First Principles · AI Economy · ROI Analysis

Are We in an
AI Bubble?

The answer isn't vibes — it's ROI. We track actual revenue generated by AI against the cost of building the infrastructure beneath it. The numbers tell a surprising story.

First Principles Analysis5-Layer Economic ModelSources: Goldman Sachs, McKinsey, Gartner, Deloitte, IBM
Statistics refreshed 2026-08-31 · Updates daily at 6am PST

2026 AI Economy — Live Scorecard

Total Infra Capex 2026E

$340B

Hyperscaler spend

Total AI Revenue 2026E

$985B

Across all streams

Full Economy Net Surplus

$+380B

Value created vs invested

Overall ROI Multiple

1.44x

Value per $1 invested

First principles — what makes a bubble?

A bubble is when capital flows into an asset class faster than that asset class can generate returns to justify the flow. The railroad bubble of the 1870s, the dot-com bubble of the late 1990s, and the housing bubble of 2007 all share the same structure: investment outpaced demonstrated value. The honest question is not "is AI exciting?" — it clearly is. The question is: is the $700B+ being deployed annually into AI infrastructure generating $700B+ of real, measurable economic value? And critically: at which layer of the stack is the value actually appearing? Because the answer is not uniform — and that asymmetry is the whole story.

The verdict is in the math, not the narrative. We track both sides of the ledger, layer by layer.

The Question Everyone Is Actually Asking

The AI bubble question is deceptively simple. Bubbles occur when asset prices (or investment levels) disconnect from underlying economic value. Tech bubbles happen when companies spend enormous amounts of capital on infrastructure that doesn't generate commensurate returns.

For AI, the question is therefore specific: is the $700B+ being spent annually on AI infrastructure generating $700B+ of real economic value?

"The question is not whether AI is real. The question is whether the economic returns justify the investment — and at what layer of the stack that return actually materialises."

— First Principles Analysis, AI Economy Document, March 2026

To answer this properly, we need to understand what AI actually is at a physical level — because the bubble question is really a question about whether a massive physical infrastructure buildout will generate enough return.

What AI Actually Is — First Principles

Traditional software follows instructions. A calculator does exactly what the programmer told it to do — nothing more. AI is different: instead of following pre-written rules, AI learns patterns from enormous datasets and uses those patterns to make predictions, generate text, recognize images, or take actions.

That learning process — called training — requires billions of simple math operations in parallel, over and over. A single large model like GPT-4 required months of continuous computation on thousands of specialized chips running 24 hours a day.

The Chain of Dependencies

1.More intelligence → more computation
2.More computation → more chips
3.More chips → more data centers
4.More data centers → more power
5.More power → energy becomes the foundation of everything

This chain is exactly what Jensen Huang described as his "five-layer cake" at Davos, January 2026. [6]

Jensen's Five-Layer Cake — Layer by Layer ROI [6]

The most important insight in understanding the AI bubble question is that ROI is not uniform across the stack. Some layers are structurally profitable. Others are structurally loss-making but strategically essential.

L5
ApplicationsValue layer
Cost: $90B
Rev: $680B+
+$590B
L4
AI ModelsClosing fast
Cost: $80B
Rev: $60B
−$20B
L3
Cloud InfrastructureCloud surplus
Cost: $200B
Rev: $230B
+$30B
L2
Chips & ComputingInternalizing
Cost: $240B
Rev: $165B
−$75B
L1
EnergyPure cost
Cost: $90B
Rev: $0
−$90B

Source: Jensen Huang, NVIDIA / WEF Davos 2026 [6] · All figures 2026E estimates

The Infrastructure Buildout — Scale and Cost [7][10]

Five companies — Amazon, Microsoft, Alphabet, Meta, and Oracle — are driving a pace of capital investment with no historical precedent. The cumulative 5-year investment of $920B is more than double what was spent 2022–2024 alone. Goldman Sachs projects $1.15 trillion in hyperscaler capex across 2025–2027 alone. [1]

Category20212022202320242025E2026E5-yr Total
Energy & power$10B$15B$22B$30B$50B$90B$217B
Chips & accelerators$20B$28B$47B$110B$195B$240B$640B
Data center construction$28B$35B$55B$80B$120B$200B$518B
Networking & storage$17B$25B$36B$50B$90B$125B$343B
Software & other$8B$12B$15B$20B$30B$45B$130B
TOTAL CAPEX$83B$115B$175B$290B$485B$340B$920B

Sources: Amazon, Microsoft, Alphabet, Meta, Oracle Q4 2025 earnings · Goldman Sachs [1][7]

What's Actually Coming Back — AI Revenue [1][7]

Revenue from AI flows from four distinct streams. The most explosive is Generative AI (SaaS + API) — projected at $288B in 2026, up from near-zero in 2022. Cloud AI revenue reaches $315B. Total AI revenue of $985B tracks within ~3% of total infrastructure capex — the key statistic that defines whether this is a bubble.

Revenue stream20212022202320242025E2026E5-yr Total
Traditional AI software$36B$55B$75B$110B$155B$205B$636B
Cloud AI (AWS/Azure/GCP)$22B$35B$55B$95B$160B$230B$597B
Generative AI (SaaS+API)$1B$4B$13B$36B$80B$134B
AI hardware / chip sales$12B$18B$47B$90B$125B$165B$457B
TOTAL AI REVENUE$70B$109B$181B$308B$476B$985B$2450B
Revenue vs. Capex ratio: 0.84× (2021) → 0.95× (2022) → 1.03× (2023) → 1.06× (2024) → ~0.97× (2025E) → ~0.97× (2026E). The slight dip in 2025–2026 reflects a surge in capex ahead of the application revenue wave — consistent with infrastructure build cycles, not bubble dynamics.

Enterprise ROI — The Real Answer [2][3][4][5]

The hyperscaler supply-side tells only half the story. The other half is what enterprises — the buyers — are actually getting back. This is the demand-side story and it is the most important data in the entire analysis.

$525B

Enterprise Spend 2026E

$892B

Value Generated 2026E

$+367B

Net Enterprise ROI 2026E

79%

Orgs with AI in Prod

Metric20212022202320242025E2026E5-yr
AI software/SaaS spend$12B$18B$28B$52B$95B$140B$345B
Cloud AI/API consumption$8B$12B$22B$45B$85B$120B$292B
Internal build + consulting$21B$28B$37B$50B$75B$77B$288B
Labour/ops cost savings$14B$22B$38B$80B$170B$280B$604B
Revenue uplift from AI$7B$12B$22B$45B$95B$170B$351B
Ops efficiency gains$13B$22B$37B$75B$140B$225B$512B
TOTAL VALUE GENERATED$34B$56B$97B$200B$405B$892B$1,467B
Value per $1 spent$0.83$0.97$1.11$1.36$1.59$2.00$1.59 avg

Key Enterprise ROI Findings [2][3][4][5]

$4.20average ROI per $1 invested for early adopters
$13.50ROI per $1 for top-performing organizations
74%of organizations report AI initiatives meeting or exceeding ROI expectations
31%average productivity gain for 2026E across all enterprise deployments

Crossover year: 2023 — enterprises first generated more value from AI than they spent on it.

The Master Scorecard — Full AI Economy

Combining hyperscaler infrastructure, hyperscaler revenue, and enterprise value gives the complete picture. The cumulative net economic surplus from 2021 to 2026 is $+622B — meaning the AI economy as a whole has created more value than it has consumed.

Dimension20212022202320242025E2026E
Total AI infrastructure capex$83B$115B$175B$290B$485B$340B
Total AI revenue$70B$109B$181B$308B$476B$985B
Hyperscaler net position−$13B−$6B+$6B+$18B−$9B$-55B
Total enterprise AI spend$41B$58B$87B$147B$255B$525B
Total value generated$34B$56B$97B$200B$405B$892B
Enterprise net ROI−$7B−$2B+$10B+$53B+$150B$+367B
TOTAL INVESTMENT$124B$173B$262B$437B$740B$865B
TOTAL VALUE CREATED$104B$165B$278B$508B$881B$1245B
NET ECONOMIC SURPLUS−$20B−$8B+$16B+$71B+$141B$+380B
Overall ROI (value/invest.)0.84×0.95×1.06×1.16×1.19×1.44x

So — Are We in a Bubble?

First principles requires us to separate where we look from what we conclude. The evidence points to three distinct layers of truth:

NOT a bubble at the application layer (L5)

Layer 5 — applications — generated $680B+ in revenue against $90B in cost. That's a +$590B surplus. Drug discovery, AI copilots, ad targeting, logistics optimization: these are generating provable, measurable ROI. Enterprises that crossed over to net-positive AI ROI in 2023 are now generating $+367B in net surplus in 2026.

SPECULATIVE at the infrastructure layers (L1–L4)

Layers 1–4 are collectively loss-making on a standalone basis. The $700B+ annual capex is not fully covered by infrastructure-layer revenue. This is not irrational — it follows the exact pattern of the internet buildout (1993–2000), mobile (2004–2010), and cloud (2010–2015). Infrastructure gets built before applications generate returns. The question is timing.

The risk: application returns must scale faster than capex

If application-layer revenue growth plateaus while capex continues accelerating, the overall ROI multiple compresses. The current 1.44x overall multiple is healthy but thin. The bubble risk is not that AI doesn't work — it clearly does. The risk is that the rate of return growth doesn't keep pace with the rate of capital commitment growth.

The Net Honest Answer

No, we are not in a classic bubble. A bubble requires assets priced above their fundamental value with no path to justified returns. AI has a documented, measurable path to returns that is already positive at the application and enterprise layer.

What we are in is a classic infrastructure boom — the same dynamic as railroads in the 1870s, electricity grids in the 1920s, the internet in the 1990s, and mobile in the 2000s. In every case: massive upfront capital, years of negative net return, followed by an application explosion that dwarfed the infrastructure cost by an order of magnitude.

The real question isn't "bubble or not?" — it's "which companies will capture the application layer surplus, and how fast?" That's the only bet worth making.

0 views
0 likes

Start a Critical Discussion

These questions don't have consensus answers. Share one to LinkedIn or X and see what your network actually thinks.

"The data shows AI is net-positive ROI at the application layer but structurally loss-making at L1–L4. Is that a bubble or just a build cycle?"

"Goldman Sachs says $1T in AI capex. McKinsey says $3.70 return per $1 spent. Both can be true — so who's actually wrong?"

"DeepSeek proved frontier AI can be built for $6M. Does this collapse the hyperscaler ROI thesis or accelerate it?"

Share this analysis

If this changed how you think about something, share it. The AI workforce conversation needs more data and less hype.

References

All sources accessed March 2026

Live data sources: Goldman Sachs Global Investment Research, McKinsey Global Institute, Gartner IT Spending Forecast 2Q26, Deloitte State of AI 2026, IBM Institute for Business Value, NVIDIA FY2027 Q2 Earnings, Bloomberg Intelligence AI Primer, Microsoft 2026 Annual Report

We use cookies

Essential cookies keep the platform running (authentication, session). We also use analytics cookies to improve your experience. EU/UK users: non-essential cookies require your explicit consent under GDPR Art. 6(1)(a) and the ePrivacy Directive. See our Privacy Policy for details.