Reliance Industries LimitedReliance Intelligence · AI CloudConfidential — internal
GPU Cloud Billing Cockpit
User guide · full walkthrough of every page and tab

User guide

A walkthrough of the GPU Cloud Billing Cockpit

Every page and all seventeen cockpit tabs, in order: what the screen is, what is on it, how to read it, the formulas behind the numbers, and the decisions it is built to support. It ends with the working routines — the monthly close review, the pricing review, and a routing table from question to tab.

1. Orientation — how the app is put together

The app has six top-level pages. One of them, the Cockpit, holds seventeen tabs. Everything else is either a source document rendered as a web page, or a download. All numbers come from a single deterministic dataset, so a figure quoted on one screen is the same figure everywhere else.

The map

Six pages, one of which contains the whole operating picture

The Cockpit (/) is the operating surface: seventeen tabs that between them cover the full commercial mandate — P&L ownership, metering and billing, pricing architecture, revenue assurance, go-to-market, marketplace, FinOps, org and the fit case.

The P&L primer, Strategy & GTM, Compute futures and Platform spec pages are long-form documents. They carry the reasoning and the worked arithmetic that the cockpit tabs compress into tiles and charts. Documents is the download centre for the PDF and Word originals plus the generated monthly CEO report.

Two conventions hold everywhere. First, every tab opens with a Key insight card: the headline claim, the number carrying it, and why the tab exists. Second, every tile and chart title carries an ⓘ tooltip whose text is the same string shown in the Glossary tab — the two cannot disagree.

What you see

Header nav
Cockpit · P&L primer · Strategy & GTM · Compute futures · Platform spec · Documents, plus the theme toggle (auto / light / dark).
Tab strip
Seventeen cockpit tabs, ordered from the mandate index outward: scorecard, financials, operations, commercial, economics, org, then reference.
Key insight card
Top of every tab. Headline, the metric, and the decision the tab supports.
ⓘ tooltips
On tiles, chart titles and table headers. Definition, formula and any caveat.
Assistant
Floating panel, grounded in the primer, strategy, futures, levers paper, spec bundle and every cockpit dataset. Ask it to compute; it shows the arithmetic.

How to read it

  • Read top to bottom in a tab: insight card → KPI tiles (the state) → charts (the trend or distribution) → tables (the exceptions to act on).
  • Chart/table toggles sit on most cards. The table view is the printable, quotable form; the chart view is for spotting shape.
  • Deltas under a tile compare the closed month to the prior month unless the tile says TTM.

Decisions it supports

  • Which tab to open for a given question — the Mandate scorecard tab is the index and links straight through.

The data model behind every number

One deterministic dataset, twelve closed months

The app models a GPU fleet sold two ways on the same silicon: reserved capacity under take-or-pay contracts, and tokens served from multiplexed inference endpoints. Twelve closed months of synthetic-but-consistent data drive every tile.

The economic spine is: $2.05/GPU-hr realized on committed capacity, roughly $0.485/GPU-hr cash operating cost, and about $1.01/GPU-hr of depreciation on a five-year straight-line schedule for GB300-class hardware. Everything else — margins, paybacks, crossovers — is derived from those three numbers plus utilization.

What you see

Capacity ledger
The classification layer: every installed GPU-hour lands in exactly one of six classes, and they sum to installed hours with no residual.
Meter-to-cash chain
metered → rated → invoiced → collected. Each link has a yield, and the product of the yields is monetisation.
Price book
List, floor and realized price per SKU, with the discount waterfall between them.
Cost stack
Power, cooling, network, storage, staff and platform, expressed per GPU-hour and per million tokens.

How to read it

  • Cash margin excludes depreciation; EBITDA-after-depreciation is shown separately wherever the distinction changes the decision. On a fleet this capital-intensive it usually does.
  • Utilization means sold hours ÷ available hours. Density (token tab) means the fraction of an endpoint's peak throughput actually in use — a different quantity, and the two are never mixed.

Formulas

Gross margin(realized revenue − cash cost to serve) ÷ realized revenue
EBITDArevenue − cash opex (depreciation excluded)
Paybackcluster capex ÷ annual EBITDA
Monetisationcollected ÷ metered value = yield(rating) × yield(invoicing) × yield(collection)

Decisions it supports

  • Whether a margin claim is cash or post-depreciation before it is quoted externally.

Back to top ↑

2. The Cockpit — tab by tab

Seventeen tabs. Each entry below states what the tab is, every element on it, how to read it, the formulas involved and the decisions it supports.

2.1 Mandate scorecard

The index: every line of the commercial mandate mapped to an owning metric

This tab turns the Chief Commercial Officer mandate into a scorecard. Each mandate line — own the end-to-end P&L, set pricing architecture, own billing and metering, build enterprise and partner GTM, establish the marketplace model, own cost and token economics, partner with the platform teams, build the leadership team — is given exactly one owning metric, a target, a current value, a trend and the tab that evidences it.

It exists so nothing in the remit is asserted without a number behind it, and so a reviewer can traverse from a claim to the evidence in one click.

What you see

Coverage tiles
How many mandate lines are on target, at risk, or off target this month.
Mandate table
Mandate line · owning metric · target · current · trend · evidencing tab (clickable).
Status chips
Green on target, amber watch, red off target — thresholds are the target column, not judgement.

How to read it

  • Amber means inside 10% of target on the wrong side; red is beyond that. A line with no metric would be a gap in the design, so there are none.
  • The trend arrow compares to the three-month average, not last month, to suppress single-month noise.

Decisions it supports

  • Where to spend the next management cycle: pick the red line, click through, act on the exception table there.
  • What to put on a board slide: this table is the one-page answer to 'is the commercial function delivering'.

2.2 P&L cockpit

Margin is made in the cost stack, not the rate card

The end-to-end profit and loss for the capacity and token businesses in one place: revenue by product line, the margin bridge from gross revenue to EBITDA, fleet and token utilization, and the depreciation drag that separates a healthy cash margin from a thin accounting one.

What you see

Revenue (closed month)
Total recognised revenue with month-on-month delta and a twelve-month sparkline.
EBITDA margin
Cash margin after operating cost, before depreciation and interest.
Achieved $/GPU-hr (blended)
All revenue ÷ all sold GPU-hours, including token revenue converted to a GPU-hour equivalent. This is the single most useful pricing number in the app.
Fleet utilization
Sold hours ÷ available hours across the fleet.
Pre-tax margin
After depreciation and interest — the number that services debt.
Revenue by product line
Stacked series across reserved capacity, tokens, endpoints, fine-tuning and marketplace.
Margin bridge waterfall
Revenue → cost to serve → gross margin → opex → EBITDA → depreciation → pre-tax.
Fleet & token utilization chart
The two utilization series plotted together; divergence is the story.

How to read it

  • Compare achieved $/GPU-hr against $2.05. Above it, token multiplexing or premium endpoints are carrying the mix; below it, discounting or idle capacity is.
  • If EBITDA margin is healthy but pre-tax margin is thin, the problem is capital intensity, not commercial execution — that pushes you to the Compute futures tab, not to pricing.
  • A widening gap between fleet and token utilization means silicon is being held for capacity buyers who are not consuming; check the take-or-pay drawdown on Customers & contracts.

Formulas

Achieved $/GPU-hrtotal revenue ÷ (capacity GPU-hours + token GPU-hour equivalents)
Cost to servecash opex per GPU-hr ≈ $0.485 (power, cooling, network, storage, staff, platform)
Depreciation≈ $1.01/GPU-hr at 5-year straight line, GB300-class

Decisions it supports

  • Whether this month's margin move came from price, mix, utilization or cost — the bridge answers it directly.
  • Whether to escalate a cost line to FinOps (the FinOps & capacity tab holds the joint sign-off).

2.3 Meter to cash

Metered is not billed, and billed is not collected

The order-to-cash funnel, from raw meter events through rating, invoicing and collection, with the drop at each step quantified. It exists because unbilled consumption is pure margin loss that never appears in a revenue chart — it is simply absent.

What you see

Meter-to-cash yield
Collected value ÷ metered value for the closed month. The headline efficiency of the whole billing chain.
Rated ÷ metered
The rating stage yield. Anything under 100% is either legitimate suppression or a defect.
Unbilled usage
Rated value sitting in no invoice, in dollars, with aging.
Collected ÷ invoiced
The cash conversion stage.
Funnel chart
Each stage as a bar, with the absolute and percentage drop between stages.
Yield trend
Twelve-month series for the composite yield.
Unbilled aging table
By customer and age bucket (0–30, 31–60, 61–90, 90+ days).

How to read it

  • Read the funnel right to left when diagnosing: a collection problem is a credit problem, an invoicing problem is a process problem, a rating problem is a platform defect.
  • Unbilled usage older than 60 days should be treated as at risk of becoming unbillable — many contracts have billing windows.

Formulas

Composite yieldrated/metered × invoiced/rated × collected/invoiced
Unbilled agingrated_line_item value with no invoice_id, bucketed by rating date

Decisions it supports

  • Which stage owns this month's leak, and therefore which team gets the action.
  • Which customers to chase before the billing window closes.

2.4 Revenue assurance

Leakage is recoverable; perished capacity is not

The control layer. Recovering one to three percent of revenue leakage carries no incremental cost of goods, which makes these controls a direct, high-margin lever on EBITDA — arguably the highest-return work in the commercial function.

The tab is built around the daily automated three-way match (control C-01), a set of BLOCKING controls that halt billing closes, code releases or general-ledger postings when they fail, and a set of operational controls that close the softer leakage channels.

What you see

Leakage rate (TTM)
Leaked value ÷ rated revenue over twelve months. Target under 0.5%.
Billing accuracy
Share of rated lines that survive the match without adjustment. Target above 99.5%.
Recovery rate
Recovered ÷ identified leakage for the month. Target above 70%.
SLA credit ratio
Service-credit value as a share of invoiced revenue.
DSO — enterprise
Days sales outstanding on the enterprise book.
Leakage by mode
Where value is lost: mis-rating, missing meters, un-invoiced usage, credit notes, SLA credits.
Identified vs recovered
Twelve-month series; the gap is the write-off.
Three-way match panel
The C-01 backbone: the three legs matched, the tolerance on each, and the current state.
BLOCKING controls table
Control · what it blocks · mechanism · data source · formula/test · state.
Operational leakage controls
C-03, C-05, C-06 and siblings, with the no-incremental-COGS note.
RA KPI targets
Completeness >99.8%, billing accuracy >99.5%, recovery >70%, dispute rate <1%.

How to read it

  • A BLOCKING control in a failed state means the close is halted by design — it is not a warning, it is a stop. Treat any failed blocking control as the top item of the day.
  • The Data source and Formula columns are there so the control can be re-run by hand against the warehouse; a control you cannot recompute is a control you cannot audit.
  • Identified-but-not-recovered leakage compounds: the older it is, the lower the recovery probability, which is why the recovery rate is measured monthly rather than cumulatively.

Formulas

Three-way matchmeter events ↔ rated line items ↔ invoiced lines, matched on usage window and resource ID, within tolerance
Leakage rateleaked value ÷ rated revenue
Recovery raterecovered value ÷ identified leakage, same period
Completenessmeter events landed ÷ meter events expected from fleet inventory

Decisions it supports

  • Whether the month can close (blocking controls) and whether the GL posting is safe.
  • Which leakage mode to fund an engineering fix for — rank by TTM value, not by count of incidents.
  • Whether SLA credits are a service problem or a commercial-terms problem; the incident-ID linkage settles it.

2.5 Customers & contracts

Renewals are won nine months before they expire

Concentration, contract value, take-or-pay drawdown and expiry timing in one view, so the desk opens renegotiation while it still has leverage rather than discovering a lapsed anchor contract at quarter end.

What you see

TTM revenue
Twelve-month revenue across the customer book.
Top-1 and top-3 concentration
Share of revenue from the largest and three largest accounts.
Backlog
Contracted-but-unconsumed value remaining.
ACV up for renewal (12 mo)
Annual contract value with an expiry inside the next twelve months.
Revenue by customer
Ranked bar chart.
Take-or-pay drawdown
Contracted hours versus consumed hours per customer — the unconsumed portion is billed regardless, and is a renewal risk.
Contracted backlog
Runoff of remaining contract value by month.
Renewal timeline
Contracts bucketed by expiry window with status chips (open, in negotiation, at risk, closed).
Customer economics table
Per customer: revenue, achieved rate, drawdown, expiry, action.

How to read it

  • Low drawdown on a large contract is a warning, not a win: the customer is paying for hours it is not using and will negotiate hard at renewal. Start those conversations first.
  • Top-1 concentration above roughly a third makes the whole P&L a single-counterparty story; that changes the risk framing on the Compute futures tab.
  • The nine-month rule: any contract inside the nine-month window with no negotiation status is an exception.

Formulas

Drawdownconsumed hours ÷ contracted hours
ACV up for renewalsum of annualised contract value where expiry ≤ 12 months
Concentrationcustomer revenue ÷ total revenue, TTM

Decisions it supports

  • Renewal sequencing and the opening position for each negotiation.
  • Where concentration risk needs a mitigating hedge or a new anchor.

2.6 Capacity ledger

Every installed GPU-hour is classified exactly once

The reconciliation from installed hours to billed hours with no unexplained residual. It exists because the difference between installed and sold is the largest source of silent margin loss on a depreciating fleet — the meter runs whether or not anyone is paying.

Six mutually exclusive, collectively exhaustive classes: pre-acceptance, faulted, productive, contracted headroom, leaked and perished.

What you see

Monetisation
Productive hours ÷ installed hours.
Availability
1 − (pre-acceptance + faulted) ÷ installed.
Leaked — recoverable
Consumed but never billed. Recoverable, and it flows to Revenue assurance.
Perished — foregone
Idle and unsold. Not recoverable, ever. This is the number the futures hedge exists to address.
Ledger waterfall
Installed hours stepped down through each class to billed hours.
Class table
Hours, share of installed, dollar value at reference rate, and owner for each class.
Invariant check
The six classes sum to installed hours — displayed as a pass/fail, and unit-tested in the codebase.

How to read it

  • Leaked and perished look similar in a chart and are opposite in economics. Leaked hours are a billing defect worth chasing; perished hours are gone and can only be prevented next month.
  • Contracted headroom is not waste: it is capacity a take-or-pay customer has already paid for and chose not to use. It should not be resold without an overbooking policy.
  • If the invariant fails, no other number on the tab is trustworthy.

Formulas

Invariantpre-acceptance + faulted + productive + headroom + leaked + perished = installed
Monetisationproductive ÷ installed
Perished valueperished hours × reference rate ($2.05/GPU-hr)

Decisions it supports

  • Whether the fleet problem is technical (faulted, pre-acceptance), commercial (perished) or systems (leaked).
  • How much residual capacity the RL simulation should be allowed to price.

2.7 GTM & SKU mix

Sell certainty first, then monetize the remainder

Revenue and margin by SKU across the seven-segment portfolio — reserved capacity, token-as-a-service, private endpoints, fine-tuning, telecom edge inference, agent runtime, marketplace revenue share — sequenced into three waves, plus the granular initiative set for each segment.

What you see

Committed share of revenue
Take-or-pay revenue ÷ total. The target band is 60–75% before any merchant hour is priced.
Wave-1 SKU revenue
Revenue from the first-wave SKUs against the seven-SKU total.
Blended gross margin
Margin across the mix — the number that mix decisions move.
Take-or-pay drawdown
Fleet-level consumption against commitment.
NRR / marketplace GMV
Net revenue retention and marketplace gross merchandise value.
Revenue by SKU
Ranked bars with the wave each SKU belongs to.
Segment coverage
Contracted versus consumed by customer segment: captive, anchor labs, enterprise/BFSI, developers, government.
Strategic initiatives
Per segment: trends, use cases, sales motions, named-type target customers, partners and the ROI posture.
SKU ROI table
Gross margin band, capital intensity, payback and the primary risk per SKU.

How to read it

  • Margin and capital intensity move in opposite directions across the portfolio: reserved capacity is 30–45% margin and very capital-hungry; agent runtime is 70%+ at almost no incremental capital but with speculative demand. A portfolio needs both ends.
  • Committed share below 60% means the depreciation schedule is unfunded; above 75% means you are giving away the upside of a tight market.
  • Coverage gaps by segment are the actual pipeline targets — the chart names them.

Formulas

Committed sharetake-or-pay revenue ÷ total revenue
Blended marginΣ(SKU revenue × SKU margin) ÷ Σ(SKU revenue)
NRR(starting ARR + expansion − contraction − churn) ÷ starting ARR

Decisions it supports

  • Which wave-2 SKU to fund next, given margin, capital intensity and payback.
  • Which segments to over-index the sales motion against this quarter.

2.8 Pricing & packaging

The discount waterfall, not the price book, sets realized price

Price architecture by SKU and packaging model — reserved take-or-pay, per GPU-hour, per million tokens, per endpoint, per training job, outcome-based per resolved task, revenue share — with list, floor and realized prices, the list-to-realized waterfall, realization by segment, and the governance log of every price change.

What you see

Realization tiles
List value, contracted value, realized value and the erosion between them.
Discount waterfall
List → volume discount → commitment discount → non-standard concession → realized.
Price realization by segment
Realized ÷ list per segment; the outlier is where governance failed.
Price book table
SKU, packaging model, unit, list, floor, realized, margin at realized.
Governance log
Every price change: date, SKU, old, new, approver, rationale.
Outcome vs metered crossover
Where outcome-based pricing beats metered pricing for the same workload.

How to read it

  • Non-standard concessions are the line to watch: the design target is under 3% of list. Volume and commitment discounts are earned; non-standard concessions are conceded.
  • A floor price breach in the price book table is a hard exception — the floor is set at cost-to-serve plus the minimum acceptable contribution.
  • Segment realization below the fleet average with no strategic rationale is a discipline problem, not a market problem.

Formulas

Realized priceinvoiced value ÷ billed units
Realization raterealized ÷ list
Floorcost to serve + minimum contribution margin

Decisions it supports

  • Whether to tighten the approval threshold on non-standard concessions.
  • Which SKUs should move from metered to outcome-based pricing.

2.9 Billing platform (OSS/BSS)

Billing is an SLA-bound system, not a back office

The health of the order-to-cash platform: onboarding, entitlement, metering, rating, invoicing, collections and settlement, each with a service-level objective, plus disputes by root cause and credit-note volume. A metering or rating defect is a revenue defect, and it surfaces here days before it reaches the P&L.

What you see

Stage health
Each order-to-cash stage against its SLA, with current state.
Platform SLOs
Ingestion lag, rating accuracy, invoice-on-time rate, DSO, dispute rate, unbilled aging.
Disputes by root cause
Ranked: metering gap, rate-card mismatch, contract terms, service credit, customer error.
Credit notes
Volume and value, trended.

How to read it

  • Ingestion lag is the leading indicator of everything else: if events land late, rating is late, invoicing slips, and DSO follows about six weeks later.
  • Dispute root causes that are internal (metering gap, rate-card mismatch) are engineering backlog items; external ones are a collections conversation.

Formulas

Invoice on timeinvoices issued by contractual date ÷ invoices due
Rating accuracyrated lines matching recomputation ÷ rated lines sampled
Dispute ratedisputed invoice value ÷ invoiced value

Decisions it supports

  • What goes on the platform engineering roadmap, ranked by revenue exposure rather than ticket count.

2.10 Marketplace & partners

Take rate only counts after partner settlement

Two-sided marketplace economics: gross merchandise value, take rate, net revenue, partner tiers and revenue-share terms, the settlement flow and its aging, and the cold-start and attach metrics that say whether the marketplace is actually two-sided yet.

What you see

GMV
Gross merchandise value transacted through the marketplace.
Take rate
Platform share of GMV, by tier.
Net revenue
What is recognised after partner settlement — the only marketplace number that belongs in the P&L.
Partner tiers
Tier, rev-share terms, partner count and GMV contribution.
Settlement flow and aging
Amounts owed to partners, by age.
Attach and cold-start metrics
Attach rate to core compute, supply-side and demand-side growth.

How to read it

  • Quote net revenue, never GMV, in any P&L context. GMV is a scale metric for the marketplace team only.
  • Settlement aging that stretches is a partner-retention risk long before it is a cash benefit.
  • Cold start is only broken when both sides grow in the same period; one-sided growth reverts.

Formulas

Net revenueGMV × take rate − settlement adjustments
Attach ratecustomers using ≥1 marketplace SKU ÷ core compute customers

Decisions it supports

  • Whether to invest in supply or demand next, and whether a tier's take rate is defensible.

2.11 FinOps & capacity

Cost per GPU-hour and per million tokens is the shared scoreboard

The cost-to-serve stack per GPU-hour and per million tokens, unit-cost efficiency against batching density, and commitment-versus-available capacity by region. This is the joint sign-off surface with AI Cloud Platform, Model Serving, Compute & Hardware, Data Center Infrastructure and Finance.

What you see

Cash cost stack
Power, cooling, network, storage, staff, platform — per GPU-hour, summing to about $0.485.
Cost per million tokens
The same stack expressed on the token unit, which moves with density.
Unit cost vs batching density
Cost per million tokens falls as density rises; the curve is the argument for multiplexing.
Capacity by region
Committed versus available, so commitments are never sold against capacity that does not exist.
Joint sign-off panel
Which partner function owns which cost line and what they have signed.

How to read it

  • Power and cooling dominate the cash stack; platform and staff are the lines commercial decisions actually move.
  • A region where committed exceeds available is an oversell — that is a stop, not a stretch target.
  • Unit cost per token is only comparable across months at equal density; the chart is there so the comparison is made correctly.

Formulas

Cost per GPU-hrΣ cash cost lines ÷ available GPU-hours
Cost per M tokenscost per GPU-hr ÷ (tokens per GPU-hr at density d) × 1e6

Decisions it supports

  • Where the next capacity commitment can safely be sold.
  • Which cost line to attack for the largest unit-economics gain.

2.12 Token economics

Above 21.3% traffic density, tokens beat renting the node

The multiplexing arbitrage made explicit. A node sold as reserved capacity earns the contracted rate no matter what runs on it. The same node serving tokens earns per million tokens, so its revenue scales with how densely the endpoint is loaded. Below a threshold density, renting wins; above it, tokens win by a wide margin.

With the worked node economics used here, the crossover sits at roughly 21.3% density: $16.80 of node cost against $78.84 of peak node revenue. At 60% density the same silicon earns about $5.91 per GPU-hour at roughly 76% gross margin, against roughly 33% sold as reserved capacity.

What you see

Scenario sliders
Token price per million, node cost, peak throughput and density — all tiles recompute live.
Crossover chart
Token gross margin and effective $/GPU-hr against density, with the reserved-capacity line drawn across it. The intersection is the crossover.
Density ladder table
Density step by step with revenue, cost, effective $/GPU-hr, margin and the decision (rent / serve tokens).
Margin tiles
Effective rate and margin at the current slider position versus the reserved benchmark.

How to read it

  • Margin rises from roughly 29% at 20% density to above 75% at 60%+ density. The curve is steep in the middle, which means small changes in routed traffic have large margin consequences.
  • The crossover is a floor for a decision, not a target. Serving tokens at 25% density technically beats renting, but with no safety margin against a demand dip.
  • Density is not utilization. A fully utilized node can be running at low density if requests are small and poorly batched.

Formulas

Crossover densitynode cash cost ÷ node revenue at peak density ≈ 16.80 ÷ 78.84 ≈ 21.3%
Effective $/GPU-hr(tokens served per hour ÷ 1e6) × price per million tokens ÷ GPUs in node
Token gross margin1 − (node cost ÷ node revenue at density d)

Decisions it supports

  • Which slice of the fleet should serve tokens at all, and at what minimum routed density.
  • The floor price per million tokens that still clears the reserved-capacity opportunity cost.

2.13 Compute futures

Hedging converts a perishable asset into bankable cash flow

GPU capacity is perishable: an unsold hour is destroyed revenue, and the hardware depreciates over four to six years regardless. This tab models a 1,024-GPU cluster (scalable with the slider) across the $2.10 → $1.70 per GPU-hour price path, unhedged versus hedged with a futures overlay, and shows the effect on EBITDA and debt payback.

What you see

Scenario sliders
Utilization, cluster size, price-path start and end, hedge ratio and futures strike.
Cluster capex
Capital at the chosen cluster size, GB300-class, five-year depreciation.
Sold GPU-hours per year
Derived from cluster size and utilization.
EBITDA at path start and end
The unhedged sensitivity, with the hedged figure alongside.
Payback at path end
Years to recover capex, unhedged and hedged, with the spread.
Hedge overlay tile
Ratio of sold hours hedged and the strike.
Sensitivity chart
EBITDA and payback across the price path, unhedged versus hedged.

How to read it

  • Shorting at $2.10 into a market that clears at $1.70 returns about $0.40 per hedged GPU-hour, which is what compresses the payback spread.
  • The hedge does not raise the ceiling; it raises the floor. In a rising market the hedged line underperforms — that is the premium being paid for financing certainty.
  • Payback beyond the depreciation life is the failure condition. Watch the point on the slider where that happens; that is the real risk boundary, and it is usually a utilization number, not a price number.

Formulas

Annual revenueGPUs × 8,760 × utilization × price
EBITDArevenue − (sold hours × $0.485)
Hedge P&Lhedged hours × (strike − spot)
Paybackcapex ÷ EBITDA

Decisions it supports

  • What hedge ratio to run against the merchant book, and at what strike.
  • Whether a financing structure survives the downside price path.

2.14 RL simulation

Pricing residual capacity is sequential, not a rate card

An agent-based environment for the non-committed slice of the fleet. Four customer cohorts each have their own elasticity, willingness to pay and patience; demand that is priced out defers and comes back, so today's discount destroys tomorrow's price. Reward is revenue-weighted utilization net of cash operating cost and a perishment penalty. Take-or-pay hours are a hard constraint served before the agent acts.

Four policies are compared: the static price book, greedy fill, a threshold heuristic, and an epsilon-greedy contextual bandit trained in the browser.

What you see

Environment controls
Cohort mix, elasticity, patience, demand volatility, episode length, exploration rate.
Policy comparison
Revenue-weighted utilization, realized rate, perished hours and reward per policy.
Learning curve
Bandit reward across training episodes.
Uplift tiles
Each policy against the static price book baseline.
Validation panel
Hold-out backtests, invariant pass/fail checks, and a parameter-sensitivity tornado chart.

How to read it

  • Greedy fill almost always wins on utilization and loses on realized rate — that is the trap the reward function is designed to expose.
  • Never read the uplift without the validation panel. A result that fails an invariant is not a small error, it is a broken model.
  • The tornado chart ranks which assumption the conclusion depends on. If the top bar is an assumption you cannot observe in production, the result is not yet decision-grade.

Formulas

Revenue-weighted utilizationrevenue ÷ (capacity hours × reference rate)
Rewardrevenue − cash opex − perishment penalty
Uplift(policy reward − static reward) ÷ static reward

Decisions it supports

  • Whether to pilot a dynamic pricing policy on the residual book, and with what guardrails.
  • Which parameters need real-world instrumentation before the policy is trusted with revenue.

2.15 Commercial org

A P&L this size is owned by a team, not a leader

The organisation scorecard: five functions — billing platform, enterprise sales, marketplace, FinOps and revenue assurance, customer success — each with an owner, headcount, objectives, succession depth and talent metrics. Org capability is the constraint on every other tab's plan.

What you see

Function cards
Owner, headcount, current objectives and status per function.
Succession depth
Ready-now and ready-later cover for each leadership seat.
Talent metrics
Open roles, time to fill, regretted attrition.

How to read it

  • A function with no ready-now successor is a single point of failure on a revenue-critical process; that is a risk to state alongside financial risk, not below it.
  • Objectives here should map one-to-one to mandate lines on the scorecard tab. Where they do not, either the objective or the mandate line is wrong.

Decisions it supports

  • Hiring sequence and where to build bench strength first.

2.16 Why Reliance / Why Ilan

The mandate, matched requirement by requirement

The case in two halves: three reasons this token and capacity P&L is worth building at Reliance specifically, three reasons the operator's record matches the mandate, then a requirement-by-requirement evidence map against the role criteria and a first-90-days plan.

What you see

Why Reliance
Three bullets on captive demand, integrated infrastructure and the sovereign-scale opportunity.
Why the operator
Three bullets with the record behind each.
Evidence map
Each 'what you'll bring' criterion against the artefact in this app that demonstrates it.
First 90 days
Sequenced plan with what is closed by day 30, 60 and 90.

How to read it

  • Every row of the evidence map links to a live tab or document. The claim and the artefact are never separated.

Decisions it supports

  • It is the closing argument of the whole app — read it after the cockpit, not before.

2.17 Glossary

One definition per term, used everywhere in the app

Every ledger class, invariant and KPI formula defined once, searchable and filterable by category. The same strings power the ⓘ tooltips throughout the cockpit, so a definition can never drift from its tooltip.

What you see

Search box
Matches term, definition, formula and note.
Category filters
Ledger, funnel, pricing, margin, contract, platform, simulation and so on.
Entry cards
Term, definition, formula in monospace, and any caveat.

How to read it

  • If two people disagree about a number, settle the definition here before arguing about the value.

Decisions it supports

  • It is the reference layer; it supports every other decision rather than making one of its own.

Back to top ↑

3. The document pages

Four long-form pages carry the reasoning behind the cockpit, and one page is the download centre.

3.1 P&L primer

The economics from first principles

The full P&L primer rendered as a web document: executive summary, the business in one picture, pricing for capacity and tokens, the cost stack, two worked profit-and-loss statements on the same fleet (a 1,024-GPU cluster and a token endpoint), metering, rating and billing, settlement and revenue recognition, the revenue-assurance control framework with its KPI table, and the assembled income statement.

What you see

Sticky section nav
Jump between the eleven sections.
Price tables
Capacity and token price points as real tables.
Worked P&Ls
Line by line, so every cockpit tile can be traced to an arithmetic step here.

How to read it

  • Read sections 4 and 5 (cost stack, worked P&Ls) before quoting any margin number from the cockpit — they define what is in and out of each margin.
  • Section 9 is the source for the revenue-assurance control framework shown on the RA tab.

Decisions it supports

  • It is the reference the cockpit compresses; use it to defend a number under challenge.

3.2 Strategy & GTM

SKU sequencing, segments and the three waves

The go-to-market strategy: the seven-SKU portfolio, the three-wave sequencing, demand by segment, sales motions and the ROI posture per SKU. The GTM & SKU mix cockpit tab is the instrumented version of this document.

What you see

Wave sequencing
Which SKUs launch when, and what has to be true first.
Segment demand
Captive, anchor labs, enterprise/BFSI, developers, government.
ROI model per SKU
Margin band, capital intensity, payback, primary risk.

How to read it

  • Wave order is a capital-allocation argument, not a preference: wave 1 funds depreciation, wave 2 raises blended margin, wave 3 buys optionality.

Decisions it supports

  • What to build and sell next, and what to explicitly not do yet.

3.3 Compute futures

The hedging memo, with the live model

The memo on compute futures as a hedging instrument for a depreciating fleet: why compute is perishable, hedging revenue risk, de-risking capital financing, the 60/70 capacity overlay, and the basis, liquidity and margin risks a hedge programme carries.

What you see

Sensitivity model
The same interactive model as the cockpit tab, with more explanation around it.
Risk section
Basis risk, liquidity, margin calls — the reasons a hedge can hurt.

How to read it

  • The 60/70 overlay means roughly 60% of capacity contracted and up to 70% of the merchant remainder hedged; the two percentages are different things and are frequently confused.

Decisions it supports

  • Whether to stand up a hedging programme at all, and how to govern it.

3.4 Platform spec

The engineering contract for metering, billing and settlement

The build artefacts: JSON Schemas for usage-event envelopes and payloads, price book, rated line item and invoice; an OpenAPI 3.1 surface; Postgres billing DDL and ClickHouse event DDL; reconciliation SQL for meter → rated → invoiced → collected; a conformance checklist and golden test vectors. Also the India GST / e-invoice (IRN) treatment and the Ind AS 115 revenue-recognition position.

What you see

Schema browser
Each JSON Schema with its fields.
OpenAPI surface
Endpoints for ingest, rating, invoicing and settlement.
DDL
Postgres ledger tables and ClickHouse event tables.
Reconciliation SQL
The queries behind the three-way match on the RA tab.
Conformance vectors
Golden inputs and expected outputs a vendor implementation must reproduce.

How to read it

  • The reconciliation SQL and the RA tab's Formula column are the same logic in two forms. If they ever diverge, the SQL is authoritative.

Decisions it supports

  • Whether a vendor or internal build meets the bar — the conformance vectors are the test.

3.5 Documents

Download centre

PDF and Word originals of the primer, the compute futures memo and the platform specification; the pricing-levers paper as a PDF; and the monthly CEO report, generated live from the closed-month data at the moment you click.

What you see

Monthly CEO report
A board-ready PDF: month at a glance, top four insights with commentary, value drivers and cluster ROI, mandate, pricing, cost stack and portfolio sections, and ranked next best actions.
Source documents
Three documents, each in PDF and Word, with an in-app reading link.
Optimal pricing levers paper
The four-lever optimisation framework as a downloadable PDF.

How to read it

  • The CEO report is generated from the same dataset as the cockpit, so its narrative cannot drift from the tiles.

Decisions it supports

  • What to circulate outside the app.

Back to top ↑

4. Working routines

Suggested reading paths for the recurring questions this app is built to answer.

4.1 The monthly close review

Roughly forty minutes, in this order

A repeatable sequence that moves from control to cause to action.

What you see

1 · Revenue assurance
Confirm no BLOCKING control is failed. If one is, stop and fix — nothing downstream is trustworthy.
2 · Capacity ledger
Check the invariant passes, then read leaked and perished.
3 · Meter to cash
Locate the stage that dropped, and its dollar value.
4 · P&L cockpit
Read the margin bridge to attribute the month: price, mix, utilization or cost.
5 · Customers & contracts
Work the renewal window and any low-drawdown anchor.
6 · Mandate scorecard
Confirm what changed colour, and generate the CEO report.

How to read it

  • Do not start at the P&L. Starting at the P&L means arguing about a number whose quality has not yet been established.

Decisions it supports

  • The month's action list, ranked by dollar exposure.

4.2 The pricing review

Four levers, in the order they bind

Pricing is optimised across four interacting levers rather than one rate card: the cost floor, the take-or-pay anchor, token multiplexing, and the residual book priced dynamically or hedged.

What you see

FinOps & capacity
Establish the cost floor: $0.485/GPU-hr cash, plus about $1.01 depreciation.
Customers & contracts + GTM
Set the take-or-pay anchor at 60–75% of capacity around $2.05/GPU-hr.
Token economics
Route the 15–25% slice that clears the 21.3% density crossover into tokens.
RL simulation + Compute futures
Price and hedge the remaining 10–15% residual.
Pricing & packaging
Govern the waterfall so realized price does not leak the gains back out.

How to read it

  • The target blended outcome is roughly $2.40–$2.70 per GPU-hour equivalent at 82–88% revenue-weighted utilization. Each lever is bounded by the one before it.

Decisions it supports

  • The quarter's price book, discount thresholds and hedge ratio, as one decision rather than four.

4.3 Where to look for a given question

Quick routing table

What you see

"Why did margin move?"
P&L cockpit → margin bridge, then FinOps for the cost lines.
"Are we billing everything we serve?"
Meter to cash, then Revenue assurance, then Capacity ledger.
"What is at risk this year?"
Customers & contracts → renewal timeline and drawdown.
"Should we sell this node as capacity or tokens?"
Token economics → density ladder.
"Can we finance the next cluster?"
Compute futures → payback under the downside path.
"Is a discount defensible?"
Pricing & packaging → waterfall and governance log.
"Is the marketplace real revenue?"
Marketplace & partners → net revenue, not GMV.
"What does this term mean?"
Glossary, or the ⓘ next to the number.

Back to top ↑

GPU Cloud Billing Cockpit — user guide. Every figure in this guide is the same figure the app computes; definitions match the Glossary tab exactly.