MTS

Wiki/Demand

Demand and the Revenue Gap

A primary-source-focused look at whether realized enterprise and consumer revenue from generative AI can justify the capital expenditure being deployed. The thesis spans Sequoia's "$200B → $600B question," Goldman's June 2024 skeptic note, Bain's $2T framing, and Apollo's bubble warnings — measured against actual lab disclosures, hyperscaler call-outs, and enterprise spending surveys.

A note on units: "ARR" (annualized run-rate revenue) is a peak-month or peak-quarter figure multiplied by twelve. It is not realized GAAP revenue and not what falls into a 10-K. Where possible, both ARR and realized revenue are listed below.

1. Foundation Model Lab Revenue — Disclosed vs. Run-Rate

OpenAI

Period Metric Source
FY2023 realized ~$2B revenue CFO Sarah Friar (Jan 2026 retro)
FY2024 realized ~$3.7B revenue, ~$5B net loss The Information, Sep 2024
End-2024 ARR ~$6B Friar, Jan 2026
Mid-2025 (Aug) $13B annualized Friar via CNBC, Sep 2025
End-2025 ARR $20B+ Friar; corroborated by Sherwood/Bloomberg
Feb 2026 ARR ~$25B Industry reporting (The Information)
2026 projected loss ~$14B Internal projections leaked Nov 2025
2025 inference cost $8.4B Internal docs
Implied gross margin ~33–48% Epoch AI / Exponential View analysis of GPT‑5 margins

Enterprise has become a meaningful share of mix — OpenAI's CFO disclosed in early 2026 that enterprise represented ~40% of revenue, on track to reach parity with consumer by end of 2026. ChatGPT consumer subscriptions remain the single largest line.

The "consumed by labs themselves" point matters here: a non-trivial slice of OpenAI revenue is API consumption by partners that are themselves funded via OpenAI equity rounds or strategic round-tripping (see §4 on the hyperscaler money cycle).

Anthropic

Period Metric Source
Jan 2024 ARR $87M Sacra; reporting by The Information
Dec 2024 ARR ~$1B The Information
Mid-2025 $4–5B ARR Bloomberg, July 2025
End-2025 ARR ~$9B Bloomberg, Jan 21 2026; corroborated by Anthropic's own retrospective tweets
Q1 2026 ARR $14B SaaStr, Mar 2026
Mar 2026 ARR ~$20B (approaching) Bloomberg, Mar 3 2026
Late 2026 ARR $30B Anthropic's own disclosure (X/Twitter)

Anthropic's mix is sharply enterprise/API-skewed. Menlo Ventures' 2025 State of GenAI report (Dec 2025) found Anthropic at 40% of enterprise LLM spend (up from 24% the prior year, 12% in 2023), with OpenAI at 27% (down from 50% in 2023) and Google at 21%. In coding specifically, Anthropic commands ~54% share vs. OpenAI's 21%, driven by Claude Code.

xAI

  • April 2025: ~$1B revenue run-rate disclosed to investors.
  • July 2025: $200B valuation discussed.
  • Sep 2025: $10B raise at $200B post-money confirmed (CNBC).
  • Series E ultimately upsized to ~$20B. Total primary funding ~$22B + $5B debt facility (Morgan Stanley).
  • Consumer ARPU mix: SuperGrok ~$30/mo, SuperGrok Heavy ~$300/mo; "consolidated" mix with X advertising puts combined run-rate >$3.3B by year-end 2025, but standalone xAI model revenue remains opaque.

Mistral, Cohere, AI21

  • Mistral: ARR ~$16M end-2024 → ~$312M Dec 2025 → ~$400M Jan 2026 (Sacra). €1.7B Series C at €11.7B valuation, Sep 2025. Targeting $1B ARR by end of 2026.
  • Cohere: ~$35M ARR early 2025 → $240M ARR end-2025 (beat $200M target). >50% QoQ growth through the year. Total funding >$1.5B.
  • AI21: No 2025 ARR disclosure surfaced; remains the smallest of the western frontier labs.

Chinese labs

Lab Signal
Doubao (ByteDance) >100M DAU in China; peak load 63.3B tokens/min on New Year's Eve 2025
Qwen (Alibaba) 200M+ MAU on Qwen chatbot app by Jan 2026
DeepSeek V3 disclosed $5.576M marginal training cost (2.79M GPU-hrs × $2/hr H800 rental). SemiAnalysis estimated true infrastructure investment ~$1.6B in servers, ~$500M in GPUs, ~$944M opex. R1 disclosed at $294K incremental over V3
Moonshot (Kimi) K2.5 launched Jan 2026; within 20 days, cumulative revenue exceeded all of CY2025. Overseas revenue now exceeds domestic. Valuation $20B (May 2026 round)
MiniMax >70% of 2025 revenue from international markets. M2.5 led OpenRouter token usage at 2.45T tokens

Chinese open-source models accounted for >60% of token consumption on OpenRouter in early 2026 — a direct pricing-pressure vector on closed-model margins.

2. Hyperscaler AI Revenue Disclosures

Microsoft

  • Oct 30 2024 (FY25 Q1 call): Satya Nadella disclosed Microsoft AI business at $10B annualized run-rate — fastest division to that milestone in company history. Crucial framing: "It's all inference." Microsoft was turning away training requests due to inference demand.
  • FY26 reporting: AI annualized revenue at $37B, +123% YoY.
  • M365 Copilot: 15M paid seats by Feb 2026 (+160% YoY in adds). Several customers with deployments >35,000 seats; one cited at 95,000.
  • GitHub Copilot: 4.7M paid subscribers by Jan 2026; ~20M total users including free/trial; ~75% YoY growth in paid seats. Moving to usage-based billing in 2026.

Google

Q3 2025 (call Oct 2025): Google Cloud revenue $15.2B (+34% YoY). Sundar Pichai disclosed:

  • Revenue from products built on generative AI models grew >200% YoY.
  • Nearly 150 Google Cloud customers each processed ~1T tokens with Google models over the trailing 12 months.
  • 1.3 quadrillion monthly tokens processed across all surfaces (>20x YoY).
  • Gemini consumer app: 650M MAU (3x QoQ).
  • Google Cloud signed more >$1B deals through Q3 2025 than in the prior two years combined.

AWS

AWS does not break out Bedrock as a line item, but:

  • Bedrock reported at "multi-billion dollar" annual run-rate by end of 2025.

  • 100,000 customers running Claude on Bedrock.

  • AWS custom-silicon business (Trainium, Inferentia) generating >$20B annual revenue.
  • Anthropic-AWS deal: up to 5 GW of capacity, ~1 GW Trainium2+3 by end-2026. Anthropic currently using >1M Trainium2 chips.
  • Wolfe Research estimate: Anthropic contribution to AWS revenue ~$3.9B in 2025 → ~$25B by 2027.

3. Enterprise Spending Surveys

Menlo Ventures — 2025 State of GenAI in the Enterprise (Dec 2025)

  • Total enterprise GenAI spend: $37B in 2025, up 3.2x from $11.5B in 2024.
  • Applications: $19B (>6% of total SaaS market).
  • Infrastructure: $18B.
  • LLM API spend specifically: $8.4B.
  • Vendor share at the API layer: Anthropic 40% / OpenAI 27% / Google 21%. Combined 88%.

Andreessen Horowitz — 100 Enterprise CIOs (2025)

  • Enterprise LLM spend growing ~75% YoY.
  • "Innovation budget" share of LLM spend collapsed from 25% (2024) → 7% (2025); the remainder shifted to centralized IT / business unit operating budgets — the explicit signal that GenAI moved from POC to line-item.
  • 78% of surveyed CIOs run OpenAI in production (either direct or via hyperscaler).
  • 44% run Anthropic in production, 63% including testing — fastest share gain of any frontier lab (+25 pts since May 2025).
  • 37% now deploy 5+ models in production (vs. 29% prior year).

McKinsey — State of AI (Mar 2025 + Nov 2025)

  • 88% of orgs regularly use AI in at least one function.
  • 72% use generative AI specifically (up from 33% in 2024).
  • 39% report any EBIT impact at the enterprise level; of those, most say AI accounts for <5% of EBIT.
  • Out of ~2,000 respondents, only ~5.5% (109) report >5% of EBIT attributable to AI. McKinsey calls these "AI high performers."
  • Nearly two-thirds report their organization has not yet begun scaling AI across the enterprise.
  • Workflow redesign is the single biggest differentiator (high performers 2.8x more likely to have done it).

MIT NANDA — The GenAI Divide: State of AI in Business 2025 (Aug 2025)

The "95% fail" study. Method: 52 executive interviews, 153 leader surveys, 300 public deployments.

  • 95% of GenAI pilots delivered no measurable P&L impact.
  • 5% capture essentially all of the value.
  • Despite $30–40B in enterprise GenAI spend, ROI is concentrated in a tiny minority.
  • "Shadow AI economy": only 40% of companies have official LLM subscriptions, yet 90% of workers report daily use of personal tools (ChatGPT, Claude).
  • External vendor tools succeed 2x as often as internal builds.
  • Investment bias: budgets concentrated in sales/marketing despite higher ROI being available in operations and finance.

Gartner

  • 2025 GenAI spending forecast: $644B, +76% YoY (Mar 31 2025 release).
  • Worldwide AI spending all-in (Sep 2025 release): $1.5T in 2025, >$2T in 2026.
  • End-user spend on GenAI models specifically: $14.2B in 2025 (note the gap vs. $644B — most spend is infra/services).
  • Data center spending +35% in 2024, +15.5% in 2025.
  • Gartner separately observed CIOs culling internal GenAI projects in favor of commercial off-the-shelf in 2025 — directly relevant to the POC graveyard pattern MIT NANDA documented.

4. The Revenue Gap Thesis

Sequoia / David Cahn

Sep 2023 — "AI's $200B Question":

  • Math: $1 of GPU revenue → $1 of data-center energy/buildout → 50% margin requirement → $4 of end-customer revenue needed for every $1 of GPU.
  • At Nvidia's then ~$50B run-rate of GPU revenue, the industry needed ~$200B in end-customer revenue to break even.
  • Most generous bottoms-up tally Cahn could construct: ~$75B. Gap: $125B.

Jun 2024 — "AI's $600B Question":

  • Nvidia GPU run-rate roughly doubled; required end-customer revenue therefore ~$600B.
  • Most generous bottoms-up: OpenAI ~$3.4B (up from $1.6B late 2023), plus a handful of startups <$100M each. Gap widened to ~$500B.
  • Cahn's framing: "How much of this CapEx build-out is linked to true end-customer demand, and how much is being built in anticipation of future demand?"

Goldman Sachs — Gen AI: Too Much Spend, Too Little Benefit? (Top of Mind, Jun 25 2024)

Lead skeptic: Jim Covello, Head of Global Equity Research. Key quotes:

"AI bulls seem to just trust that use cases will proliferate as the technology evolves, but eighteen months after the introduction of generative AI to the world, not one truly transformative — let alone cost-effective — application has been found."

"If AI technology ends up having fewer use cases and lower adoption than consensus currently expects, it's hard to imagine that [a bust] won't be problematic for many companies spending on the technology today."

Covello pegged the all-in cost of developing and running AI at ~$1T and argued the cost curve may not deflate the way bulls expect because the critical input (advanced GPUs) is a monopoly bottleneck, not a commodity. He also noted that genuinely transformative technologies (internet, mobile) enabled low-cost solutions to disrupt high-cost incumbents — the inverse of AI's situation today. (Note: Goldman's research head partially walked this back in early 2026, conceding "the shovels are fully priced in," while staying constructive on cloud.)

Bain — Global Technology Report 2025 (Sep 2025)

The "AI's $2 Trillion Question." Key claims:

  • By 2030, AI compute demand → ~200 GW globally (half in the US).
  • Capex required: ~$500B/year by 2030.
  • Revenue required to fund that capex profitably: ~$2T/year.
  • Even crediting all plausible AI-driven savings in sales/marketing/customer support/R&D, the world is $800B short.
  • AI compute demand is outpacing semiconductor efficiency at >2x Moore's Law — i.e., the deflation lever is structurally insufficient.

Apollo / Torsten Sløk

  • Jul 2025: "The difference between the IT bubble in the 1990s and the AI bubble today is that the top 10 companies in the S&P 500 today are more overvalued than they were in the 1990s." Top 10 contributed 54% of S&P 500 returns since Jan 2021.
  • 2026 outlook: still expects the AI cycle to continue into 2026, but flags that "any rollover would have material negative consequences for data center investment, the Magnificent 7, and broader consumer sentiment."
  • Productivity caveat: "We're not seeing AI productivity gains yet" — Sløk argues macro productivity data has not picked up corresponding to AI investment, undermining the demand-side justification.

Ed Zitron

Blog: Where's Your Ed At. Persistent critique of unit economics, citing OpenAI's $5B 2024 loss on $3.7B revenue, gross-margin compression from compute, and the "round-tripping" pattern (see below). His "Premium: OpenAI Burned $4.1 Billion More Than We Knew" and "Why Everybody Is Losing Money On AI" are the most-cited pieces.

5. Pro-Side Responses

Mary Meeker / BOND — Trends — Artificial Intelligence (May 2025, ~340 pp)

  • ChatGPT 800M WAU reached in 17 months — 5.5x faster than Google's historical growth.
  • 50% of S&P 500 companies regularly discuss AI on earnings calls.
  • Big Tech capex +63% YoY to $212B.
  • AI inference costs down ~99% over two years.
  • Developer communities around NVIDIA/Google +6–7x YoY (Google's AI dev community: 1.4M → 7M between May 2024 and May 2025).
  • Argues AI's adoption curve is genuinely different from past technologies on speed, capital intensity, and concurrency of consumer + enterprise + developer adoption.

Bessemer — State of the Cloud / State of AI (2025)

  • Cloud 100 aggregate value $1.117T (+36% YoY).
  • "AI Supernovas" hit ~$40M ARR in year 1, ~$125M in year 2 — fastest in Cloud 100 history. But gross margins ~25%.
  • "AI Shooting Stars": $3M ARR in year 1, 4x YoY, ~60% gross margins.
  • The gross-margin bifurcation is the implicit acknowledgement: the supernova revenue ramps look real, but unit economics on token-pass-through businesses look much closer to colocation than to classic SaaS.

Anthropic Economic Index

Ongoing release of aggregated Claude usage data:

  • Computer/mathematical tasks: ~33% of Claude.ai conversations, ~50% of API traffic. The single most prevalent task in Nov 2025 — "modifying software to correct errors" — was 6% of all usage.
  • 49% of all O*NET jobs have at least 25% of their tasks done on Claude (rising from 36% in Jan 2025).
  • Augmentation (52%) overtook automation (45%) in mid-2025 — the data suggests current usage is mostly co-pilot, not autonomous replacement, which has implications for both upside (large addressable market) and downside (reseller margin, not labor substitution).

a16z

Beyond the CIO survey (§3), a16z's "LLMflation" thesis (Nov 2024 and updated) argues inference cost-per-million-tokens for a fixed performance level is dropping ~10x/year, expanding the application surface faster than it shrinks per-call revenue. The pro version of the demand argument: cheaper inference creates more use, more than enough to compensate.

6. Real Adoption Signals (Hard Numbers)

Metric Value Date Source
ChatGPT WAU 400M Feb 2025 OpenAI
ChatGPT WAU 500M Mar 2025 OpenAI
ChatGPT WAU 700M Aug/Sep 2025 OpenAI
ChatGPT WAU 800M Oct 6 2025 (DevDay) Altman
ChatGPT WAU 900M Feb 27 2026 OpenAI funding announcement
OpenAI API tokens >6B tokens/min late 2025 OpenAI
M365 Copilot paid seats 15M Feb 2026 Microsoft
GitHub Copilot paid 4.7M Jan 2026 Microsoft
GitHub Copilot all users ~20M Jul 2025 GitHub
Gemini consumer app MAU 650M Q3 2025 Alphabet
Google tokens processed (all surfaces) 1.3 quadrillion/mo Q3 2025 Alphabet
Doubao (ByteDance) DAU >100M 2025 ByteDance
Qwen chatbot MAU >200M Jan 2026 Alibaba
Claude on Bedrock customers >100,000 2025 AWS
Fortune 100 with Copilot ~90% 2025 Microsoft

7. Inference vs. Training Economics

Cost-per-million-tokens deflation

Model Date $/M input tokens $/M output tokens
GPT-4 (8k) Mar 2023 $30 $60
GPT-4 Turbo Nov 2023 $10 $30
GPT-4o May 2024 $5 $15
GPT-4o-mini Jul 2024 $0.15 $0.60
GPT-3.5-equivalent (open) Oct 2024 ~$0.07 (combined)
  • Equivalent-performance cost down ~280x in roughly two years from Nov 2022.
  • a16z and others observe the empirical curve at ~10x/year for fixed-performance.
  • Mary Meeker pegs inference costs down ~99% over two years.

Implications

  • For closed-model labs: pricing pressure from Llama, DeepSeek, Qwen and Kimi (60%+ of OpenRouter token consumption in early 2026) is real. Closed-model gross margins must rely on capability premium, latency, tooling, and integration — not on raw token economics.
  • For hyperscalers: inference workloads look more like a utility than a software business. Microsoft's "it's all inference" line is a tell: the AI business at $10B → $37B ARR is fundamentally a compute-resale business with capacity constraints, not a high-margin software business.
  • For Cahn's math: deflating token prices means more tokens consumed per dollar, but it also means GPU revenue per query falls. The relevant question is whether aggregate dollar revenue keeps pace with capex — and Bain's $800B 2030 shortfall says no, even crediting deflation.

DeepSeek-V3 / R1 (Jan 2025)

  • V3 disclosed: 2.79M H800 GPU-hours × $2/hr = $5.576M marginal training cost.
  • R1: $294K incremental on top of V3.
  • These are marginal rental-cost numbers, not all-in. SemiAnalysis estimated DeepSeek's true infrastructure spend at ~$1.6B in servers, ~$500M in GPUs alone, ~$944M opex.
  • Even discounting for the hype, V3/R1 demonstrated that a frontier-competitive model could be trained at a tiny fraction of the budgets US labs were publicly anchoring to — collapsing one of the main bull cases (insurmountable training-capex moats).

8. Underexplored Mechanics

ARR vs. realized revenue

Almost every public AI number is ARR — the most recent month or quarter annualized. OpenAI's "$20B ARR end of 2025" translates to ~$13B realized for the calendar year (consistent with Friar's mid-year disclosure). Anthropic's "$9B end-2025" is similar — realized 2025 revenue is materially lower than headline ARR. When stacking lab revenue against capex, the relevant denominator is realized, not annualized.

Round-tripping / the money cycle

The hyperscaler-to-lab loop is structurally important:

  • Microsoft has invested >$13B in OpenAI (largely as Azure compute credits and cash that flows back as Azure consumption).
  • Amazon has committed up to $8B to Anthropic + multi-GW Trainium capacity contracts.
  • Google has invested >$3B in Anthropic and is also a major compute supplier.
  • A meaningful portion of OpenAI and Anthropic "revenue" recognized by Microsoft / AWS / Google as cloud revenue is paid out of equity rounds those hyperscalers led. This inflates the headline revenue numbers at both ends.

The labs themselves are also consuming a large share of their own compute spend for internal R&D (training next-generation models, evaluations, RL). When OpenAI reports $20B ARR but $14B projected losses for 2026, the gap is dominated by training compute that does not generate end-customer revenue.

POC graveyard

MIT NANDA's 95% figure is the cleanest data point, but it's corroborated by Gartner's observation that CIOs are killing internal GenAI projects in favor of vendor tools, and by McKinsey's finding that ~95% of orgs see <5% EBIT impact. The implication: the durable revenue base is much narrower than top-of-funnel spend suggests — Copilot/Cursor/ChatGPT Enterprise/Claude Code seats that actually stick, rather than the long tail of pilots.

Open-source pricing pressure

Llama (Meta), DeepSeek-V3/R1, Qwen, and Kimi K2.5 collectively account for >60% of OpenRouter consumption by early 2026. For closed-model APIs to defend margin, they must continually re-establish a capability lead — which feeds back into the training capex demand the bears are calling unsustainable.

9. Synthesis — Scoring the Gap

Stacking the bottoms-up revenue (annualized, end-2025/early-2026) against bear-case capex:

Source Annualized revenue
OpenAI ARR end-2025 ~$20B
Anthropic ARR end-2025 ~$9B (→ $30B by late 2026)
xAI (model-only) ~$1–2B
Mistral/Cohere/AI21 <$1B combined
Microsoft AI business $37B (FY26)
Google Cloud GenAI products "billions per quarter" (>$8B implied annually)
AWS Bedrock + Trainium AI Multi-billion + $20B custom silicon
Chinese labs (excl. consumer ads) Single-digit billions
Bottoms-up frontier ecosystem ~$80–100B annualized

Against:

Capex / required revenue benchmark $
Big Tech AI capex 2025 (Meeker) $212B
Sequoia $600B question $600B
Bain 2030 required revenue $2T
Bain 2030 shortfall even crediting savings $800B

The gap on Cahn's framing is still on the order of $500B/year of "missing" end-customer revenue if you take Nvidia GPU run-rate at face value. The gap on Bain's framing extends through 2030. The bull rebuttal: adoption velocity (Meeker's 800M WAU in 17 months, Menlo's 3.2x enterprise spend YoY) makes a $2T 2030 number more plausible than a $200B 2024 number looked at the time of Cahn's first post.

The genuinely open empirical question — surfaced cleanly by MIT NANDA and McKinsey — is whether the spend is converting to EBIT. Until that 5% number meaningfully rises, the gap narrative remains intact regardless of how impressive the ARR slopes look.

Sources

Lab disclosures and revenue tracking

Hyperscaler disclosures

Enterprise spending surveys

The revenue gap thesis

Pro-side

Adoption signals