.

This is the same mechanism behind the recent hyperscaler debt binge and buyback cuts, the financing channels that funded AI capex so far are hitting real limits, concentration risk, market saturation, issuer caps, not unlimited appetite for hyperscaler debt.

One precision note: Goldman's slide talks about financing capacity constraints, not capex outpacing physical construction. That second claim is the post's own inference, not a quote.

@rohanpaul_ai If they can’t find the credit to build it how are they expecting to find customers that can buy the service?

@rohanpaul_ai How much profit is “AI” actually making to cover such mega expenditure

I can’t see it being profitable without a step change in the quality of models

@rohanpaul_ai $5.3 trillion through 2030 and goldman is already flagging issuer concentration risk in the bond market. nvidia, alphabet, amazon, meta, oracle, and salesforce can't keep tapping the same investors indefinitely without pricing pushback

@rohanpaul_ai Surprising take from him .. Capex boom then was telco .. and Cisco etc were booming due to that

@rohanpaul_ai The tell is in your last line: capex estimates rising faster than construction. The binding constraint was never dollars — capital is fungible. It's megawatts and interconnect queues that clear in 2029. A $5.3T number is really a power-and-permitting story in a finance costume.