Day01 | The Global AI Compute Value Chain: Profit Pools, Bottleneck Rankings and Geopolitical Fractures

1. Core judgment in one sentence

By October 2026, profits are spreading toward memory and model providers, but control over deliverable compute still determines their allocation.

Falsification conditions: this judgment should be weakened if subsequent comparable accounts show that memory windfalls quickly disappear, model improvements mainly reflect accounting adjustments, and power or delivery constraints no longer affect contracts, pricing or utilization. Those outcomes should change the conclusion, not prompt another untested “picks and shovels” narrative.

2. Industry boundary and revisions

Frozen boundary: saleable compute, not every business associated with AI

The scope includes GPU, ASIC, NPU and FPGA design; leading-edge fabrication and packaging such as CoWoS/SoIC; HBM and AI-server DRAM/NAND; servers, power supplies, liquid cooling and connectivity; data-centre grid access, substations, cooling and land; hyperscalers and NeoClouds; and model training/inference services. The central question is how contracts and asset ownership allocate risk and profit between chip production and a billable output. End-user applications, traditional IT services and data centres unrelated to AI are excluded from the research denominator. Financial results of mixed businesses are proxies, not entirely attributable to AI.

Boundary revision 1: include networking, optical interconnects, substrates and power management. A GPU count is not a measure of usable cluster capacity. Interconnects, substrates and power components can prevent manufactured chips from becoming operational systems. Broadcom's AI-semiconductor revenue includes networking and must not be treated wholly as ASIC revenue.[E35][E37][E52]

Boundary revision 2: include financing, depreciation, utilization and contract risk. These are not additional hardware stages, but they determine whether rental economics become profit. The gap between CoreWeave's adjusted EBITDA and operating income materially changes the claim that cloud rental is exceptionally profitable.[E15]

Framework revision: do not assume a linear chip → electricity → memory → model migration. Stock-price leadership, order shortages, rising gross margins and growing absolute profit pools are separate observations.

Definitions and research limitations

• The cutoff is 2026-10-25. Full-year 2026 is not an actual result. Latest reported quarters differ by issuer; forecasts do not fill unpublished periods.
• CY means calendar year; FY means issuer fiscal year. NVIDIA FY2026 ended 2026-01-25 and Dell FY2026 ended 2026-01-30. They serve as near-CY2025 annual proxies, not exact calendar-2025 results. Micron retains its August/September fiscal year-end.[E01][E07][E09]
• Gross margin (GM) = gross profit/revenue; operating margin (OM) = operating income/revenue. An absolute profit pool requires a common period and accounting layer. Financing revaluations, interest and taxes affect net income; it cannot substitute for operating earnings. Segment earnings and consolidated GAAP operating income are not necessarily comparable either.
• Research cross-checks issuer IR/SEC material, specialist research, professional reporting and regulatory texts. [E-number] references locate ledger entries. Reproducing an issuer's number in several newspapers does not create independent underlying measurements. Unsupported same-definition double sourcing is disclosed and rated Medium or lower.
• Paid full texts, undisclosed product-level profits, consistent training/inference shares and regional lead-time distributions remain gaps. Unsupported ranges are not inserted simply to fill a table. Export-control discussion describes the evidence at the cutoff, not authorization for any specific transaction.

3. Verdicts on the four hypotheses

Hypothesis A: upstream picks-and-shovels suppliers capture profits, stably

Verdict: large realized upstream profits are supported; “all the money” and permanent stability are weakened. Product, period and accounting definitions matter.

Three different tests are needed. NVIDIA retains a very large absolute operating-profit base. Memory-company margins rose sharply in 2026. Several reports indicate improving model-level unit economics, but bridging that improvement to recurring, fully costed operating profitability remains difficult. Asset 1 provides annual and quarterly comparators.[E01][E02][E06][E08][E18]–[E21]

Did HBM-specific margins exceed GPU margins? The public evidence does not support an unconditional yes. Micron's Cloud Memory Business Unit is not pure HBM; SK hynix also mixes products. TrendForce instead found server DDR5 per-wafer economics overtaking HBM from Q1 2026. “Memory windfall” therefore cannot simply mean “HBM windfall.” Annual HBM contracts can lag more rapidly repriced DRAM.[E23]

Improving model-provider bargaining power could reduce the premium captured from hardware performance gains, generic GPU rental rents or cloud distribution fees. It need not cause chip profits to fall in the same quarter. Model providers may reinvest incremental revenue in training, service levels and expansion. Multiplying inference margin by ARR does not calculate distributable cash.

Conflict record: [冲突E16-vs-E17] OpenAI “margin”

• Source A: Reuters, 2026-02-20, relaying a 33% adjusted gross margin for 2025.
• Source B: The Decoder's 2025-12-21 reporting chain, approximately 70% paid-user compute margin in October 2025.
• Nature: different user populations, cost boundaries and measurement windows.
• Ruling: neither disproves the other; neither establishes company-wide operating profitability in 2026.[E16][E17]

Conflict record: [冲突E19-vs-E20] Anthropic inference margins

• Source A: SemiAnalysis, 2026-05-01, inference-infrastructure gross margin above 70%.
• Source B: FT reporting relayed by Reuters, 2026-09-13, above 80% before partner revenue share and training costs.
• Nature: different dates and cost boundaries; both may be true without being matching GAAP observations.
• Ruling: directionally supportive of improvement, not an audited two-point margin time series.[E19][E20]

Conflict record: [冲突E21-vs-E18] Profit improvement versus enormous losses

• Source A: Reuters Breakingviews indexed text dated 2026-10-15 reports 85% nine-month GAAP gross margin and 11% adjusted operating margin; directly opening the page returned only a limited summary.
• Source B: Reuters, 2026-10-20, reports a first-half net loss of $80.7bn, including $72bn of non-cash financing revaluation charges.
• Nature: different periods, gross/operating/net layers and adjustments.
• Ruling: net losses do not disprove all inference economics; adjusted earnings do not establish full-year GAAP or cash profitability. Without the reconciliation, the bridge remains unresolved.[E18][E21]

Conflict record: [冲突E22-vs-E19] Who captures the profit pool?

• Source A: GSAM, 2026-05-07, uses a roughly 90% upstream narrative based on a listed-company next-12-month EBIT sample as of 2026-02-26.
• Source B: SemiAnalysis, 2026-05-01, emphasizes value shifting to model labs.
• Nature: different samples; forecasts versus actuals; public versus private companies; value capture versus accounting profit.
• Ruling: neither alone supplies the six-stage shares of realized global AI operating income. The GSAM denominator is especially unsuitable as a substitute for this report's boundary.[E22][E19]

Confidence: Medium. Disclosed issuer profit magnitudes are more secure than global AI-only shares, HBM product margins or model-profit sustainability.

Hypothesis B: GPUs/HBM are still the only binding bottleneck in 2026

Verdict: weakened as a universal claim. Semiconductor and physical-infrastructure constraints coexist; location and project stage determine priority.

IEA's regional analysis and equipment-supply evidence show why new buildings do not automatically create energized capacity. Grid connections, transformers, engineering and permits can prevent purchased chips from producing revenue on schedule. The emergence of those constraints does not automatically loosen HBM, packaging or leading-edge fabrication supply.[E25]–[E30][E35]

An expansion inside an energized facility differs from a greenfield campus. The former is more likely to hit accelerator, memory and cooling-retrofit constraints first. The latter is more likely to encounter a power/engineering critical path. Financing can prevent even a technically viable project from starting.

Conflict record: [冲突E31-vs-E32] Multiple US power-shortfall figures

• Source A: Morgan Stanley's February 2026 public article forecasts 74 GW of US data-centre demand and a 49 GW access shortfall by 2028.
• Source B: its 2026-02-11 report, page 44, discusses a 10%–20% shortage for 2025–2028 construction after mitigation such as onsite generation.
• Nature: total versus incremental-build denominators, mitigation assumptions and model boundaries; the full bridge cannot be reproduced.
• Ruling: do not divide 49 by 74 to “disprove” a differently defined estimate, or average them into a new number. The direction is supported; a precise net shortfall remains unresolved.[E31][E32]

Confidence: Medium; US/European critical-path conclusions are stronger than precise China/Gulf rankings. Asset 2 presents falsifiable analytical rankings, not a measured global league table.

Hypothesis C: NVIDIA dominance is simply a market-share story

Verdict: weakened. Its advantage combines deliverable systems, software switching costs, networking and supply allocation. Shares with different denominators are not interchangeable.

This table asks whether four different share measures support the same conclusion; missing data are not replaced by marketing statistics.

Conflict record: [冲突E39-vs-E40] China: 34% versus 8%

• Source A: IDC via Reuters, 2026-10-13, estimates 34% of actual H1 shipments.
• Source B: Bernstein via AP, 2026-06-29, forecasts approximately 8% for full-year 2026.
• Nature: actual estimate versus forecast, half-year versus full-year, potentially differing value/unit definitions.
• Ruling: do not present 8% as an achieved result. Preserve the 34% shipment definition and await full-year evidence.[E39][E40]

Conflict record: [冲突E36-vs-E37] ASIC-server share: 27.8% versus about 27%

• Sources A/B: TrendForce releases of 2026-01-20 and 2026-04-15.
• Nature: a forecast revision within one institution, not two independent sources or conflicting actuals.
• Ruling: the later forecast supersedes the earlier one for its forecast vintage; neither is completed October shipment data.[E36][E37]

Substitution asymmetry follows workloads, not a rule that training cannot be replaced while inference can. Stable, high-volume services allow ASIC design and migration costs to be amortized over more usage. Rapidly changing model architectures and time-critical training value flexibility, reliability and development ecosystems. TPU/Trainium are counterexamples to a blanket training prohibition; Meta recommendation training is not frontier generative-model training.[E41]–[E43]

Observable weakening of CUDA lock-in means upper-layer frameworks and inference engines reach heterogeneous hardware, not that CUDA disappears. Google's TPU support for vLLM/SGLang lowers some service-migration friction; kernels, communication, diagnostics and fleet operations still carry costs. OpenAI/Broadcom deployment targets are not delivered substitute capacity.[E44][E45]

Confidence: Medium. Ecosystem competition is observable; global training/inference shares and the distribution of migration costs remain unquantified.

Hypothesis D: the supply chain is one unified global market

Verdict: weakened. A permission-constrained network of interdependence is emerging, not two fully independent worlds.

The January 2026 move to case-by-case review of certain advanced-chip exports to China combines with Chinese purchasing/import choices and localization incentives to produce different legally available products, channels and costs. July 2026 UAE concessions apply to specified approved recipients, not unconditional access across the Gulf. The original 2025 AI Diffusion country map is not a valid standalone map of 2026 law.[E46][E47][E53]

This table separates specific fractures from continuing dependence.

Chinese substitution contains both policy selection and commercial adaptation. Neither excludes the other. Competitiveness should be evaluated using operational stability, effective throughput, electricity, migration labour, downtime and service guarantees. Peak specifications, announced cluster sizes and shipment counts are not complete TCO evidence. The CXMT yield model flags conversion risk between wafer input and qualified HBM stacks; it is not an official production-yield disclosure.[E49]

Subsidies and controls coexist rather than replace each other sequentially. Subsidies can alter private costs, financing thresholds and geography, but do not automatically remove construction, talent or yield disadvantages in higher-cost locations. GAO verifies award structures, not that every grant was disbursed or every factory entered production.[E48]

Confidence: Medium. Rule/recipient differentiation is documented; cross-regional same-service effective prices and comparable domestic-substitution costs remain insufficiently established.

4. Asset 1: profit map

Annual and latest-quarter results: margins, scale and accounting together

This table identifies reported profits without treating mixed businesses as pure AI companies. Dollar amounts are USD billions; Korean won amounts are separately identified.

The 2024 baseline also undermines permanent upstream stability: NVIDIA FY2025 operating income was $81.453bn, TSMC CY2024 OM was 45.7%, Micron FY2024 OP was only $1.304bn, and Google Cloud CY2024 segment OP was $6.112bn. These are changes within issuer reporting definitions, not calendar-aligned increments to AI-only profit.[E01][E03][E07][E13]

Conflict record: [冲突E08-vs-E55] Micron: approximately 81% or 87%?

• Issuer release, 2026-09-30: Q4 gross profit 47.047/revenue 54.229 = 86.76%; operating income 43.751/revenue 54.229 = 80.68%.
• Tom's Hardware, 2026-10-15, calls approximately 81% gross margin; its 2026-10-02 article uses approximately 87%.
• Nature: an apparent media metric-label mismatch; 81% is also close to annual GM, so the period must be checked.
• Ruling: use directly calculated Q4 GM 86.76% and OM 80.68%; do not label 81% as Q4 GM.[E08][E55]

Conflict record: [冲突E04-vs-E56] TSMC: OM 60.3% versus 58.7%

• An indexed page excerpt shows 60.3%; the 2026-10-15 earnings release/management report assigns 58.7% to Q3 and 60.3% to Q2.
• Nature: a cross-quarter extraction mismatch.
• Ruling: use the original Q3 table's 58.7%, not an average.[E04][E56]

Global six-stage shares: what is identifiable?

No reproducible denominator for realized global AI operating profit across all six stages was obtained. The approximately 90% estimate found is forward EBIT for a listed-company sample. Its remaining 10% cannot be assigned to private model labs. Fabricating six percentages would be less credible than retaining the gap.[E22]

This table shows identifiability and the inputs required for an estimate, rather than an unsupported pie chart.

The estimating identity is: stage AI operating profit = sum of each issuer's AI revenue × AI operating margin. Stage share = stage profit/six-stage net operating-profit total. If the total is near zero, a stage is negative, or allocations are biased, shares can exceed 100% or be negative. A pie chart should not hide this.[E54]

This draft does not pretend a range is identified. Without AI-revenue splits and fully costed private accounts, an arbitrary “20%–50% AI allocation” is not an evidence-based estimation interval. One conclusion is secure: higher margin does not mean the largest absolute pool. Micron's latest quarterly OM exceeds NVIDIA's, but its reported quarterly OP remains below NVIDIA's latest reported OP. The quarters are not identical, so this is a scale illustration, not a synchronized share calculation.[E02][E08]

Who captures rents? Contracts, not labels

• Chip/system suppliers earn premiums for performance and lower deployment risk, jointly delivered by hardware and software. If customers retain all next-generation efficiency gains, volume growth need not deliver proportional unit-profit growth.
• Fabrication/memory suppliers capture scarce-capacity, qualification and repricing rents. New capacity and customer de-specification can reverse them. Long-term HBM agreements improve visibility but can delay repricing.[E23][E24]
• Power/site owners can monetize energized locations and reliable commissioning dates. Regulation, upfront construction contributions and long-term contracts determine whether shareholders retain that value. Physical scarcity does not automatically mean high profit.
• Rental/model providers divide the residual according to utilization, SLAs, differentiation and switching costs. Depreciation, financing and retraining can absorb seemingly attractive cash margins.[E15][E18]

Hooks for public-facing content

• Does “memory earns more than GPUs” mean gross margin, operating profit or product mix?
• Does the denominator behind “upstream captures 90%” contain model labs at all?

5. Asset 2: bottleneck urgency rankings

Definition and global baseline

These are analytical rankings as of 2026-10-25 for new large-scale commercially operable AI campuses, not survey statistics. Criteria, in order, are whether a constraint blocks commissioning, availability of substitutes, time to resolution, and breadth of exposure. Adjacent ranks do not imply measurable gaps. Expansions in already energized sites are treated separately.

The table covers the requested eight stages and explains each position.

Legal accelerator access, land/permits, finance and optical/networking supply are treated as preconditions rather than squeezed into the requested eight-stage ranking. For some Chinese projects, legally available GPUs or qualified domestic substitutes are the first gate.

Different versions for the US, China, Europe and the Gulf

The table gives full ordinal rankings across the eight stages. All are scenario-based judgments, not regional population statistics.

US/European rankings draw on multiple grid/equipment sources. China emphasizes eligible advanced accelerators and qualified HBM; inexpensive western-region electricity is not equivalent to low-latency service in eastern cities. Gulf rankings assume chip permissions are satisfied. Otherwise, permission precedes every item in the table.[E26]–[E35][E46][E47][E49][E50]

Confidence: Medium for US/Europe, Low for China/Gulf. Chinese urban sites, power-constrained campuses and western training clusters should not be averaged together. Gulf heat, water, onsite power and delivery conditions require project-specific evidence. Without it, “cheap energy” must not become “fastest delivery.”

For an energized facility with power already contracted, the alternative ordering is HBM/qualified accelerators → packaging/leading-edge nodes → rack cooling/power adaptation → assembly and staffing. Power connection moves down because this particular project has cleared it, not because the wider problem has disappeared.

Lead times, shortage rates and queues are different variables

IEA's 20% is the risk of delays to planned capacity, not a 2026 chip shortage rate. The median interconnection time of generation projects is not the average wait of data-centre load customers. Transformer delivery is not an energization date. Long-term allocation or “sold out” does not directly establish a shortage percentage.[E25][E28][E30]

For HBM, CoWoS, leading-edge foundry, cooling, AI assembly and labour, comparable four-region shortage rates and lead-time distributions were not obtained. Greenfield fabs take years, while equipment additions and yield improvements can release supply quarter by quarter. Grid construction, permits and equipment form overlapping critical paths; adding every maximum lead time mechanically is incorrect.[E26][E35]

Hooks for public-facing content

• A purchased GPU cannot energize a campus: the scarce product may be a firm power-on date.
• The same chip faces different bottlenecks in an existing facility and a greenfield campus.

6. Asset 3: the chronology of profit migration

Replace the relay race with overlapping layers of rent

The timeline separates realized earnings, constraint signals and market narratives, rather than treating equity performance as profit migration.

An exclusive sequence of “chips in 2023, power in 2024, memory in 2025, models in 2026” is unsupported. A better mechanism is scarce-equipment sales first, competition for delivery locations/long-term contracts next, and realized utilization, monetization and efficiency determining retained returns. One additional unit of downstream demand can raise several stages' absolute profits while their shares move in opposite directions.

Identifying true profit migration requires a common period, sample and accounting definition, plus evidence that changes are not driven by added sample members, foreign exchange, non-AI businesses or impairment reversals. NVIDIA can grow absolute profit while losing share of newly supplied compute.[E02][E38]

Different realization windows

• Design/memory suppliers realize earnings around shipment, acceptance and repricing; customer procurement can precede site operation.
• Equipment/engineering suppliers move through orders, delivery, installation and collection. Backlog is not earned profit.
• Cloud/NeoCloud capacity generates operating cash only when energized, operable and billable. Depreciation and interest do not stop when GPUs are idle.
• Model providers can see lower request costs before better margins; R&D, free usage, partner share and capital costs still determine final profitability.

These are accounting/project mechanisms, not assertions about undisclosed recognition terms at particular issuers.[E54]

Hooks for public-facing content

• AI profits did not change owners in a single handover: several tollbooths appeared at once.
• Why is a year of stock-price leadership not necessarily a year of dominant operating profit?

7. Asset 4: counterfactuals and stress tests

These are scenarios, not forecasts. Horizons run forward from the research cutoff; numerical inputs are expressly assumed or calculated from disclosed figures.[E54]

Scenario 1: HBM supply loosens materially in 2027–2028

Transmission: more qualified bits → better buyer bargaining/inventory → pressure on ASPs or contract repricing → lower memory earnings → cheaper accelerator BOM/rental inputs → contractual allocation of savings among GPU vendors, clouds and model providers.

A reproducible unit model tests whether volume offsets price. Start with revenue 100, variable costs 30 and fixed operating costs 20, producing operating profit 50. Assume ASP falls 30%, volume rises 20% and variable cost per unit is unchanged. Revenue becomes 84, variable costs 36 and fixed costs 20; profit falls to 28, a 44% decline. This is not any company's cost structure. It demonstrates why more units need not offset falling price.[E54]

The most exposed suppliers have high costs, weak yields, short contracts and expansion debt. Leading suppliers retain qualification/manufacturing advantages, but cannot guarantee peak-cycle margins. The most plausible beneficiaries are system/cloud buyers with demand and effective cost pass-through. If competition forces retail prices down, users of models may ultimately capture the savings.

The counter-chain is that larger HBM4/4e dies and more complex stacking absorb added wafer capacity, so qualified bit supply grows less than nameplate capacity. TrendForce's September 2026 view still forecasts tightness in 2027: loosening is a stress test of that view, not an asserted base case.[E24]

Track qualified bits, certification, unit cost and cancellation rights, not factory groundbreaking alone. Separate HBM, conventional DRAM and NAND; they need not become oversupplied together.

Scenario 2: inference ASIC/NPU substitution accelerates over 12–18 months

Transmission: more portable workloads → custom-silicon utilization exceeds its economic threshold → alternative supply pressures incremental GPU orders/prices → NVIDIA discounts or increases system value → profits respond to volume, price, cost and mix together.

Custom chips are not free. Design, software migration, supply assurance and failure risk must be amortized over sufficient usage to beat full-stack GPU TCO. Stable inference is a natural candidate, not a universal rule: TPU/Trainium also train models, while rapidly changing models, communication-intensive tasks and short iteration windows may favour general-purpose GPUs.[E41]–[E45]

The table is a mechanical NVIDIA pricing illustration, starting from its FY2027 Q2 75% GM and normalizing unit price to 100/cost to 25. It is not a predicted discount.[E02][E54]

If only 30% of sales are exposed to a 20% price reduction, with volume/cost/mix unchanged, total revenue falls 6% and gross profit 8%. Additional volume loss in the exposed portion increases the impact. The 30% and 20% are assumptions, not share estimates. Product upgrades, cost reductions or more networking/software value can offset the effect.[E54]

Beneficiaries may include scale-qualified custom-chip buyers and their foundry, packaging and networking suppliers, not every ASIC startup. Replacing a GPU need not reduce HBM or leading-edge demand: purchasing power may simply move to another customer.

Scenario 3: the US tightens China compute-export restrictions further

Transmission: more restricted SKUs/entities → less legal supply → domestic substitution, lower specifications or project delays → heavier HBM/packaging/software constraints → potentially higher effective token cost initially → guaranteed domestic demand encourages learning/investment → selected costs may fall only later.

Short term, scarce eligible hardware can earn higher rents, but energy and migration costs may reduce buyers' effective purchasing power. Supply excluded from China does not automatically become cheaper elsewhere: it may already be allocated. If outside demand is insufficient, inventory and discount risks increase instead.[E39][E46][E49]

Medium term, more certain demand helps domestic vendors amortize development, but does not guarantee process, HBM yield or software breakthroughs on policy schedules. Effective compute cost should use usable throughput × stable operating time. Comparing peak specifications of restricted imports with domestic system sticker prices is invalid.[E49][E50]

A stress illustration: with unchanged facility cost, a decline in effective billable output from 100 to 80 raises fixed cost per output by 25%, not 20%. Extra power/software expenses add further pressure. Conversely, software optimization can improve utilization and lower service cost without a chip upgrade.[E54]

Third countries also bear compliance, traceability and end-use costs. An approved-recipient UAE pathway is not permission to reexport to China. The plausible long-run outcome is partially fragmented product/software stacks with continuing manufacturing/material dependencies, not two neatly self-sufficient markets.[E47][E48]

Hooks for public-facing content

• If HBM becomes cheaper, who benefits first: chip makers, clouds or token buyers?
• Why can a 20% GPU price cut reduce gross profit by more than 20%?
• How can a cheaper domestic card produce a more expensive effective token?

8. Asset 5: evidence gaps and unresolved questions

Claims that must not be presented as settled

The table identifies missing evidence, why it matters and what would change the verdict.

What this draft completes, and what it does not

This draft supplies hypothesis verdicts, a six-stage profit map, four-region rankings, chronology, counterfactuals and an evidence ledger. It does not meet independent, same-definition double sourcing for every key number, and does not establish credible intervals for global six-stage profit shares. Gaps are explicit rather than filled with syndicated reporting or assumptions disguised as observations. This is a source-controlled research base, not an audited global profit census.

Single-disclosure financial facts retain reproducible calculations but receive reduced evidence confidence. Unreconciled private-company profits and regional inferences are further downgraded. Subsequent publication must retain these qualifications and must not turn “not established” into “did not happen.”

Hooks for public-facing content

• The recurring AI-numbers trap: treating revenue, ARR, gross profit and cash as the same thing.
• Good research does not fill every cell; it knows which cells cannot responsibly be filled.

9. Evidence ledger

The table records dates, provenance and verification limits. Financial data are reproducible from issuer disclosures, but a single underlying disclosure is rated Medium under the requested rule. Syndication is not independent corroboration. Confidence labels use 高 / 中 / 低 / 暂不裁决 (High / Medium / Low / Unresolved) to preserve the stipulated schema.

10. References by source tier

All sources accessed 2026-10-25. Publication/data dates are in the ledger; dynamic pages are not immutable snapshots.

Primary financial data

• [E01] NVIDIA, Fourth Quarter and Fiscal 2026 Results — FY2026 ended 2026-01-25; published 2026-02-25.
• [E02] NVIDIA, Second Quarter Fiscal 2027 Results — FY2027 Q2 ended 2026-07-26; published 2026-08-26.
• [E03] TSMC, 2025 Annual Report — CY2025/CY2024; 2025 annual report.
• [E04] TSMC, 3Q26 Earnings Release — 2026 Q3; 2026-10-15published.
• [E05] SK hynix, FY25 Financial Results — CY2025; 2026-01published.
• [E06] SK hynix, 2Q26 Financial Results — 2026 Q2; 2026-07-29published.
• [E07] Micron, Fourth Quarter and Full Year Fiscal 2025 Results — FY2025 ended 2025-08-28; 2025-09-23published.
• [E08] Micron, Fourth Quarter and Full Year Fiscal 2026 Results — quarter ended 2026-09-03; 2026-09-30published.
• [E09] Dell, Fourth Quarter and Full-Year Fiscal 2026 Results — FY2026 ended 2026-01-30; 2026-02-26published.
• [E10] Dell, Second Quarter Fiscal 2027 Results — FY2027 Q2 ended 2026-07-31; 2026-09-01published.
• [E11] Amazon, Q4 2025 Earnings Release — CY2025; 2026-02published.
• [E12] Amazon, Second Quarter Results — 2026 Q2; 2026-07-30published.
• [E13] Alphabet, 2025 Form 10-K — CY2025/CY2024; 2026-02disclosed.
• [E14] Alphabet, Q2 2026 Results / SEC Exhibit 99.1 — 2026 Q2; 2026-07-22published.
• [E15] CoreWeave, Quarterly Results archive — CY2025 and 2026 Q2; 2026-02-26/2026-08-11published.
• [E51] AMD, 2025 Results and Q2 2026 Results — CY2025 and 2026 Q2.
• [E52] Broadcom, Third Quarter Fiscal 2026 Results — quarter ended 2026-08-02; 2026-09published.
• [E56] TSMC, 3Q26 Management Report — 2026-10-15.

Industry, sell-side and specialist research

• [E19] SemiAnalysis, AI Value Capture – The Shift To Model Labs — early 2026 to 2026-05-01.
• [E22] Goldman Sachs Asset Management, Investing in the Architecture of AI’s Future — FactSet snapshot 2026-02-26; article2026-05-07.
• [E23] TrendForce, Tight DRAM Supply Gives Suppliers Greater Pricing Power in HBM — 2026 Q1; 2026-06-02published.
• [E24] TrendForce, HBM Supply Constraints Persist — 2026-09-29forecast for 2027.
• [E25] IEA, Energy and AI – AI and energy security — 2025 report; planning period 2025–2030.
• [E26] IEA, Key Questions on Energy and AI — 2026-04report.
• [E27] Rystad Energy, Grid Equipment Market Outlook — 2026-05-28.
• [E30] Goldman Sachs, Navigating a System of AI Bottlenecks — 2026-10-19article; some inputs are older observations.
• [E31] Morgan Stanley, Powering AI: Markets Race to Invest — 2026-02article, forecast for 2028.
• [E32] Morgan Stanley, AI Rate of Change — 2026-02-11report, p.44.
• [E33] Uptime Institute, Global Data Center Survey Results 2026 — 2026survey; as of 2026-10-25accessed.
• [E34] IEA, Powering Artificial Intelligence — 2026-10-14report.
• [E35] TrendForce, AI Competition Turns into a Supply Chain Arms Race — 2026-04-30.
• [E36] TrendForce, Global AI Server Shipments Forecast — 2026-01-20forecast for CY2026.
• [E37] TrendForce, Extended Component Lead Times — 2026-04-15forecast for CY2026.
• [E38] Epoch AI, AI chip market share by designer — 2026 Q2; 2026-10-21updated.
• [E49] SemiAnalysis, China’s CXMT Is Set to Challenge DRAM Incumbents — 2026-06-23model.

Professional media and investigative reporting

• [E16] Reuters, OpenAI expects compute spend of around $600 billion through 2030 — 2025/2024; 2026-02-20reporting.
• [E17] The Decoder, OpenAI reportedly improved compute margins — 2025-10; 2025-12-21reporting.
• [E18] Reuters via CNA, Anthropic logged net loss of over $80 billion — CY2025/2026 H1; 2026-10-20reporting.
• [E20] Reuters via MarketScreener, Anthropic second straight quarterly profit — 2026-09-13reporting; Q2 actual / Q3 expectation.
• [E21] Reuters Breakingviews, Anthropic’s $2 trln dream faces deflation trap — first nine months of 2026; 2026-10-15, 10-16updated.
• [E28] Reuters, US power companies scramble to secure equipment — 2026-07-09reporting, respectively 2026 Q1/2025 H2.
• [E29] Reuters, Power grid delays challenge Amazon’s expansion in Europe — 2026-02-03.
• [E39] Reuters, Nvidia’s China AI-chip share falls — 2026 H1; 2026-10-13reporting.
• [E40] AP, Nvidia’s AI chip sales in China stall — 2026-06-29reporting, forecast for full-year 2026.
• [E55] Tom’s Hardware, Elon Musk memory pricing report — 2026-10-15.

Company announcements and regulatory documents

• [E41] Google Cloud, Ironwood TPUs and new Axion-based VMs — published 2025; accessed 2026-10-25.
• [E42] Amazon, Fourth Quarter Results / Trainium deployment — February 2026 earnings release.
• [E43] Meta Engineering, MTIA 300 — 2026-08-24.
• [E44] OpenAI and Broadcom, Strategic collaboration — 2025-10-13announcement; additional Jalapeño announcement June 2026.
• [E45] Google Developers, New vLLM and SGLang Inference Engines on TPU — 2026-10-20.
• [E46] BIS, Revised License Review Policy for China — 2026-01-13.
• [E47] Federal Register, Enhanced Favorable Treatment for UAE — 2026-07-10effective; 2026-07-14published.
• [E48] GAO-26-107882, Semiconductor Projects — July 2025 award snapshot; 2025-12report.
• [E50] 中国网信网，东数西算：贵州算明白了三件事 — 2026-09-01; project status end-July 2026.
• [E53] BIS, Rescinds Biden-Era AI Diffusion Rule — May 2025 announcement; checked against 2026 rules.

Additional cross-checks and exclusions

• CoreWeave Q2 2026 full issuer release via Business Wire — 2026-08-11; accessed 2026-10-25.
• NVIDIA Q2 FY2027 10-Q — 2026-08; accessed 2026-10-25.
• TSMC 3Q26 Management Report — 2026-10-15; accessed 2026-10-25.
• Yonhap, SK hynix record Q2 profit — 2026-07-29; accessed 2026-10-25.
• Tom’s Hardware, Micron 88% consumer-memory margin — 2026-10-02; accessed 2026-10-25.
• Tom’s Hardware, Elon Musk memory pricing / 81% margin wording — 2026-10-15; accessed 2026-10-25.
• AMD Q2 2026 results — 2026-08-04; accessed 2026-10-25.
• Google TPU architectural research — 2026-06; accessed 2026-10-25.
• Broadcom/OpenAI Jalapeño announcement — 2026-06-24; accessed 2026-10-25.
• ChosunBiz, CXMT HBM yield — 2026-09-09; accessed 2026-10-25.
• BIS UAE press release — 2026-07-10; accessed 2026-10-25.
• GAO follow-on CHIPS review — 2026-08; accessed 2026-10-25.
• J.P. Morgan AI capex outlook — 2026-10; accessed 2026-10-25.
• SK hynix IR release calendar — accessed 2026-10-25; accessed 2026-10-25.
• Gartner, Companies to Beat in AI semiconductors — 2026-07-16; accessed 2026-10-25.

Forum reposts, search rank, inaccessible paid material and unverified “industry averages” are not independent evidence. No paid report was purchased or paywall bypassed.

This is the Day01 research draft. All data are governed by the time anchors in the evidence ledger; unresolved conflicts must not be presented as settled facts in published work.