[OpenXSys] Ding. System activated.

A structural report · About 18 minutes to read

Intelligence at the surface,
profits in the bedrock.

The AI industry sells scarcity. Drill through eight layers: where money comes from, where it accumulates, and which loop is losing pressure.

Data as of 2026-11-16 · Market data as of the November 13 close · Sources and definitions at the end

Doubao Work × DeepSeek v4.1 Flash · Generated from one sentence

Prompt (English translation):

/open-x-sys I want to research the entire artificial intelligence industry!

About this run and the repository edition →

2026 Q3 · Operating profit of the three memory leaders$179B
Same quarter · Capital expenditure of four cloud providers$209B
OpenAI + Anthropic · Annualized revenue$131B

The first figure is profit; the second is spending. A year's revenue from the two leading model companies is less than three months of memory profits.

Click a method to see what it examines here. This report is the result of following that route.
Descend 70 meters
DiscoveryFirst, locate the profits

A year's revenue from the two leading model companies is less than one quarter of profit at three memory makers.

This reveals the distribution of profit across the chain: closer to “intelligence,” the business looks like a subsidy; closer to what is hard to replicate, it looks like collecting rent.

In Q3 2026, Samsung DS, SK hynix and Micron generated combined operating profit of approximately $179 billion171819; in the same quarter, four cloud providers spent approximately $208.7 billion12on capital expenditure. Memory profits equaled 86% of cloud capex.

Meanwhile, OpenAI and Anthropic had combined annualized revenue of approximately $131 billion934. This compares annualized revenue with quarterly profit. Although the measures differ, their scale illustrates the point: making intelligence earns less than selling memory.

Measurement note:Quarterly operating profit and annualized revenue use different measures; this comparison illustrates scale only. Samsung DS includes some foundry operations, SK hynix is converted at KRW 1,400/USD, and Micron uses FY26 Q4, ended September 3.
One quarter vs. one year2026 Q3 / 2026 annualized
Three memory leaders · Quarterly operating profit$179B
Samsung DS $76B + SK hynix $60B + Micron $44B
OpenAI + Anthropic · Annualized revenue$131B
OpenAI $50B + Anthropic $81B
Bars use one scale. Memory figures are quarterly operating profit; model-company figures are annualized revenue. Sources: company reports, S-1 filings and media disclosures.
For every $52 spent on infrastructure, $1 goes to model softwareGartner 202610
Generative model software: $28.3BAI infrastructure: $1,480B
Each cell represents about $2.8 billion. Gartner's “AI infrastructure” includes servers, data centers, networking and devices; it differs from cloud capex. “Model software” also differs from model-company revenue. The ratio illustrates where spending concentrates.
0 mSurface · Applications

Applications are closest to the customer, yet receive only $77.3 billion.

Enterprise AI deployment spending grew from $33 billion to $176.8 billion in one year, but most money went to models and infrastructure; independent applications accounted for 44%.

Under Menlo Ventures' 2026 definitions, total enterprise AI deployment spending was $176.8 billion, 5.4 times the 2025 level; infrastructure accounted for $99.5 billion, applications $77.3 billion, and training data and services $15.3 billion11. Within applications, coding received $34 billion and enterprise chatbots $25 billion—a combined 76%.

Frontier labs captured $122.7 billion, or 69% of enterprise deployment spending. They supply models and develop applications; independent app companies must consider what happens when their model supplier builds the same product.

Meanwhile, 72% of enterprises use AI in production, yet nearly 80% of deployments have brought no net-profit improvement, and 95% of generative AI pilots have produced no direct financial return32. Failed pilots and successful deployments coexist. The difference is whether AI is embedded in workflows and produces verifiable results.

Where application spending goes 76% Coding and enterprise chatbots take three quarters of application spending. Healthcare, customer service, legal and financial back offices remain early.
Enterprise AI deployment spending: $176.8BMenlo Ventures 2026
Infrastructure $99.5BApplications $77.3BTraining data & services $15.3B (separate scope)
Of $77.3 billion in application spending, coding accounts for $34 billion, chatbots $25 billion, healthcare $3.7 billion, customer service $2.7 billion, and the remainder roughly $11.9 billion. Menlo's categories overlap and do not sum to the total; this compares category scale.
Concentration of application spending2026
Coding
$34B
Enterprise chatbots
$25B
Healthcare
$3.7B
Customer service
$2.7B
Other
$11.9B
“Other” is the remainder after subtracting listed categories, rather than one category. Coding and chatbots share verifiable outputs, frequent use and low switching costs.
10 mModels

The smartest layer has a margin of just 10%.

Model gross margins resemble software; expense ratios resemble heavy industry. These companies sell tokens, pay for compute, sign long-term commitments and face their first public-market valuation.

OpenAI reported 2025 revenue of $13.07 billion, total costs of roughly $34 billion and operating losses of $20.92 billion. Payments to Microsoft Azure were $17.2 billion, more than its full-year revenue8. In H1 2026, revenue was $12.4 billion and operating losses $21.6 billion; Q2 alone brought $6.7 billion in revenue and $12.3 billion in operating losses9。

Anthropic followed a different path: H1 2026 revenue was $16.286 billion, with operating losses of just $693 million. Q2 operating profit reached $974 million, a margin of about 8%; but Q3 revenue of $17 billion and expenses of $17.5 billion brought it back to $567 million in operating losses534. Profitability is a choice rather than a settled condition.

Both trade the future for the present: OpenAI's nominal ceiling on chip and cloud contracts is approximately $710 billion, with planned negative free cash flow for 2026–2030 of $278 billionand compute and infrastructure spending of $856 billion9; Anthropic's long-term compute commitments are $90.09 billion, including $451.6 billion due before 20305。

Key difference:OpenAI combines consumer subscriptions, Azure bills and owned compute; Anthropic combines enterprise APIs, higher pricing and more restrained spending. The former has greater scale, the latter better unit economics. Both bet on continued exponential token growth.
Spending per $1 of revenue2026
$2.22OpenAI · Total cost / revenue
2026 Q1, adjusted
$1.04Anthropic · Total expenses / revenue
2026 H1
OpenAI
2.22×
Anthropic
1.04×
Break-even
1.00×
Full bars represent 2.4×; 1.00× is break-even. OpenAI uses Q1 adjusted operating losses/revenue; Anthropic uses H1 total expenses/revenue. The measures differ, but point in the same direction: the model layer still trades profit for scale.
Enterprise model API shareMenlo, September 2026
AnthropicOpenAIGoogleOther 5%
Enterprise APIs are the model layer's most profitable entry point; consumer subscriptions are a separate market. Menlo Ventures' enterprise survey measures share within its sample, rather than total market revenue.
The first public-market valuation 11-18 Anthropic plans to list on Nasdaq as ANT at $70–78 per share, issuing 825 million shares. Basic market capitalization would be about $916 billion–$1.02 trillion, or $1.3–1.4 trillion fully diluted. Seven technology companies committed $26 billion in private placements: Google $10B, Nvidia $10B, AMD $2.5B, Broadcom $1B, Micron $1B, Microsoft $1B and TSMC $0.5B.67
MetricOpenAIAnthropic
Annualized revenue$50B(2026-09)$81B(2026-10)
2026 H1 operating resultsRevenue $12.4B; operating loss $21.6BRevenue $16.29B; operating loss $0.69B
Gross margin39% (2026 Q1); 33% in 202585%(2026 H1)
Latest quarterly operating profitQ2 approximately -$12.3BQ2 +$0.97B;Q3 -$0.57B
Cash on handApproximately $73B (2026 Q1)$131.8B(2026 Q3)
Long-term compute commitments$710B chip/cloud contract ceiling; $1.2T infrastructure plan$900.9B; including $451.6B due before 2030
Valuation / IPO$852B (2026-03); target $1.4T$965B (2026-05); IPO basic market cap $0.92–1.02T
Note:OpenAI figures follow unaudited media disclosures; Anthropic figures follow its S-1 and amendments. Revenue recognition, compute commitments and stock-based compensation differ. Individual figures do not support a direct ranking; the comparison examines structure.
Discovery 02Why deeper layers earn more

Cheaper intelligence,
costlier bedrock.

Every model cost reduction sends orders to the physical layer. The price of a unit of intelligence falls, while calls, tasks and infrastructure demand grow faster.

Price per unit of intelligenceAbout 50× cheaperper year, at fixed capability
Global inference tokens14×year on year; about 30 trillion/month in mid-2026
Enterprise AI spending+35%About $407 billion in 2026
A direction rather than an equation: faster price declines bring faster growth in use and new tasks; the latter determine total spending. This is the AI industry's version of the Jevons paradox.

In 1865, Jevons observed that more efficient steam engines increased British coal consumption: cheaper power enabled new uses. AI follows a similar curve. Cheaper tokens enable agents, long contexts, multimodality and real-time interaction, ultimately returning the bill to memory, packaging, power and capital33。

Model efficiency gains therefore redistribute value toward capacity that is hardest to expand. The harder the model layer works, the more memory, packaging and power can earn.

20 mCloud & capital

Cloud providers are AI's bankers.

In 2026, four cloud providers will spend about $730 billion on capex—a sum their own free cash flow can no longer cover.

The four North American cloud providers spent approximately $208.7 billion, up 84.2% year on year; combined full-year guidance is approximately $720–744.5 billion, versus $413 billion in 202512. Including Oracle's FY2027 guidance brings the total to about $835 billion. For 2027, sell-side consensus points to $1.0–1.2 trillion13。

Where does the money come from? Hyperscalers issued approximately $182 billionof bonds in 2025, versus $92 billion in 2024. Alphabet's debt rose from about $16 billion to about $100 billion; in June 2026 it also announced an $80 billion equity raise14. JPMorgan estimates that only about 15% of large AI companies' debt and lease commitments appear on their balance sheets13。

And returns? Capex/EBITDA is approximately 90%in 2026 and approximately 94%in 2027; incremental EBITDA yield on the previous year's capex is only 20–40%13. With GPU lifetimes of 5–6 years, a low-return asset approaches retirement just as it pays back.

The question is who finances AI.Who advances the money, and for how long? Once capex exceeds operating cash flow, interest rates, credit spreads and share prices set the pace alongside demand.
Capital expenditure of four cloud providers2023–2027E, USD billions
2023/2024 are industry estimates for the four providers; 2025/2026 use company guidance; 2027 uses sell-side forecasts (Capital Economics: $1.0–1.2T). Bars start at zero; forecast bars are hatched.
2026 Q3 free cash flowUSD billions
Alphabet
-$9.1B
Meta
-$2.1B
Microsoft
+$18.6B
Amazon
-$42.4B
Amazon uses trailing-12-month free cash flow; the others use one quarter. Microsoft's figure is positive but down 27% year on year.
Free cash flow = operating cash flow − capital expenditure. Three of four are negative: capex has exceeded current cash generation.
Five-year recovery per $100 of capexIncremental EBITDA yield: 20–40%
012345 years $100$200$0 20%/year30%/year40%/year GPU lifetime: 5–6 years
Illustrative curves, rather than actual company returns. At low yields, cumulative recovery only reaches cost near the end of an asset's life; at high yields, payback takes 2.5 years. Utilization and pricing determine the difference.
90% → 94%Capex / EBITDA, 2026 → 2027E. Capital expenditure consumes almost all EBITDA.
15%Share of large AI companies' debt and lease commitments on balance sheet. The rest sit off balance sheet, in joint ventures or in supplier financing.
30 mChips

Nvidia collects a tax on compute.

FY2027 Q2 revenue was $96.2 billion, with data centers contributing 92.5% and gross margin at 75%; Q3 guidance was $108 billion. Its constraints are supply, governance and the cycle.

FY2026 revenue was $215.9 billion (+65%), with a 71.1% gross margin. FY2027 Q2 revenue, for the quarter ended 2026-07-26, was $96.221 billion(+106% year on year, +18% sequentially), including data-center revenue of $89.023 billion(+117% year on year, 92.5% of total); GAAP gross margin was 75.0%, GAAP operating profit about $63.8 billion and GAAP net income $59.688 billion, including $7.78 billion in equity gains1。

Q3 guidance $108 billion(±2%), with a 74% gross margin and no China data-center revenue. The company said customer forecasts suggest potential business could double next fiscal year, but chip, memory, power and construction constraints limit confident growth to approximately 70%. This is a supply story.

The market prices it cyclically: market cap $3.92 trillion, trailing P/E 20.35× and share price $162.85, down from an October high of about $243 by 33%; the November 5 decline alone was 9.3%331. At 20× earnings for a $3.9 trillion company, the market already prices in slower growth.

Governance risk In November, the founder resigned as CEO; the board launched an independent investigation into related allegations; CFO Colette Kress became interim CEO. At least 10 analysts downgraded the stock to Hold, while Goldman Sachs, JPMorgan and RBC cut targets. The FY2027 Q3 report, due after the November 17 close, is the first earnings report after the event3. This is a governance stress test for a $3.9 trillion company and an example of key-person risk.
Data-center revenue: Hyperscale vs. ACIEUSD billions, FY26 Q2–FY27 Q2
Hyperscale cloudACIE: AI cloud / industrial / enterprise / sovereign
ACIE grew 138% year on year, faster than Hyperscale's 102%; it approached half of revenue in Q1, before Hyperscale reaccelerated in Q2. Demand is broadening but still depends on cloud providers.2
$108BFY2027 Q3 revenue guidance (±2%), excluding China data centers
75.0%Q2 GAAP gross margin; Q3 guidance: 74%
20.3×Trailing P/E; market cap $3.92 trillion
-33%Decline from the October high; November 5 alone: −9.3%
Supplier commitments, guarantees and receivable days2026
Supply-chain commitments
$279B
Maximum AI-cloud guarantee
$108.5B
Days sales outstanding
60 days
Supply commitments jumped from $119 billion to $279 billion; DSO stretched from 45 to 60 days. As suppliers finance customers, demand signals become less independent. This leads into the loop examined next.
40 mMemory

HBM is AI's oil: three suppliers, 80% gross margins, annual contracts.

This layer has no consumer brand or developer ecosystem—just capacity, yields and qualification cycles. It captures the thickest slice of profit in the chain.

SK hynix reported Q3 2026 revenue of KRW 104.62 trillion and operating profit of KRW 83.33 trillion, an operating margin of 80%; net income was KRW 39.78 trillion. More than 50% of 2027–2028 capacity is locked in, with 14 customers signing five-year contracts; the company expects shortages through 203017。

Samsung's Q3 2026 revenue was KRW 195.3 trillion, with operating profit of KRW 107.4 trillion, making it the first technology company to exceed KRW 100 trillion in quarterly operating profit. DS revenue was KRW 147.1 trillion and operating profit KRW 106.5 trillion, a 72% margin; memory revenue was KRW 140.7 trillion. HBM4 represented about 60% of HBM sales, and 2027 HBM capacity was sold out at target prices18。

Micron's FY2026 revenue, for the year ended September 3, was $133.19 billion, with GAAP operating profit of $99.34 billion, an operating margin of 80.7%; Q4 operating profit was $43.75 billion. HBM4 36GB 12-high entered volume production in Q1 2026, and all 2026 HBM capacity was sold out19。

Definitions:Samsung DS includes some foundry operations. SK and Samsung use 2026 Q3; Micron uses FY26 Q4. KRW figures are converted at 1,400/USD. Combined quarterly operating profit is approximately $179 billion.
HBM stacks: each generation costs morePrice per stack
Substrate / interposer 12-Hi HBM4 36GB / stack $550 HBM3E: about $300 HBM3: about $200 Around one GPU typically sit 8 stacks Price per GB rises with each generation
Schematic, rather than an engineering drawing. HBM4 36GB 12-high costs about $550 per stack; HBM3E about $300 and HBM3 about $200. Supply-chain and media estimates vary with contracts and yields.
HBM revenue share2026 Q2,Counterpoint
SK held about 64% a year earlier. Samsung regained share through early HBM4 volume production; Micron joined the leading tier with HBM4 36GB 12-high. Shares measure revenue rather than capacity; rounding brings the sum slightly above 100%.20
Company / scopeOperating profitOperating margin
SK hynix
2026 Q3
KRW 83.33 trillion
About $59.5B
80%
Samsung DS
2026 Q3, including some foundry operations
KRW 106.5 trillion
About $76.1B
72%
Micron
FY26 Q4, ended September 3
$43.75B80.7%
Combined total: approximately $179B in quarterly operating profit. This is scarcity rent: long HBM qualification cycles, slow capacity expansion and customers willing to sign long contracts. It is also a cyclical peak, rather than a permanent moat.
AI data centers consume 70% of global memory capacity2026, IDC / supply chain
AI data centersConsumer electronics & other
HBM consumes 3–4× as much wafer capacity per GB as ordinary DRAM. AI competes for both memory and wafers. IDC forecasts 2026 phone shipments down 5% and PC shipments down 9%.21

This transmits the “memory tax”: AI accepts 80% HBM gross margins, so memory makers move wafers from DDR5 to HBM, reducing consumer-electronics supply and raising prices. Your next phone may help pay the data-center bill.

HBM revenue: 2027 is a forecastUSD billions
2025 is back-calculated from 2026 growth; 2026 is about $54.6B (BofA, +58%). TrendForce forecasts 2027 ASP growth of 122% to $3.7/Gb and revenue growth of 230%, exceeding $200B. Sell-side forecasts vary widely.
2026 · ShortageCapacity sold out, prices rising, long contracts locked in.
2027 · PeakASP forecasts rise, revenue surges and capex accelerates.
2028 · Capacity arrivesNew capacity comes online, supply catches up and prices soften.
2029 · DownturnThe cycle turns down, margins normalize and weaker players exit.
Key judgment:An 80% gross margin is scarcity rent. Qualification cycles and oligopoly are real; so is cyclical pricing. Treating a cyclical peak as perpetual is this layer's biggest misreading.
50 mPackaging

CoWoS is AI's on-ramp: without it, GPUs cannot leave the line.

TSMC has turned packaging from a back-end cost into a source of profit. The constraint is how many GPUs and HBM stacks can fit into one package.

TSMC reported Q3 2026 revenue of $46.72 billion(+16.2% sequentially, +41.2% year on year), gross margin of 66.4%, operating margin of 58.7%and net margin of 51.1%, with net income of TWD 763.06 billion; HPC contributed 68%. 2026 capex was $60–64 billion, including $13.84 billion in Q3; Q4 revenue guidance was $52.3–54.3 billion, with roughly 44% full-year USD revenue growth15。

Monthly CoWoS capacity rose from approximately 70,000 wafers in 2025 to 120,000–140,000 wafersby year-end 2026. Adding OSAT partners' 50,000–60,000 gives industry capacity of roughly 170,000–200,000 wafers/month. Customers still face a shortage of 10–20%; TSMC raised CoWoS prices by 10–20% in October 2026. C. C. Wei said packaging capacity is limiting customer growth16。

Reticle sizes are growing: 3.5× is in volume production; 5×/5.5× entered volume production in 2026; 9× and 14× are in development. The 14× target for 2028 could integrate about 10 large compute dies and 20 HBM stacks. CoWoS wafer ASP approaches $10,000, with margins above TSMC's average and packaging revenue expected to exceed 10% of total.

This means:Packaging has become part of the advanced-process system. Access to CoWoS determines who can define next-generation systems. Foundries, HBM makers and packaging providers are becoming three faces of one bottleneck.
Inside a CoWoS packageSchematic, rather than engineering drawing
Substrate / package Silicon interposer · CoWoS GPU Compute die HBM HBM 5.5× reticle · Volume production in 2026 One GPU + 8 HBM stacks; the 14× target for 2028 is about 10 compute dies + 20 HBM stacks
CoWoS places the GPU and HBM on one silicon interposer, connected by TSVs and microbumps. Larger packages fit more HBM and bandwidth, but make yield harder. Manufacturing capacity is the bottleneck.
Monthly CoWoS capacityThousand wafers/month, equivalent
2025 · TSMC
70k
2026 · TSMC
120–140k
2026 · + OSAT
170–200k
Customers still face a 10–20% shortfall; October prices rose 10–20%.
Capacity figures are supply-chain and media estimates in equivalent monthly capacity; consumption varies by product and reticle size. They show the expansion rate rather than a precise supply-demand gap.
3.5× · 2023–24In production; the H100-era packaging baseline.
5× / 5.5× · 2026In production; larger packages for Blackwell / Rubin.
9× · 2027–28In development; more compute dies and HBM.
14× · 2028Target; about 10 large compute dies + 20 HBM stacks.
66.4%TSMC Q3 2026 gross margin; operating margin: 58.7%.
10–20%CoWoS customer shortfall; packaging now limits growth.
60 mPower & the grid

AI's next wall is electricity.

You can build a data center in 18 months and wait five years for electricity. Power is the one input money cannot supply immediately.

Under IEA definitions, global data-center electricity use in 2024 was 415 TWh(about 1.5% of the global total), about 485 TWh in 2025 and approximately 945 TWhin 2030, approaching Japan's national consumption. AI-optimized servers account for roughly 15% of data-center electricity use, growing about 30% annually22。

US data-center load: about 23 GW in 2023, 31 GW in 2025, 41–42 GW in 2026, 66 GWin 2027 and 134 GW23in 2030. Grid-connection queues exceed 2,060 GWacross about 8,200 projects. The median wait for projects commissioned in 2025 exceeded five years; only 13% of capacity proposed in 2000–2020 was built.

Prices speak: PJM capacity prices rose from $28.92/MW-day to 269.92(+833%), hit the $329.17 ceiling in 2026/27, and reached $325 in 2028/29, with a 6,831 MW shortfall24. Large-transformer lead times are about 128 weeks, step-up transformers about 144. GE Vernova heavy-duty gas turbines are sold out through 2028, with commitments extending to 2031; prices rose about 300% in three years to roughly $600/kW. Around 50% of planned 2026 capacity may be delayed25。

Time is the constraint.Global data-center investment exceeded $550 billion in 2025 (+17%), surpassing upstream oil and gas investment. Power-purchase agreements reached 45 TWh in H1 2026 (+20%). Nuclear agreements cover 9.8 GW, 13 projects and seven buyers, but most supply starts after 2030.
US data-center loadGW,2023–2030E
2023–2027 use Goldman and the median of several forecasts; 2030 is a forecast. Nearly 6× in five years: a step change.
From decision to power: how long?Years
Data-center construction
1.5 years
Transformer delivery
2.6 years
Grid-connection queue
5+ years
Heavy-duty gas turbines
5 years
Nuclear power
7–10 years
Illustrative typical times from investment decision to first power. The slowest input determines the whole: a data center can be built in 18 months, while power takes five years.
PJM capacity-auction clearing priceUSD / MW-day
0100200350 2024/252025/262026/272028/29 $28.92$269.92$329.17$325 +833%6,831 MW shortfall
PJM is one of the largest US power markets. Capacity prices pay for availability, rather than electricity itself; rising prices indicate a premium for reliability.
Global data-center investment $550B+ 2025 (+17%), exceeding upstream oil and gas investment; major operators' 2026 capex exceeds $700 billion.
After the November 2026 US midterms 40 / 41 local data-center proposals were rejected. Opponents won 40 of 41 referendums; some elected officials favor moratoriums or limits. The November 5 sell-off priced in this political constraint for the first time26。

Permits also constrain supply. A data center needs land, grid access, air permits, water and local political consent. Each can delay a project; treating electricity only as engineering understates the constraint.

The bottleneck moves2023–2027+
2023–24CoWoS packaging: the first gate for GPU output.
2024–25HBM: memory bandwidth and capacity become the next bottleneck.
2025–27Power and grid access: the constraint money cannot accelerate.
2027+Demand: if end-market revenue lags, the bottleneck returns to the wallet.
Goldratt's constraint: system output is set by its slowest link. Bottlenecks move; investment should watch the next one.
65 mPut the layers in one table

Profit follows irreplaceability.

Compare Q3 2026 operating profits: memory captures almost all, with chips and foundries next. The model layer's best performer earns around $1 billion; its worst loses $12.3 billion in a quarter.

The model layer creates demand for the industry and bears its research and customer-acquisition costs. On the income statement, it looks more like an intermediate toll point than the final destination.

Physical-layer profits come from slow-to-replicate capacity, long qualification cycles and concentrated supply. Model profits come from scale, data and distribution. The former are structural; the latter remain contestable.

Definitions:Memory figures use Q3/FY26 Q4 operating profit; Nvidia uses FY27 Q2 GAAP operating profit; TSMC uses Q3 net income; Anthropic uses Q2 operating profit; OpenAI uses Q2 operating losses. Measures differ; compare scale only.
2026 Q3 profit poolUSD billions
Three memory leaders
$179B
Nvidia
$64B
TSMC
$24B
Anthropic
≈$1B
The model layer's largest quarterly loss -$12.3B OpenAI's 2026 Q2 operating loss. It buys future compute, users and model capabilities, but its income statement is negative.
Full bars represent $180B; Anthropic's $974 million is almost invisible at this scale. TSMC uses net income, while Nvidia uses operating profit.
Layer2026 Q3 profitMarginWho collects the rent / why
Memory$179B (three combined)72–81%Three suppliers, HBM qualification cycles and sold-out capacity: scarcity rent rather than permanent technological monopoly.
Chips$64B(Nvidia)66%Systems, ecosystem and supply combined; customers are also shareholders and financing recipients.
Foundry / packaging$24B (TSMC net income)51%Advanced processes and CoWoS capacity; packaging becomes central to manufacturing.
Models+$0.97B / -$12.3B8% / -170%Scale, data and distribution, but higher compute and customer-acquisition costs.
AppsRevenue $77B; profits dispersed—Workflows, distribution and verifiable results; vertically squeezed by model providers.
In one sentence:The smartest layer earns the least; the heaviest layer earns the most. Margins reward being impossible to bypass.
70 mBedrock · The loop

Demand is only part of this machine's fuel.

When suppliers are also shareholders, customers and guarantors, demand signals become interdependent. Circular financing postpones independent validation of demand.

OpenAI has a nominal ceiling of approximately $710 billionin chip and cloud contracts. Nvidia, Microsoft and Amazon are both suppliers and shareholders. Nvidia's proposed $100 billion OpenAI investment was never signed; its actual investment was about $30 billion at an $852 billion valuation. Nvidia also owns about 9% of CoreWeave, commits to buying unsold cloud capacity (initially $6.3 billion), and acts as supplier, shareholder and customer4。

Nvidia guarantees some leases for OpenAI's Ohio data center; AMD reportedly offers equity-for-orders arrangements to OpenAI / Meta. Anthropic's IPO makes the pattern more direct: Google and Nvidia subscribe $10 billion each, AMD $2.5 billion, Broadcom, Micron and Microsoft $1 billion each, and TSMC $0.5 billion, totaling $26 billion—with suppliers also becoming shareholders7。

These arrangements are not inherently illegal or proof of artificial demand. They finance capacity built ahead of demand. But when supplier and customer money comes from one loop, independent demand validation weakens. Supply commitments of $279B, guarantees of $108.5B and DSO of 60 days are its gauges4。

Minsky's ladder:Build from cash flow → build from the balance sheet → build from valuations. The industry is on the second rung; marginal projects are on the third. The third depends on new financing to service old obligations.
Physical layer TSMC · HBM · Power Slowest, costliest, hardest to replicate Nvidia Chips + systems Supplier / shareholder / guarantor Cloud / models OpenAI · Anthropic · Oracle CoreWeave · neoclouds End users Enterprises · Consumers Sovereign customers Wafers GPU token Money Money Money Investment / guarantees Equity / IPO subscriptions
Goods / servicesMoneyCapital / commitments
Without the capital loop, one value chain remains: physical layer → chips → cloud/models → end users. Money returns along the same path. Switch to “The full loop” to see supplier investments, guarantees and equity turn the chain into a cycle.
Long-term commitments / annualized revenuetimes
OpenAI
14.2×
Anthropic
11.1×
Nominal contract ceilings / current annualized revenue. Commitments differ from debt but still entail rigid spending. Revenue must grow for commitments to become economies of scale rather than fixed costs.
Loop gaugesReadingWhy watch it
Supply-chain commitments$279BUp from $119B; suppliers lock in future capacity for customers.
Maximum AI-cloud guarantee$108.5BSuppliers back customers' leases/capacity; failed demand returns to suppliers' balance sheets.
Days sales outstanding60 daysUp from 45 days; the gap between revenue recognition and cash collection is widening.
Customer free cash flowThree negativeCapex exceeds operating cash flow; capital markets begin setting AI's pace.
Supplier IPO subscriptions$26BSeven technology companies are Anthropic shareholders, suppliers and customers.
Commitments / revenue11–14×Revenue must keep growing rapidly to absorb fixed commitments.
Reflexivity.Capex creates revenue expectations, which lift valuations, which support more capex. On the way up, this appears to validate demand; on the way down, the same mechanism contracts balance sheets.
Build from cash flowCapex < operating cash flow. Steadiest but slow; end demand must arrive first.
Build from the balance sheetCapex > operating cash flow. Debt, equity and leases fund it; the industry is here today.
Build from valuationsFuture valuations and supplier financing support capex; marginal projects are here. They break first when the loop reverses.
The faultNovember 2026

On November 5, the market's question changed: from “How big can it grow?” to “When does it pay back?”

The pricing framework shifted: capex changed from proof of demand to cash-flow risk; valuation changed from revenue multiples to unit economics.

That day, the Nasdaq 100 fell 6.4%, the S&P 500 fell 4.4%and the Philadelphia Semiconductor Index fell 5.23% (down 11% intraday). About $2.3 trillionin market value vanished. Nvidia fell 9.3%, AMD 17.3%, Micron 17% and Tesla 10.8%; the Magnificent Seven lost about $1.04 trillion26。

Politics appeared to trigger it: after the midterms, 40 of 41 local data-center proposals were rejected, and some officials favored limits or moratoriums. Financial pressures supplied the tinder: capex above operating cash flow, negative free cash flow, off-balance-sheet debt, supplier financing, questioned AI returns and crowded trades.

“This is repricing.” Cash-rich giants remain. Marginal projects, off-balance-sheet leverage, high valuations and financing-dependent companies are being reassessed. Nvidia is about 33% below its October high but still trades at 20× earnings; Anthropic IPO expectations fell from $2T to $1.4T; OpenAI's latest round values it at about $1.25T, below a $1.4T target36。

-6.4% / -4.4%Nasdaq 100 / S&P 500, November 5
$2.3TMarket value lost that day; Magnificent Seven about −$1.04T
-5.23%Philadelphia Semiconductor Index; down 11% intraday
-9.3%Nvidia that day; AMD −17.3%, Micron −17%
90-day signal calendar2026-11 to 2027-04
11-17

Nvidia FY2027 Q3 earnings

First report after the CEO change. Watch data-center revenue above $100B, Q4 guidance above $120B and gross margin below 73%.

11-18

Anthropic IPO pricing

Range: $70–78. Above $78 indicates strong demand; below $70 suggests public markets are repricing the model layer.

11-19

ANT's first trading day

Watch first-day turnover and buying support: the model layer's first continuous public-market pricing.

2027-01/02

Cloud Q4 earnings and 2027 guidance

Combined 2027 capex above $900B means further acceleration; below $750B means digestion. This is the most important single signal.

2027-Q1

Quarterly HBM contract prices

Flat or higher sequentially means shortages persist; −20% marks a cyclical turn. Memory income statements respond before share prices.

2027-Q1

PJM capacity auction

Watch clearing prices and shortfall in MW. Power prices are another measure of data-center costs.

2027-04

Nvidia FY2027 Q4 / FY2028 guidance

The company says roughly 70% growth is achievable. Guidance below 50% indicates changing supply constraints or demand expectations.

Dates follow public information; company announcements govern earnings dates. Thresholds are monitoring lines rather than forecasts.
The market's question2025 – 2026 H1After November 2026
GrowthWith growth, losses do not matter.Growth needs a calculable payback period.
Capital expenditureCapex proves demand; more is bullish.Capex creates cash-flow risk; more tightens constraints.
LossesLosses are investment in future share.Losses imply financing dependence: who advances the cash?
ValuationRevenue multiples, vision and TAM.Unit economics, contracted capacity and free cash flow.
FinancingEquity and debt are easy; the IPO window opens.Credit spreads, IPO pricing and supplier financing set the pace.
In one sentence:As capex funding moves from operating cash flow to balance sheets and then valuations, market language shifts from growth to solvency. That is the report's reading of November 5.
AnticipateFour futures

Four futures, defined by conditions.

Two variables shape the future: end-demand growth and capital availability. Two variables, four combinations.

Acceleration needs demand to keep beating expectations and willing capital markets. Digestion needs demand to catch supply as capital becomes selective. A bubble has continuing funding but lagging revenue. An air pocket sees capital withdraw before demand proves itself.

These scenarios can form a path: acceleration → digestion → air pocket → bubble → renewed acceleration. The order depends on policy and credit cycles rather than the technical roadmap.

Subjective probabilities:Digestion 40%, acceleration 25%, air pocket 20%, bubble 15%. These prioritize monitoring rather than predict outcomes. Two signals turning together trigger a scenario change.
Capital availability: tight ↓ / loose ↑
End-demand growth: low ← / high →

Digestion · Demand catches supply

What must be true

Capex growth falls from +80% to +20–30%; HBM contract prices flatten; cloud free cash flow improves; applications move from pilots to production.

Early signals

Cloud providers stop raising 2027 guidance; HBM prices are flat sequentially; token usage keeps doubling; enterprises shift from buying models to buying results.

Winners and losers

Winners: applications, cash-generative cloud providers and undervalued second-tier memory makers. Losers: highly leveraged neoclouds and financing-dependent model companies.

Another stackAI in China

China's AI bottleneck is advanced manufacturing.

The US lacks power and permits; China lacks advanced logic and HBM. Different constraints create different opportunities.

Huawei's H1 2026 revenue was RMB 467.82 billion (+9.6%), R&D represented 25.9% (about RMB 121.4 billion), and attributable net profit was RMB 23.43 billion (down over 30%). H1 shipments of AI chips in China reached 572,000 units, a share of 28%, second only to Nvidia's 34% (IDC)28. The 1,024-card Ascend Atlas 900 supernode costs hundreds of millions of yuan and is commercially available, with DeepSeek, ByteDance and Meituan buying. Huawei's computing R&D spending exceeded wireless R&D for the first time.

DeepSeek ordered 160,000 Ascend 950DT chips from Huawei for inference; training still mainly uses Nvidia29. Kimi K3, a 2.8-trillion-parameter open model, approaches the closed frontier. CAICT estimates China's daily token consumption in June 2026 at about 500 trillion, nearly four times December 2025. China's normalized average token price in 2026 was RMB 13.46 per million, up about 98% from RMB 6.8 in H2 2025, versus roughly RMB 117.8 abroad27。

Power is not an immediate hard limit: Chinese data centers used 196 TWh in 2025, with 300–700 TWh projected for 2030, representing <2%of national electricity use. State Grid's 2026 investment is about RMB 82.5 billion; the east-to-west computing initiative continues30. The constraints are advanced logic, HBM, high-end equipment and ecosystem migration.

A counterintuitive signal:Global token prices are falling while China's are rising. Demand is growing faster than supply, and market structures differ; the price war has not pushed prices to marginal cost.
Normalized average token priceRMB / million tokens
China 2025 H2
6.80
China 2026
13.46
Outside China 2026
117.8
China's 2026 average is about 98% above H2 2025, yet roughly 11% of the overseas average.
The normalized average adjusts different models, contexts and cache-hit rates to one scope. It indicates direction rather than a company's quoted price.
28%Huawei's share of China's AI chip market; Nvidia: 34%
500TDaily token consumption in China, June 2026; 4× in four months
160,000 chipsDeepSeek's Huawei Ascend 950DT order for inference
<2%Data centers' share of China's electricity use; power is not the immediate bottleneck
DimensionUS stackChina stack
ConstraintsPower, grid access, permits and local politics.Advanced logic, HBM, high-end equipment and ecosystem migration.
CapitalOperating cash flow, debt and public markets; capital availability is tightening.Industrial policy, state capital and some public markets; more patient capital with lower return requirements.
ModelsClosed frontier models and enterprise APIs; token-based pricing.Open models and vertical use cases; more outcome- and project-based pricing.
ComputeNvidia + TSMC + HBM; packaging and power constrain supply.Huawei Ascend and domestic substitution; process nodes and HBM constrain supply, with inference first.
PricesDeflation: roughly 50× cheaper annually at fixed capability.Inflation: average token prices up 98% in 2026; still about 11% of overseas prices.
Overseas marketsGlobal markets and allied networks; affected by export controls.Global South and Belt and Road; price and sovereign AI provide entry points.
Judgment:China follows a different growth curve through cheaper inference, more open models and denser grids. Advanced manufacturing sets the ceiling; application demand supports the floor.
EnergyStrength: robust grids, low electricity prices and data centers using under 2% of national power.
Domestic computeBottlenecks: advanced processes, HBM and high-end equipment; inference first, training catching up.
Open modelsStrength: DeepSeek, Qwen and GLM; availability across all 22 MaaS platforms.
Vertical applicationsOpportunity: manufacturing, government, healthcare and education; charge for results.
Overseas marketsOpportunity: Global South and sovereign AI; openness and pricing win market access.
LeverageWhere to stand

Three positions, three ways to act.

An observation checklist, rather than investment advice. Each position has one recommendation, one trade-off and one low-cost first step.

Capital allocators

Remove capex growth from the investment thesis.

Watch return on capital and contracted capacity. Capex growth tells a story; free cash flow and payback tell the accounts. Memory's 80% gross margins are cyclical rents.

First step:Build a six-line monitor: capex/EBITDA, FCF, HBM contract prices, supplier-financing share, credit spreads and PJM prices. Update weekly; record rather than forecast.

Trade-off: you may miss a melt-up, but avoid being passive at the turn.

When to change:Turn defensive when capex guidance cuts, falling HBM prices and widening credit spreads occur together.

Product builders

Go to applications. Sell outcomes.

Applications receive $77.3 billion versus $1.5 trillion for infrastructure; coding and chatbots take 76%. The next opportunity lies in verifiable results in healthcare, legal work and financial back offices.

First step:Choose a workflow with measurable results and charge for outcomes. Start with 10 real customers rather than 100 demos. Treat the model as a supplier.

Trade-off: lower margins than infrastructure; distribution and trust are scarcer than model capability.

When to change:When model companies bundle your features, deepen proprietary data, compliance and workflow integration; avoid a price contest with the model layer.

People building AI in China

Advanced manufacturing is the constraint.

Open models, domestic inference, vertical applications and the Global South create opportunities. Rising token prices signal demand above supply, but per-card performance, power efficiency and ecosystem migration remain hard constraints.

First step:Run your own workload on a domestic inference stack. Measure throughput, latency and unit cost. Begin with inference and fine-tuning.

Trade-off: export controls, per-card performance and power consumption, and software migration can reduce the short-term experience.

When to change:Reassess owned training clusters when HBM/advanced-packaging availability improves or open models reach the frontier.

Scope:These observations are not investment advice or company recommendations. Scenarios are conditional; conclusions change with conditions. Take away what to monitor.
01Explore memory on its own: where is HBM's 2027 turning point?
02Make a China–US comparison of the power constraint.
03If 2027 capex falls 20%, which layers suffer first?
04Open a system for me: where should I focus within the AI industry?
SystemWhat this run leaves behind

What this run leaves behind.

This page is the deliverable; the card alongside it is the system. The page stays in the conversation. The system can follow you if you give it a home.

This card is the starting point for the next run. It records how the industry works, where the tightest constraint is and what to watch next. Say “Continue” and I start here.

The system uses an entry, two understanding records, a practice record and a watch list. These are plain-text files you can read, edit and delete.

Not saved yet.On first use, choose a location. No system files have been written yet. The choice is at the bottom; reply 1 or 2.
Essence in one sentence

Intelligence at the surface, profits in the bedrock. The AI industry sells scarcity.

Eight layers

Applications → models → cloud/capital → chips → memory → packaging → power → the loop.

Four toll points

Memory, packaging, power and capital: all resist rapid replication.

Three loops

Jevons (efficiency → usage), the memory cycle (shortage → expansion → downturn), and financing (capex → valuation → capex).

One moving bottleneck

CoWoS → HBM → power → demand. Watch the next link.

Six signals

Anthropic pricing, Nvidia earnings, cloud capex guidance, HBM contract prices, PJM prices and local referendums.

Data as of

2026-11-16; market data as of the 2026-11-13 close.

This run's route

Open → X-ray → Flow → Loops → Bottleneck → Leverage → Make.

The system has not been saved.

First choose a location. No system files have been written yet; this page is the current deliverable and system summary.

① Follow me:~/.openxsys/, across projects and AI tools. Recommended.
② This folder only:AI-Industry-Research/.openxsys/, for this project.

Reply 1 or 2; next turn I write SYSTEM.md, understanding and practice records, and a watch list. The system stores no passwords, keys or sensitive personal information. You can inspect, edit and delete every file.

ShareTake it with you

One card, three drafts.

Publish as-is or rewrite in your own words. A small “OpenXSys · 2026-11-16” credit is included; remove it if you prefer.

EvidenceSources & definitions

Sources and definitions.

Key figures link back to their sources. Inferences, forecasts and media accounts are labeled separately. Private-company data follows public disclosures, with unaudited items identified.

Four limitations 1. OpenAI is private; financial figures come from media disclosures and leaked documents and are unaudited. Anthropic figures follow its S-1, whose forward-looking statements are conditional. 2. Quarterly memory profits are at a cyclical peak; 80% gross margins will not last forever. Valuing cyclical stocks on peak profits is a common error. 3. 2027 capex, HBM pricing and power loads are forecasts with wide sell-side disagreement. Read their stated conditions. 4. Market caps, share prices and P/E ratios use the 2026-11-13 close. Later changes are excluded. All judgments are conditional and do not constitute investment advice.
  1. 1
    Nvidia,“NVIDIA Announces Financial Results for Second Quarter Fiscal 2027”,2026-08-26。LinkOfficial earnings
  2. 2
    Nvidia,“Revenue by Market Qtrly Trend Q127”,2026。LinkOfficial data
  3. 3
    Tencent News, “Nvidia Q3 earnings date and CEO change,” 2026-11-16.LinkMedia
  4. 4
    AInvest,“The Crises That Price the Nvidia Selloff”,2026-11。LinkMedia analysis
  5. 5
    Anthropic, “Prospectus / S-1,” 2026-10-26, with subsequent amendments.LinkProspectus
  6. 6
    Invezz,“Anthropic IPO terms”,2026-11。LinkMedia
  7. 7
    Nasdaq,“Nvidia and Alphabet buy $10B each of Anthropic stock”,2026-11。LinkMedia
  8. 8
    Safeguard,“OpenAI revenue, losses and cash burn”,2026。LinkMedia compilation
  9. 9
    ValueAdd VC,“OpenAI revenue 2026”,2026。LinkMedia
  10. 10
    Gartner,“Worldwide AI spending 2026”,2026-09-16。LinkResearch forecast
  11. 11
    Menlo Ventures,“2026: The State of Enterprise AI”,2026。LinkResearch
  12. 12
    Company earnings compilation,Microsoft / Alphabet / Amazon / Meta 2026 Q3,2026-10。Official earnings
  13. 13
    TradingKey,“US stock selloff / AI bubble / S&P 500 2027”,2026-11。LinkMedia analysis
  14. 14
    Alphabet, June 2026 equity financing / debt, 2026-06.LinkCompany announcement
  15. 15
    TSMC,“3Q 2026 Earnings Release”,2026-10。LinkOfficial earnings
  16. 16
    TSMC,“3Q26 Transcript / CoWoS reticle”,2026-10。LinkOfficial earnings call
  17. 17
    SK hynix, Q3 2026 earnings, 2026-10.LinkOfficial earnings
  18. 18
    Samsung Electronics, Q3 2026 earnings, 2026-10.LinkOfficial earnings
  19. 19
    Micron, FY2026 Q4 earnings, 2026-09.LinkOfficial earnings
  20. 20
    Counterpoint / TrendForce / BofA, HBM share and 2027 forecast, 2026.LinkResearch forecast
  21. 21
    IDC, AI data-center memory use and phone/PC output cuts, 2026.LinkResearch forecast
  22. 22
    IEA,“Energy and AI”,2026。LinkOfficial research
  23. 23
    LBNL / Goldman Sachs, grid-connection queues and US data-center load, 2026.LinkResearch
  24. 24
    PJM, capacity-auction results, 2026.LinkOfficial
  25. 25
    GE Vernova / Wood Mackenzie, gas-turbine and transformer lead times, 2026.LinkCompany / research
  26. 26
    36Kr, “US market sell-off on November 5,” 2026-11-06.LinkMedia
  27. 27
    CAICT, token data, 2026-10.LinkOfficial research
  28. 28
    Huawei H1 2026 report; IDC, China's AI chip share, 2026.LinkOfficial / research
  29. 29
    Caixin, “DeepSeek's 160,000 Ascend 950DT chips,” 2026-10-31.Media
  30. 30
    China Energy News, Chinese data-center electricity use, 2026-11-02.Media
  31. 31
    Wind / Motley Fool, market capitalization and prices, 2026-11-13.Market data
  32. 32
    McKinsey, “95% of pilots lack direct financial returns,” 2025.LinkResearch
  33. 33
    Jevons,“The Coal Question”,1865。LinkPrimary source
  34. 34
    Anthropic S-1 / Sina Finance, Q3 financial breakdown, 2026-10-27.LinkOfficial / media
  35. 35
    PitchBook,“Q2 2026: An Eye on OpenAI”,2026。LinkResearch
  36. 36
    Economic Daily, “The AI bubble and 92% of GDP,” 2025-11-19.LinkMedia