ORCL has transformed from a sleepy database compounder into a leveraged AI-infrastructure landlord, with backlog up roughly 4x but free cash flow gone negative as capex tripled. The 62% trailing ROIC reflects the old Oracle; the new one earns its keep on commodity GPU rentals.

Plain English

Oracle sells the database that runs most of the world's biggest companies — boring, sticky, and very profitable. Recently it promised hundreds of billions of dollars of AI computing power to a few enormous customers and is borrowing heavily to build the data centers. The old business is wonderful. The new business is a giant bet that may or may not pay off, and it consumes all the cash the old business produces. At today's price you're paying for the bet to work. Wait for a much cheaper price, or own a much smaller position than you would for a pure database company.

Thesis

Oracle is two businesses stapled together. The first is a near-monopoly relational-database and ERP franchise (Oracle Database, E-Business Suite, NetSuite, Fusion) that historically produced a 10-year average ROIC of 62.1% and 110.8% FCF conversion (per scorecard) on negligible incremental capital. Switching costs here are among the deepest in software: a Tier-1 bank does not 'rip and replace' Oracle DB. The second business, only fully visible since 2024, is Oracle Cloud Infrastructure (OCI) plus a slate of multi-hundred-billion-dollar AI training-capacity contracts (OpenAI/Stargate-class deals). As of 28 Feb 2026, remaining performance obligations were $552.6 billion (10-Q). Of that, only ~12% recognizes in the next 12 months; ~66% beyond month 36. To deliver, Oracle has spent $39.2B of capex in nine months (vs. $12.1B prior year) against $17.4B of operating cash flow, with ~$4.5B of unpaid capex on top. The result: net debt that the scorecard still flags at -1.86x EBITDA flips materially positive on a forward basis once leases and committed capex are layered on, and 5-year ROIIC has already collapsed to 2.9%. At $171.83 the stock trades at 40.6x earnings, 50.6x EV/FCF, and a px/IV ratio of 0.87 against an IV-base of $197.68. Reverse DCF implies ~5.1% perpetual growth — modest in isolation, demanding in the context of the capex ramp. Owning it makes sense only well below the IV-low of $132.6, where the price compensates for the ROIIC reset. Above $200 the bull-case is fully embedded.

Moat

Oracle's moat is a tale of two franchises with different durability profiles. Each must be evaluated separately, because they have different competitive structures and erosion paths.

1. Switching costs (legacy database & applications) — the historical engine. Damodaran identifies switching costs as the most reliable software moat: 'Microsoft recognized earlier than most firms that the most significant barrier to entry in the software business is the cost to the end-user of switching from one product to a competitor' [1][2]. Oracle Database is the canonical example outside Microsoft. A Fortune 500 ERP migration off Oracle takes 3-7 years, costs hundreds of millions, and risks audit-trail and regulatory continuity. The same applies to Fusion ERP/HCM and NetSuite for mid-market. Stress test: even if a $10B competitor (think Snowflake, Databricks, AWS Aurora) emerged tomorrow, the installed base behavior over the last decade — Oracle's revenue compounded slowly but operating margins expanded — shows the cost of switching is actually rising, not falling, because customers have layered application logic, custom PL/SQL, regulatory certifications, and integrations on top. The 10-year ROIC average of 62.1% is the financial fingerprint of this moat. Erosion risk: PostgreSQL/Aurora at the new-workload margin is real; existing workloads are sticky for another decade. Verdict for legacy: WIDE.

2. Cost advantages / scale economics (OCI hyperscale) — the new bet. Damodaran's framework: 'In businesses where scale can be used to reduce costs, economies of scale can give bigger firms advantages over smaller firms' [2]. OCI is attempting to insert itself as the fourth hyperscaler. Oracle's advantages are non-trivial: it owns the Exadata stack, can co-locate database with compute (a genuine product differentiator), and has signed multi-decade RPO with concentrated AI customers. But this is fundamentally a commodity GPU rental business with the same cost curve as AWS, Azure, and GCP — three competitors who each spend $70-90B/year of capex versus Oracle's $50B/year run-rate. Scale here is reflexive: the leader gets the cheapest power, fiber, and Nvidia allocations. Oracle is #4, not #1. The $552.6B RPO disguises this — most of that backlog is hyperscale capacity at margins below the legacy 80%+ database business, with 6-figure GPU-cluster contracts whose unit economics depend on yet-to-be-deployed liquid-cooled data centers. Stress test: if a $10B + 5-year competitor enters? Already happened: AWS, Azure, GCP, plus CoreWeave, Lambda, Crusoe. Verdict for OCI as standalone: NARROW at best, possibly NONE depending on whether Oracle can keep utilization above 70% post-2027 customer concentration cliff.

3. Pricing power. On legacy database: meaningful. Oracle has historically raised license maintenance 4-8% annually with single-digit churn. On cloud: zero. AI capacity is priced to clear, with hyperscaler price wars already visible in 2025-2026 GPU-hour pricing. Damodaran [3]: 'managers who take over a valuable brand name and then dissipate its value, will reduce the values of the firm substantially.' The risk is using cash flow from the moated database business to subsidize a non-moated infrastructure business.

4. Intangibles / brand. Oracle's brand is respected but not loved (compare to Apple's brand exclusivity argument [4]). It is a procurement-driven, CIO-driven brand. Useful for selling more SKUs into the existing base, not for pulling new logos.

5. Network effects. Limited. Oracle's ecosystem partner network is real but not self-reinforcing the way AWS Marketplace or Salesforce AppExchange are.

The 5-year ROIIC of 2.9% versus 10-year ROIC of 62.1% is the single most important moat datum in the scorecard: incremental capital is no longer earning the legacy moat's economics. Oracle is still a wide-moat franchise overall because the legacy still throws off owner-earnings at scale, but the marginal dollar of capital is being deployed where the moat is narrow.

Moat verdict: NARROW (composite — wide legacy stitched to narrow/none cloud, with capital allocation now flowing predominantly to the narrow side).

Management

Oracle is founder-controlled. Larry Ellison owns ~40% of shares outstanding and is Chairman/CTO; Safra Catz is CEO. This is both the bull case ('owner-operator with skin in the game') and a key risk ('one person decides $300B of capex'). Evaluating the five capital-allocation choices:

1. Reinvestment. Capital expenditures ran $39.2B in the first nine months of FY26 versus $12.1B a year earlier — a tripling — with $4.5B of unpaid capex layered on top. The 9-month operating cash flow was $17.4B. Oracle is consuming all internally generated cash plus issuing debt to build AI training capacity for a small number of large counterparties. The 5-year ROIIC of 2.9% (scorecard) tells us the marginal return on this reinvestment is currently far below cost of capital. Either (a) ROIIC dramatically inflects upward as RPO converts to revenue at high incremental margins from 2027+, or (b) Oracle is destroying capital at the margin. Management is making the bet a wide-moat operator should make only if the unit economics are clearly differentiated; the jury is genuinely out.

2. Acquisitions. The Cerner deal (~$28B in 2022) is the most relevant data point and remains controversial — integration is multi-year, healthcare cloud has been slow to monetize. The Ampere divestiture in November 2025 (cash proceeds $4.3B, $2.7B realized gain) is a positive: management exited a non-core silicon bet at a reasonable price. The TikTok USDS JV minority (15%) is small but reflects Ellison's geopolitical comfort.

3. Debt. Total senior notes were $43B as of February 28, 2026 (10-Q). Net debt to EBITDA of -1.86x in the scorecard reflects the cash position pre-capex commitments. With ~$300-500B of contracted future capex implied by RPO conversion, the forward debt picture is materially worse. Interest coverage of 4.96x is healthy today but compresses if EBITDA dilutes from the AI mix shift.

4. Buybacks. Share count has shrunk only 4.6% over 10 years (scorecard), with most retirement happening pre-2020. With shares now at 40.6x earnings versus a 10-year average P/E of 44.6x, Oracle is no longer cheap to itself, and rightly the buyback has been throttled to fund capex. Average P/IV on legacy buybacks (2014-2020 era) was reasonable — Oracle did not chase. Grade for past buybacks: B+.

5. Dividends. Steady and growing, though modest as a percent of cash flow. Reasonable.

Communication quality. Mixed. Ellison's commentary on Q1/Q2 FY26 calls about the RPO ramp has been bombastic ('we will be the largest cloud provider'), with selective disclosure about customer concentration and gross margin progression at OCI. Catz's CFO-style discipline partially offsets. The 10-K and 10-Q remain dense and comprehensive, but segment economics for OCI are not separated cleanly enough for an outside analyst to underwrite the marginal-return question.

Applying Buffett's filter [6]: 'Partner with high integrity leaders who understand their customers and act like owners.' Ellison qualifies on owner-alignment but the scale and concentration of the AI infrastructure bet is the opposite of 'concentrate our capital in a few high conviction ideas' applied to durable advantage businesses; this is concentrating capital in a non-durable business. Track record on legacy capital allocation is strong; track record on the current bet is unproven and too large to fail-safely.

Capital allocator: B (downgrade from historical A- because the marginal capital is being deployed where the moat is weakest, with ROIIC already showing the strain).

Industry

Oracle competes in two industry structures simultaneously, which is part of why this is hard.

Legacy enterprise software (database + ERP + apps).

• Threat of new entrants: LOW. Mission-critical DB takes a decade-plus to build and certify; ERP requires deep vertical knowledge.
• Bargaining power of buyers: MEDIUM. Large enterprises have audit leverage on Oracle but cannot realistically switch quickly.
• Bargaining power of suppliers: LOW. Oracle largely controls its own software stack; Java is a wholly-owned asset.
• Threat of substitutes: MEDIUM and rising. PostgreSQL, Aurora, Snowflake, Databricks for analytics workloads, plus SaaS-native ERP (Workday, SAP S/4HANA cloud) for new logos.
• Industry rivalry: MEDIUM. Stable oligopoly with SAP, Microsoft, IBM. Pricing discipline mostly holds. Value pool: Excellent today; gradually shrinking on a same-store basis as cloud-native alternatives gain new-workload share. Verdict for legacy: GOOD.

Hyperscale cloud + AI training capacity.

• Threat of new entrants: LOW for full hyperscalers (capital intensity), MEDIUM for AI-specialist neoclouds (CoreWeave, Crusoe, Lambda, Nebius) with similar Nvidia partnerships.
• Bargaining power of buyers: HIGH. AI training customers (OpenAI, Anthropic, Meta, xAI) are sophisticated, have option value across providers, and exhibit extreme concentration. A small number of customers control a large share of the bookings.
• Bargaining power of suppliers: HIGH. Nvidia controls GPU allocations, pricing, roadmap. Power utilities and grid interconnect queues are increasingly the binding constraint, and they are not Oracle's.
• Threat of substitutes: HIGH. Custom silicon (Google TPU, AWS Trainium, Microsoft Maia), and at the workload level inference may shift to edge or specialized providers.
• Industry rivalry: HIGH. AWS, Azure, GCP, Meta in-house, plus a dozen neoclouds. Capacity glut risk by 2027-2028 if AI training demand plateaus. Value pool: Currently expanding rapidly in absolute dollars; per-unit economics already deteriorating. Long-term commodity dynamics likely. Verdict for cloud: AVERAGE to POOR depending on whether Oracle's database+cloud co-location story drives differentiated unit economics or whether GPU rental dominates the mix.

Where the value pool is going. The legacy pool is stable and migrating slowly toward subscription pricing — net-neutral to slightly positive for Oracle. The cloud pool is growing fast but with worse unit economics and customer concentration; the durable share of those profits accrues to (a) Nvidia, (b) power generators, (c) the top 1-2 hyperscalers with cheapest cost-of-capacity. Oracle is not in either of those positions.

Canon comparison is illustrative [3]: Damodaran warns that 'the competitive advantage that comes from exclusive licensing or a legal monopoly is a mixed blessing.' Oracle's database moat was built on functional monopoly; its cloud moat is being attempted in a structurally commoditizing market.

Industry Verdict: Average. The blended structure looks acceptable but is being pulled toward the worse half by Oracle's own capital deployment.

Inversion

I am now a short-seller. I want to make the strongest credible bear case for Oracle at $171.83.

1. The single event that kills this. A material customer concentration disclosure or counterparty wobble. The $552.6B RPO is reportedly anchored by a small number of mega-deals, including hyperscale AI training contracts whose counterparties are themselves cash-burning, equity-funded, or both. If even one $50-100B contract is renegotiated, deferred, or its counterparty's funding becomes constrained, the entire RPO narrative — which is the basis for the multiple expansion from ~$120 in mid-2024 to $200+ in late 2025 — unwinds in a single quarter. Oracle's contracts may include take-or-pay protection, but enforceability against an impaired counterparty is a recovery issue, not a revenue issue. The single event: an AI infrastructure customer announces a strategic pivot that defers $30B+ of committed Oracle spend. The stock re-rates to legacy-multiple territory in days.

2. Why the moat is narrower than bulls think. Bulls implicitly assume Oracle's database switching-cost moat extends to OCI because OCI runs Oracle Database better than anyone. This is partially true for existing Oracle workloads but largely irrelevant for the actual driver of recent growth, which is GPU-hour rental for non-Oracle-DB workloads (LLM training/inference). Those workloads have zero Oracle-specific lock-in. Once a model is trained, it can be served from any hyperscaler. The training contracts are multi-year, which provides revenue visibility, but renewal economics will reflect the actual switching costs, which are low. Meanwhile the legacy moat itself is slowly eroding at the workload-acquisition margin: net-new applications are increasingly built on PostgreSQL/Aurora/Snowflake stacks. Composite moat: narrow, and the new business has no moat at all once the contracted period ends.

3. Why management is worse than it appears. Ellison's track record is in the 1990s-2010s database era, not in commodity infrastructure operating economics. The Cerner integration has been poor. Capital discipline at the marginal dollar is now objectively weak: 5-year ROIIC of 2.9% (scorecard) versus 10-year ROIC of 62.1% means the last five years' incremental capital has earned roughly nothing relative to legacy. Adding the FY26 capex surge and locking in $300-500B of forward capex against ~$25-30B of forward run-rate operating cash flow (before further debt) is a bet-the-company decision being made by a chairman who is 81 years old. Succession planning beyond Catz is unclear. Compensation continues to feature large stock grants. Buybacks have been throttled at exactly the time the business needs capital, so good — but it confirms that Oracle is no longer self-funding its strategy.

4. What bulls are extrapolating that won't hold. Bulls extrapolate (a) RPO recognizes at legacy software gross margins, (b) AI training demand grows linearly through 2030, (c) Oracle keeps utilization above 80% throughout. All three are heroic. (a) AI infrastructure gross margins are ~20-35% steady-state versus 80%+ for license; the consolidated gross margin will compress every year for the next five years. (b) Training compute may saturate as model scaling laws bend; inference compute redistributes geographically and to specialized silicon. (c) Customer concentration means utilization is binary — if the anchor customer pulls back, you don't get to 70%, you get to 30%. Reverse DCF implied growth of 5.1% looks modest, but it's 5.1% of the new, lower-margin revenue base, which means owner earnings need to grow much faster than that — and the capex consumes the cash before it can compound.

5. Valuation trap (multiple compression / regime change). P/E TTM of 40.6x is roughly in line with the 10-year average of 44.6x, which sounds fine until you note that the 10-year average reflects a software-margin business and the next five years are an infrastructure-margin business. Comparable infrastructure providers trade at 15-22x EBITDA / 18-28x earnings. EV/FCF of 50.6x is genuinely expensive when forward FCF is negative for at least 2-3 years. The IV-low of $132.6 is not a worst case — it's a reasonable case if AI capex doesn't compound. The realistic bear-case IV is below the IV-low: a 25x earnings multiple on $5-6 of legacy-only owner earnings per share gives ~$125-150, and a stress case (multiple compression to 18x, EPS hit from depreciation surge) implies $90-110.

If I am right, the stock could be worth $95-110 within 2-3 years.

Lollapalooza Bias Check

Active biases in me as I write this:

Recency bias — strong. Oracle's stock has roughly doubled in 18 months on the OpenAI/Stargate-scale RPO disclosures. The narrative of 'Oracle as the fourth hyperscaler' is novel and exciting, and recent price action is anchoring my sense of what is possible. Counter-discipline: I am forcing myself back to the 5-year ROIIC of 2.9% as the more reliable signal than the 5-year stock chart.

Anchoring — moderate. The IV-base of $197.68 is anchoring my sense of fair value, but that IV is itself produced by a model that takes RPO conversion at near-legacy margins. If I anchor instead on owner-earnings of $19.45B at a 25x multiple (legacy-software fair value), I get a market cap closer to $480B versus the current ~$480B (this matches roughly, but the legacy alone earned that). The anchoring is actually serving me here, but I should notice it.

Authority/social-proof — material. Larry Ellison's track record (a quarter-trillion of personal net worth, decades of correct big calls) makes me reluctant to call this a capital-allocation mistake. Many sell-side and buy-side analysts I respect are bullish. I am partially deferring to authority and partially fading it. The Buffett/Munger discipline says: never override your own framework because the founder is famous.

Confirmation — moderate. I came in expecting 'Too Hard or Hold' as the most likely answer. I therefore over-weight evidence of capital-intensity risk and under-weight evidence that OCI's database co-location really does create a structural cost edge for Oracle Database workloads (which is a credible claim).

Commitment-and-consistency — mild. Once I write 'Hold' or 'Trim' here it's hard to switch. To counter this I'm leaving room in the recommendation framework for the price to do the work — meaningful margin of safety below $135 changes the answer.

Deprival super-reaction — present. Oracle's stock is up significantly off the lows; there is the felt sense of 'this is the one big AI re-rating you're going to miss.' This is exactly the bias Munger calls the most dangerous in late-cycle environments.

Incentive bias (mine, structurally) — I am paid to write differentiated views. A 'Too Hard' is the least differentiated answer. I should resist the pull toward false precision.

Net: my biases push me modestly toward bullishness (recency, authority, deprival). The discipline is to weight the scorecard's ROIIC and the FCF math more heavily than the narrative.

10-Year Outlook

Same fundamental business model in 10 years? Partially. The legacy database/ERP business will look very similar — Oracle will still be charging maintenance and selling Fusion Cloud Apps to large enterprises. The AI infrastructure business is genuinely unpredictable: it could be 50%+ of revenue, or capacity could be re-purposed, or customer concentration unwound. The blended business of 2036 is not the same shape as 2026.

Customer base larger? Yes for legacy (slow growth in count, faster growth in seats per customer); uncertain for cloud where the customer count is small and concentrated.

Profit per customer higher? Mixed. Legacy yes (Fusion ARR per customer compounds at 8-12%). Cloud likely no — pricing pressure is the dominant force in commodity infrastructure once the supply tightness ends.

Moat wider? No. The legacy moat will be approximately as wide; the marginal capital is going into a business where the moat is narrow. Composite moat is narrowing.

Single biggest threat? AI capacity glut by 2028-2030. If model-scaling returns to compute saturate, or if inference shifts decisively to custom silicon and edge, Oracle has $300-500B of depreciating data-center assets earning sub-cost-of-capital returns. Secondary threat: PostgreSQL/Aurora encroachment on net-new database workloads. Tertiary: Ellison succession.

The Munger 4-test result: I can explain Oracle to a 12-year-old (database company that's also building AI factories), but the second test — 'same fundamental shape 10 years ago, 10 years forward' — is genuinely failed by the second business. The fourth test — 'does it require predicting tech adoption curves' — is a clear yes for OCI: I would have to forecast AI capex absorption, GPU obsolescence cycles, and customer concentration outcomes.

Buffett's principle [6] is to 'invest in businesses we thoroughly understand, with durable advantages and long-term economic prospects.' The legacy passes; the consolidated entity at the current capex run-rate does not, with high confidence.

CONFIDENCE: medium

Position Guidance

• Recommendation: Hold
• Conviction: medium
• Target buy price: $135 (just above IV-low of $132.6, where the price compensates for the ROIIC reset and AI capex risk)
• Target trim price: $230 (above IV-high of $250.93 means full bull case is priced; trim begins ~10% below to lock in gains)
• Position sizing: If owned, no more than half of what a pure legacy-database compounder would warrant — treat the OCI build as an unpaid call option, not a cash-flow engine. Avoid initiating new positions above $180. Below $135, sizing can step up; below $115, becomes interesting on legacy alone.