Long-Term Thesis
Long-Term Thesis — One Variable Decides a Decade
Oracle is two companies sharing a balance sheet, and the 5-to-10-year underwriting question collapses to a single variable. Underneath sits a 48-year-old, wide-moat database-and-support annuity where "substantially all license support customers renew their support contracts with us upon expiration" [2] — a roughly $19.5 billion, near-100%-margin stream inside a 63%-margin cloud-and-license business [3]. On top, management is staking the firm's entire incremental capital — and its valuation — on renting AI compute, a layer Oracle's own 10-K concedes has "low barriers to entry" [5], funded by debt and equity after free cash flow swung to negative $23.7 billion in FY2026 [4].
The bridge between the two is a number with no precedent in enterprise software: total remaining performance obligations (RPO) ended FY2026 at $638 billion, up 363% year-over-year — roughly nine times annual revenue [1]. The entire long-term thesis reduces to one underwriting question: does that backlog convert to collected cash at a return on the ~$70–90 billion-a-year capex that clears Oracle's cost of capital — before the leverage, the counterparty concentration, or the margin math breaks? Everything else is detail. This page frames what must be true for the answer to be yes, and the multi-year evidence that would prove the thesis is working or breaking.
The four dials
Thesis Strength
Durability (core vs new bet)
Reinvestment Runway
Evidence Confidence
Source: analyst assessment synthesizing the multi-year primary record and the upstream Industry, Financials, Moat, History, People and Bull/Bear tabs; the evidence behind each dial is cited inline below.
The split read is deliberate. The durability of the legacy annuity is high-confidence — direct, quantified, and unbroken across five 10-Ks. The durability of the AI-infrastructure bet is low-confidence — the demand is contracted but the cash returns are unproven, and the capital is voluntarily flowing to the one layer with the thinnest moat. The runway is genuinely large but it is bought, not free: it requires capital Oracle no longer generates internally.
What has to be true: the five load-bearing conditions
A long-term thesis is only as strong as the conditions it silently assumes. Below are the five things that must hold over the next 5-to-10 years for Oracle to be a superior investment, each with the current evidence and the single observation that would falsify it. The order matters: condition 1 funds the bet; conditions 2-4 are the bet; condition 5 governs who captures the upside.
Sources: renewal language, FY2025 10-K p.18 [2]; RPO $638B and 12% conversion, Q4 FY2026 call [1]; no single customer over 10%, FY2025 10-K p.109 [14]; 30-40% AI margin target, Q2 FY2026 call p.6 [15]; FY2027 funding plan, Q4 FY2026 call p.5 [21]; founder stake and pledge, 2025 Proxy [25] [26].
The conditions are sequenced by dependency, not by importance. Condition 1 is the safest and it is what makes the rest underwritable: the annuity is the floor that funds and de-risks the bet. The thesis lives or dies on conditions 2 and 3 — whether contracted demand turns into cash at an acceptable return. Condition 4 sets the time limit on getting there.
Condition 1 — The annuity floor must hold
This is the part of the thesis with the most evidence and the least controversy, and it is what separates Oracle from a pure AI-infrastructure speculation. The database-and-support franchise is the textbook high-switching-cost moat, and Oracle quantifies it: support is priced as a percentage of license fees, "substantially all" customers attach it, and "substantially all license support customers renew their support contracts with us upon expiration" [2]. The switching cost is not a fee; it is the operational and compliance risk of migrating the mission-critical data the business runs on — anchored to "the world's most popular enterprise database" [9].
The clearest proof it is a moat and not just a good product is the shape of the revenue: license support generated about $19.5 billion in FY2025, essentially flat, yet sits inside a cloud-and-license business earning a 63% total margin [3]. A flat, near-100%-renewing, very-high-margin stream is exactly what a switching-cost moat looks like: it does not grow fast, but it does not leave.
Source: cloud and license offerings detail (cloud services $24,506M, license support $19,523M, license $5,201M), FY2025 10-K p.74 [3].
Why this matters for a decade, not a quarter. Two durability facts give the floor unusual confidence. First, the renewal language is a constant — it appears in every 10-K from FY2021 through FY2025, surviving the on-prem-to-cloud transition, the $28.2 billion Cerner integration, a near-zero-growth stretch, and a CEO transition. Second, the moat now demonstrably travels: Oracle's database runs inside Azure, AWS and Google Cloud, and that multicloud revenue grew 404% year-over-year in Q4 FY2026 [10]. Customers who chose a rival's cloud still pay to run Oracle inside it — decisive proof the lock-in is the data, not captive hosting.
The slow-erosion risk to watch. The annuity is mature and the growth premium no longer rides on it: infrastructure (OCI) reached 56% of cloud services and license support revenue in FY2025, overtaking applications [11]. The multi-year tail risk is not the base shrinking but new AI-era workloads defaulting to cloud-native stores (PostgreSQL, Snowflake), slowly starving Oracle of new license seats while the old base renews. Renewal rates would mask that for years — which is why the single most important sentence Oracle publishes is the "substantially all renew" language, and the first place to look for cracks.
Condition 2 — The backlog must convert to cash, not just bookings
The $638 billion RPO is the fact that reframes Oracle's valuation, its capex, and its risk all at once. It is the hero chart of the decade — flat-ish for years, then a near-vertical move once AI-training contracts landed.
Sources: RPO of $137.8B at May 31, 2025 — FY2025 10-K p.107 [12]; FY2026 RPO of $638B — Q4 FY2026 call p.2 [1]; FY2021–FY2024 levels as reported in company segment data.
The slope is the whole story: RPO compounded steadily from $41B to $138B over five years, then multiplied roughly 4.6x in a single year. Management frames it as "exceptional visibility into our future revenue growth, all supported by long-term contractual customer commitments" [1]. The bull case is in that sentence. The bear case is in the same fact read differently — and three structural features decide which reading wins over a decade:
- It is back-end loaded. Only about 12% of the $638B is expected to convert within twelve months [1] — versus ~33% of the smaller, pre-AI FY2025 backlog [12]. Most of the cash is years out, which means most of the capital must be spent before it arrives.
- It is concentrated. The backlog leans on "the who's who of AI, including OpenAI, xAI, Meta, NVIDIA, AMD, and many others" [13]. Independent estimates put one counterparty (OpenAI) near half of RPO. A backlog this concentrated is a contract on a few sophisticated buyers who can multi-source compute — not a diversified SaaS book.
- The 10-K's reassurance predates the ramp. Oracle still discloses that no single customer reaches 10% of revenue [14] — but that was written before the largest AI contracts begin converting to revenue. It is a line to re-check every year.
The disconfirming signal is observable and near. RPO declining sequentially, a credit reserve taken against a marquee counterparty, or any renegotiation/delay would convert the thesis from "largest contracted backlog in software history" to "concentrated counterparty credit" overnight. The proof the bull needs is narrower: RPO still rising while collected cash visibly tracks the contracted economics.
Condition 3 — AI capacity must earn its cost of capital
This is the crux, because it is where the capital and the valuation premium go and where the moat is weakest. The old Oracle scaled software at near-zero incremental cost; the new Oracle rents depreciating hardware. Management guides AI workloads to "the 30% to 40% range over the life of a customer contract" [15] — well below the legacy 80%-plus software margins — and consolidated gross margin already "stepped down around 5 points" during the FY2026 ramp [1]. Leaked figures reported externally suggest mid-teens margins on early Nvidia rentals; management counters that delivered-capacity margin is already in the low-30s. The truth is genuinely contested and will only be settled by reported segment economics as data centers fill.
The capital intensity is the part that is not contested. Capex went from a steady ~$2 billion a year through FY2021 to $21.2 billion in FY2025 [16], and the company simultaneously switched off the buyback machine — repurchases collapsed from ~$21 billion in FY2021 [24] to $0.6 billion in FY2025 [16].
Sources: FY2025 capex $21.2B and buybacks $0.6B, FY2025 10-K p.102 [16]; FY2021 buybacks ~$21B, FY2021 10-K p.57 [24]; FY2026 net capex from the June 2026 earnings presentation p.5 [6]; FY2026 buybacks shown nil for scale.
Two quality-of-return caveats temper the headline economics and belong in any 5-to-10-year underwrite. First, reported profit is being flattered by a server useful-life extension from five to six years, which increased FY2025 net income by about $573 million [17] — Oracle is lengthening depreciation precisely as some peers shorten it for faster AI obsolescence. Second, the capital base is larger than the cash-flow statement shows: Oracle disclosed $43.4 billion of additional data-center lease commitments expected to commence through FY2028, off the FY2025 balance sheet [18]. The honest verdict: no return on the OCI capex should be assumed; it must be earned contract by contract, and the financials do not yet prove it has been. This is the single lowest-confidence link in the chain.
Condition 4 — The balance sheet must survive the build
For forty years Oracle's defining virtue was self-funding cash flow. That virtue has been deliberately spent down. Free cash flow swung from roughly +$11.8 billion in FY2024 to negative $23.7 billion in FY2026 [4], as a record $32 billion of operating cash flow [6] was dwarfed by the capex outlay. And the hole gets deeper before it gets better: management guided FY2027 to roughly $70 billion of net capex, funded by ~$40 billion of fresh debt and equity, including a $20 billion at-the-market equity issuance [21].
Sources: FY2021–FY2025 operating cash flow and capex, FY2025 10-K p.102 [16]; FY2026 OCF $32B and FY2026 FCF negative $23.7B [4]; FY2027 net capex ~$70B and ~$40B funding, Q4 FY2026 call p.5 [21]. FY2027 OCF/FCF are analyst illustration, not company guidance.
The leverage backdrop: notes payable and other borrowings stood at $92.6 billion at May 31, 2025 [19] and have climbed past $120 billion since, while the buyback decade left an accumulated deficit of $15.5 billion and a thin equity base [20]. The mitigants are real: maturities are long-dated rather than a near-term wall, the investment-grade rating has been affirmed, and the $2.7 billion gain on exiting the Ampere stake added one-time cushion [30]. The credit risk is not today's leverage ratio (~3x net debt/EBITDA is elevated but workable for this margin and backlog); it is the trajectory — another ~$40 billion of funding planned, FCF guided to stay negative, and equity-linked paper now in the mix from a company that once retired a third of its shares. The thesis breaker here is a rating action off investment grade or a financing pull; the FY2027–FY2028 window is when capex and funding collide hardest with the cash drain.
Condition 5 — Stewardship and the capital-allocation track record
Over a decade the question is not only whether the bet pays but who captures it. Oracle is a founder-controlled company: Larry Ellison owns 40.6% of the stock — about 1.16 billion shares — with no super-voting structure, so his vote equals his economic interest [25]. That alignment is genuine, and the team's credibility on near-term delivery is earned — in FY2024 the CFO told investors her own multi-year targets "might prove to be too conservative" [29], and two years later the backlog printed at $638 billion exactly as pointed.
The capital-allocation record is best read as three chapters: a decade of aggressive buybacks that retired a third of the shares, the $28.2 billion debt-funded Cerner acquisition in 2022 [22] that has underperformed on EHR share, and now the all-in AI build. The track record is therefore mixed: superb per-share compounding in chapter one, a questionable healthcare deal in chapter two, and an unproven mega-bet in chapter three.
The two stewardship watch-items for a long-term owner:
- The pledge. Ellison has pledged 346 million shares against personal loans — a carve-out from a policy that prohibits pledging for every other director and officer [26]. A pledge of ~30% of the controller's stake is a structural tail risk shareholders cannot diversify away — forced selling would be most damaging precisely in the drawdown scenario the bear case contemplates.
- The new pay regime. On promotion, the co-CEOs received option grants to purchase $250 million and $100 million of stock, largely time-vesting [27]. This is the opposite signal from FY2025, when the Compensation Committee voluntarily reduced earned NEO bonuses to zero, redirecting the cash to capex [31]. How these grants perform and whether the board can hold a two-CEO structure accountable beneath an 81-year-old founder is the governance swing factor.
The reinvestment runway — large, but bought not free
The bull's strongest structural argument is the size of the opportunity. The industry is in a once-in-a-generation demand shock, and Oracle has more contracted forward demand than any enterprise-IT peer. Management has laid out an OCI revenue ramp from $18 billion in FY2026 to $144 billion by FY2030 [7], underpinning a long-term framework of a 31% revenue CAGR and 28% non-GAAP EPS CAGR through FY2030 [8], with FY2027 revenue guided to roughly $90 billion, up 34% in constant currency [23].
Source: OCI ramp $18B to $144B, Q1 FY2026 call p.2 [7]; long-term 31% revenue / 28% EPS CAGR, June 2026 presentation p.15 [8]. Forward figures are company guidance.
Two features make the runway better than a pure commodity-compute story. The highest-quality slice is capital-light: Oracle reported $75 billion of "bring-your-own-hardware or prepaid" contracts with no margin degradation, where the customer funds the hardware [10]. And multicloud — the database traveling into rival clouds, growing 404% — is a distribution channel funded by someone else's balance sheet. To the extent OCI demand is pulled by the database (run it here because the data is here), the infrastructure franchise inherits a sliver of the real moat.
But the runway is bought, not free — and the binding constraints are now physical, not financial. Oracle's own risk factors flag "rising costs for data center energy demands" and dependence on third-party suppliers for advanced AI accelerators [28]. Power, GPU allocation, and construction timing — not demand — are the ceiling on this runway. That is why "reinvestment runway" rates High while the return on that reinvestment rates only Medium: the opportunity is real, the capacity to fund it exists, but the cash-on-cash payoff is still unproven.
Scenario frame: what the variable is worth
The same facts support a wide outcome distribution because the central variable — backlog conversion at an acceptable return — is genuinely unresolved. The three paths below are not forecasts; they are the value of resolving condition 2 and condition 3 in different directions, anchored to management's own guided FY2027 non-GAAP EPS of roughly $8.05 [21] and consensus FY2028 EPS near $11.
Sources: scenario anchors — FY2027 guided non-GAAP EPS ~$8.05, Q4 FY2026 call p.5 [21]; long-term 28% EPS CAGR framework, June 2026 presentation p.15 [8]. Indicative values synthesize the upstream Bull and Bear tabs; FY2028 consensus EPS is per analyst estimates, as reported.
The spread is the thesis: a roughly 2.5x gap between the bear and bull outcomes, driven almost entirely by one unresolved question. That is unusually binary for a $500B-plus company, and it is why this is a "lean long, wait for the cash" rather than a "back up the truck" — the contracted demand earns conviction; the cash that proves it has not yet appeared.
The multi-year watch dashboard — separating thesis evidence from noise
These are the durable signals that confirm or break the thesis over years, deliberately distinct from quarterly beats. The discipline for a PM: a revenue or bookings beat is noise relative to these; a move in any one of these is signal.
Sources: signals derived from the cited evidence above — renewal language, FY2025 10-K p.18 [2]; the cash gap and FY2027 funding, Q4 FY2026 call p.5 [21]; gross-margin step-down and RPO conversion, Q4 FY2026 call p.2 [1]; multicloud growth, Q4 FY2026 call p.3 [10]; founder pledge, 2025 Proxy p.32 [26].
The underwriting verdict
What has to be true for Oracle to be a superior 5-to-10-year investment is narrow and knowable. The wide-moat annuity must keep renewing (high confidence), the $638 billion backlog must convert to collected cash on schedule (contracted but unproven), the AI capacity must earn a return above the cost of the ~$70-90 billion-a-year capital it consumes (the lowest-confidence link), the balance sheet must survive the build without a downgrade (a trajectory risk, not yet a level risk), and a founder-controlled board must allocate the windfall to owners. Conditions 1 and 5 are largely in hand; the entire equity premium rests on 2 and 3.
The single most important driver is backlog-to-cash conversion at an acceptable margin: it is the variable on which the cash drain, the leverage, and the multiple all resolve at once. The single most dangerous failure mode is a marquee AI counterparty (OpenAI/Stargate, xAI) renegotiating, delaying, or wavering on funding — which would expose the backlog as concentrated counterparty credit, strand debt-financed depreciating capacity, and break the financing math simultaneously.
Lean long, underwrite the cash, not the promise. The contracted backlog and the durable database annuity are stronger, harder evidence than the de-rating bear case — but commit conviction only as OCI gross margin troughs and stabilizes while operating cash flow visibly narrows the gap to capex, with RPO still rising and no counterparty crack. The day the "substantially all customers renew" language weakens, or the day a marquee counterparty's funding wobbles, the thesis converts from lean-long to avoid. Until the cash prints, this is a great franchise making a great bet — not yet a proven one.