This tool produces a strategic intelligence report on where competitive advantage in AI is migrating — and why raw model capability may matter less than power infrastructure and deployment reliability by 2028. It delivers a quantitative cost-performance comparison of leading Chinese and U.S. frontier models, a projected inference-cost decay curve through 2028 with confidence intervals, a ranked analysis of ten emerging moats by durability and value capture, and a scenario-based verdict on which business models survive near-free inference.
Copy the text below and paste the prompt into your preferred AI conversational search interface.
# AI Market Analysis: Infrastructure vs. Intelligence Moats
**Prompt revision date:** August 2026
---
## 0. Operating Rules (read before anything else)
**Nothing in this prompt is a fact.** Every model name, price, benchmark, URL, and figure below is a hint about where to look, not evidence. Treat the entire document as stale until verified. If a named entity does not appear in search results, say so and move on — do not reconstruct it from memory.
**Date-stamp everything.** Every retrieved figure carries the date the source published it and the date you retrieved it. Pricing older than 90 days is provisional and must be labeled as such.
**Search budget.** This is a broad-retrieval task. Expect 25–50 retrievals. Prioritize in this order: (1) vendor pricing and model-card pages, (2) regulatory filings and government energy data, (3) independent leaderboards, (4) analyst and press reporting. Do not stop early to save calls; do not pad with redundant queries.
**Failure is reportable.** If a search returns nothing, write "no result found for [query], as of [date]." Absence of results is not evidence of absence, and it is never grounds for estimation.
**Refuse to guess.** `[Insufficient Data]` is a valid and expected output. A report with twelve honest gaps is more useful than one with twelve invented numbers.
---
## 1. Core Thesis
> "Raw model intelligence is undergoing a commoditization hyper-cycle. The moat is shifting from **model weights** (rapid leakage, distillation, open-weight release) toward **infrastructure** (gigawatt-scale power, grid interconnection, and the industrial supply chain that energizes it) and **systemic agency** (reliable, verifiable multi-step execution). The US retains advantages in frontier reasoning, but China's industrial power base, component manufacturing position, and deployment speed create a structural cost-floor advantage that could materially impact long-run ROIC."
Produce an evidence-based strategic analysis that supports or falsifies this thesis. The thesis has three separable legs — **commoditization**, **infrastructure primacy**, and **China cost floor**. Evaluate each independently; report which legs survive contact with the data and which do not. A verdict of "two of three" is a legitimate and probably more informative outcome than a global yes or no.
---
## 2. Deliverable Specifications
- **Format:** Structured Markdown report
- **Length:** 5,000–7,000 words (quality over volume — do not pad; if a section has no data behind it, cut it and say why)
- **Tone:** Analytical, politically neutral, evidence-based
- **Tables:** 8–12
- **Forecast horizon:** end of 2029
- **Index base year:** 2026 = 100
---
## 3. Data Integrity Rules
### 3.1 Source Hierarchy
Prefer sources that are both **primary** (original data) and **authoritative** (the author or institution bears legal, professional, or reputational accountability).
| Tier | Type | Examples |
|------|------|----------|
| 1 | Regulatory and legal filings | SEC EDGAR (10-K, 10-Q, S-1), court documents, patent filings, Commerce/BIS rules, export-control notices |
| 2 | Official government and IGO data | EIA, FERC, IEA, NBS China, BLS, Eurostat, State Grid disclosures |
| 3 | Peer-reviewed research | Journal DOIs; cite the paper, never the press release |
| 4 | Vendor primary sources | Official pricing pages, model cards, system cards, earnings calls, investor decks, engineering blogs |
| 5 | Preprints and technical reports | arXiv, MLCommons, Epoch AI — label explicitly as unreviewed |
| 6 | Primary institutional research | Brookings, NBER, IMF/World Bank, RAND, CSIS, SemiAnalysis — note known editorial or client orientation |
| 7 | Established journalism | Reuters, Bloomberg, FT, The Information — for reported facts citing named primary sources |
| 8 | Analyst notes and independent leaderboards | Bank research, ARK, Artificial Analysis, LMArena, BenchLM, OpenCompass |
| 9 | Social media, vendor marketing claims, unverified benchmarks | X, forums, self-reported scores with no independent replication |
**Rules:**
- **Do not cite aggregators that repackage without adding analysis** (Statista, Macrotrends, SEO "statistics" roundups, model-comparison content farms). If an aggregator is the only path to a number, trace it to the stated origin and cite that. If the origin cannot be verified, flag the limitation instead of presenting the figure as established.
- Flag all Tier 8–9 sources in the Appendix with the specific claim each supports.
- Vendor-published benchmark scores are Tier 9 until independently replicated. Say which they are.
### 3.2 Claim Transparency Labels
Every quantitative claim carries exactly one:
- `[Sourced: URL, published DATE, retrieved DATE]`
- `[Estimated: method]` — derived via a stated, reproducible proxy; show the arithmetic
- `[Insufficient Data]` — no reasonable basis exists; do not guess
### 3.3 Standardization
- **Token prices:** USD per 1M tokens, listed input and output separately, plus a blended figure at 3:1 input:output. State the ratio used every time. Note cached-input and batch pricing separately where offered — these now differ by an order of magnitude at some vendors and blended-only comparison is misleading.
- **Reasoning tokens:** count as output tokens. Where a vendor bills them differently, say so.
- **Dates:** Month YYYY
- **Parameters:** billions active (MoE: active, not total; report total in parentheses)
- **Power:** GW capacity, TWh consumed
- **Currency:** USD; state the CNY rate and date used for any conversion
### 3.4 Conflict Resolution
When sources disagree:
1. Show the range: `[Source A (Tier n): X] to [Source B (Tier m): Y]`
2. **Prefer the higher-tier source.** Use the midpoint only when sources are of equal tier and comparable methodology. Do not average a 10-K against a blog post.
3. Note dispersion as `±Z%` and state which resolution rule you applied.
### 3.5 Proxies
v1 included a training-cost proxy (active params × $/B) and a China-pricing proxy (US price × 0.3–0.7). **Both are withdrawn.** The first has no defensible basis across architectures, hardware generations, or training-token budgets; the second is unnecessary because Chinese vendors publish list prices. Where training cost is unavailable, report `[Insufficient Data]` or cite a published compute estimate (e.g., Epoch AI) with its stated uncertainty. Do not manufacture a number to fill a cell.
---
## 4. Phase 1: Model Landscape (as of retrieval date)
### 4.1 Discover the Current Leaderboard
Search and fetch, in this order:
1. `https://arena.ai/leaderboard/text` — record the vote-cutoff date shown on the page, the total vote count, and whether any re-baselining note is displayed. Arena has re-baselined at least once in 2026; a re-baseline makes cross-period Elo comparison invalid, so check.
2. Artificial Analysis intelligence and price/performance rankings.
3. OpenCompass or an equivalent China-domestic leaderboard — Western leaderboards under-sample Chinese models, and several major Chinese releases are absent or stale on Arena.
4. A general web search for frontier releases in the current and prior month, to catch models too new to be ranked.
Identify:
- **Top 5 US-origin frontier models** by Elo and by independent composite index (these lists will differ — report both and note the divergence)
- **Top 5 China-origin frontier models** by the same two measures
- **Open-weight status** for each: fully open weights, restricted license, or closed API only
**Reference hints only — confirm or discard by search.** As of mid-2026, names in circulation included, on the US side, Claude Opus 5 / Fable 5 / Sonnet 5, GPT-5.5 and GPT-5.6 variants, and Gemini 3.x Pro; on the China side, DeepSeek V4 and V4-Pro, Qwen3.6/3.8-Max, Kimi K2.6 and K3, GLM-5.2, Doubao-Seed-2.0-Pro, MiniMax M2.x, and ERNIE. Version numbers in this space move monthly. **If a name here does not appear in your search results, do not use it and do not fabricate data for it.**
### 4.2 Model Comparison Table
| Model | Origin | Release (Mo YYYY) | Weights | Arena Elo (as of) | Active Params (B) | Total Params (B) | Context | Input $/1M | Output $/1M | Cached In $/1M | HLE (%) | GPQA-D (%) | SWE-bench Verified (%) | Source |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
Mark `[Insufficient Data]` where unavailable. **Do not estimate benchmark scores under any circumstance.**
### 4.3 The Denominator Problem
Earlier methodologies specified GPQA Diamond as the capability denominator. **This is no longer defensible.** GPQA Diamond is at or near saturation — frontier models cluster in the high 80s and low 90s, and the community has migrated to HLE, ARC-AGI-2, SWE-bench Verified/Pro, and Terminal-Bench for frontier discrimination. A saturating denominator systematically compresses measured capability differences and therefore *overstates* the cost-efficiency of cheaper, weaker models. Since the thesis turns on exactly that comparison, the choice of denominator is not incidental.
Use instead:
```
Capability-Adjusted Cost = (Blended $/1M tokens) / (Capability Score)
```
Compute it **three times**, with three different denominators:
- **HLE %** — frontier knowledge/reasoning, unsaturated
- **SWE-bench Verified %** — agentic coding, the closest available proxy for the "systemic agency" leg of the thesis
- **GPQA Diamond %** — retained solely for continuity with v1; report it and note the saturation caveat
If the three rankings disagree, that disagreement is a finding. Report it rather than picking a favorite.
### 4.4 Cost Per Completed Task
Price per token is the wrong unit for comparing 2026 models. Reasoning models consume wildly different token volumes for the same task; a model at one-tenth the per-token price that burns fifteen times the tokens is more expensive, not cheaper.
Search for published token-consumption or cost-per-task data — Artificial Analysis publishes cost-to-run-index figures, agentic benchmark harnesses (SWE-bench, Terminal-Bench, GAIA) increasingly report token or dollar cost per task, and some vendors disclose average reasoning-token counts.
| Model | Tokens/task (benchmark, source) | $/task | vs. cheapest US model | vs. cheapest China model | Source |
|---|---|---|---|---|---|
If task-level cost data cannot be found, mark `[Insufficient Data]` and state explicitly that the per-token comparison in 4.3 may misstate real economics by an unknown factor. Do not silently let per-token figures stand in for task economics.
### 4.5 China Discount Factor
```
China Discount Factor = (US avg capability-adjusted cost) / (China avg capability-adjusted cost)
```
Present under each of the three denominators from 4.3, and — where data allows — on a per-task basis from 4.4.
| Denominator | Scenario | US price assumption | China price assumption | Discount factor | Interpretation |
|---|---|---|---|---|---|
| HLE | Optimistic for China (floor prices) | | | | |
| HLE | Base (median) | | | | |
| HLE | Conservative for China (ceiling) | | | | |
| SWE-bench V | ... | | | | |
| GPQA-D | ... | | | | |
**Interpretation thresholds:** >1.5 strong structural advantage · 1.0–1.5 moderate · <1.0 none.
**Required caveats to address explicitly:**
- Are listed Chinese prices sustainable, or subsidized/promotional? Check vendor pricing pages for announced increases and note any.
- Do the prices reflect comparable serving quality (throughput, latency, uptime, rate limits, context)?
- Do open-weight releases make the API list price economically irrelevant for large buyers who self-host? If so, the discount factor understates the effect, and self-hosted cost per token becomes the right comparison.
---
## 5. Phase 2: Cost Trajectory & Margin Analysis
### 5.1 Historical Pricing Baselines
Document launch vs. current pricing for at least six model families, spanning both countries:
- GPT-4 (Mar/Apr 2023) → current OpenAI frontier
- Claude (first API offering, 2023) → current Anthropic frontier
- Gemini Pro (Dec 2023) → current Google frontier
- DeepSeek (first public API) → current
- Qwen (first Max-tier API) → current
- One additional open-weight family with hosted-inference pricing (e.g. Llama-class or GLM), to capture the self-host price floor
| Model Family | Launch price ($/1M blended) | Launch date | Current price | Current as-of | Months elapsed | Implied halving period | Source |
|---|---|---|---|---|---|---|---|
**Quality adjustment (required):** the current model is not the launch model. State plainly that this table measures the price of the *frontier tier*, not the price of constant capability, and note the direction of bias. Where possible, add a second series: price of the *original capability level* over time (e.g. what does GPT-4-class performance cost today from any vendor). That series is the one the commoditization thesis actually predicts.
### 5.2 Decay Model
1. **Fit** `Cost(t) = Cost₀ × e^(−λt)` by log-linear regression across all observations gathered, not by averaging two-point halving periods. Report λ, its standard error, R², and n.
2. **State the limitation:** frontier pricing moves in step functions driven by discrete releases and competitive responses, not smooth exponential decay. An exponential fit to fewer than ten observations per family is descriptive, not predictive. Say so, and report the residual pattern.
3. **Project forward** using the median λ with 80% confidence intervals derived from the observed cross-family variance in λ.
4. **Report a floor.** Inference price cannot fall below the marginal cost of energy plus amortized silicon. Estimate that floor from published power costs and accelerator pricing, or mark `[Insufficient Data]`. A projection that crosses zero is a broken model.
| Year | Training cost index (2026=100) | Inference $/1M (median) | 80% CI low | 80% CI high | Est. marginal cost floor | Gross margin estimate |
|---|---|---|---|---|---|---|
| 2026 | 100 | [sourced] | — | — | | |
| 2027 | | | | | | |
| 2028 | | | | | | |
| 2029 | | | | | | |
5. **Crossover predictions** (as ranges, never point estimates):
- When does GPT-4-level reasoning fall below $0.10/1M blended?
- Below $0.01/1M blended?
- When, if ever, does *frontier-tier* pricing fall below $1.00/1M blended? (Distinguish clearly from the two above — the thesis conflates them and the data may not.)
### 5.3 Capital Structure & Financing
Cover in 400–600 words:
- Disclosed AI capex and depreciation schedules for major US hyperscalers, from 10-K/10-Q filings. Note any changes in useful-life assumptions for AI hardware — these materially affect reported margin and are a known area of analyst dispute.
- Circular or vendor financing arrangements (chip vendors investing in customers, compute-for-equity deals). Describe the structures found; do not characterize their prudence.
- Where disclosed, revenue run-rates and gross margins of the major model labs. Mark `[Insufficient Data]` for private companies without credible primary disclosure — do not use press-reported figures as though they were audited.
- Chinese-side equivalents: state-directed capital, provincial compute subsidies, and disclosed capex from listed firms (Alibaba, Baidu, Tencent filings).
**This section determines whether the price war is a cost advantage or a funding advantage.** They have very different durations, and the thesis depends on which it is.
### 5.4 Business Model Implications at Near-Free Inference
300–400 words:
- What happens to API-first revenue models when inference approaches marginal cost?
- Where does value capture migrate — distribution, memory, orchestration, verification, proprietary data, or the physical layer?
- Does agent-driven token consumption grow fast enough to offset per-token deflation? Model the two curves explicitly: if tokens/task grows faster than $/token falls, total spend rises even as unit price collapses, and the commoditization leg of the thesis is directionally right but economically inverted.
- **Historical analogy:** cite one comparable commoditization cycle (long-haul bandwidth 1998–2002, cloud object storage, SMS, DRAM) with specific figures, and draw explicit parallels *and* disanalogies. Name at least one way the analogy fails.
---
## 6. Phase 3: Moat Rankings (2029 Horizon)
### 6.1 Framework
Rank on two dimensions:
**A. Durability:** years until the moat loses 50% of its competitive effectiveness
**B. Value capture:** standalone contribution to EBITDA premium over a commodity provider, as High/Medium/Low plus a qualitative % estimate
> Do not force percentages to sum to 100%. Moats stack and overlap. Note synergies explicitly.
### 6.2 The Moats
1. Gigawatt-scale power access — generation, PPAs, behind-the-meter
2. **Grid interconnection and electrical supply chain position** *(new — separated from #1)*: transformers, switchgear, HV cable, turbines, battery storage. In 2026 the binding constraint on US buildout has repeatedly been interconnection queues and component lead times rather than generation capacity, and much of the component supply chain is Chinese. Treat as a distinct moat with distinct owners.
3. OS-level distribution (defaults, pre-installation, bundled enterprise suites)
4. Regulatory compliance (sovereignty, safety mandates, audit infrastructure)
5. Security and authentication (agent trust, identity, payment rails for agents)
6. Agentic reliability (verified multi-step execution, low failure compounding)
7. Persistent memory (long-term personalization, switching costs)
8. Synthetic data pipelines (post-internet training data, RL environments)
9. Model weights (raw reasoning capability)
10. **Open-weight ecosystem and developer mindshare** *(new)*: the strategic position built by releasing weights — downstream fine-tune lock-in, default status in the self-host stack, and the price floor it imposes on closed competitors. Chinese labs currently dominate open-weight release volume, which makes this directly load-relevant to the thesis.
11. Real-time data loops (live sensors, transactions, feeds)
12. Static proprietary data (licensed corpora)
### 6.3 Rankings Table
| Rank | Moat | Durability (yrs) | Value capture | Who currently holds it | Key synergies | Justification |
|---|---|---|---|---|---|---|
Each justification is 2–3 sentences and **must** reference at least one of:
- Energy or infrastructure economics with a cited figure
- A historical commoditization precedent
- A specific regulatory instrument or trajectory
- A deployment timeline or adoption curve with a date
Add a **"Who currently holds it"** column — a moat with no identifiable current holder is a hypothesis, not a moat, and should be labeled as such.
### 6.4 Deep-Dive: Top 3
For each of the top 3, address: what would have to be true for this moat to fail by 2029, and what observable would signal that failure early.
---
## 7. Phase 4: Strategic Synthesis
### 7.1 Evidence-Based Verdict (5–7 paragraphs)
Ground every claim in Phase 1–3 data. Address:
1. **Leg-by-leg verdict.** Does the data support commoditization? Infrastructure primacy? A structural China cost floor? Score each leg supported / partially supported / falsified, with the specific evidence.
2. Is China's cost advantage **structural** (durable input-cost differences in power, labor, capital, and industrial supply chain) or **cyclical** (subsidy, promotional pricing, regulatory arbitrage, funding-driven loss leadership)? The Phase 5.3 capital-structure evidence is decisive here.
3. **Policy is now bidirectional.** US export controls have constrained Chinese access to leading-edge accelerators; Chinese component manufacturing and export policy constrain US datacenter energization. Both directions have produced concrete 2026 effects. Analyze the *net* compute-deployment consequence, not one direction only.
4. Which 3 moats dominate by 2029, and what in the data points there?
5. What business models survive near-free per-token inference, and what new ones emerge?
6. Second-order effects: capital flows, geopolitical response, labor displacement patterns, energy prices for non-AI consumers.
### 7.2 Quantitative Scenarios
> Scenarios are coherent world-states, not mutually exclusive events. Assign each a probability that it becomes the *dominant* dynamic shaping market structure by end of 2029. Probabilities must sum to exactly 100%.
Four scenarios (v1 had three, which did not span the outcome space — there was no world in which the thesis is simply wrong):
---
**SCENARIO 1: Infrastructure & Energy Dominance**
*China's industrial power base and electrical supply chain position create a durable cost floor US providers cannot match within the forecast period.*
```
Probability: X%
Observable triggers (check by Q4 2027):
✓ [Metric + threshold + named data source + update frequency]
✓ [Metric + threshold + named data source + update frequency]
✓ [Metric + threshold + named data source + update frequency]
Key trigger: China AI-dedicated power capacity exceeds [X] GW per [named source]
2029 outcomes:
- Market share by inference revenue: US X%, China Y%, EU Z%, RoW W%
- Frontier inference cost: $[X]/1M (US) vs $[Y]/1M (China)
- Global AI compute demand: [Z] TWh/year
- Top 5 firms by AI-attributable market cap: [ranked]
Strategic implications:
• Capital flows
• Geopolitical response
• Industry structure
• Labor/workforce effects
```
---
**SCENARIO 2: Distribution Consolidation**
*2–3 platform firms lock in AI defaults across devices and enterprise software; distribution, not model quality, determines revenue.*
*(Same block structure. Key trigger: 3 or fewer firms control >75% of on-device AI defaults.)*
---
**SCENARIO 3: Regulatory Fragmentation**
*Incompatible US, EU, and China standards bifurcate the market, rewarding compliance infrastructure over raw capability.*
*(Same block structure. Key trigger: materially incompatible governance regimes enforced in all three jurisdictions, with at least one documented market exit or model withdrawal attributable to compliance.)*
---
**SCENARIO 4: Capability Discontinuity — Thesis Falsified** *(new)*
*Agentic reliability or frontier reasoning fails to commoditize. A persistent capability gap at the top re-establishes model weights as the primary moat, and infrastructure becomes a cost center rather than a differentiator.*
```
Probability: X%
Observable triggers (check by Q4 2027):
✓ [Metric + threshold + named data source + update frequency]
✓ [Metric + threshold + named data source + update frequency]
✓ [Metric + threshold + named data source + update frequency]
Key trigger: frontier-to-open-weight gap on an unsaturated agentic benchmark
widens rather than narrows over four consecutive quarters
2029 outcomes: [same fields]
Strategic implications: [same fields]
```
---
**Probability check:** must sum to exactly 100%.
### 7.3 Falsification Criteria
Six criteria — at least one per thesis leg, plus one that would falsify your own scenario weighting.
```
CRITERION #N:
Thesis leg: [commoditization / infrastructure / China cost floor / scenario weighting]
Observable: [specific, measurable metric]
Threshold: [precise value that disproves]
Data source: [URL or named publication + update frequency]
Check by: [Quarter YYYY]
What this shows: [why this invalidates the leg]
Alternative theory: [what would better explain the world if observed]
```
### 7.4 Pre-Registered Predictions (new)
State 5 falsifiable predictions with dates and thresholds, in a form that can be scored without interpretation. Include for each your confidence (%) and the single observation that would resolve it. This section exists so the next revision of this analysis can be scored against the last one rather than rewritten from scratch.
| # | Prediction | Resolves by | Resolution source | Confidence |
|---|---|---|---|---|
---
## 8. Final Output Structure
```
# AI Market Analysis: Infrastructure vs. Intelligence Moats
## 1. Core Thesis, Methodology, and Retrieval Log
(including: retrieval date range, leaderboard vote-cutoff dates,
known data gaps, and any methodology deviations from this prompt)
## 2. Model Landscape (Month YYYY)
### 2.1 Leading Models Comparison
### 2.2 Capability-Adjusted Cost (three denominators)
### 2.3 Cost Per Completed Task
### 2.4 China Discount Factor
## 3. Cost Trajectory & Margin Analysis
### 3.1 Historical Pricing Baselines (frontier tier and constant capability)
### 3.2 Decay Model & 2026–2029 Projections
### 3.3 Capital Structure & Financing
### 3.4 Business Model Implications
## 4. Moat Rankings (2029 Horizon)
### 4.1 Full Rankings
### 4.2 Deep-Dive: Top 3
## 5. Strategic Synthesis
### 5.1 Leg-by-Leg Verdict
### 5.2 Quantitative Scenarios (sum to 100%)
### 5.3 Falsification Criteria
### 5.4 Pre-Registered Predictions
## 6. Appendix A: Tier 8–9 Sources & Data Gaps
## 7. Appendix B: Failed Searches and Unverifiable Claims
## 8. Sources (all Tier 1–7 citations with URLs and dates)
```
---
## 9. Pre-Delivery Checklist
- [ ] Every quantitative claim carries `[Sourced]`, `[Estimated]`, or `[Insufficient Data]`
- [ ] No model, product, or figure appears that was not confirmed by search
- [ ] No benchmark score estimated for any model
- [ ] Retrieval dates recorded; leaderboard vote-cutoff and re-baselining status noted
- [ ] Conflict ranges shown with tier-preference resolution stated
- [ ] Capability-adjusted cost computed under all three denominators, with disagreements reported
- [ ] Cost-per-task attempted; explicit statement if unavailable
- [ ] Decay model reports λ, SE, R², n; step-function and quality-adjustment limitations stated
- [ ] Projection includes a marginal-cost floor
- [ ] Capital structure section distinguishes cost advantage from funding advantage
- [ ] Moat table includes current holder; unheld moats labeled as hypotheses
- [ ] Four scenarios present, including the thesis-falsifying one; probabilities sum to exactly 100%
- [ ] Six falsification criteria, at least one per thesis leg
- [ ] Five pre-registered predictions with resolution sources
- [ ] Tier 8–9 sources flagged in Appendix A; failed searches logged in Appendix B
- [ ] Historical commoditization analogy includes at least one stated disanalogy
- [ ] Word count 5,000–7,000
---
**BEGIN ANALYSIS**