canonical/modules/M073-ai-accelerators-and-compute-analyst-playbook.md

<!-- Generated loss-aware reference mirror from God_Level_Public_Company_Financial_Analyst_Job_Guide_V6_99_ALL_SUB70_FIXED.docx. Canonical source remains the bundled DOCX. -->

<!-- Module: 073 | Title: AI Accelerators and Compute Analyst Playbook -->

## PART XV - SECTOR PLAYBOOKS | MODULE 073

# AI Accelerators and Compute Analyst Playbook

> Mission. Build a sector-specific research system for AI Accelerators and Compute that converts operating data into financial outcomes, highlights the accounting areas most likely to distort comparability, and selects valuation methods that reflect the sector's economics.

## Economic engine and binding constraints

Trace accelerator units, ASP, HBM content, packaging capacity, rack power, networking content, lead times, customer capex, and utilization. Distinguish supply-constrained shipments from sustainable end demand and test concentration risk.

## Primary KPI stack

| KPI | Construction / analyst control |
| --- | --- |
| accelerator shipments | AI accelerator units or systems shipped and recognized in revenue during the period, separated by product generation and form factor where disclosed. Validation: Tie physical/operating units to company disclosures or source-system data; reconcile beginning/ending populations where applicable and test scope, ownership, and period consistency. |
| ASP | Revenue attributable to the relevant product family divided by units sold/shipped, adjusted for rebates, mix, and channel treatment. Validation: Recalculate price/cost from underlying dollars and physical units; test mix, rebates, FX, timing, and unit-definition effects; reconcile to reported revenue or expense. |
| HBM content | HBM memory capacity or dollar content attached per accelerator/system multiplied by accelerator shipments; distinguish generations and supplier mix. Validation: Recalculate independently from cited source data; verify definition, period, units, scope, signs, and any reconciliation to reported financial or operating totals. |
| rack density | IT/accelerator power load or compute capacity per rack, typically kW or accelerators per rack, using deployed rather than theoretical maximum configuration. Validation: Reperform the count from the defined population, inspect every material exception, and confirm the denominator/universe did not change between periods. |
| power per rack | Average or design electrical demand per deployed rack in kW, including the defined IT load and clearly stating whether cooling/overhead is excluded. Validation: Recalculate independently from cited source data; verify definition, period, units, scope, signs, and any reconciliation to reported financial or operating totals. |
| lead time | Elapsed time from order acceptance to shipment/installation for the relevant product or capacity, reported as median/range and compared with normal cycle levels. Validation: Verify start/end timestamps or periods from source records, use a consistent calendar/business-day convention, and test outliers rather than averaging them away. |
| cloud capex | Capital expenditures by hyperscale/cloud customers attributable to compute, networking, data-center buildings, and power infrastructure; normalize definition across companies. Validation: Tie the dollar measure to filed statements/footnotes; reconcile classification adjustments, one-time items, acquisitions/FX, and period consistency before using it analytically. |
| utilization | Actual productive output or occupied capacity divided by practical available capacity after planned downtime, yield loss, and maintenance constraints. Validation: Recalculate from same-scope numerator and denominator; confirm period, units, cohort/geography, and issuer definition; reconcile material differences to filings or operating data. |



## Sector-specific accounting and comparability traps

- Supply allocation: reconcile issuer treatment with peer treatment and quantify the effect on reported growth, margin, cash flow, capital, or valuation before comparing outputs.

- Customer concentration: reconcile issuer treatment with peer treatment and quantify the effect on reported growth, margin, cash flow, capital, or valuation before comparing outputs.

- Rapid obsolescence: reconcile issuer treatment with peer treatment and quantify the effect on reported growth, margin, cash flow, capital, or valuation before comparing outputs.

- Capitalized software: reconcile issuer treatment with peer treatment and quantify the effect on reported growth, margin, cash flow, capital, or valuation before comparing outputs.

- Warranty and returns: reconcile issuer treatment with peer treatment and quantify the effect on reported growth, margin, cash flow, capital, or valuation before comparing outputs.

## Valuation frameworks

- EV/sales with margin bridge: enterprise value divided by revenue plus an explicit path from gross margin to EBIT/FCF, including SBC, capex, working capital, and terminal margin assumptions.

- EV/EBIT: enterprise value divided by normalized operating profit after depreciation; useful where depreciation is economically meaningful and capital intensity differs.

- DCF: forecast FCFF from operating drivers, discount at a capital-structure-consistent WACC, model terminal growth/ROIC coherently, and bridge enterprise value to common equity.

- Reverse DCF: solve for the growth, margin, reinvestment, ROIC, and duration assumptions required for the current market price, then compare those expectations with evidence.

## Sector diligence questions

- What is the most important leading indicator for AI Accelerators and Compute, and how many months does it lead reported revenue or cash flow?

## Sector stress and falsification

- Stress accelerator shipments and ASP together in the direction most likely to break the equity story; flow the result through working capital, capex, liquidity, financing, dilution, and valuation.

- Explicitly test supply allocation. Determine whether it can make the reported sector comparison look better or worse without equivalent economic change.

## 99-point standalone execution extension

### Model architecture and forecast chain

Model accelerator units, ASP, HBM content, networking, rack density, power per rack, software attach, cloud utilization, and customer concentration. Constrain shipments by packaging, memory, foundry, power, and deployment capacity.

### Leading-indicator dashboard

Track hyperscaler capex, backlog, lead time, CoWoS or advanced-packaging capacity, HBM supply, rack deliveries, power procurement, model efficiency, utilization, and custom silicon adoption.

### Primary-source map

SEC filings; hyperscaler capex disclosures; foundry and advanced-packaging supplier filings; memory-vendor filings; U.S. Commerce/BIS export-control releases; public power/interconnection disclosures.

### Accounting normalization test

Supply allocation, customer prepayments, concentrated demand, rapid product transitions, warranty, capitalized software, and channel inventory can make short-term revenue nonrepeatable.

### Valuation implementation

Reverse DCF should solve for unit growth, ASP/content, mature margin, and duration. Test value under custom silicon, efficiency gains, and lower accelerator intensity per workload.

### Worked numerical mini-case

> Illustrative supply-constrained case.

Demand supports 1.20m accelerators at $28k ASP, but packaging/HBM capacity limits shipments to 0.90m. Revenue is constrained to $25.2bn before networking/software attach. A 15% capacity expansion raises the ceiling to 1.035m, not to unconstrained demand.

Value the bottleneck duration explicitly and stress custom silicon, utilization and compute-efficiency improvements.

### Monitoring and falsification cadence

Thesis breaks include demand concentration reversing, model-efficiency reducing compute intensity faster than workload growth, customer vertical integration, or supply scarcity rents normalizing sooner than expected.

At every quarterly update, rebuild the driver bridge from operating units to revenue, margin, cash flow and valuation; compare leading indicators with the prior forecast; record definition changes; and precommit the threshold that would trigger a thesis reset rather than a cosmetic estimate change.

## Sector exit standard

The AI Accelerators and Compute work is complete only when the analyst can explain the business in its native operating units, reproduce the KPI history, identify the binding growth constraint and marginal price setter, normalize sector-specific accounting, quantify a coherent adverse case, and translate the current market price into the operating expectations that must be met or exceeded.