From Metal to Agentic Revenue
*Blog by Jitin Bhandari, EVP, NeoCloud Platform and Forward Deployment Engineering, Mavenir
The Agentic Economy Playbook: Architecture, Unit Economics and Go-to-Market for CSPs and NeoClouds
Where the series has taken us
Paper 1 made the market case: agentic demand is larger and less predictable, and the durable opening sits above raw capacity. Paper 2 defined the platform: seven capabilities, built on selective ownership, that turn heterogeneous AI resources into trusted agentic services.
This paper addresses the question every board will ask next: how do we turn that into a business without making an uncontrolled capital or product bet?
Estimates of the prize vary with scope, and that range is itself instructive:
| Estimate | Figure | Scope |
|---|---|---|
| Analysys Mason | $21B (2024) → $134B (2030) | Total GPU-as-a-service revenue, all providers[1] |
| Analysys Mason | ~$9.6B by 2030 | GPUaaS revenue captured by telecom operators[2] |
| McKinsey | $35–70B by 2030 | GPUaaS market addressable by telcos[2] |
The gap between what is addressable and what operators are forecast to capture is the central point. Capacity alone will capture a modest share; the rest depends on whether providers can move up to governed services, platforms and ecosystems. And execution risk is real: as Paper 1 noted, Gartner expects more than 40% of agentic AI projects to be canceled by the end of 2027.[3]
| The winners will not be the providers with the largest announced capacity. They will be those that convert accepted capacity into trusted, billable and repeatable agentic outcomes fastest. |
A vendor-neutral, staged path: decide where to play, build what to sell, scale with discipline.
Step 1: Choose the strategic position
| Position | What is sold | Capital intensity | Primary competitor | Key risk |
|---|---|---|---|---|
| Capacity provider | GPU-hours, reserved clusters | Very high | Hyperscalers, scaled NeoClouds | Price cyclicality, utilization, refresh |
| Managed AI platform | Inference, models-as-a-service, developer tools | Medium | NeoClouds, inference specialists | Undifferentiated catalog |
| Sovereign Agentic Service Cloud | Governed agents and workflows with evidence | Medium; can be asset-light | Hyperscaler sovereign regions | Demand aggregation |
| Marketplace and ecosystem orchestrator | Agent marketplace, settlement, partner revenue | Low to medium | Hyperscaler marketplaces | Ecosystem critical mass |
The capacity position is the most cyclical. On-demand H100 rates fell from above $7 per GPU-hour in early 2024 to a median near $3 by mid-2026, while one-year committed rates rose about 40% between October 2025 and March 2026 as inference absorbed supply. Owners benefit when markets tighten and absorb the losses when they loosen. One independent analysis estimates hardware amortization at about 28% of a NeoCloud H100-hour price at 70% utilization. The remainder covers power, facilities, networking, operations, cost of capital and margin, and the margin portion is the part most exposed to competition. Capacity returns depend on timing, utilization and refresh cycles, not on demand alone.[7]
TM Forum points to the other end of the spectrum: a CSP- or consortium-owned plane for agent lifecycle, policy, execution and settlement, rather than CSPs participating only as consumers in hyperscaler-governed marketplaces. Most providers will combine positions over time, but should enter through one.[4]
Step 2: Select the first market wedge
| Wedge | Why it works | Evidence |
|---|---|---|
| CSP internal operations and autonomous networks | Own data, own budget, measurable ROI; builds credibility | TM Forum Level 4 validations in service assurance (Telkomsel, AIS, DNB) and energy (TDC NET)[4] |
| Regulated enterprise workloads | Evidence and trust outweigh unit price | Regulated buyers writing residency and operator location into AI procurement |
| Sovereign and public-sector programs | Operator is a natural national counterparty | 18 telcos across five continents launched sovereign AI factories in 18 months (NVIDIA)[5] |
| Communications agents | Operator controls the session, the verified subscriber identity and real-time media in the network path | AI voice agents, real-time translation, fraud and spam screening, and number intelligence. All of these depend on in-network position, not just compute |
| Vertical agentic services | Domain depth beats catalog breadth | NeoClouds specialized by industry, language or geography |
Telenor AI Factory illustrates an operator-led model: GPU-as-a-service and models-as-a-service with more than 190 open models hosted in Norway, internal network use cases alongside external customers, and an open platform stack chosen explicitly for portability and freedom from lock-in.[6]
| For CSPs | For emerging NeoClouds | |
|---|---|---|
| Lowest-risk first wedge | Internal autonomous operations — the operator is its own first customer | A vertical, geography or inference specialization with a named anchor customer |
| Most differentiated wedge | Communications agents built on the network session, identity and real-time media | Inference specialization or a vertical where a named anchor customer shares its domain data |
| Second wedge | Regulated enterprise and sovereign programs through existing channels | Partnering with a CSP or SI for regulated and public-sector access |
| Wedge to avoid first | Undifferentiated GPU rental against scaled NeoClouds | Broad horizontal catalog against hyperscalers |
Step 3: Build an asset-light supply model
As Step 1 showed, spot and term markets are moving in opposite directions, and a tight term market raises the cost of leasing. That is why the case for an asset-light model does not rest on falling prices. It rests on three risks a provider should not take before demand is proven: demand risk (capital committed before consumption is contracted), refresh risk (accelerators losing value as new generations arrive) and capital-allocation risk (balance sheet consumed by hardware rather than by the software and services layers where differentiation sits). Asset-light is a risk-allocation choice, not a bet on falling prices. In practice, that means a portfolio:
| Source | When to use | Watch-out |
|---|---|---|
| Owned | Anchored demand, firm power, sovereignty requirement | Depreciation, refresh and utilization risk |
| Partner (OEM, colocation, operator) | Fast time to first token; shared risk | Contract terms, acceptance and support boundaries |
| Leased (NeoCloud, marketplace) | Burst, validation, geographic reach | Counterparty concentration, jurisdiction, and rising term prices in tight markets |
- Match commitment to demand: commit term capacity against contracted demand, use partner and operator-owned facilities where they exist, and reserve spot capacity for validation and burst
- Multi-source portability: workloads move between owned, partner and leased capacity without re-engineering.
- No capital commitment before validated demand: expand owned capacity only against contracted or observed consumption.
- Counterparty resilience: public filings from scaled NeoClouds disclose heavy customer concentration; diversify suppliers and customers from the start.
Step 4: Define the initial service catalog
| Service family | Example SKUs | Primary buyer |
|---|---|---|
| Capacity products | Reserved and on-demand GPU, dedicated clusters | AI-native builders, research |
| Managed inference | Model APIs, dedicated endpoints, provisioned throughput | Developers, enterprise IT |
| Governed models and agents | Approved model catalog, agent hosting, tool gateway | Enterprise AI teams |
| Vertical workflows | Service assurance, AI voice and messaging agents, real-time translation, fraud screening, clinical documentation, citizen services | Line-of-business leaders |
| Sovereignty and evidence profiles | Residency, operationally bounded, customer-controlled keys, national enclave | CISOs, risk and compliance |
| FDE and managed operations | Lighthouse delivery, managed agent operations | Enterprises without in-house capacity |
Suggested starting rule (author’s recommendation): launch with two or three families and expand only as blueprints become repeatable.
Step 5: Create the commercial control plane
Pricing is moving beyond seats. Bain finds that most AI software companies are layering new meters on top of seats rather than replacing them, and that capacity-based commitments with overages are emerging as the preferred model — predictable budgets for customers, stable revenue for providers. Bain distinguishes three meters: effort (tokens, agent-hours), output (records processed) and outcome (issues resolved).[9]
- Entitlement and pre-use authorization: confirm a tenant may run a workflow and reserve budget before execution.
- Wallets, commitments, quotas and budgets: prepaid, postpaid and committed-use constructs with hard and soft limits.
- Effort, output and outcome pricing on one rating engine.
- Supplier-cost reconciliation: every token and workflow traced to its capacity source and cost.
- Partner and reseller settlement: revenue share, wholesale rates and dispute handling
In effect, this is online charging applied to agents: reserve budget before execution, re-authorize as the workflow runs, and release it on completion. Operators have run real-time consumption control at national scale for decades. What is new is rating tokens, workflows and outcomes on one engine, and settling across multiple model, agent and capacity providers within a single transaction. Providers that already run real-time charging start with the hardest part of this problem solved.
| Metering after the fact tells you what happened. Authorization before the fact decides what is allowed to happen. Agentic economics needs both. |
Step 6: Establish real unit economics
Token prices are falling fast: a16z estimates the price of a fixed capability level has declined about 10x per year, and Epoch AI finds declines of 9x to 900x per year depending on the task. Yet total spend keeps rising,because consumption per task is growing faster than unit prices fall in many agentic workloads.[10] GPU-hour and headline token prices are therefore incomplete measures.
Cost per useful output = (capacity + platform + operations + supplier cost) Ă· outputs that met quality and latency targets
The business model matters as much as the cost base. Selling capacity by the hour and selling governed services earn different margins, not just different revenue. The services model adds platform, serving and operating cost, but it captures value that raw capacity cannot: governance, evidence, routing and outcome-based pricing. Whether it pays depends on cost per useful output, sustained utilization and how much traffic is served from cache or smaller models. Because token prices are falling quickly, the services model has to win on margin per useful output, not on headline token price.
| Lever | Effect |
|---|---|
| Utilization | Fixed capacity and platform cost spread over fewer useful outputs; margin falls faster than utilization |
| Caching | Cached prompts already dominate agentic token volume (see Paper 2) |
| Routing and right-sizing | Routine steps to smaller open models; frontier models reserved for hard steps |
| Failure and retry | Failed runs and retries consume tokens but produce no billable output |
| Price decline | Token prices falling about 10x a year erode revenue per output unless cost per output falls at least as fast |
Revenue and margin must be visible by tenant, model, workflow and capacity source. CloudZero’s survey of 260 finance leaders found that 75% of organizations unable to measure AI outcomes had held back investment, compared with 38% of those that could.[11]
Step 7: Use Forward-Deployed Engineering as the launch mechanism
As Paper 2 noted, demand for forward-deployed engineers has surged, reflecting that deployment, not models, is the constraint. For a CSP or NeoCloud, FDE is how the first customers become references:
| Milestone | Exit criterion |
|---|---|
| 1. Lighthouse design | Agreed workflow, data access, sovereignty profile and success metric |
| 2. Controlled evaluation | Baseline vs. agentic performance measured on real data |
| 3. First production outcome | Workflow in production with bounded autonomy and human oversight |
| 4. First reconciled invoice | Consumption authorized, metered, rated and reconciled to supplier cost |
| 5. Capability transfer and blueprint | Customer team operating; pattern packaged for the next customer |
The fourth milestone is the one most often skipped – and the one that proves a business rather than a pilot.
Step 8: Design the GTM ecosystem
| Partner group | Role |
|---|---|
| Infrastructure and silicon | Capacity, reference architectures, financing, supply assurance[8] |
| Platform and open source | Serving, orchestration, agent frameworks, tool protocols |
| Systems integrators | Delivery capacity, industry access, public-sector contract vehicles |
| Operator channels | Enterprise and SMB relationships, billing and customer ownership |
| Vertical software and agent providers | Domain workflows and agents published into the marketplace |
An operator that becomes an AI provider brings its enterprise, public-sector and wholesale relationships with it. Channel rules – who owns the account, who sets price, how revenue is shared — must be defined before the first partner-sourced deal, not after.
Step 9: Apply disciplined stage gates
| Gate | Pass condition | Stop or rethink if |
|---|---|---|
| Evidence of demand | Named accounts with budget and a defined workflow | Interest without a budget owner |
| Open platform seat | Customer’s control plane not already locked in | A hyperscaler or incumbent owns the seat |
| Capacity readiness | Accepted, not just announced or contracted, capacity | Power, delivery or acceptance dates slip |
| Acceptable unit economics | Positive contribution margin per useful output | Margin depends on unproven utilization |
| Repeatability | Second customer live on the same blueprint | Every deployment is bespoke |
| Declining customization | Engineering effort per new customer falling | FDE effort flat or rising |
Indicative roadmap (author’s recommendation, to be adapted per provider): Foundation (lighthouses committed), Prove (repeatable outcomes and reconciled revenue), Scale (multi-tenant and partner-sourced growth), Monetize (marketplace and ecosystem revenue) — each with a measurable kill-gate.
| Author’s perspective At Mavenir, our forward-deployed engineering practice is built around exactly these gates: a lighthouse is not complete until an outcome is in production and consumption has been authorized, metered and reconciled end to end. Two areas of our experience apply directly. Real-time policy and charging, built for operators running consumption control at national scale, is the foundation the commercial control plane in Step 5 requires. Voice, messaging and IMS, the systems that carry the communications session, are where communications agents will run. Applying this operating discipline to agentic services is how capacity becomes a business rather than a stranded asset. |
Note:Â Frameworks, thresholds and phase timings in this series are the author’s recommendations, not industry benchmarks.
A 12-question readiness framework
A discussion framework for CSP and NeoCloud executives (not a validated maturity model). Score each question 0 (no), 1 (partly) or 2 (yes). As a starting heuristic, a score below 12 suggests beginning with internal or partner-led lighthouses before external commitments.
| # | Area | Question |
|---|---|---|
| 1 | Strategy | Have we chosen one entry position and one first wedge? |
| 2 | Strategy | Do we have a structural right to win in that wedge? |
| 3 | Demand | Do we have named accounts with budget and a defined workflow? |
| 4 | Supply | Can we serve lighthouse demand without new owned capacity? |
| 5 | Supply | Can workloads move across owned, partner and leased capacity? |
| 6 | Platform | Which of Paper 2’s seven capabilities do we own, and which are we implicitly outsourcing? |
| 7 | Trust | Can we produce sovereignty evidence on demand? |
| 8 | Trust | Can every agent be identified, authorized and stopped at runtime? |
| 9 | Economics | Do we measure cost per useful output, not just cost per GPU-hour or token? |
| 10 | Economics | Can we authorize spend before execution and reconcile cost to revenue after? |
| 11 | Delivery | Do we have an FDE motion that produces reusable blueprints? |
| 12 | GTM | Are channel, partner and settlement rules defined before the first partner deal? |
Closing: an invitation
Market opening → Platform design → Commercial execution
The Agentic Economy is creating demand that CSPs and NeoClouds are structurally placed to serve. Capturing it does not require building a hyperscaler or making an undisciplined capital bet. It requires a clear position, a bounded first wedge, a supply model that matches capital commitment to proven demand, a platform that owns the right control points, real unit economics, and a delivery motion that turns each lighthouse into a repeatable blueprint.
If you are exploring a lighthouse, a sovereign AI program or an agentic marketplace, I would welcome the conversation. Take the 12-question assessment, and let’s compare notes on where your organization stands.
References
- Analysys Mason, GPUaaS worldwide forecast 2025–2030 (Oct 2025)
- McKinsey, AI infrastructure: a new growth avenue for telecom operators (2025)
- Gartner press release, June 25, 2025 (agentic project cancellations)
- TM Forum, The Agentic AI Marketplace Opportunity (June 2026); TM Forum Autonomous Network validations
- NVIDIA, Telcos Across Five Continents Are Building Sovereign AI Infrastructure (May 2025)
- Telenor AI Factory; Telenor and Red Hat announcement (MWC 2026)
- AIMultiple Cloud GPU Rental Price Index and SemiAnalysis H100 1-year rental index, as reported (2026)
- Silicon Analysts, GPU rental premium decomposition (July 2026)
- Bain & Company, AI Pricing: A Reality Check on Effort, Usage, and Outcomes (Aug 2026)
- a16z, Welcome to LLMflation (Nov 2024); Epoch AI, LLM inference price trends (Mar 2025)
- CloudZero 2026 AI ROI survey
