Goldman Sachs just projected $1.2 trillion in AI infrastructure capital expenditure by 2027. Buried inside that headline number is a technical assumption almost nobody is questioning: that the Scaling Law holds, uninterrupted, for another three years. And a physical constraint that makes the financial projection look almost conservative: energy.
I've spent the last month auditing hyperscaler capex disclosures and pulling power interconnect data from regional grid operators. Here's what the sell-side reports are not telling you.
The Capex Structure Nobody Decomposes
Strip away the branding and the $1.2 trillion is a hardware commitment, not a software one. Based on my audit of public disclosures from Microsoft, Google, Amazon, and Meta, accelerator silicon accounts for 40% to 60% of system cost. The rest is servers, data center shell, networking, and โ critically โ power and cooling.
That math puts roughly $500 billion to $700 billion flowing directly into semiconductor fabrication and AI accelerators. Implicit in that figure: a deployment in the range of 20 million GPU-class accelerators by 2027.
That is not a capex forecast. That is a structural vote of confidence in the Transformer architecture and the Scaling Law. If the industry believed model architecture was about to pivot to something non-Transformer โ state space models, linear attention, whatever comes next โ you would not see $700 billion committed to the current silicon paradigm.
I watched a version of this in 2017. During the Ethereum Homestead upgrade, I deployed testnet nodes and tracked gas optimization in real time. The lesson from that sprint was simple: infrastructure commitments telegraph technical assumptions long before anyone says them out loud. The $1.2 trillion AI number is doing exactly that.
Energy Is Not a Side Note. It Is the Constraint.
Here is where the report gets interesting, and where it stops short.
A single H100 draws roughly 700 watts. A B200 crosses 1,000 watts. Rack power density is now exceeding 150 kilowatts โ beyond what traditional air cooling can handle. Energy stopped being an operating cost line item and became a design constraint.
The report says energy is the key bottleneck. That is correct, but it undersells the implication. When energy becomes the binding constraint, it forces a technology roadmap whether the industry wants one or not.
Liquid cooling moves from optional to mandatory. Optical interconnect replaces copper inside the rack. Power conversion efficiency becomes a competitive moat measured in basis points, not percentages. I've seen this transmission chain before โ bottleneck forces efficiency, efficiency forces new commercialization windows.
Here's the data point I keep returning to. According to EIA estimates, U.S. data center electricity consumption climbs from roughly 150 TWh in 2024 to somewhere between 300 and 400 TWh by 2030. That is a load growth curve the grid has not seen in decades. Interconnection queues in major markets now run three to five years. In some regions, data center buildout is gated by substation capacity, not by chip supply.
The bottleneck is not the GPU. It is the transformer on the pole.
The Commercial Contradiction
Now the part that should worry anyone holding AI infrastructure exposure.
Hyperscaler capex is growing faster than AI revenue. The report flags 'pressure on revenue growth' at the hyperscalers. That is a polite way of describing a widening gap.
Run the math. The four major U.S. hyperscalers spent roughly $220 billion in capex in 2024. If the Goldman trajectory holds, that annualizes above $450 billion by 2027. AI-related revenue across cloud services, subscription products, and advertising incrementality was somewhere in the $50 to $100 billion range in 2024.
For that revenue to justify the asset base by 2027, you need a compound annual growth rate north of 50%. In some scenarios, closer to 100%.
That is possible. It is not guaranteed. And the report does not give us the revenue-side number that would make the $1.2 trillion a rational bet rather than a faith-based one.
There is a second pressure point. Free cash flow at these companies is strong but finite. Moving from $220 billion to $450 billion-plus in annual capex pushes free cash flow toward zero or negative. The gap gets financed with debt. The report acknowledges rising debt dependency. Capital structure, not chip supply, may end up being the real constraint on the AI buildout by 2027.
I learned this lesson the hard way during DeFi Summer 2020. I rushed into Yearn vaults chasing yield, and when withdrawals froze during a gas war, I documented the block-by-block congestion in real time. Speed without a read on the underlying structure is fatal. The AI capex story has the same shape โ impressive velocity, under-examined plumbing.
The Unasked Question: Who Pays for the Power?
The energy bottleneck has a social dimension nobody is pricing.
When data centers pull hundreds of terawatts onto local grids, utilities have to build transmission and distribution capacity. Those costs get socialized across the rate base. Residential and small commercial customers absorb the bill.
Expect 'compute versus households' to become a public policy fight in the U.S. and parts of Europe within 24 months. It is already happening in Northern Virginia and Ireland. The report does not touch this, and it is the single most likely source of regulatory friction against the $1.2 trillion number.
There is a geographic reshuffle coming too. Regions with surplus hydro or nuclear โ the Nordics, the U.S. Southeast โ gain a structural advantage in siting. Data center location stops being about latency and fiber and starts being about megawatts. The map of AI infrastructure is about to be redrawn by power availability, not network topology.
The Arms Race Trap
One more angle the sell-side misses entirely.
The hyperscalers are simultaneously the largest customers of AI model labs โ through investments in OpenAI, Anthropic, and others โ and the largest suppliers of AI compute. They sit on both sides of the table. Capex forecasts in this structure are partly self-fulfilling: spending creates the moat, the moat justifies more spending.
That is not necessarily irrational. But it has a predictable failure mode. When every major player expands simultaneously, relative advantage cancels out. The logic of 'spend or die' can push industry-wide returns on capital down even as absolute spending rises. The $1.2 trillion could be defensible in isolation and value-destructive in aggregate.
The real question is not whether the money gets spent. It is whether the return on that money survives the arms race.
What I'm Watching Next
Three signals will tell us whether the $1.2 trillion is a buildout or a bubble inflating.
First, the training-versus-inference split inside that number. If inference is taking a larger share, chip selection shifts away from top-end training silicon and the power model changes entirely. Nobody has published this breakdown.
Second, liquid cooling penetration and in-house ASIC share by 2027. Both are proxies for how seriously the industry is treating the energy constraint rather than just financing around it.
Third, the revenue number. Every capex forecast should be published next to the revenue it assumes. Until we see both, the $1.2 trillion is a headline, not an analysis.
The cheetah does not chase every gazelle. It picks the one it can catch. Right now, the AI infrastructure trade is moving faster than the data that would justify it. That is exactly when you slow down and count the watts.