The Token Business: Turning India's Sovereign AI Factory into a Durable Revenue Engine

Reliance Intelligence is building one of the largest AI compute platforms in the world. The commercial mandate is to convert that capacity into contracted, metered, collected revenue faster than it depreciates.

The asset base is extraordinary: 120 MW of AI compute at Jamnagar targeted by end-2026 on NVIDIA GB300 systems (75,000+ H100-equivalents on inference, scaling past 200,000), renewable power from Kutch that structurally lowers cost per token, a nationwide Jio edge network, captive demand across the Reliance group, and a ₹10 trillion seven-year infrastructure commitment. Meta's 168 MW built-to-suit lease proves the anchor-tenant model works at Jamnagar.

GPU infrastructure is a depreciating asset with a 4 to 6 year useful life; every idle GPU-hour is perishable inventory, like an empty airline seat. The commercial strategy therefore optimizes one master metric — revenue-weighted utilization — and is built on five principles:

• Sell certainty first — Anchor the P&L in multi-year reserved and committed-capacity contracts (target 60 to 70% of capacity) before optimizing spot and on-demand token yield on the remainder.
• Price in tokens, cost in GPU-hours — Customers buy outcomes (tokens, fine-tunes, agent runs); the platform's cost basis is GPU-hours, power, and network. The pricing architecture must arbitrage that translation profitably, model by model.
• Captive demand is the launch customer, not the business plan — Jio, Retail, JioStar, and the five Jio AI services (JioBharatIQ, Vyapar, JioHealthIQ, JioLearnIQ, JioKrishiIQ) de-risk utilization in year one and serve as public reference architecture, but external enterprise, SaaS, and marketplace revenue is what makes the token business a standalone P&L.
• Billing is a product — A token business lives or dies on metering integrity, invoice trust, and revenue assurance. The OSS/BSS stack is treated as a first-class product with an SLA, not back-office plumbing.
• Sovereignty is the moat — Data residency under the DPDP Act, Indian-language models across 22 languages, and government-grade compliance are differentiators the hyperscalers cannot easily match at Indian price points on Indian soil.

India's AI compute market is supply-constrained and price-sensitive at the same time. The IndiaAI Mission has anchored subsidized GPU access for startups and academia; hyperscalers (AWS, Azure, GCP) serve the premium enterprise tier from limited Indian regions; domestic providers (Yotta, E2E, Tata Communications, CtrlS, Sify) compete on price and availability. Demand from Indian enterprises, GCCs, SaaS builders, and government programs is growing faster than any domestic supply response except Reliance's.

Where Reliance wins

• Scale and cost — gigawatt-trajectory capacity plus captive renewable power yields a structural cost-per-token advantage over both hyperscalers (imported margin stacks) and subscale domestic players.
• Distribution — Jio's enterprise relationships, telecom channel, and 500M+ consumer reach constitute a GTM asset no competitor holds. Every Jio enterprise account is a warm token-business lead.
• Sovereignty and language — sovereign hosting plus first-class Indic-language model serving addresses the two hardest requirements of Indian government, BFSI, and healthcare buyers.
• Edge — telecom edge inference across the Jio network enables latency-tiered products (real-time voice, vision, agentic workflows) that centralized clouds cannot price competitively in India.

Where discipline is required

• Price-sensitive demand — Indian on-demand GPU pricing runs well below US list; chasing spot volume at Indian street prices without committed anchors erodes margin. The answer is contract mix, not price leadership everywhere.
• Hyperscaler counterattack — expect aggressive India-region expansion and credits. Defense: sovereignty, edge latency, rupee-denominated committed pricing, and group distribution.
• Capacity credibility — selling ahead of commissioned megawatts destroys trust. Commercial commitments must be gated to the delivered-capacity curve agreed with Infrastructure leadership.

Seven launch SKUs, sequenced in three waves against the Jamnagar commissioning curve. Wave 1 maximizes certainty and utilization; Wave 2 builds margin-rich differentiated products; Wave 3 builds the ecosystem flywheel.

Sequencing logic: reserved capacity and token serving can be sold today against the end-2026 commissioning date (contracts signed now, revenue recognized at delivery). Fine-tuning, private endpoints, and edge follow within two quarters of first capacity. Agent runtime and marketplace launch once the platform has reference customers and settlement rails proven at scale.

Four commercial models, one discipline: every price traces to cost per token, which traces to GPU-hour cost, power, utilization, and model efficiency. The pricing council reviews monthly as GB300 performance data and market rates move.

• Packaging — three enterprise tiers (Launch, Scale, Sovereign) bundling tokens, fine-tuning credits, endpoint throughput, and support SLAs; a self-serve developer tier with transparent card-billed pricing to seed the funnel.
• Price governance — a deal desk owns discount authority above thresholds; every non-standard deal carries a margin floor at cost per token plus target return on the underlying reserved capacity.

Segment 1: Captive Reliance businesses (utilization floor)

Jio network AI, JioBrain workloads, JioStar GenAI Media Studio, Retail demand forecasting, and the five Jio AI services migrate onto internal committed-capacity contracts at arm's-length transfer pricing. Target: 25 to 35% of Wave-1 capacity under internal MOUs before external launch, priced to recover full cost plus a modest internal margin so external pricing integrity is never undermined.

Segment 2: Anchor external tenants (the Meta motion, repeated)

Five to eight named-account pursuits for multi-year reserved capacity: global hyperscalers needing India capacity, frontier labs seeking inference distribution in India, large GCCs (banks, retailers, pharma) with sovereign-AI mandates, and Indian AI startups with national-scale ambitions. This is a CEO-to-CEO sales motion supported by the CCO with bespoke structuring: take-or-pay with utilization credits, expansion options priced as capacity calls, and co-marketing rights.

Segment 3: Enterprise and BFSI (the margin engine)

Top 200 Indian enterprises and GCCs, led by regulated industries where sovereignty commands premium pricing. Land with a private endpoint or fine-tuning pilot scoped to a 30 to 45 day decision, expand into committed-spend contracts. Field organization of enterprise AEs paired with solution architects, comped on committed ARR and net revenue retention, not bookings.

Segment 4: SaaS, startups, and developers (the volume flywheel)

Self-serve token tier with transparent pricing, generous free tier for Indic-language models, IndiaAI Mission alignment for subsidized startup access, and a partner program for SaaS platforms embedding the API. This funnel feeds the marketplace and generates the usage data that sharpens cost-per-token management.

Segment 5: Telecom edge and government

Edge inference sold through Jio's existing enterprise telecom channel as an attach to connectivity contracts; government and PSU demand pursued through sovereign-cloud framework agreements where residency, auditability, and Indic-language capability are the spec, and price competition is secondary.

• Listing economics — revenue share tiered by exclusivity and integration depth (approximately 80/20 for standard listings, improving toward 70/30 platform share for premium placement, fine-tuned variants, and bundled distribution through Jio channels).
• Settlement mechanics — monthly settlement to sellers with full metering transparency; the platform's billing system is the single source of truth for both sides, with seller-facing dashboards mirroring buyer-facing invoices.
• Telecom partners — edge-inference revenue shared with carrier partners where their spectrum and sites participate in delivery; modeled on interconnect settlement, a discipline the Jio organization already runs at scale.
• Curation — a quality-gated catalog, not an open bazaar: every listed model carries benchmark cards, safety attestations, and Indian-language capability scores, protecting enterprise trust in the platform.

This is the discipline most token businesses get wrong, and the one my CRO background treats as home ground. Revenue leakage in usage-based businesses typically runs 1 to 3% of revenue; at $100M+ ARR that is a leadership-team-sized sum lost annually. The cockpit's revenue-assurance view is how this gets run week to week. The plan:

• Metering integrity — token and GPU-hour metering instrumented at the serving layer with immutable event logs; daily automated reconciliation between metering, rating, and invoicing, with variance thresholds that page the FinOps team — the same control philosophy as trade-settlement breaks on a derivatives desk.
• Order-to-cash ownership — one pipeline from onboarding, KYC, and credit assessment through rating, invoicing, collections, and dunning; enterprise net-terms risk managed with credit limits and prepay conversion triggers.
• Revenue assurance function — a dedicated team inside the CCO org running leakage analytics: unbilled-usage detection, rate-plan drift, discount-policy compliance, marketplace settlement audit, and fraud/abuse monitoring (prompt-loop abuse, token-farming, resale violations).
• FinOps for customers — customer-facing cost dashboards, budgets, and anomaly alerts. Counterintuitive but essential: helping customers avoid bill shock is the single strongest driver of renewal and expansion trust in consumption businesses.
• Build/buy posture — rate and mediate on a proven usage-billing engine rather than building from scratch; differentiate in the metering layer and the reconciliation controls, and keep the system auditable to Big-4 standard from day one, because anchor-tenant contracts will demand it.

• Master metric — revenue-weighted utilization of commissioned capacity, reviewed weekly with Platform and Infrastructure leadership; target trajectory from 50% at commissioning to 80%+ within four quarters of each tranche.
• Cost per token — managed as a living model per model-class: GPU-hour cost (depreciation, power at Kutch solar rates, network, facilities) divided by tokens per GPU-hour (batching, quantization, GB300 inference efficiency). Every serving-stack optimization the platform team ships is commercial margin; the CCO org funds and prioritizes that roadmap jointly.
• Load shaping — off-peak and batch pricing moves elastic workloads (training, batch inference, embeddings) into utilization valleys; edge pricing captures premium for latency instead of leaving it on the table.
• Capacity gating — a single capacity ledger, jointly owned with Infrastructure, gates every committed-capacity contract against the commissioning curve plus a reserve buffer, so the sales organization can sell aggressively without ever selling capacity that does not exist.

Five pillars under the CCO, each with a named leader by day 90 and succession depth by month 12:

Culture: a build-from-scratch commercial organization with desk-level P&L discipline — weekly revenue-weighted-utilization reviews, monthly pricing council, quarterly renewal and margin reviews with Finance, and playbooks written down from the first deal onward so the organization scales on documented practice rather than heroics.

Governance cadence: weekly utilization and pipeline review; monthly pricing council and revenue-assurance report; quarterly board pack covering contracted coverage, margin trajectory, and renewal health — the same rhythm I ran as a desk head and CRO, applied to a token P&L.

Every element of this plan is drawn from businesses I have already run: consumption pricing and committed-capacity contracts from a $250M+ ARR GPU and foundation-model business at AWS; client-franchise building and multi-million dollar desk P&L ownership from 11 years at Morgan Stanley; settlement, controls, and revenue assurance from two CRO seats and a derivatives career; build-from-scratch commercial construction from founding ventures and scaling functions in an emerging market. The token business at Reliance Intelligence is the convergence of those disciplines at national scale, and I would be honored to build it.

Modeled estimates built from public market rates and standard cloud unit economics as of mid-2026, not Reliance internals. These are opening hypotheses for diligence, stated with their assumptions so each one can be pressure-tested and replaced with actuals in the first 30 days.

Cost basis assumptions

• All-in GPU-hour cost — $1.40 to $1.80 per H100-equivalent GPU-hour, comprising GB300 capex amortized over 5 years, power at Kutch solar transfer rates, facility (PUE assumption 1.2 to 1.3), network, and operations. Renewable power is the structural advantage: power is typically 15 to 25% of all-in cost, and captive solar compresses it.
• Utilization — margins quoted at 70% revenue-weighted utilization for always-on SKUs; reserved capacity is utilization-independent by construction (take-or-pay).
• Pricing references — public India and global rates for on-demand GPU-hours, per-token serving, and sovereign-cloud premiums observed across hyperscaler India regions and domestic providers.

ROI by SKU

Portfolio read

Reserved capacity carries the worst margin but buys the certainty that pays for the asset. Private endpoints are the best risk-adjusted return and the SKU where India-specific advantage (sovereignty, DPDP compliance, Indic-language serving) is hardest to compete away. Token-as-a-service shows good headline margin but is most exposed to commoditization. Agent runtime and marketplace are inexpensive options on the future, not year-one P&L.

Claims to validate in the first 30 days

• Cost floor — the $1.40 to $1.80 all-in GPU-hour figure depends on actual PUE, solar transfer pricing, and GB300 delivered performance; validate against Infrastructure's cost model.
• Sovereignty premium — the 30 to 50% private-endpoint premium should be benchmarked against what Yotta, hyperscaler India regions, and domestic providers actually charge BFSI and government buyers.
• Serving yield — tokens per GPU-hour is a function of batching, quantization, and model mix; it can swing token-SKU margin 20 points in either direction and is the single highest-leverage number in the model.