Skip to the content.

What Did AI Actually Revolutionize?

$1.2T Capex Has Flowed In. Where’s the Real Output?

June 1, 2026 — Industry Analysis

The user’s question, paraphrased: Most AI investment came from Mag7’s traditional cloud/ads profits. The biggest winners are Nvidia and its suppliers (SK Hynix, Samsung, Micron — all now trillion-dollar companies). But actual AI usage is still quite limited. What did AI actually revolutionize?

TL;DR — Short, honest answer:


⚠️ Protocol Notice

Applies the Two-Step Research Protocol from .github/copilot-instructions.md. Section 1 = fact-base. Section 2 = Step 1 draft answering “what was revolutionized?”. Section 3 = Step 2 audit. Section 4 = honest synthesis + open questions.


Section 1 — Fact-Base

1.1 Capex Flowing IN (the spending side)

Hyperscaler AI infrastructure capex, by year:

Year Total ($B) Microsoft Google Amazon Meta
2024 ~$197B ~$70-80B ~$70-80B ~$75B ~$40B
2025 $300-380B $80B $75-93B $125B $60-65B
2026 $700B+ $180-190B $180-190B $180-200B up to $145B

Cumulative 2024-2026: ~$1.2 trillion from just 4 companies, plus Oracle, plus pure-play infrastructure (CoreWeave, Crusoe, etc.).

Critical financial signals (per CreditSights, Allianz Research, Futurum):

Source: Visual Capitalist, valueaddvc, CNBC, Futurum, AICerts, CreditSights, Allianz, epoch.ai.

1.2 Profit Captured by Picks-and-Shovels

Nvidia

Memory (SK Hynix + Samsung + Micron) — all now $1T+ market cap

Company Q1 2026 Revenue Operating Margin HBM Share Market Cap
SK Hynix ₩52.6T ($38B) — +198% YoY 72% (record) 50-59% $1.06T
Samsung (DS division) ₩81.7T ($60B+) 65%+ ~22% $1T+
Micron $23.86B (FQ2’26); ~$33.5B forecast Q3 Gross 81% ~21% $1T+

Components (PCB, MLCC, networking, optics, power)

Estimated total value-chain capture by suppliers: ~$650-750B annualized of value created (revenue, not profit) — roughly matching the 2026 hyperscaler capex on the supply side.

1.3 What End-Customers Are Generating (the demand side)

Pure-play AI-software ARR (mid 2026)

Company Jan 2025 ARR Dec 2025 ARR Mid 2026 ARR Growth
OpenAI ~$13B ~$25B ~$33B (Q2’26) ~2.5x
Anthropic ~$1B ~$9B ~$45B (May’26) 30x in 15 months — overtook OpenAI
Combined ~$14B ~$34B ~$78B ~5.5x

Embedded AI revenue inside hyperscaler stacks

Product Paid users / Revenue Note
Microsoft Copilot 15M paid seats, $30/user/month ~$5.4B ARR from this alone; plus Azure OpenAI
Google Gemini Enterprise Bundled in Workspace; specific seat count undisclosed Significant but opaque
AWS Bedrock Infrastructure (no seat count); ~$10B+ ARR for foundation model services Multi-model platform
Meta AI-driven advertising Q4 2025 ad rev $58B (+24% YoY); Advantage+ at $60B+ ARR; AI attribution at $1B+ rapid run rate Massive: AI directly attributable to several percentage points of ad lift

Estimated total AI software + AI-driven revenue mid-2026: $150-200B annualized

The math gap

Metric Value
2026 hyperscaler capex ~$700B
2026 AI software revenue (broad) ~$150-200B
Gap ~$500B annually
Gap funded by Debt (+$230B 2026) + FCF cushion + existing cash

Capex/AI-revenue ratio: ~3.5-4.7x. Even if revenue triples by 2028 ($450-600B), it would only catch up to the 2026 level of capex — and 2027/2028 capex is projected even higher.

1.4 Productivity Reality (the “did it work?” side)

McKinsey 2025 + BCG 2026 + multiple surveys:

Metric Result
% of orgs using AI in ≥1 function 88%
% scaled enterprise-wide 33%
% reporting any measurable EBIT impact 39%
% “AI high performers” (5%+ EBIT from AI) 6%
Productivity gain (early adopters) 15-22%
BCG-measured task-specific gains 30-90%
Top performer ROI multiples 5-10x

Translation: The revolution exists for the 6%. For the median company, AI is a 15-22% productivity tool with limited bottom-line impact — better than spreadsheets in the 1980s, but not on the scale the capex implies.


Section 2 — STEP 1 Draft Answer

Core Conclusion (answering “what did AI actually revolutionize?”)

One knowledge-work category has been revolutionized at scale: software engineering. Two adjacent categories — advertising and customer service — have been meaningfully improved but not transformed. Everything else (legal, healthcare clinical, manufacturing, education, general knowledge work) is in the pilot-to-augmentation phase, generating modest productivity gains for the 6% of “AI high performers” and noise for the other 94%. The financial geometry the user described is correct: ~$1.2T of cumulative capex has flowed in 2024-2026, captured first by Nvidia (~$100B+ net income) then by memory (~$300B+ revenue, $1T each in market cap), then by component suppliers (~$150B+ revenue). The end-customer AI revenue is $150-200B annualized — a real number growing fast (100-300%/yr at pure-plays), but 3-4x smaller than 2026 capex. The 2026-2028 window will reveal whether revenue scales into the capex (justifying the supplier supercycle) or the gap widens (triggering a capex pullback and memory/Nvidia profit pool compression). Currently the evidence points toward partial validation, not full justification — coding is the proof-of-concept that the revolution is possible; the breadth question is unresolved.

3 Supporting Points (The Revolution IS Real Where It’s Real)

S1. Software engineering has been demonstrably transformed at scale.

S2. AI-native software revenue is growing at unprecedented speed.

S3. Existing services are getting AI-driven incremental boosts at scale.

2 Opposing Points (The Revolution Has NOT Broadly Arrived)

O1. Median enterprise sees modest-to-zero impact; only 6% are “AI high performers”.

O2. The capex-to-revenue ratio is unprecedented and partially circular.

Explicit Unknowns

  1. Will Anthropic-style 30x revenue growth continue or saturate in 2027? If the next 18 months adds another 5-10x, the math gap closes. If it slows to 2-3x, the gap widens.
  2. Will the 94% of enterprises catch up to the 6%? The lag could be (a) early-cycle normal, or (b) genuine limit on AI’s economic utility for most use cases.
  3. Will productivity gains show up in MACRO statistics (TFP, GDP per hour worked)? As of 2026, the answer is “not yet visible at the aggregate level” — a meaningful caution flag.
  4. Are current LLM architectures hitting scaling limits? Some signals from GPT-5/Claude 4 progression suggest diminishing returns per training-dollar; capex math depends on continued capability scaling.
  5. Will hyperscaler capex actually hit $700B+ in 2026, or will some pull back as gap widens? Q3-Q4 2026 capex guidance revisions are the key tell.
  6. Coding revolution: does it generalize to other knowledge work, or is software special? Code is structured, verifiable, modular — properties that don’t apply to many other workflows.

Section 3 — STEP 2 Strict Peer Review

3.1 Facts Still Needing Verification

# Claim Best primary source
1 “Anthropic ARR $45B May 2026” Anthropic IR / secondary market filings
2 “OpenAI ARR $33B Q2 2026” OpenAI revenue press / SoftBank reporting
3 “Hyperscaler 2026 capex $700B+” MSFT/GOOGL/AMZN/META latest Q1 2026 calls
4 “Memory companies all $1T+” Bloomberg market data
5 “Microsoft 15M paid Copilot seats” Microsoft Ignite / Q1’26 earnings
6 “Meta Advantage+ $60B ARR” Meta Q4 2025 / Q1 2026 transcripts
7 “McKinsey 6% AI high performers” McKinsey State of AI 2025 PDF direct
8 “GitHub Copilot 20M users” GitHub Universe 2025/2026
9 “FCF -95% at Amazon” Amazon Q1 2026 10-Q
10 “$230B sector debt 2026” CreditSights tech debt tracker

3.2 Logical Leaps

  1. “Software engineering = revolution; everything else = augmentation” — binary framing; reality is a spectrum. Customer service is being meaningfully restructured (significant headcount changes at e.g., Salesforce, Klarna). Could argue customer service is also “revolutionized”.
  2. “Math gap of 3.5-4.7x is unsustainable” — assumes 5-7 year amortization is required for justification. If GPU/asset lifespans extend (re-use across model generations) or if revenue compounds at 100%/yr, the gap could close.
  3. “Anthropic 30x growth scales to revolution” — extrapolating from a 17-month period that includes coding-tool launch. Could be a one-time enterprise capture, not sustained.
  4. “Capex-revenue ratio looks like telecom/utilities” — the analog isn’t perfect; AI infrastructure has both networking (commodity) and strategic (cognitive) characteristics.
  5. “Only 6% high performers = mostly hype” — same 6% statistic was true of early SaaS adoption in 2005-2010, which became universal by 2020. Inflection-point dynamics could repeat.

3.3 Missing Counterexamples

  1. Coding may NOT generalize: code has unique properties (structured, verifiable, modular, syntactically constrained). The fact that AI works for code doesn’t necessarily mean it’ll work for legal briefs, financial analysis, or operations management.
  2. The “94% laggards” may include AI-skeptical industries (manufacturing, government, agriculture) that simply won’t be AI-led. The 6% being concentrated in tech/finance/consulting may be a feature, not a bug — and the absolute revenue from that 6% could be substantial.
  3. Productivity gains may be captured by employers, not creating new revenue. If AI makes accountants 30% faster, firms may just hire 30% fewer accountants — that shows up in labor costs, not AI revenue.
  4. Energy bottleneck: hyperscaler capex now constrained by power availability, not chip availability. This could naturally slow capex even if revenue keeps growing — reducing the gap from the supply side.
  5. AGI / scaling-law uncertainty: if AI capability plateaus (GPT-6 not meaningfully better than GPT-5), the entire revenue thesis weakens. If it accelerates (genuine AGI), entire industries get displaced and revenue catches up.
  6. Compute prices falling: GPU lifespan, model efficiency, and inference cost-per-token are all improving 2-4x/year. The “same dollar of capex” produces more output each year.

3.4 Most Important Primary Sources to Add

  1. McKinsey State of AI 2025 PDF (direct, not summaries)
  2. BCG AI Maturity Survey 2026 (full report)
  3. Anthropic latest funding deck / IPO S-1 (when filed)
  4. OpenAI revenue confirmed numbers (SoftBank / Microsoft cross-check)
  5. Hyperscaler Q1 2026 earnings call transcripts (capex guidance)
  6. GitHub Octoverse 2026 (developer metrics)
  7. Federal Reserve TFP / productivity data 2024-2026
  8. SemiAnalysis HBM market reports
  9. Allianz / CreditSights AI capex sustainability papers
  10. Independent academic studies on coding-tool productivity (CACM article)

3.5 Sentences That Remain Speculation

# Statement Status
1 “The 2026-2028 window will reveal whether revenue scales into capex” Conditional forecast
2 “Anthropic 30x growth, if it continues, scales into capex within 2-3 years” Linear extrapolation
3 “Only 6% are AI high performers means revolution hasn’t broadly arrived” Interpretive claim
4 “Capex-revenue ratio is unsustainable” Depends on amortization period + growth rate assumptions
5 “Coding may not generalize to other knowledge work” Untested hypothesis
6 “Software engineering is the only fully revolutionized category” Pattern recognition; arguable cutoffs
7 “Meta AI lift = several percentage points of ad rev” Estimated from disclosed metrics
8 “$1.2T cumulative capex” Summed from individual disclosures
9 “Free cash flow -95% at Amazon” Quarterly figure; could revert
10 “2027 is inflection point” Subjective forecast

Section 4 — Honest Synthesis + Open Debate

What the data actually tells us

The bull case (steelmanned):

The bear case (steelmanned):

My honest synthesis:

The user’s framing — “trillion-dollar memory companies + Nvidia took the profit; actual AI usage is limited” — captures the balance-sheet truth of mid-2026 but understates the growth derivative. The pure-play AI revenue growth (100-300%/yr) is structurally different from past tech bubbles where revenue growth slowed BEFORE the capex slowed (telecom 2001, internet 2000). Here, revenue growth is still accelerating in mid-2026. That said, the math gap is genuinely unsustainable as a 3-year trajectory — if revenue doesn’t 3-5x from here by 2028, expect capex pullback and supplier-stock multiple compression.

The single most important question: Will the AI-coding pattern (genuine 50% productivity gains, broad adoption) replicate in 2-3 other major knowledge-work categories by 2028? If yes — full revolution, capex justified, supplier oligopolies durable. If no — partial revolution like internet (real but narrower than hype), capex pullback, supplier multiple compression but underlying technology persists.

Bull/bear handicap (subjective): 55% bull, 35% middle, 10% bear. The Anthropic 30x growth is the single biggest argument against the bear case — that pace doesn’t happen in fake revolutions.

What I’d debate with you

  1. You said “AI actual usage is still quite limited” — I’d push back. 900M weekly ChatGPT users + 20M Copilot seats + 70% of F100 using Anthropic = NOT limited usage. What IS limited is broad enterprise EBIT impact (only 6% see >5%). Usage ≠ economic transformation. Maybe the distinction you mean is “transformation”?
  2. You attributed the value capture to Mag7 → Nvidia → memory/components — true, but you might be UNDERVALUING the Meta AI advertising contribution. Meta’s $60B+ Advantage+ ARR and 24% YoY ad growth is AI value that doesn’t show up as separate AI revenue — it shows up as core business revenue lift. The same is happening more subtly at Google and Amazon.
  3. You implied this is one-way circular — but Anthropic just had its first profitable quarter and is at $45B ARR. That’s not a hyperscaler subsidy; that’s real customer payment by 70% of Fortune 100. The circularity argument was strong in 2023-2024; it’s weakening fast.
  4. Where I 100% agree with you: the trillion-dollar memory market caps look stretched relative to the “revolutionized” footprint. If I had to short something here, it would be memory at the next sign of capex moderation — historically the most cyclical part of the value chain.

Open questions I’d find genuinely uncertain


Primary Sources

Capex

  1. Visual Capitalist — The Rise of AI Hyperscaler Spending: https://www.visualcapitalist.com/the-rise-of-ai-hyperscaler-spending/
  2. CNBC — How much Google, Meta, Amazon and Microsoft are spending on AI: https://www.cnbc.com/2025/10/31/tech-ai-google-meta-amazon-microsoft-spend.html
  3. CNBC — Tech AI spending approaches $700 billion in 2026, cash taking …: https://www.cnbc.com/2026/02/06/google-microsoft-meta-amazon-ai-cash.html
  4. AI Certs — Hyperscaler Capex Surge Redefines 2026 Budgets: https://www.aicerts.ai/news/hyperscaler-capex-surge-redefines-2026-budgets/
  5. Allianz Research — AI capex cycle: war-proof for now: https://www.allianz.com/content/dam/onemarketing/azcom/Allianz_com/economic-research/publications/specials/en/2026/march/2026_03_25_AI.pdf
  6. CreditSights — Hyperscaler Capex 2026 Estimates: https://know.creditsights.com/insights/technology-hyperscaler-capex-2026-estimates/
  7. Futurum — AI Capex 2026: The $690B Infrastructure Sprint: https://futurumgroup.com/insights/ai-capex-2026-the-690b-infrastructure-sprint/
  8. Introl — Hyperscaler CapEx Hits $600B in 2026: https://introl.com/blog/hyperscaler-capex-600b-2026-ai-infrastructure-debt-january-2026

Memory / HBM trillion-dollar club

  1. Morningstar — SK Hynix, Micron Join Trillion-Dollar Club: https://www.morningstar.com/news/dow-jones/202605271274/sk-hynix-micron-join-trillion-dollar-club-update
  2. US News — SK Hynix Joins $1 Trillion Club After Samsung, Micron on AI Chip Boom: https://money.usnews.com/investing/news/articles/2026-05-26/sk-hynix-joins-1-trillion-club-after-samsung-micron-on-ai-chip-boom
  3. AlphaPilot — Micron and SK Hynix Join Trillion-Dollar Club: https://www.alphapilot.tech/discover/micron-and-sk-hynix-join-trillion-dollar-club-as-ai-memory-demand-surges
  4. CurrentAffair — HBM Memory Chip Supercycle 2026: https://www.currentaffair.today/blog/finance-9/hbm-memory-chip-supercycle-2026-how-sk-hynix-micron-and-samsung-formed-the-newest-1-trillion-club-747
  5. KoreaInvestInsights — SK Hynix HBM Market Share 2026: https://koreainvestinsights.com/post/sk-hynix-hbm-market-share-ai-memory-demand-2026/
  6. MomoView — HBM Three-Way War: https://momoview.com/blog/en/posts/hbm-industry-analysis-sk-hynix-samsung-micron-2026-ai-memory-supercycle-investment-thesis/

AI software revenue

  1. OpenTools — Anthropic Revenue Hits $45B ARR: https://opentools.ai/news/anthropic-revenue-surpasses-openai
  2. NerdLevelTech — Anthropic $30B ARR How Claude Overtook OpenAI: https://nerdleveltech.com/anthropic-30-billion-arr-surpasses-openai
  3. IDLEN — Anthropic Passes OpenAI October 2026 IPO: https://www.idlen.io/news/anthropic-overtakes-openai-30-billion-revenue-april-2026/

Productivity / ROI

  1. McKinsey — State of AI 2025 PDF: https://www.mckinsey.com/~/media/mckinsey/business%20functions/quantumblack/our%20insights/the%20state%20of%20ai/november%202025/the-state-of-ai-2025-agents-innovation_cmyk-v1.pdf
  2. BCG via n5r — AI Agents Deliver 30-90% ROI Gains: https://n5r.com/en-us/blog/bcg-ai-agents-mcp-a2a-enterprise-roi-agent-economy
  3. Welcome.ai — AI Adoption Insights from McKinsey’s 2025 Global Survey: https://www.welcome.ai/content/ai-adoption-insights-from-mckinseys-2025-global-survey
  4. Punku — State of AI 2025: 78% Adoption, 74% See ROI: https://www.punku.ai/blog/state-of-ai-2024-enterprise-adoption

Coding tool revolution

  1. GitHub — Quantifying Copilot impact in enterprise with Accenture: https://github.blog/news-insights/research/research-quantifying-github-copilots-impact-in-the-enterprise-with-accenture/
  2. SecondTalent — GitHub Copilot Statistics & Adoption Trends: https://www.secondtalent.com/resources/github-copilot-statistics/
  3. JetBrains — Which AI Coding Tools Do Developers Actually Use at Work? (Apr 2026): https://blog.jetbrains.com/research/2026/04/which-ai-coding-tools-do-developers-actually-use-at-work/
  4. CACM — Measuring GitHub Copilot’s Impact on Productivity: https://cacm.acm.org/research/measuring-github-copilots-impact-on-productivity/

Meta AI advertising

  1. PPC Land — Meta’s ad business hits record $58B as AI drives conversion gains: https://ppc.land/metas-ad-business-hits-record-58b-as-ai-drives-conversion-gains/
  2. FifthPerson — Meta Q4 2025: AI driving engagement, ads and growth: https://fifthperson.com/meta-q4-2025/
  3. Facebook About — 2026: AI Drives Performance: https://about.fb.com/news/2026/01/2026-ai-drives-performance/

Bubble / circular financing

  1. AI2.Work — Hyperscalers Pledge $725B in AI Capex While Revenue Returns Lag Behind: https://ai2.work/blog/hyperscalers-pledge-725b-in-ai-capex-while-revenue-returns-lag-behind
  2. Cresset — 2026 Outlook: Is AI a Bubble?: https://cressetcapital.com/articles/market-update/market-update-12-17-25-2026-outlook-is-ai-a-bubble/
  3. PAASA — AI Infrastructure Investing: How Hyperscaler Capex is Collapsing Free Cash Flow: https://paasa.com/blog/ai-capex-supercycle

Microsoft Copilot adoption

  1. Stackmatix — Microsoft Copilot Adoption Statistics & Trends 2026: https://www.stackmatix.com/blog/copilot-market-adoption-trends
  2. EPC Group — Microsoft Copilot vs Google Gemini Enterprise Comparison 2026: https://www.epcgroup.net/microsoft-copilot-vs-google-gemini-enterprise-comparison

Generated using the Two-Step Research Protocol. Not investment advice. Personal view: 55% bull / 35% middle / 10% bear on whether the current AI capex super-cycle proves justified by revenue scaling 2026-2028. Open to update on new data.


Companion 中文版: report_cn