I'm LongbridgeAI. Here is the key information I extractedBelow is Dolphin Research's summary of$Oracle.US FY27 Investor Day (2026) minutes.
Key takeaways:
1. One-and-done financing: This cycle requires only an additional $20 bn, split evenly between equity and debt. There will be no further net new financing thereafter.
2. Cash flow inflection with a timetable: CapEx peaks in FY27–FY28, then declines. FCF turns positive in 2H FY28 and exceeds $30 bn in FY29, enabling self-funding.
3. Long-term targets with a '+': FY30 revenue $225 bn+ and Non-GAAP EPS $21+, with OPM 40%+. The revenue and EPS targets are unchanged from a year ago, and the '+' signals higher confidence.
4. Lower capital intensity via financing mix: A blend of supplier financing, BYOH (bring your own hardware), and customer prepayments cuts upfront cash needs by roughly half in this cycle. The 7 GW slated for the next two years are all contracted, and the financing mix is largely set.
5. Attractive project-level returns: A 1 GW project carries about $3.7 bn in CapEx, with run-rate revenue of about $1.6 bn at full output. Project OPM is around 34%, with ROI in the high-20% range, and most contracts are cost-plus, allowing cost pass-through.
6. Faster and more diversified delivery: About 850 MW delivered in FY27 Q1 across 38 projects in 11 countries. Total capacity reaches roughly 10 GW by end-FY28, and delays can be offset by reallocations across projects.
7. Usage-based pricing at the application layer: Fusion Claw and AI units introduce a new revenue layer beyond seats. The current guide bakes in only a modest contribution.
8. Supply-demand remains tight: Management expects GPU supply to trail demand for at least 3–5 years. Memory and wafer capacity, more than power, will govern deployable compute.
I. Core financial recap
1. Financing cap and shareholder returns
a. The current investment cycle, through the point of self-funding, needs $20 bn of fresh capital, split $10 bn equity and $10 bn debt, to be opportunistically executed as early as calendar 2026. Oracle credit will not be tapped for other structured financing, and data centers will remain on a lease model.
b. Capital allocation priorities: first, maintain investment-grade ratings. Second, return capital via dividends. Third, fund attractive business opportunities.
c. Balance sheet: after adding $10 bn of new debt, total debt should be broadly flat over the next two years with normal refinancing. Net debt/EBITDA is guided to fall below 3.5x by FY29.
2. Mid- to long-term guidance
a. Revenue: FY27 growth guided at 34% YoY. FY30 revenue $225 bn+, implying a 4-year CAGR of 35% and $158 bn+ of incremental revenue, with FY28 and FY29 growth roughly steady.
b. Profit and cash flow: FY30 Non-GAAP EPS $21+. FCF turns positive in 2H FY28 and exceeds $30 bn in FY29, with self-funding from FY29 onward.
c. Margins: FY26 Non-GAAP OPM at 43%. Mix and ramp will pressure margins in the next few years, but OPM during the FY28–FY29 peak ramp 'will not be meaningfully below' 40%, and returns to 40%+ in FY30.
3. CapEx and capacity
a. About 7 GW of new capacity over the next two years, all contracted, taking total capacity to ~10 GW by end-FY28. CapEx peaks in FY27–FY28 and declines in FY29.
b. Five large campuses sum to roughly 5 GW, each around 1 GW with several hundred MW variance, delivering from FY26 to mid-FY29. Abilene is near completion, and the other four begin initial deliveries in calendar 2027, with guidance allowing up to two-quarter delays for any single campus.
4. 1 GW unit economics (based on current RPO average project; assumes leased shells and power)
a. Inputs: CapEx of about $3.7 bn, slightly under 50% of total contract value. In self-funded mode, all CapEx is paid in the first two years.
b. Outputs: At full run-rate, annualized revenue of about $1.6 bn, up from ~$1.0 bn last year mainly due to cost-plus pass-through of higher memory and other costs. GPM around 35% and project-level OPM around 34%, with ROI in the high-20% range, and 100% conversion of post-tax OP plus D&A to cash at full output.
c. Lowering capital intensity (1 GW example, standalone and non-additive): In self-funded mode, the $3.7 bn CapEx is entirely paid in the first two years. With 70% supplier financing repaid over five years, the first two-year cash need drops to $1.8 bn, a reduction of roughly 50%. BYOH reduces the cash need to $600 mn or less, with revenue lower but higher margins and ROI.
d. Contract structure: Large AI infra deals are broadly cost-plus or similar. Revenue may vary, but margins and ROI are preserved, with renewals expected at equal or better profitability, and Q1 seeing 4-year GPUs sold or renewed at a 20% premium to original contract price.
5. Biz mix and backlog
a. RPO at $664 bn, with 90%+ from infrastructure. The first major FY26 contract was concentrated in a single client, and RPO then grew each quarter and became more diversified.
b. Three biz models: SaaS at double-digit growth, low capital intensity, and very high GPM. Cloud DB growing 50%+, mid capital intensity, and very high GPM, while infra has lower GPM and higher capital intensity but lean opex, narrowing the OPM gap meaningfully.
c. The serviceable TAM carries a 20%+ CAGR. The company expects to grow well above that pace.
II. Investor Day details
2.1 Exec highlights
1. Full-stack AI strategy and three value pillars
a. Three pillars: differentiated growth, strong margins/FCF/EPS, and disciplined capital allocation. All internal investments must align to these.
b. Three industry shifts: AI is moving from answering to acting, raising the bar on security and permissions, while AI is becoming the UI, as seen in NetSuite Next and Fusion Companion. Customers are shifting from experimentation to ROI-proof deployments and push back on indiscriminate token stacking.
c. Full-stack advantage: infra, database, AI Data Platform, and apps stack to deliver end-to-end scenarios such as healthcare, robotic surgery, and mining. Others need to stitch multiple vendors.
2. Fusion app AI adoption
a. In FY27 Q1, customers invoked embedded Fusion AI over 150 mn times, with usage up 42% from Jun to Aug. This signals accelerating adoption.
b. In Sept alone, Fusion AI consumed nearly 500 bn tokens, up 22.7% MoM. Token consumption continues to climb.
c. Live agents rose from just over 1,000 in May to 3,500 in Sept, with DHL, Avis, Orange, and ADT expanding usage. Adoption is broadening across verticals.
d. A new automated migration tool supports competitors' on-prem clients as well. About half of ERP clients remain on-prem, and moving to cloud multiplies revenue.
3. OCI infra: power and data centers
a. Power access: once power and shells are ready, builds can go live in 30 days. In 2024 at Abilene, behind-the-meter gas turbines were adopted for backup power, the second TX GW-scale campus uses VoltaGrid reciprocating engines, and the New Mexico campus is pivoting to GW-scale fuel cells.
b. Grid-friendly operations: rewriting UPS software and firmware smooths AI load spikes so the grid sees a profile like '1,000 households'. A 10 MW AI load simulator was installed in Shackelford, TX.
c. Community strategy: 200+ local hires and training at Abilene, with closed-loop liquid cooling to conserve water, including for fuel cells. Transparency programs disclose usage of air, water, heat, and light.
d. Build approach: civil works and fit-outs are combined and run in parallel, with modular, repeatable design and compatibility across compute architectures. GW campuses deliver over 1.5–2 years in phases.
4. OCI infra: supply chain and delivery speed
a. A GB300 rack has ~6,000 parts with multi-sourcing at each step. For Vera Rubin racks, Oracle took over the manufacturing line software to ensure racks are plug-and-play on arrival.
b. Capacity doubled over the past year, with GPU rack shipments tripling without delays. The supply chain is scaling reliably.
c. FY27 Q1 deliveries were ~850 MW across 38 projects in 11 countries, with Abilene just one of them. That equates to ~8.5 MW per day, 4x last year's pace.
d. From room entry to revenue takes roughly 21 days. At Abilene, OpenAI can complete acceptance in 24 hours, and availability exceeds 99%.
5. OCI infra: fungibility and resilience
a. A 512-GPU cluster can be securely sliced into 32 customers within minutes, and capacity was reallocated from Client A to B within hours last quarter. AI infra now serves 1,000+ customers.
b. CPU and network racks have moved to liquid cooling with density on par with accelerator racks and can swap across any DC. For air-cooled sites, the sidecar solution OCI L2Arc supports up to 1 MW racks per unit, with 100+ MW already deployed at a US DC.
c. General CPU biz: a recent client committed to ~1 mn CPUs, 99% of which are older platforms launched between 2017–2021. Contracts carry 18–48 months of 100% committed consumption.
d. BYOH: leveraging SmartNICs and off-box virtualization, customer-owned racks can connect to OCI and access the full OCI service suite.
6. Database and multicloud
a. On average, AI infra customers spend 13% on non-accelerator services and use over 40 additional OCI services. Cross-sell remains substantial.
b. Machine-generated data will rise from 30% two years ago to 50%–80%. The next 'bn users' of databases will be agents, and Oracle launched the Deep Trust AI Architecture and the AI Data Platform's AI Discovery agent that auto-reads app code and semantic models to build business ontologies.
c. Multicloud DB is now in about 71 regions QTD, up over 209%, making it one of the fastest-growing businesses. Clients like Zema Global and Conduent have migrated rapidly.
7. Fusion apps: from 'systems of record' to 'systems of outcomes'
a. The app layer is moving from AI-assisted to AI-autonomous work after 18 months of the former. Models, guardrails, and observability are provided by OCI, including gpt-oss models, with cross-system data via the AI Data Platform.
b. Fusion Claw executes expert workflows previously done in planning or reconciliation tools or Excel. AI handles reasoning and planning, deterministic compute executes, and it reconciles 1–2 mn records and runs 1,000 scenarios without token usage, with a single execution engine spanning thousands of use cases.
c. This release adds 75 agentic apps, 25 of which support Claw, and offers both no-code and professional environments via VS Code/CLI. Developer velocity is improving.
d. Fusion Companion and Fusion Agentic Home provide a unified entry point across Fusion, with an OpenAI plug-in already available. Fusion counts ~40–50 mn users, many frontline, who need native agent entry points.
8. Fusion apps: business model changes
a. AI units introduced usage-based billing for tokens and compute, adding a revenue layer beyond seats. The more AI does, the more revenue accrues without adding seats.
b. With inference and compute decoupled, AI units carry strong GPM and are not dependent on token costs. This enhances unit economics.
c. Agents ran 5.7 mn times last quarter, and AI Studio certified developers now exceed 100k. About 7,000 customers and 3,600–3,700 agents are in production.
9. Industry apps and NetSuite
a. Growth thesis: on-prem to cloud lifts revenue 3–4x, with cross-selling a second app tripling again and a third app driving another step-up. Most customers in many industries run only one Oracle app, with 60% of North American financials using one app, and cross-sell opportunities exceed $5 bn in North America alone.
b. Customer cases: MGM Resorts runs Fusion Financials and OPERA across 16 properties and 43k rooms, integrating bookings to the GL. Canada’s CAMH runs finance, HR, supply chain, CX, plus Oracle EHR, while Saudi Arabia’s largest residential builder uses Fusion and Primavera Cloud, improving payment speed by 50% and cutting workload by 50%.
c. NetSuite: 44k midmarket customers and a broad set of AI and industry solutions across healthcare, restaurants, construction with Textura, and local governments via SuiteApp, with SuiteExpress accelerating implementations.
d. New agent apps: a delivery reconciliation agent for Burger King Italia built in weeks and now sold broadly. New apps include utilities, construction, agriculture via Oracle Agriculture Intelligence, government, financials for cash management and retail lending, hospitality, and retail invoice matching.
e. Embedded finance: ACH and card payments in the US, UK, and Canada, expanding to supply chain payments, wallets, payroll, cross-border payments, and T&E. TAM expands from IT budgets into overall operating budgets.
2.2 Q&A
Q: CapEx peaks in FY27 and FY28, but does FY28 total CapEx hold flat vs. FY27, or step up with many new go-lives?
A: The 1 GW example is provided for your own modeling. Directionally, some GW ramps occur this year, and the ramp could be larger next year.
Q: To confirm, the $10 bn equity plus $10 bn debt totals $20 bn in FY27, with no additional financing in the following two years?
A: Yes, that is the firm commitment today. With supplier financing or similar arrangements, BYOH, and prepayments, we are confident in capping needs at $20 bn. We do not expect other structured financings that rely on Oracle credit, and we will continue to lease data centers.
Q: Pricing differs significantly across 5–6 year contracts, 1–3 year deals, and spot. How do you manage capacity sales?
A: It is a portfolio choice that depends on risk appetite. Shorter terms fetch higher early-life pricing but carry more future uncertainty, and we balance upfront cash, payback timing, partners, and how customers allocate capital. When cash is ample and risks tolerable, we take more risk for higher returns, while longer terms suit lower-risk needs, alternative capital, or locked-in yields.
The market underestimates that DCs and hardware going live this year will be far cheaper than new builds five years out. That is why we showed GPU renewals, as the industry fears hardware life is overestimated while we see the opposite. There is also more than one pricing event, as a 4-year deal can be renewed for 2–3 years or sold into spot, like airlines and hotels requiring highly dynamic pricing. We also act as distribution for clients, for example Fusion Claw aggregates some model clients whose outputs we can monetize again, a full-stack advantage beyond selling raw compute.
Q: With inflation, indexation disputes, and higher rates last year, FY30 targets are largely unchanged. Was there buffer in the original, or did factors offset?
A: The target is now '$225 bn+' instead of '$225 bn'. Large AI infra deals are some form of cost-plus, and CPU contracts are signed closer to go-live, giving confidence to cover CPU or GPU cost inflation during builds. The $158 bn+ four-year revenue add is likely underappreciated, and after a year we are more confident in $225 bn revenue and $21 EPS. The '+' reflects that confidence.
Q: Beyond this cycle, how do you view subsequent cycles? Can FCF keep growing after $30 bn in FY29, and what is your appetite and size for new commitments?
A: FCF of $30 bn+ leaves room to grow, and all three biz models are strong FCF franchises. If FCF far exceeds $30 bn in FY29–FY30, we will be in a different position and able to sustain appropriate growth post-2030 based on market opportunities.
The world is dynamic and we want flexibility, but proportions matter. Going from 1 GW to 10 GW is a very steep curve, and it is not feasible to jump from 10 GW to 100 GW in 2030–2034, so funding mix will naturally differ then.
Q: What changed to reduce financing needs so sharply this cycle?
A: Our message has been consistent, as supplier financing, BYOH, and prepayments are not new and have been discussed for several quarters. This time we quantified scope and scale, while external models differed before. Mechanisms have now been implemented rather than just envisioned, giving us the confidence to make today's commitments.
Q: For the 7 GW going live in the next two years, how should we think about the mix of supplier financing, BYOH, and prepayments? Would mix shifts require external financing, or rejecting some customers to preserve the mix?
A: The 7 GW are contracted and the three-way mix is essentially set; otherwise we would not commit to $20 bn. We have high confidence in the plan across current customers and suppliers. Incremental growth is mostly after FY28, and what goes live between now and FY28 is largely this slate.
Q: If supplier financing grows more important and the company relies heavily on a single chip vendor, are those commitments locked?
A: That is an external inference. We have not stated any specific financing with any vendor or disclosed size and quantity, and such conclusions should not be drawn. We have reasons to be confident in multiple financing avenues.
Q: How meaningful is hosting Google TPU for Oracle? What does it imply for revenue, backlog, and margins?
A: TPU is an excellent accelerator, and our strategy is to offer whatever accelerators customers want. Not every customer will move to TPU, and that is fine, just as we offer AMD, Intel, NVIDIA ARM, and Ampere ARM on CPUs. We keep integration costs for ourselves and partners extremely low to reach more customers and prove our architecture is flexible, which underpins confidence in integrating new hardware 5–8 years out. Some customers want TPU on OCI, many want Vera Rubin, and Cerebras and Broadcom Jalapeño are also drawing interest, while demand for all accelerators will far exceed supply for several years.
Q: With DC permitting challenges, how much fungibility and buffer do you have if delays occur?
A: Large builds are inherently complex, and routine issues hit the headlines because the industry is in the spotlight. Civil works are a small part of capital, while racks, networking, fiber, cabling, chillers, and UPS are movable, and the immovable elements like site prep and interconnects are a small share. If one project is delayed, components can accelerate another. With ~40 concurrent projects per quarter, we keep optimizing the system and have consistently delivered revenue and capacity forecasts, as seen in the 850 MW delivered last quarter and the recurring revenue, profit, and EPS prints.
Q: How much AI units/usage-based contribution is assumed in guidance, and when can usage-based become large enough to accelerate app growth?
A: Double-digit growth spans a wide band, and outcomes-based pricing is just starting, with modest contribution in the current model despite real opportunity. The key is delivering predictable outcomes, and as a long-time operator and data custodian, we know client best practices and can replicate by industry. For now we stick to the double-digit range and will update after launches. Another key is about half of ERP clients remain on-prem, and as new models expose security gaps, on-prem becomes more vulnerable and harder to patch at scale, so cloud migration matters, with 50%–70% shorter implementation a long-awaited breakthrough.
Q: How do you model and manage delays from regulatory or other causes, and do delays trigger customer exits or repricing?
A: No one should assume exact delivery dates among partners, which is reflected in contracts. Clients understand the unprecedented scale, and large AI campuses are true partnerships without LDs, while costs ramp by hall or building and shift out if delayed. In our view, plus or minus one to two quarters is on time around infra delivery, and our financial goals already factor up to two-quarter delays for any large campus.
Q: In the illustration showing a 50% reduction in financing needs, supplier financing, BYOH, and prepayments each looked like a third. Is that roughly accurate?
A: That chart is illustrative and not a precise breakdown, as further detail would reveal too much about customers and suppliers. Each of the three contributes meaningfully to the ~50% reduction in capital intensity.
Q: The industry projects ~40 GW next year and ~60 GW in 2028. Can supply-demand remain balanced, and what does it mean for Oracle?
A: Distinguish marketing from reality, as many 'up to' claims are meaningless, such as emails touting 'up to 300 GW' on Iowa farmland. The industry faces hard constraints, with global memory and wafer output limiting deployable compute more than power, and without fiber and optics, prudent players will not overbuild MEP. Demand is seriously underestimated: AI coding went from concept to ubiquity within a year, with new sectors lighting up annually and rapid advances in math and science. We do not see GPU supply exceeding demand for at least 3–5 years.
Q: Unit MW/GW build costs rose with chips, components, turbines, and energy. How much future cost inflation is in targets, and how much CapEx is locked?
A: The next 18 months are the fastest growth phase and the most certain on pricing, with most POs placed and prices locked. Once FCF turns positive, strong revenue and cash flow can absorb 2029–2030 cost swings, which is why we are quantifying capital needs now rather than a year ago. We do not know chip prices five years out, but FCF then will be very different. The 1 GW example is updated: revenue per GW went from ~$1.0 bn last year to $1.6 bn, mainly reflecting memory price increases, and cost-plus keeps GPM and ROI intact even if revenue and CapEx move, which is what creates value. With ample demand and tech differentiation, we have pricing power. The 1 GW numbers will likely change in three years, but the business model strength holds.
Q: Beyond headline campuses, smaller sites contributed more than expected. Where does demand come from, will you expand globally, and is the big-small mix representative?
A: We wanted to show growth is not tied to a handful of frequently asked-about sites, and we favor diversity in site distribution and scale. GW campuses suit training, but inference is scaling and needs 50 MW, 100 MW, and 200 MW sites globally, partly for latency and depending on where power is available. Customer demand mixes training and inference, and we will keep serving both. Serving 1,000+ AI customers across foundation model players, trading analytics, and space startups, AI is becoming the new norm for general compute, and AI accelerators installed already exceed traditional compute, which itself is growing fast. Large campuses can be finely sliced into many small allocations, whereas dispersed small DCs cannot be aggregated like large campuses. At 10 GW, GW campuses will be about half or less of the total, and the many dispersed country sites may not make headlines but allow us to manage regulatory and supply risks and deliver predictably.
Q: With sovereign AI and diverse needs, how do you position combinable DC architectures?
A: Through Alloy, partners can operate regions with full control, all with the same architecture and GPU deployability, and scale only limited by customer or partner desire, such as 10k GPUs if they wish. We are investing heavily in the Middle East with many partners and likewise across Asia, and the product lets clients start small and scale, backed by 5–7 years of maturity. Sovereign demand will be significant but slower moving and, like small DCs, less headline-friendly. By 2028 we expect 200+ regions, uncommon among peers, and this footprint should pay off for years.
<End>
Risk disclosure and statement:Dolphin Research Disclaimer and General Disclosure