AI Investment & Strategic Diligence · Investment Committee Memo · April 2026
Why the GPU refresh cycle, customer concentration across NVIDIA / AMD / OpenAI, and a flattering useful-life policy put Oracle's fair value at $75 against $164.52 today.
Bottom Line
Methodology & Sources
The findings below are not management commentary repackaged. The valuation is built from public-disclosure sources read across four companies and structured so every assumption can be re-checked against the document it came from. Four methodology pillars hold the case together.
Oracle's most recent 10-K, three sequential 10-Q filings, and the September 2025 8-K. NVIDIA FY26 Q2 10-Q. AMD FY25 10-K and the October 2025 8-K. Microsoft FY23, Alphabet 2023, Meta 2024, and Oracle FY24 10-Ks for cross-company useful-life policy. No private data, no proprietary surveys.
The valuation is bracketed by two fully specified scenarios. The conventional view applies consensus assumptions (90% revenue realization, 30% margins, capex fading to legacy baseline). The new normal view applies the four evidence pillars (80% realization, 27% margins, 15% steady-state capex, 10% WACC). Continuing value is where the scenarios diverge.
Hyperscaler peer accounting (Microsoft, Google, Meta, Oracle) read together, calibrated against the empirical 3-to-4-year GPU generation cycle (Hopper to Blackwell to Rubin). The audit surfaces an industry-wide pattern that single-company filings normalize away.
Oracle, NVIDIA, AMD, and OpenAI mapped as a single AI infrastructure system. Customer concentration, accounts receivable concentration, future purchase commitments, and warrant economics tracked across all four names so single-name fundamentals cannot hide the correlated counterparty risk.
The Situation
Oracle has become the most aggressive narrative beneficiary of the AI infrastructure build cycle. Between February and August 2025, Oracle's Remaining Performance Obligation (RPO) grew 250% to $455.3B, anchored by a small number of named hyperscaler and AI-lab contracts. The headline reads as a generational demand event for OCI. Management's preview implies a $32B FY27 OCI revenue print and a multi-year capex commitment to support it.
The investor question is whether that backlog is the leading edge of durable, diversified cloud demand, or whether it is concentrated, longer-dated, and structurally entangled with the same handful of capital-constrained AI labs that drive revenue at NVIDIA and AMD as well. The answer matters because the AI capital cycle requires Oracle to depreciate hardware that empirically refreshes every 3 to 4 years against a GAAP useful life of 6 years. The accounting and the economics are pulling in opposite directions.
This memo treats Oracle not as a company in isolation, but as a node in a four-name AI infrastructure system (Oracle, NVIDIA, AMD, OpenAI) where customer concentration and counterparty overlap create correlated risk that does not show up in single-name fundamentals. The thesis is built on four independent pieces of public-disclosure evidence, calibrated against a cross-company useful-life audit, and translated into a two-scenario DCF where continuing value is the entire valuation question.
Key Findings
Each pillar is independently sourced from public disclosures. The thesis does not require any single pillar to be correct in isolation; it requires the cross-company picture to be coherent, which it is.
Between February and August 2025, the longer-dated bucket (months 37+) absorbed nearly all of the increase. Implied next-12-month revenue rose only $5B (from $40.3B to $45.5B) even as the backlog expanded by $325B. The headline backlog flatters near-term cash optics, and the longer-dated portion is exactly the cash flow that drives terminal value: it is the most important number in the case and the most uncertain.
In FY2026 Q2, two direct customers were 39% of NVIDIA revenue (versus 25% prior year). Three direct customers held 56% of accounts receivable. NVIDIA carries $46B of total future purchase commitments and took a $4.5B excess-inventory charge on H20 products after the April 2025 China export controls. A small number of hyperscale buyers fund the entire build-out, and supplier-side write-down risk has been demonstrated in a single fiscal quarter.
AMD issued OpenAI a warrant for 160 million shares at a $0.01 strike, roughly 9.8% of AMD's outstanding share count. At AMD's January 30, 2026 close of $236.73, the underlying market value is approximately $37.9B against $1.6 million of cash proceeds. Anchor AI demand is being secured only with substantial economic concessions. OpenAI sits across all three names as customer (Oracle), indirect customer (NVIDIA), and warrant counterparty (AMD); concentration on a single capital-constrained name flows through the entire system.
Microsoft extended server useful life from 4 to 6 years in FY23 (~$3.7B/year margin benefit). Google extended to 6 years in 2023 (~$3.9B/year). Meta extended to 5.5 years in 2024. Oracle extended from 5 to 6 years in FY24. The empirical GPU cycle is 3 to 4 years (Hopper to Blackwell to Rubin in roughly 36 months). The accounting direction is economically inconsistent: GAAP lives lengthened on the books precisely as economic obsolescence accelerated on the GPU subset, so reported margins overstate economic profitability and reported D&A understates the perpetual reinvestment requirement.
Valuation
The conventional view applies consensus assumptions: 90% of management's $190B FY30 revenue trajectory, 30% operating margins, capex fading from 37% of revenue to 6% (Oracle's pre-AI cloud baseline), 8.5% WACC, 2.5% terminal growth. The new normal view (preferred) applies the four evidence pillars and produces a fundamentally different picture.
| Driver | Conventional View | New Normal (Preferred) |
|---|---|---|
| FY30 revenue ($B) | 171 | 148 |
| FY30 EBIT margin | 30% | 27% |
| FY30 capex / revenue (steady state) | 6% | 15% |
| WACC | 8.5% | 10.0% |
| Terminal growth (g) | 2.5% | 2.0% |
| Enterprise value ($B) | 640 | 298 |
| Implied value / share | $192 | $75 |
| vs current $164.52 | +17% | (54%) |
Risk & Catalyst Map
Four catalysts that would compress the gap between current price and our preferred fair value, ranked by conviction and time horizon.
Management's OCI revenue preview implies $32B in FY27. Realization meaningfully below this level validates the demand-quality discount and accelerates price discovery. This is the highest-conviction catalyst in the memo and the one most likely to reset the multiple.
12 to 18 month horizonIf even 10% of long-dated RPO is renegotiated or scope-reduced as customer financing and AI training economics evolve, FY2028 to FY2030 revenue falls meaningfully below either scenario, with the largest impact in the years where DCF value is most concentrated.
24 to 36 month horizonOpenAI drives revenue at Oracle, NVIDIA, and AMD simultaneously. Any delay or scope-down to OpenAI's spending creates correlated revenue impact across the entire AI infrastructure stack and is not diversifiable within the four-name complex.
Event-drivenA hyperscaler or Oracle reversal of useful-life extensions, or audit-driven impairment of GPU asset balances, forces a step-down in reported margins and validates the economic-depreciation thesis directly. Probability rises with each new GPU generation that empirically obsoletes the preceding one.
12 to 36 month horizonDisconfirming Evidence
The thesis is bracketed honestly. We would revisit the Underweight call if any of the following four conditions were met. Each is a specific, observable disclosure event, not a qualitative shift in narrative.
Recommendation
At $164.52 the stock trades at roughly 2.2x our preferred fair value. The risk is asymmetric: upside requires consensus assumptions to hold (full management revenue trajectory, capex fading to legacy baseline, useful-life extensions justified by genuine efficiency gains), while downside requires only the empirical GPU refresh cycle to behave as it has for three generations.
Implementation by Mandate
Position sizing should reflect that the central catalyst (FY27 OCI conversion) is 12 to 18 months out. Avoid building exposure on intra-quarter strength.
Long $130 puts, short $85 puts, 18 to 24 month expiry. Captures most of the move with defined risk. We do not recommend an outright equity short given squeeze risk in AI mega-caps and multi-quarter catalyst horizons.
Short ORCL, long SAP or Workday. Isolates AI-quality-of-revenue from broad software beta. Best fit for accounts that can carry single-name short exposure and want the thesis stripped of sector beta.
Uncertainty. The thesis depends on the empirical GPU obsolescence cycle continuing to outpace GAAP useful lives, and on the industry eventually reflecting that in either accounting changes or sustained capex elevation. The central catalyst (FY27 OCI conversion) is 12 to 18 months out; position sizing and derivative expiry should accept that timing risk.
What FutureInSites Brings
Investment-grade analysis of AI businesses and AI's impact on incumbent economics. The Oracle case is the proof point for a complementary practice area that sits alongside operator-side AI strategy work: rigorous diligence on AI investment theses, with auditable assumptions and explicit downside bracketing.
GPU economic obsolescence vs. GAAP useful lives. Capex / D&A trough modeling. Refresh-cycle reinvestment requirements. The cost-side analysis that conventional cloud DCFs miss when they fade capex to legacy baselines.
Cross-company concentration analysis when revenue, accounts receivable, and warrant economics overlap on a small number of capital-constrained names. The diligence move that single-name fundamentals don't surface.
Where the valuation question is concentrated entirely in continuing-value assumptions, the case is built around making those assumptions explicit. Conventional view, new normal view, and the precise driver list that reconciles the two.
Useful-life policy changes, segment margin disclosures, and revenue recognition timing read across hyperscaler peers to surface industry-level inconsistencies that single-company filings normalize away.
Conviction translated into specific, sized positions. Put spread expiry windows aligned to catalyst horizons. Pair-trade construction that isolates the thesis from broad sector beta.
Every forecast assumption ties to a specific public-disclosure exhibit: a 10-K, 10-Q, 8-K, or earnings release line item. Two-scenario DCF, cross-company useful-life audit, and continuing-value bracketing are the work product, not a slide-deck summary.