McKinsey 3 Horizons of Growth

Anchor the growth arithmetic before naming any horizon

Start with where the profit actually comes from, because the headline revenue mix is misleading. Retail is roughly 83 percent of sales but North America runs at only about a 6 percent operating margin and International at roughly 3 percent, so the two retail segments together contribute under half of group operating income. AWS is about 17 percent of revenue and roughly 58 percent of operating income. The immediate insight is that Amazon's profit engine today is a business that did not exist in 2005 and was not disclosed separately until 2015 — so any judgement about the next decade has to be a judgement about the pipeline, not about retail.

Map the actual businesses onto horizons by maturity, not by age

Horizon 1 is retail marketplace, Prime, first-party retail, third-party seller services, AWS, and now advertising — all profitable at scale. Horizon 2 is the set with real customers and real revenue but no settled economics for Amazon: healthcare (One Medical, acquired for roughly 3.9 billion dollars in 2023, plus Amazon Pharmacy), physical and online grocery including Whole Foods, Buy with Prime and logistics sold as a service to third parties, Prime Video advertising which only switched on in January 2024, and India and other emerging-market retail where demand is proven but profitability is not. Horizon 3 is Project Kuiper satellite broadband, Zoox robotaxis, delivery drones, and the newer agentic-AI layers above Bedrock. Note the discipline of classifying by maturity: Alexa is over a decade old but as a standalone business it never proved economics, so it sits in Horizon 2 at best, not Horizon 1.

Trace the migrations — this is the diagnostic that matters

AWS is the textbook full journey: an internal infrastructure experiment around 2003, launched 2006 with no proven external demand (Horizon 3), a fast-growing but strategically doubted business through roughly 2010 to 2014 (Horizon 2), and the group's profit engine by 2015 onward (Horizon 1). Advertising made the same trip faster, going from an option on marketplace traffic to roughly 56 billion dollars of high-margin revenue. Against that, the Fire Phone was killed in 2014 with a write-down of roughly 170 million dollars — a Horizon 3 bet closed at a cost that barely dented a quarter. A portfolio where options graduate on proven demand and failures die cheaply is the actual asset here; the specific bets are replaceable.

Horizon 1: what defending the core means in practice

Retail's job is not margin, it is traffic, data and cash conversion — negative working capital funds everything else, since Amazon collects from customers before it pays suppliers. The concrete Horizon 1 levers are regionalising the US fulfilment network to cut cost-to-serve per unit, growing third-party seller share of units (higher-margin fee revenue rather than inventory risk), and raising Prime attach and retention because Prime is what makes the advertising business monetisable. For AWS, Horizon 1 discipline means defending share against Azure and Google Cloud on price-performance while migrating customers up the stack into managed and AI services that carry better margin than raw compute.

Horizon 2: proven demand, unproven economics — and the structural test

Each Horizon 2 business should be judged on whether the demand question is settled and only the cost question is open. Healthcare passes that test — people obviously buy primary care and pharmacy — so the open question is unit cost per member and whether Prime distribution lowers customer acquisition cost below what a standalone clinic chain pays. Grocery is the hardest: demand is proven, but Amazon has run Fresh, Go, Amazon-branded stores and Whole Foods for years without a settled format, which is exactly what an unresolved Horizon 2 looks like. Prime Video advertising is the cleanest Horizon 2 bet because it inherits an installed base of over 200 million Prime members and an existing ad sales team, so it needs distribution scale-up rather than a new business model. The Horizon 2 discipline to insist on is separate P&Ls and adoption metrics — asking One Medical for group-level margin in year two is how incumbents strangle their own middle.

Horizon 3: where Amazon breaks the classical model

The framework says Horizon 3 should be many cheap, stage-gated options. Kuiper is not cheap: a constellation of over 3,000 satellites with publicly discussed commitments above 10 billion dollars, and no revenue until service launches. Zoox is a similar shape — acquired around 2020 for roughly 1.2 billion dollars and still burning without a commercial robotaxi at scale. That is the honest critique to raise: capital-intensive Horizon 3 bets cannot be killed cheaply, which removes the option-like quality that makes Horizon 3 rational, and it means the failure mode is sunk-cost persistence rather than a clean shutdown. The counter-argument is that both bets are defensive options on infrastructure Amazon would otherwise have to rent — connectivity and last-mile labour — which is a legitimate reason to overweight them, but it should be stated as a reason, not assumed.

Allocate and govern — and state the risk

Capex of roughly 83 billion dollars in 2024 rising toward 100 billion in 2025 is overwhelmingly Horizon 1 AWS capacity and AI infrastructure, which is the right call while cloud demand is real, but it means the classical 70/20/10 split does not describe this company at all — it is closer to a very heavy H1 with a small, concentrated H3. Governance is what makes it work: single-threaded owners with one accountable leader per bet, the written PR-FAQ that forces a demand hypothesis before spend, and an explicit stated willingness to be misunderstood for long periods, which is really just patient-capital governance for Horizon 2. The risk to flag for the fund: Horizon 1 profit is now concentrated in one business, AWS, so an AI-driven shift in where inference workloads run would hit the cash engine directly, and Horizon 2 is thin in profit terms — advertising already graduated and nothing behind it is close to a 10-billion-dollar profit pool.

Size the gap before touching the framework

Doubling PAT from 3,000 to 6,000 crore in six years is about 12.2 percent CAGR. If advances compound at a realistic 10 percent, the balance sheet reaches roughly 1.77 times today's size — but ROA will not hold at 1.2 percent, because CASA is falling, term deposits are repricing up, and the fee income that UPI destroyed is not coming back. Assume ROA drifts to about 1.05 percent: PAT lands near 3,000 × 1.77 × (1.05/1.20), roughly 4,650 crore. So the growth gap is around 1,350 crore of PAT, about 23 percent of the target — that is what Horizons 2 and 3 must deliver, and stating it first stops the answer becoming a list of digital buzzwords.

Map the bank's real portfolio and name the constraint the horizons sit inside

Horizon 1 is branch-led deposits, home loans, LAP, vehicle finance and MSME working capital — profitable and at scale. Horizon 2 candidates where demand is unquestionably proven but this bank has no scale: credit cards, unsecured personal and consumer-durable lending, gold loans, cash-flow-based MSME lending using GST and Account Aggregator data, and wealth and broking for affluent tier-2 customers. Horizon 3: embedded finance and Banking-as-a-Service APIs, ONDC-linked credit, AI-driven underwriting and collections, agri value-chain financing through FPO tie-ups, and a minority stake or two in fintechs. Crucially, a bank has a constraint a manufacturer does not: growth consumes regulatory capital. At 13 percent ROE with roughly a 20 percent payout, internal capital accrual is about 10.4 percent — barely enough to fund 10 percent RWA growth, so every rupee spent on H2 and H3 is competing with the capital the core needs to grow at all.

Diagnose the shape — this bank has the classic missing middle, plus a liability problem

The core is healthy on asset quality but is quietly weakening on the liability side: CASA at 38 percent and drifting down is the real disease, because it directly sets NIM and therefore ROA. Horizon 2 is close to empty — there is a credit card product, but a sub-scale one, and MSME lending is done the old way on collateral and balance sheets rather than on cash-flow data. Horizon 3 is scattered: a few fintech pilots run by the IT department with no stage gates and no hypothesis. The diagnosis to say out loud: this is a bank with an adequate Horizon 1, an empty Horizon 2 and an unmanaged Horizon 3, so the answer is at least as much about structure and governance as about which product to launch.

Horizon 1: defend the deposit franchise and take out cost, because it funds everything

Three concrete levers. First, CASA: deepen salary accounts and current accounts in the 1,200 tier-2 and tier-3 branches where the bank has genuine relationship density that a digital-only competitor cannot replicate — every 100 basis points of CASA mix is direct NIM. Second, cost-to-income from 52 percent to about 45 percent by shifting servicing to digital channels and converting branches from transaction counters to sales-and-advisory nodes, which on a 14,000-crore revenue base is roughly 1,000 crore of pre-tax benefit and is the single largest identifiable contributor to closing the gap. Third, rebuild fee income that UPI destroyed by distributing third-party insurance and mutual funds to an existing customer base — capital-light fee revenue, no RWA consumed. Flag the trap: cutting the branch network hard to hit cost-to-income would destroy the CASA moat, which is exactly the over-milking of Horizon 1 the framework warns about.

Horizon 2: build the missing middle with ring-fenced structure and capital-aware design

Prioritise by risk-adjusted yield and capital consumption, not headline yield. Gold loans are the best first bet — proven demand in exactly the client's tier-2 geography, high yield, short duration, fully secured and therefore capital-efficient, and buildable through the existing branch footprint. Cash-flow-based MSME lending using GST returns and Account Aggregator consent is the highest-value bet because it lets the bank underwrite the customer it already banks without demanding property collateral. Unsecured personal loans and cards must be built through co-lending and partnership rather than on balance sheet, because RBI's November 2023 increase in risk weights on unsecured consumer credit to 125 percent makes on-book growth capital-punitive. Structurally: one ring-fenced digital lending vertical with its own P&L, its own credit policy, market-rate hiring for data scientists, and metrics that are disbursements, activation rate and risk-adjusted yield (NIM minus expected credit cost) — not day-one ROA, which no lending business can show before its book seasons.

Horizon 3: cheap, stage-gated options with an explicit kill rule

Six to eight bets, roughly 150 to 200 crore total over six years, each capped until one hypothesis is proven. Candidates: a BaaS API stack letting fintechs originate on the bank's licence (hypothesis: partner-sourced deposits cost less than branch-sourced); ONDC-linked seller credit; an AI collections engine tested on one delinquent portfolio (hypothesis: recovery rate rises 200 basis points at lower cost-to-collect); agri value-chain financing through two FPO partnerships in one state; and a minority stake in a supply-chain finance fintech to buy visibility rather than control. Judge them on milestones and hypotheses killed, never on payback. The one bet worth over-funding is the AI underwriting and collections capability, because it improves credit cost across all three horizons rather than only its own P&L.

Allocate, govern, and set migration rules that are specific to lending

Roughly 75 percent of investable resources to Horizon 1 — mostly the regulatory capital needed to grow the core book plus the digitisation spend that delivers the cost-to-income improvement — about 20 percent to Horizon 2, and 5 percent to Horizon 3. That is deliberately more H1-heavy than the 70/20/10 benchmark, and you should justify it: a bank cannot grow its balance sheet without capital, so the benchmark from unregulated industries does not transfer. The MD personally chairs the digital lending review, because a new unsecured business will lose every internal fight to a secured-lending head who can point to two decades of clean asset quality. Then the migration rule that separates a good answer from a great one: a lending business must not graduate from Horizon 2 to Horizon 1 on growth alone, because credit losses arrive late — set the gate at a stated book size sustained through at least one full credit cycle with risk-adjusted ROA above the core's, otherwise you will scale a book that looks brilliant for three years and then breaks.