"$700B is the greatest capital misallocation in history."
That's what AI critic Gary Marcus posted on X right after Big Tech's Q1 earnings dropped on April 30. That same day, Goldman Sachs put out an analysis noting that "$7.6T in cumulative 2026–2031 investment could shift by hundreds of billions if just a few assumptions change". Both sides are staring at the same number — and landing in completely opposite places: "bubble" versus "sustainable."
What Is It?
In 2026, Alphabet, Amazon, Meta, and Microsoft are on track to spend roughly $700B combined on AI capex. Microsoft: $190B. Alphabet: $180–190B. Amazon: $200B. Meta: $125–145B. That's about 0.8% of US GDP — roughly the size of Switzerland's entire annual output.
The pace is steep. From roughly $405B at the start of 2025, that's a 62%+ jump in a single year. Goldman Sachs Global Institute's baseline model projects $765B in 2026 rising to $1.6T annually by 2031 — $7.6T cumulative over six years.
The hit to earnings is just as striking.
- Amazon Q1 free cash flow: $1.2B (down 95% YoY)
Capex ate nearly every dollar of operating cash in a single quarter. - Alphabet Q1 free cash flow: $10.1B (down 47%)
Still profitable — but cut nearly in half. - Meta: issuing $20–25B in additional bonds (on top of last year's $30B)
They're borrowing because cash isn't keeping up. CreditSights already rates Meta bonds at underperform. - All four companies: "capacity-constrained"
Every earnings call included some version of "we can't handle all the demand coming in." Datacenter and chip supply is the bottleneck — they'd build more if they could.
Why Does Marcus Call It a Misallocation?
Gary Marcus makes four arguments.
(1) No company is generating meaningful profit from AI. (2) There's no real moat — LLMs are commoditizing. (3) Price wars are inevitable. (4) Customers aren't seeing big ROI either. His conclusion: $700B poured in, and what you get is something that'll soon be a free commodity.
One more data point reinforces the argument. According to an American Affairs analysis, at the current pace, hyperscalers need to generate $2T+ in revenue by 2030 to justify planned capex. With OpenAI and Anthropic combined at roughly $20B ARR in 2025, that means 100x growth in five years.
So What's Goldman's Take?
Goldman's answer: "bubble vs. not-bubble" is the wrong frame. The real point is that $700B isn't a fixed number — it's a highly conditional figure, deeply sensitive to four supply-side assumptions.
| Assumption | Current Baseline | Impact if It Shifts |
|---|---|---|
| Silicon useful life | 4–6 years | A swing from 3 to 7 years alone moves cumulative capex by hundreds of billions |
| Datacenter cost per MW | $15M/MW (up from $10M historically) | Moving between $11M–$19M shifts datacenter capex by trillions cumulatively |
| Chip architecture mix | NVIDIA GPU at 75%+ (75% margins) | Rising ASIC share — depending on demand elasticity, total spend shrinks or holds flat |
| Build-out elongation | Power, labor, and equipment bottlenecks = base case | If bottlenecks worsen → supply friction converts to demand-side doubt and investment pullback |
One trap Goldman explicitly flags: the gap between accounting depreciation and actual economic useful life. Even if chips are depreciated over five years on paper, if NVIDIA delivers a step-function performance jump every year, they're economically obsolete in three. The asset is still on the balance sheet — but nobody wants to buy it.
Goldman also specifies what doesn't move the total much — training vs. inference mix, memory price swings, behind-the-meter power vs. grid. Those determine who captures the margin, not the $7.6T headline number.
What Should Practitioners Actually Track?
$700B isn't a distant abstraction. Chip, power, and datacenter supply chains all feel it — and SK Hynix (HBM), Samsung (foundry), and KEPCO (power supply) all sit inside that capex graph.
Here's what to watch.
- Step 1: Know what's actually a signal
The capex announcements themselves aren't the signal. The real signals are: silicon depreciation policy changes (if Amazon extends from 4 to 6 years, short-term cash flow looks better — but it's a flag that accounting is masking obsolescence risk), HBM unit pricing (a leading indicator of capex pace), and power interconnection queue length (when that starts growing, elongation has begun). - Step 2: Track the revenue gap directly
Check the $2T hypothesis against current ARR every quarter. Whether OpenAI, Anthropic, xAI, and Google AI products are growing at 100% YoY or slowing to 30% is the most critical input variable for your own business scenarios. - Step 3: Build two AI cost structure scenarios
(a) Price war scenario: API pricing drops 10x — what happens to your cost structure? (b) Capacity-constrained scenario: prices hold or rise — can you move your dependency elsewhere? Real decisions are a weighted average of both. - Step 4: Audit your single-vendor chip dependency
If you have workloads running 100% on NVIDIA, look at migration feasibility under an ASIC shift scenario before you need to. Rising ASIC share means available models will vary by cloud provider.
Getting Started
- Step 1: Drop the bubble/no-bubble frame
You can't make any useful decision with that framing. Instead, ask: "Under what assumptions does $700B become $400B — or $1T?" - Step 2: Track four variables every quarter
(1) Free cash flow and debt issuance trends across the Big Four. (2) NVIDIA silicon depreciation policy. (3) HBM unit pricing. (4) Datacenter cost per MW. These are the core inputs to Goldman's model. - Step 3: Build two AI cost scenarios
For both the price war and capacity-constrained scenarios, model your company's P&L at 6 and 12 months out. - Step 4: Monitor the revenue gap quarterly
Track combined revenue from OpenAI, Anthropic, Google AI products, and Microsoft Copilot against the $2T hypothesis. If the gap widens past a certain threshold, the market structure itself starts to wobble. - Step 5: Map your supply chain exposure
Beyond the direct supply chain players (SK Hynix, Samsung, KEPCO, SK Innovation), identify which of your own AI-dependent workloads — AI call centers, AI search, AI recommendations — are exposed to which phase of the capex cycle.
Deep Dive Resources
Goldman Sachs: Tracking Trillions The quantitative scenario analysis behind the $700B figure — showing exactly how each assumption can move cumulative capex by hundreds of billions. Covers silicon useful life, cost per MW, ASIC mix, and build-out elongation in depth. goldmansachs.com
MarketWatch — Marcus's misallocation claim and Big Tech's free cash flow shock Post-earnings recap from April 30. Amazon -95%, Alphabet -47%, Meta bond issuance — all the numbers in one place. marketwatch.com
American Affairs — Understanding the LLM Bubble A direct challenge to the $2T revenue hypothesis. Simulates what happens if hyperscalers can't justify the spend by 2030 and how different unwind scenarios play out. americanaffairsjournal.org
