Openai Value Chain Sector IC Memo

Executive Summary & Action Plan

Verdict

NEUTRAL / STAGED — OpenAI's funded, multi-cloud and multi-silicon buildout is now producing visible supplier revenue, including live Cerebras inference, but a 15.9% four-week basket rally and only two positive house fair-value gaps leave little room for execution misses.

Loading chart

Abstract

OpenAI has turned compute diversification into a real value chain rather than a procurement slogan. Its official stack now names Microsoft, Oracle, AWS, CoreWeave and Google Cloud; NVIDIA, AMD, Trainium, Cerebras and an OpenAI-designed Broadcom chip; and Oracle, SB Energy and SoftBank data centers. Cerebras is the newest operational proof: GPT-5.6 Sol Ultrafast runs on its systems at up to 750 output tokens per second, while Q2 core cloud revenue nearly quadrupled. Demand evidence is broad—Azure grew 43%, AWS 37%, Google Cloud 82%, CoreWeave 113%, AMD Data Center 107%, and Oracle IaaS 93% in their latest reported periods. Value capture is less certain. Cash conversion is negative at Amazon, Alphabet, Oracle and CoreWeave, contract quality ranges from binding orders to term sheets and letters of intent, and the nine-name basket already gained 15.9% from July 16 through August 13. We therefore keep neutral thematic weight, stage additions around evidence gates, and cap Cerebras as optionality rather than core exposure.

Forecast Path

The scoreable proxy is an equal-dollar, no-rebalance basket of MSFT, AMZN, ORCL, GOOG, CRWV, NVDA, AMD, AVGO and CBRS, normalized to 100 at the August 13 close. It is a measurement instrument, not the recommended allocation. The expected path is nearly flat at 1M and reaches 108.94 at 12M; a short post-CBRS-IPO history makes the interval a dispersion estimate rather than a directional model.

Loading chart

Action Plan

Keep the OpenAI theme near benchmark weight; do not chase the equal-weight proxy after its 12.3 percentage-point four-week lead over SPY.
Fund 50% of the intended thematic sleeve now. Add 25% after NVIDIA on August 26 and Broadcom on September 2 only if both preserve AI demand and revision breadth remains at least five of nine Up.
Add the final 25% only after two of three operating gates clear: Cerebras' first 250MW remains on schedule, AMD's initial Helios deployment is live, and Broadcom/OpenAI Jalapeño enters commercial deployment.
In a model thematic sleeve, place 80% in higher-quality platform and incumbent exposure, 15% in ramp-sensitive suppliers, and no more than 5% in CRWV plus CBRS combined. This is separate from the equal-weight scoring proxy.
Hedge or cut the sleeve by half if the basket trails SPY by 12 percentage points over three months while forward-EPS revision breadth falls below four of nine.
Use house fair values below as decision anchors. Street targets are shown only to expose the expectations gap and receive zero weight in the house marks.
Expected 12M basket return: +8.9% | Expected excess return: about +0.9% versus an explicit +8% SPY planning assumption | P(thesis wrong): 43% | Confidence: Medium-Low

State of the Sector

OpenAI closed $122 billion of committed capital at an $852 billion post-money valuation on March 31, 2026. It reported $2 billion of monthly revenue, more than 900 million weekly ChatGPT users and API traffic above 15 billion tokens per minute. Those metrics create funded demand, but committed capital is not the same as cash already received and the infrastructure obligations are not publicly reconciled against liquidity.

An adversarial check is essential: funding and supplier demand are not independent observations. Amazon and NVIDIA invest in OpenAI while selling it compute, Microsoft owns an economic stake and receives commercial payments, and Oracle finances capacity against large contracts. Some capital can therefore circulate back into the same suppliers and inflate funding, RPO and revenue headlines without independently proving end-customer unit economics. Paid adoption and supplier cash conversion are the necessary cross-checks.

The architecture has shifted from a Microsoft/NVIDIA-centered stack to workload routing across clouds and silicon. Microsoft remains primary and products launch first on Azure, but the April amendment allows OpenAI products on other clouds and makes Microsoft's model/product IP license non-exclusive through 2032. NVIDIA remains the training foundation and majority inference platform; Cerebras and custom ASICs target workloads where latency or cost per token matters more than maximum flexibility.

The prior six-name report omitted two direct public cloud nodes and Cerebras. This report corrects the universe by including GOOG, CRWV and CBRS. SoftBank is public but remains outside the U.S.-listed scoring basket because its OpenAI exposure, currency and conglomerate structure are not comparable with the nine operating securities.

Forecast Drivers & Cycle Map

Loading diagram

Layer	Direct public exposure	Secondary / private nodes	Economic transmission	Evidence caveat
Capital and rights	MSFT, AMZN, NVDA, ORCL	SoftBank, MGX and institutions	ownership marks, revenue share, funded procurement	current OpenAI cap table is undisclosed after the March round
Cloud and data centers	MSFT, ORCL, AMZN, CRWV, GOOG	SB Energy, Vantage, Related Digital	cloud consumption, lease and capacity revenue	site energization and contract margins are not fully disclosed
Merchant compute	NVDA, AMD	CPU and rack partners	accelerators, networking and software	NVIDIA 10GW is an LOI; only AMD's first 1GW is explicitly binding
Custom / specialized inference	AVGO, CBRS, AMZN Trainium	Celestica and integration partners	lower latency or cost per token	AVGO's 10GW is a term sheet; Cerebras adds warrant and utilization risk
Foundry, memory and packaging	—	TSM, Samsung, SK hynix, MU, ASE	wafers, HBM/DRAM and packaging bottlenecks	mostly indirect; do not call these disclosed OpenAI contracts
Power and cooling	ORCL-linked projects	SBE, WEC, DTE, SO, cooling vendors	energized MW and site readiness	long-dated power is not near-term supplier revenue
Distribution	MSFT, AMZN	OpenAI partners and integrators	API, Frontier, Bedrock and enterprise adoption	distribution can shift value away from the infrastructure owner

Do not sum the headline gigawatts or dollar commitments across this map. A single OpenAI workload can appear simultaneously in a chip framework, a cloud contract and a data-center project; binding orders, options, term sheets and LOIs are not equivalent backlog.

Selection & Dispersion

The table answers four separate questions: completed August 13 price, house intrinsic fair value, the house 12-month convergence target, and the external Street target. House Base FV and the 12M decision target are the same number here; the Street column is a reference-only expectations check. No report-level OpenAI fair value is created.

Loading chart

Company	Direct lane	Close	Bear / Base FV / Bull	House upside	Street mean target	Street upside	Decision view
Microsoft (MSFT)	ownership, Azure, distribution, IP	$496.88	$410 / $525 / $675	+5.7%	$567.20	+14.2%	best risk-adjusted selection; positive FCF and broad backlog
Broadcom (AVGO)	Jalapeño, Ethernet, custom racks	$417.82	$280 / $425 / $620	+1.7%	$527.88	+26.3%	near fair value; September print must validate the ramp
CoreWeave (CRWV)	direct GPU cloud capacity	$106.29	$45 / $105 / $190	-1.2%	$143.26	+34.8%	near Base, but debt and capex keep confidence Low
NVIDIA (NVDA)	incumbent training and inference	$225.30	$125 / $221 / $320	-1.9%	$302.83	+34.4%	quality is high; August 26 resets the mark
Alphabet (GOOG)	Google Cloud capacity	$343.94	$225 / $325 / $450	-5.5%	$421.79	+22.6%	modestly above value; balance sheet offsets capex risk
AMD (AMD)	6GW framework, first 1GW binding	$483.01	$250 / $430 / $700	-11.0%	$613.33	+27.0%	Helios execution can expand Bull weight, not Base today
Amazon (AMZN)	equity, AWS, 2GW Trainium, Bedrock	$265.13	$130 / $196 / $240	-26.1%	$326.82	+23.3%	strong AWS demand; $220B capex delays cash convergence
Oracle (ORCL)	Stargate / OCI capacity	$156.22	$69 / $113 / $153	-27.7%	$247.17	+58.2%	RPO is exceptional; leverage, dilution and negative FCF dominate
Cerebras (CBRS)	750MW low-latency inference	$231.01	$60 / $95 / $135	-58.9%	$290.73	+25.9%	live proof improved; valuation still assumes near-perfect conversion

House upside is positive for two of nine and has a -5.5% median, while the Street median is +26.3%. The disagreement is not about whether OpenAI demand exists; it is about capital intensity, dilution, normalized margins and how much contract value becomes shareholder cash. Street targets receive zero weight in the house forecast.

Constituent Valuation Audit

Ticker	Three-method control	Base FV	Freshness / principal guard
MSFT	FCFF DCF / FY27 EV-EBITDA / segment SOTP	$525	rebuilt after FY26 Q4; FY27 capex and lower FCF retained
AMZN	FCFF DCF / FY27 EV-EBITDA / normalized EV-FCF	$196	Q2 price/evidence refreshed; $220B 2026 capex retained
ORCL	FCFF DCF / FY27 EV-EBITDA / normalized EV-FCF	$113	carried from July 12; no newer earnings; net debt and FY27–28 burn retained
GOOG	FCFF DCF / FY27 EV-EBITDA / Services-Cloud SOTP	$325	rebuilt after Q2; investment gains excluded and capex raised
CRWV	FCFF DCF / FY27 EV-Revenue / FY27 EV-EBITDA	$105	rebuilt after Q2; roughly $35B debt and negative FCF retained
NVDA	normalized FCFF DCF / FY27 EV-EBITDA / normalized EV-FCF	$221	control held into August 26; equity gains excluded
AMD	FCFF DCF / FY27 EV-EBITDA / FY27 EV-FCF	$430	rebuilt after Q2; Helios revenue and margin ramp probability-weighted
AVGO	FCFF DCF / FY27 EV-EBITDA / normalized EV-FCF	$425	rebuilt on current evidence; pending September print and debt retained
CBRS	seven-year FCFF DCF / normalized FY28 EV-EBITDA / FY27 EV-Sales	$95	refreshed after Q2; 301M committed-capacity diluted-share case and negative FCF retained

The 4.63% August 13 U.S. 10-year Treasury yield anchors discount rates. The CBRS Base rises from the prior $86 to $95 because Q2 revenue, margin, cash and production evidence beat the earlier control case; it does not receive full credit for $25.4 billion of RPO because pass-through economics, warrants, utilization and deployment capital remain material.

What to Watch Now

Company / node	Latest evidence	Next test	Bull threshold	Bear threshold
NVDA	Q1 revenue $81.6B, +85%; Data Center +92%	Aug 26 confirmed Q2	revenue at/above $91B and next guide rises	miss low end, GM below 74%, or guide falls
AVGO	Q2 revenue +48%; AI semiconductor +143%	Sep 2 confirmed Q3	AI revenue at least $16B and FY27 framework intact	AI revenue below $15B or annual framework cut
ORCL	Q4 IaaS +93%; RPO $638B; FY26 FCF -$23.7B	mid-September Q1, date unconfirmed	cloud +58–64% and funding plan intact	cloud below 55% or funding/dilution worsens
CBRS	Q2 core revenue $209.9M; core GM 40.6%	Q3 and 250MW gate	Q3 core revenue at least $216M, GM at least 40%, on-time 250MW	revenue below $210M, GM below 37%, or capacity delay

What to Watch Next (Technology/Structure)

The architecture is becoming workload-specific. NVIDIA should retain training and flexible inference; AMD is the merchant alternative; Trainium and Jalapeño attack cost; Cerebras attacks latency. GPT-5.6 Sol at up to 750 output tokens per second is evidence that Cerebras can serve a production-adjacent OpenAI workload, not proof that wafer-scale systems replace the GPU fleet. The decisive metric is tokens per energized megawatt at an acceptable gross margin.

OpenAI's MRC networking work with AMD, Broadcom, Intel, Microsoft and NVIDIA lowers dependence on any one fabric. This can enlarge the total compute envelope while redistributing rent toward custom silicon, Ethernet, foundry, assembly and power. Celestica is disclosed as Jalapeño's board/rack/system integrator; TSMC, HBM suppliers and utilities remain second-order exposures unless a direct OpenAI contract is disclosed.

Allocation & Selection Notes

MSFT is the best risk-adjusted constituent selection: its +5.7% house gap is modest, but it combines OpenAI economics, Azure distribution, a completed quarter, positive FCF and evidence that backlog growth extends beyond frontier labs. AVGO ranks second but awaits the September print. NVDA and GOOG are high-quality holds near fair value. AMD is ramp optionality. AMZN and ORCL need cash conversion. CRWV and CBRS should be treated as small satellites because leverage or dilution makes their equity payoff much more convex than their revenue growth.

A research-sleeve example is 25% MSFT, 20% GOOG, 20% NVDA, 15% AVGO, 7% AMD, 5% AMZN, 3% ORCL, 3% CRWV and 2% CBRS. This is a risk-budget illustration, not the scoreable equal-weight target and not personalized advice.

Cerebras Deep Dive

Cerebras is now a public operating company under Nasdaq ticker CBRS. The OpenAI agreement covers 750MW in tranches through 2028, with the first 250MW planned by year-end 2026; an additional 1.25GW is optional rather than committed. GPT-5.6 Sol Ultrafast is the first clear product-level proof after the prior report, delivering up to 750 output tokens per second in limited preview.

Q2 core revenue was $209.9 million, up 103%; core cloud revenue was $127.7 million, up 287%; and core gross margin was 40.6%. GAAP revenue was $180.1 million and GAAP gross margin only 14.2%, illustrating the effect of warrant and contract accounting. OpenAI generated $56.8 million, about 31.5% of GAAP revenue. RPO reached $25.4 billion, but Q3 core gross margin guidance of 38–40% and six-month capex of $548.9 million show that capacity arrives before mature cash economics. The company ended Q2 with $8.6 billion of cash, short-term investments and restricted cash after the IPO.

The refreshed $95 Base FV uses a committed-capacity diluted-share case and three methods; the $290.73 Street mean implies a very different conversion and scarcity-multiple path. CBRS fell 11.9% on August 13 after the result. That is described as post-results repricing, not attributed to one unverified cause. At $231.01, new money should wait for capacity, margin and dilution proof rather than equate faster inference with attractive equity value.

Catalysts, Leading Indicators & Risks

Window	Observable	Bull confirmation	Bear confirmation / risk
Aug 26	NVIDIA FY27 Q2	guide rises with at least 74.5% GM	low-end miss or declining guide
Sep 2	Broadcom FY26 Q3	$16B+ AI semiconductor revenue and intact FY27 view	ramp miss or customer-framework cut
Mid-Sep	Oracle FY27 Q1	cloud converts near 58–64% with stable financing	RPO conversion miss plus higher funding need
Q3 reporting cycle	nine-name EPS revision breadth	at least 6 Up and no more than 1 Down	3 or more Down
By Dec 31	CBRS / AMD / AVGO deployment bundle	at least two of 250MW, Helios and Jalapeño are live	two programs slip by more than one quarter
Quarterly	cloud growth versus cash conversion	Azure/AWS/Google/CRWV growth stays broad and FCF improves	growth decelerates while capex and financing remain elevated
Monthly	basket relative return	operating upgrades accompany outperformance	12pp relative drawdown plus weak revisions

Principal risks are OpenAI funding and customer concentration, duplicated headline commitments, cloud-provider substitution, model-efficiency shocks, power and site delays, export controls, custom-silicon share loss, debt-funded capex, warrant/SBC dilution and an expectations unwind after the recent rally.

Falsification & Review Rules

Breadth falsifier: if fewer than four of nine constituents have Up revisions and the basket trails SPY by 12 percentage points over a rolling three months, cut the thematic sleeve by half and move the forecast to Bear.
Cerebras falsifier: if Q3 core revenue is below $210 million, core GM below 37%, or the first 250MW is not available by December 31, remove CBRS from the investable sleeve and rebuild its fair value.
Deployment falsifier: if two of AMD's initial Helios deployment, Broadcom's Jalapeño start and NVIDIA's first Vera Rubin deployment slip by more than one quarter, remove the diversified-silicon premium.
Cash-conversion falsifier: if at least three of AMZN, ORCL, GOOG and CRWV show worsening cash conversion for two consecutive reports while cloud growth decelerates, reduce cloud/data-center exposure by one-third.
Review prices and revisions monthly, every issuer result immediately, and the entire thesis no later than 2027-02-14. Price alone can trigger risk reduction, but only operating evidence can restore Bull probability.

References

OpenAI financing and full infrastructure stack · Microsoft partnership amendment · Amazon partnership
OpenAI–Cerebras 750MW partnership · GPT-5.6 Sol Ultrafast on Cerebras · Cerebras Q2 results
Oracle Stargate capacity · CoreWeave expansion with OpenAI
NVIDIA systems partnership · AMD partnership · Broadcom collaboration
OpenAI–Broadcom Jalapeño · OpenAI MRC networking · U.S. AI supply-chain RFP
Microsoft FY26 Q4 · Amazon Q2 2026 · Alphabet Q2 filing
Oracle FY26 Q4 · NVIDIA FY27 Q1 · AMD Q2 2026
Broadcom FY26 Q2 · CoreWeave Q2 filing
Nine-name price packet · [시장 데이터 cross-check](https://시장 데이터.com/screener.ashx?v=152&t=MSFT,AMZN,ORCL,GOOG,CRWV,NVDA,AMD,AVGO,CBRS,SPY)