Goldman Sachs wants you to believe $7.5 trillion floods into AI infrastructure by 2030. That is $1.5 trillion per year — 2.5x the entire global semiconductor market today, including every laptop, car, and phone chip on the planet.
I checked the contracts instead.
Over the past 90 days, decentralized compute networks settled real GPU workloads while AI-token markets bled speculative positions. Utilization climbed. Token velocity fell. The machines got busier. The traders got bored. That divergence between narrative and settlement is where I look first.
Code does not lie. Check the contract.
The projection, surfaced through Crypto Briefing, maps a five-year capex supercycle. Assumed allocation: 50-60% into AI silicon, 20-30% into data center construction, 10-15% into networking and storage, 5-10% into software middleware. Underneath sits a single assumption — scaling laws hold. Model parameters climb from trillions to tens of trillions. Iteration stays on a six-to-twelve-month cadence. Inference demand overtakes training by 2027, directing over 60% of spend to serving models rather than training them.
None of this is impossible. None of it is verified.
Investment banks do not issue numbers like this to describe reality. They issue them to coordinate it. The actual commitments live elsewhere: hyperscaler earnings calls, foundry capacity guidance, and increasingly on-chain infrastructure markets settling compute transactions the way commodity exchanges settle grain.
This is not the first time a bulge-bracket bank produced a number too large to fail. The 1996-2000 internet build-out consumed roughly $1.5 trillion and left a generation of dark fiber. The difference matters: dark fiber lasted two decades. AI chips depreciate in three to five years. Excess AI capacity does not wait to be lit. It books an impairment in year three.
Run the arithmetic. That is where the model breaks.
Assume 50% of the $7.5 trillion reaches silicon. That is $3.75 trillion in procurement. At Nvidia B200-class pricing — roughly $30,000 per accelerator — the market must absorb 125 million units. At 20 petaflops per chip, peak theoretical capacity reaches 2,500 zettaflops. Discount cluster efficiency to 50%, and you still get a 100,000x expansion over today's largest training clusters.
Then ask who pays for the output.
Global cloud infrastructure revenue sits near $600 billion annually. Even assuming every dollar becomes AI-native, the revenue base must triple to $2-3 trillion by year five to justify a 10% return on deployed capital. Hyperscaler capex historically runs 10-15% of revenue. A $1.5 trillion annual AI spend against a $600 billion revenue base inverts that ratio entirely — capital intensity goes vertical, the exact signature of the worst telecom overbuilds.
At GPT-4-class inference prices — $0.01 to $0.03 per token — generating $3 trillion in application revenue requires processing on the order of a quadrillion tokens per year. That is not adoption. That is a physics problem.
Funding compounds the problem. If public markets price in full deployment, every earnings miss becomes a repricing event. Valuation multiples across the AI complex — including AI-token proxies — already discount perfection.
The energy constraint is worse. Five years at that scale implies 1,500 to 2,000 gigawatts of installed AI compute. Annual consumption lands between 10 and 15 trillion kilowatt-hours — over 10% of global electricity today, before a single non-AI watt is generated. Grid build-out becomes the bottleneck. Grids do not move on five-year capex cycles. They move on fifteen-year regulatory ones. Expect nuclear restarts — the US and Japan are already moving. AI, not climate policy, will be the reason.
Water is the hidden constraint. A single 100-megawatt data center consumes what a small city uses. The 500 to 1,000 hyperscale sites this plan requires do not just need electrons. They need rivers.
I have applied this concentration check before. In 2021, I scraped 50,000 CryptoPunks transactions and found 60% of volume came from 20 wallets. The same test applies to AI revenue: trace the reported cloud base to actual paying inference workloads, and the concentration of customers is dangerously narrow. A handful of frontier labs and enterprise anchors are carrying the entire utilization curve.
I ran a comparable analysis in 2026 on decentralized compute markets. Render Network and Akash Network showed a clean pattern: compute-heavy AI workloads lifted network hash rate by 200%, while speculative trading volume dropped 15%. Real workloads arrived. Speculative capital left. Utility-backed tokenomics replaced memecoin momentum in real time.
That is the on-chain version of what institutional AI capex will look like in aggregate — except institutions face an extra constraint. NVIDIA holds over 80% of training silicon. Supply concentration that extreme means the $7.5 trillion scenario is not gated by demand. It is gated by advanced packaging, HBM memory, and substation transformers. CoWoS capacity takes two to three years to expand. Power transformers take four. Direct-to-chip liquid cooling moves from niche to default, and Vertiv's order book becomes a leading indicator. The physical supply chain caps the supercycle long before the revenue side does.
Follow the smart money, not the tweets. Smart money is currently renting GPUs for inference workloads, not buying the $7.5 trillion story.
Here is what the optimists get right, and it is uncomfortable: the prediction does not need to be accurate to be effective.
Goldman's forecast is a coordination signal. Hyperscaler boards read it. Capex committees benchmark against it. Microsoft, Google, Amazon, and Meta will keep committing tens of billions annually because the market prices the narrative of scale — even if utilization lags. If each major cloud player allocates $100 billion a year to infrastructure they cannot fill with paying workloads, the $7.5 trillion becomes not a prediction but a pact.
Jevons paradox cuts in the same direction. Efficient models do not reduce total compute demand — they expand it. A non-Transformer architecture could compress training costs by an order of magnitude and still trigger more deployment, not less.
But correlation is not causation. Crypto Briefing ran this story because the AI+Web3 token complex needs a macro tailwind. The same report celebrates $7.5 trillion of infrastructure spend while ignoring the 0.5-1% of AI budgets allocated to safety alignment, and the export-control regime splitting global AI into two disconnected ecosystems. Goldman does not issue forecasts in a vacuum — the same desk may underwrite the bonds that fund the build-out. That does not make the number wrong. It makes it interested.
Liquidity leaves before the crash hits. Watch whether revenue — not headlines — shows up first.
The marker to track is not Goldman's number. It is the quarterly guide. My shortlist, in order: NVIDIA's data-center revenue growth, currently near 200% year-over-year; hyperscaler capex guidance over the next two earnings cycles; frontier benchmark progress on MMLU and GPQA; and on-chain GPU utilization on Render and Akash. If the first decelerates, the forecast bends. If the labs stall, the model breaks entirely.
Treat every prediction without a falsification trigger as marketing. The falsification trigger here is the revenue guide.
Probabilities, not absolutes. Revenue catching up to $7.5 trillion within five years: 20-30%. A narrative-driven capex bubble forming inside the same window: higher than most analysts will admit.
The infrastructure will be built. The question is who controls the revenue when the build-out ends. On-chain data will identify the answer before the earnings calls do. It always does.