Part VI — How to Challenge and Recompute the Atlas

A research framework is useful only if a reader can identify what is observed, what is inferred, what is missing, and which change would invalidate the conclusion.

6.1 Five checks before trusting a number

1. Boundary: does it describe a chip, rack, facility, company, or entire market?
2. Time: is it historical, current, guided, or a roadmap?
3. Unit: are power, currency, capacity, and performance measured on comparable terms?
4. Provenance: can the displayed value be traced to a source or formula?
5. Decision relevance: would changing the value alter the investment thesis?

A precise number that fails one of these checks can be less useful than an explicit range.

For example, “the rack uses 150 kW” is incomplete until the rack configuration, measurement boundary, product status, and source date are known. It may describe an upper design limit rather than measured operation. “The project is 1 GW” is incomplete until the reader knows whether that is gross utility supply, commissioned IT load, a phased target, or merely land and power under discussion. “The supplier has 30% share” is incomplete until the relevant product, period, geography, and revenue or unit denominator are named.

The fifth check—decision relevance—prevents false detail. If changing an estimate across its plausible range does not alter the investment conclusion, the report does not need to pretend that a precise answer exists. If a small change reverses the conclusion, the assumption should be visible and monitored.

6.2 Terms needed to read the argument

Compute and chips

• Accelerator / GPU — a chip built for the massively parallel matrix arithmetic neural networks require. A GPU is relatively general-purpose; an ASIC is optimized for a narrower workload.
• Training vs inference — training is the large, concentrated run that builds a model; inference is the recurring cost of using it. Reasoning and agents shift more compute into inference.
• Token — the unit a language model processes and typically prices; often a word fragment rather than a complete word.
• FLOPS / FP4 / FP8 — floating-point operations per second and lower-precision number formats. Lower precision raises throughput when accuracy can be preserved.
• ASIC — an application-specific chip such as Google TPU or AWS Trainium. It can deliver better cost efficiency for stable workloads with less flexibility than a merchant GPU.
• Rack-scale system / AI factory — an integrated product combining dozens of accelerators, CPUs, networking, power, and liquid cooling.

Memory, packaging, and manufacturing

• HBM — high-bandwidth memory: stacked DRAM placed beside the logic die. Capacity, bandwidth, qualification, and price are central to accelerator performance and cost.
• Advanced packaging / CoWoS — the process that combines logic, HBM, and I/O dies into one module. It determines package size, interconnect density, yield, and deliverable volume.
• Foundry — a contract chip manufacturer, including TSMC, Samsung, Intel Foundry, and SMIC.
• Process node / yield — the manufacturing generation and the share of good chips produced per wafer. A node name is not a substitute for measured yield and cost.
• EUV / High-NA — extreme-ultraviolet lithography and the next generation of higher-numerical-aperture tools. ASML is the sole EUV system supplier.
• WFE / EDA — wafer-fabrication equipment and electronic-design-automation software.

Models and software

• Scaling laws / test-time compute — improving models by enlarging training versus spending more compute on reasoning, sampling, and verification at the moment of use.
• Agent — a model system that uses tools and takes actions over an extended task. Production adoption is gated by reliability, permissions, integration, and liability.
• MoE / distillation / quantization — techniques that reduce training or inference cost.
• Open-weight vs closed — whether model parameters are published and can be deployed independently.
• CUDA / CANN — the software ecosystems around NVIDIA and Huawei accelerators.
• NVLink / UALink / InfiniBand / Ultra Ethernet — proprietary and open fabrics used within racks and across data centers.
• Co-packaged optics — moving optical engines closer to the switching silicon to ease bandwidth, reach, and power constraints.

Capital and markets

• Capex supercycle — the historic build-out of accelerators, networks, and data centers by cloud platforms and their partners.
• Circular financing — suppliers investing in customers that then purchase the supplier's products or a partner's capacity.
• RPO / backlog — contracted future obligations or revenue. It is not the same as cash already collected.
• GPU-backed debt — financing secured by accelerators and related contracts; vulnerable when equipment values fall faster than the loan amortizes.
• Depreciation life — the accounting period over which equipment cost is expensed. Extending it raises current earnings without changing economic value.
• Jevons paradox — lower unit compute cost induces enough new use that total spending rises.
• Neocloud — a specialist GPU and AI-cluster provider such as CoreWeave or Nebius.

Geopolitics

• Allied chokepoints — capabilities neither the United States nor China can quickly replace: Dutch lithography, Taiwanese foundry and packaging, Korean HBM, and Japanese materials.
• Bifurcation — one global supply chain splitting into US-led and China-led hardware, software, and procurement systems.
• Entity List / NS-CMIC / 1260H — distinct trade, ownership, and procurement restrictions. Their legal effects should never be collapsed into one label.
• 自主可控 / 信创 — China's autonomous-and-controllable localization and procurement regime.
• 国产 hierarchy — different degrees of domestic control, from self-owned architecture to licensed foreign IP or merely local production.
• 东数西算 — China's policy of moving suitable data-center workloads toward western regions with abundant energy.
• Sovereign AI — government- or sovereign-capital-funded national compute, models, and data infrastructure.
• 亿 — the Chinese unit for 100 million, a frequent source of tenfold translation errors.

6.3 How the atlas creates comparable conclusions

Fix the comparison boundary first

Cross-company and cross-country comparisons become misleading when they mix unlike objects. The atlas therefore:

• compares model capability and price at similar tasks, versions, and dates where possible;
• keeps chip, complete-system, and facility performance separate;
• models the AI-factory teardowns at accelerator-package, complete-system, and 100 MW commissioned IT-nameplate boundaries;
• separates capex from revenue, backlog from cash, and private marks from market prices;
• treats strategic importance and security valuation as different questions.

Prefer ranges to invented precision

Undisclosed price, power, supplier share, and facility cost use low, base, and high cases. The base case is a comparison point, not a claim of precision. Speculative relationships do not enter totals or security comparisons.

The common 100 MW boundary normalizes physical occupancy, capex, and operating requirements. It does not imply equal useful compute, model quality, or revenue between systems.

Work one example from disclosure to conclusion

Suppose a vendor discloses that a rack’s maximum system power is 142 kW. A 100 MW commissioned IT facility does not contain exactly 100,000 / 142 racks. Network, storage, management, and other IT systems also consume part of the nameplate. The base case must therefore allocate an explicit share of IT power to the accelerator systems before calculating rack count:

Maximum rack count under the allocation = accelerator-system IT power ÷ disclosed rack power

If the share is assumed rather than disclosed, the result is an estimate. If rack power is missing, the rack count is not calculable. A roadmap performance figure cannot substitute for the missing power input, because performance and power answer different questions.

Now consider a supplier allocation. A source may confirm that a memory company is qualified for a platform without stating its share. The atlas can name the company as a qualified supplier. It cannot assign 30% of platform memory spending to that company and include the value in an investment comparison. Qualification establishes access; allocation establishes addressable revenue.

These examples show the role of “not calculable.” It prevents one unsupported input from flowing through a formula and acquiring false authority in later charts.

Keep source quality separate from claim certainty

A high-quality product document can still describe only a roadmap. Reliable reporting can establish a relationship without proving allocation or revenue. The atlas uses five claim states:

• Confirmed — supported by formal disclosure or multiple direct records;
• Reported — supported by reputable reporting without complete first-party confirmation;
• Inferred — derived from compatible facts or an explicit formula;
• Roadmap — a stated future target not yet demonstrated in volume;
• Speculative — insufficiently supported and excluded from quantitative and investment conclusions.

A product launch page can be an excellent source for the fact that a company announced a roadmap, while remaining weak evidence that the product will ship on time or at volume. A regulator record can establish a legal restriction without revealing how customers will respond. Reputable reporting can identify a supplier relationship while leaving pricing and allocation unknown. The label describes what the evidence proves, not the prestige of the source.

Reconcile conflicts instead of averaging them

When sources disagree, preference goes to the most recent primary filing or regulator record that measures the exact entity, period, and unit. Official operating data and recognized industry bodies follow. Reputable financial reporting adds context unavailable from direct disclosure. Specialist estimates are retained only when the method is understandable and uncertainty is visible.

Material conflicts are not mechanically averaged. The analysis preserves the range and explains the disagreement; unresolved conflicts remain labeled contested.

Data hygiene

• Facts and status labels are observed through 25 July 2026; each forecast keeps its own forecast period.
• Company filings, regulator records, and specialist primary sources outrank aggregators.
• Chinese 亿, English million/billion, currencies, power, and energy units are normalized before comparison.
• Sanctions and ownership restrictions are applied to exact legal entities.
• Private valuations are treated as illiquid marks with contractual preferences, not market capitalizations.
• Prices, model rankings, capex plans, and policy status expire quickly and should not be mixed across dates without disclosure.
• Every displayed figure should trace to a public source or an explicit formula.

Keep the manuscript and outputs on one snapshot

A factual value should not be edited independently in prose, a table, and a chart. The canonical record contains the value, unit, period, source, evidence state, and formula dependency. Publication outputs should be rebuilt from that record so the English and Chinese HTML editions do not silently diverge.

For a dated research product, a synchronized snapshot matters as much as a correct formula. A security price from one day should not be combined with a later earnings release without disclosure. A policy status should not be copied forward after a license or truce changes. A roadmap product should not age into “shipping” merely because time passed; it needs new evidence.

6.4 What financial statements do not disclose directly

In the United States

“AI revenue” often combines accelerators, ordinary cloud migration, data services, and software seats. Backlog can be conditional or depend on capacity that has not yet been energized. GPU collateral, longer depreciation lives, and suppliers investing in customers can make reported growth appear stronger than independent cash demand.

Strong earnings can still produce weak returns when positioning and expectations are crowded. Allied chokepoints are not risk-free toll roads either: TSMC carries the Taiwan tail, while ASML and Japanese suppliers retain China-revenue and policy exposure.

The cash chain is a useful cross-check. Begin with the party paying from outside the AI financing loop. Follow the payment into application revenue, model or cloud cost, infrastructure orders, and the financing used to build capacity. Then compare recognized revenue with cash collection, capital spending, depreciation, customer advances, supplier financing, and related-party relationships. The purpose is not to imply that every strategic partnership is artificial. It is to identify which part of growth depends on a counterparty that must itself raise more capital.

In China

Commercial buyers may still prefer foreign accelerators while government, state-enterprise, and security-sensitive procurement must satisfy localization rules. Policy protects a slice of demand, not the entire open market.

“Domestic” should be traced through architecture, equipment, fabrication, memory, and software. A-share themes also require checks on pledged shares, government subsidies, profit excluding one-offs, state-fund selling, related-party transactions, and order evidence. System-level Ascend performance should not be compared with a rival rack without accounting for chip count, power, and network boundary.

Government support also appears in different economic forms. A cash subsidy can raise accounting profit. A cheap loan can reduce interest expense. A procurement mandate can raise revenue while leaving product economics weak. A capital injection can fund expansion but dilute other owners. These mechanisms all support industrial capacity; they do not have the same value for the listed shareholder.

6.5 Which source can support which claim

The atlas prioritizes:

• company filings, earnings, product architecture, and investor materials;
• regulator and government records, including Commerce/BIS, USGS, and Chinese authorities;
• energy and data-center research from the IEA, US Department of Energy, and Lawrence Berkeley National Laboratory;
• semiconductor research from SEMI, TrendForce, Counterpoint, and related specialist bodies;
• technical evaluations from Epoch AI, METR, official leaderboards, and model providers;
• BIS, IMF, and public debt, capex, and market data;
• Reuters, Bloomberg, CNBC, AP, and other reputable reporting where a relationship is not directly disclosed.

Third-party market shares, private valuations, future capex, and product roadmaps can all change quickly. The farther a number is from direct disclosure, the less weight it receives.

6.6 When a conclusion must stop being used

This atlas is industry and investment research, not investment advice, an offer, or a solicitation. It does not account for any reader's financial position, risk tolerance, taxes, or jurisdiction.

Public information can be incomplete and several figures are not independently audited. Private marks, forward capex, market shares, supplier relationships, and policy status are especially fluid. Recheck every ticker, corporate action, financial figure, market price, license, sanction, ownership restriction, and access rule before acting.

Forward scenarios are analytical judgments, not promises. Past performance and present strategic position do not guarantee future returns.

The essential discipline is not permanent conviction. It is rebuilding the causal chain whenever the denominator, evidence state, price, or policy changes.

A conclusion should be retired immediately when its load-bearing condition disappears. A shortage thesis expires when qualified lead times normalize and pricing rolls over. A roadmap thesis expires when qualification or production misses its gate. A policy-beneficiary thesis expires when the rule changes or when protected demand cannot be served. A valuation thesis expires when the market price moves outside the scenario range even if the industry argument remains correct.

The atlas is therefore a set of testable models rather than a permanent ranking. The reader should be able to reproduce the denominator, replace an assumption, observe which outputs change, and understand why the investment conclusion changes with them. That is the standard by which the book should be trusted.