A cell site stops being a single-purpose radio asset and becomes a distributed compute node, a small piece of an "AI factory" sitting close to where data is generated and consumed. For the people who finance, build, and operate physical network infrastructure, this is potentially the most consequential shift in the asset class since the tower-sharing model emerged.
This piece looks at what AI-RAN and shared compute actually are, why infrastructure developers stand to benefit in a specific and structural way, and the regulatory terrain they will have to cross in India, which is genuinely distinctive and, in places, unresolved.
What AI-RAN Actually Means
AI-RAN is the convergence of artificial intelligence and the radio access network onto shared, GPU-accelerated infrastructure. The industry generally describes it in three forms, and the distinction matters because each implies a different business case:
- AI-for-RAN uses machine learning to improve the network itself: better channel estimation, beamforming, interference management, and energy efficiency. This is an optimisation play; the beneficiary is the operator's own cost and quality.
- AI-on-RAN runs third-party AI applications on the same infrastructure that processes the radio signal. Inference for enterprise customers, computer vision, generative AI queries, all hosted at the edge of the network. This is where new revenue lives.
- AI-and-RAN is the architectural endgame: a single pool of compute, dynamically allocated in real time between radio processing and AI workloads depending on demand.
The technical enabler is that modern virtualised RAN runs on the same class of GPU and accelerated silicon that AI inference needs. Rather than building a separate edge-computing network (something the telecom industry has talked about for years without finding the monetisation), operators can co-locate AI on the GPUs already installed for signal processing. The radio network itself becomes the edge compute platform.
The momentum behind this is not theoretical. The AI-RAN Alliance was formed in early 2024 and now counts most major vendors as members. NVIDIA's AI Aerial platform, paired with hardware from Nokia, Ericsson, Dell, HPE, Marvell and others, has moved from proof-of-concept to field trials. SoftBank ran an outdoor 5G AI-RAN trial; T-Mobile US has tested concurrent AI and RAN workloads on a single server in its innovation lab; and at MWC 2026 in Barcelona, the framing shifted decisively, vendors stopped pitching AI merely as a way to tune the network and started pitching shared compute as a way to monetise the network beyond connectivity.
The Shared-Compute Thesis
The core economic argument rests on a simple observation: radio networks are provisioned for peak demand and sit idle much of the time. A GPU that is busy processing baseband traffic at rush hour may be lightly loaded at 3 a.m., or in the middle of a quiet rural cell. Shared compute treats that spare capacity as a sellable product.
Using techniques such as multi-instance GPU partitioning and real-time orchestration, the infrastructure can steer resources between RAN and AI on the fly, guaranteeing the radio network its quality of service while renting out whatever is left over. Estimates from platform vendors put the improvement in capacity utilisation at two to three times that of siloed, single-purpose infrastructure.
That changes the investment logic. If a site can earn both subscriber revenue and compute revenue, the capital case for densifying the network (more sites, more fibre, more silicon) improves. The pitch to operators is that they can finally justify network spending against two demand curves instead of one, and that the network can host generative AI, agentic workloads, and enterprise inference where latency and data location actually matter.
Why Infrastructure Developers Benefit
This is the part that is easy to lose in the vendor noise. AI-RAN is usually discussed from the operator's seat, but the people who develop infrastructure (tower companies, neutral-host builders, data-centre and edge developers, and the engineering and capital partners behind them) are arguably better positioned to capture the upside, for several reasons.
Asset Utilisation Is the Whole Game
The fundamental value created by shared compute is sweating an existing physical footprint harder. Whoever controls the site, the power, the cooling and the real estate captures a structural share of that value. A tower company that adds GPU-bearing edge nodes to its sites is no longer just leasing vertical steel; it is leasing compute-ready space with a second revenue line.
The Neutral-Host Model Maps Cleanly Onto AI-RAN
India and many other markets have spent years separating infrastructure ownership from service provision, shared towers, shared fibre, shared active equipment. AI-RAN extends that logic to GPUs. A neutral developer can build multi-tenant, GPU-accelerated edge infrastructure and offer it as GPU-as-a-Service to multiple operators and enterprises, the same way passive infrastructure is offered today. Orchestration blueprints have already demonstrated that GPUs can be shared safely across AI and RAN tenants in real time, precisely the technical foundation a neutral-host model needs.
Edge Proximity Is a Moat Hyperscalers Cannot Easily Replicate
Centralised cloud data centres are far from the user. AI inference that is latency-sensitive (real-time vision, robotics, agentic assistants, industrial control) wants to run close to where the data is generated. Infrastructure developers already hold the most distributed real estate footprint in the country: cell sites, aggregation points, in-building systems. That distribution is the scarce resource in edge AI, and it is hard to buy from scratch.
Capex Justification Flips from Defensive to Offensive
Historically, infrastructure developers have had to argue that a given site or fibre route would eventually pay back through connectivity demand. With AI-generated traffic projected to rival or exceed human-generated mobile data later this decade, the developer who builds compute-ready infrastructure now is positioning for both curves.
Data-Centre and Edge Developers Get a New Tier of the Market
AI-RAN does not eliminate the data centre; it creates a continuum from hyperscale core to regional edge to far edge at the cell site. Developers who can build and operate across that continuum (and stitch it together with fibre and, increasingly, subsea capacity) own the corridor along which AI workloads will flow.
None of this is frictionless. The hard problems are real: legacy operations and billing systems were built to meter minutes and gigabytes, not continuous compute provisioning; guaranteeing strict resource isolation so that a paying AI tenant never degrades carrier-grade radio quality carries its own overhead; and the power and cooling demands of GPU-dense sites strain grid connections and water resources in exactly the places infrastructure tends to be sited. These are engineering and commercial challenges, though, not reasons the thesis fails.
The India Regulatory Picture
This is where infrastructure developers should pay the closest attention, because India is simultaneously one of the most attractive markets for this model and one of the most regulatorily intricate. Several distinct frameworks intersect, and a few important questions remain genuinely open.
The Telecommunications Act, 2023 reshaped the foundation
The Act replaced the old licensing patchwork with an authorisation regime, kept spectrum assignment firmly with the Department of Telecommunications (DoT), and signalled a broader push toward convergence and lighter-touch entry for many services. Spectrum, critically, is still assigned for the purpose of providing telecommunication services, a point that becomes important the moment a network operator wants to sell third-party AI compute on infrastructure that was funded and authorised around spectrum.
The DCIP authorisation is the most relevant opening, and its boundaries are the key constraint
Acting on TRAI recommendations, the draft Authorisation for Telecommunication Network Rules, 2025 introduce a new Digital Connectivity Infrastructure Provider (DCIP) category. This is significant for infrastructure developers: the old IP-1 registration only allowed passive infrastructure, towers, dark fibre, ducts, right-of-way. DCIP is designed to let a neutral third party establish and share both passive and specified active infrastructure, including elements like RAN, on a non-discriminatory basis. That is exactly the kind of vehicle a neutral GPU host would want. The crucial caveat: DCIP explicitly excludes the core network and spectrum. A DCIP can build and share the compute-bearing active infrastructure, but the spectrum-using radio service remains the authorised operator's domain. Structuring a shared-compute business cleanly across that line is the central design problem.
The spectrum / non-telecom-use grey zone is unresolved
Selling spare GPU capacity on a RAN site as commercial AI compute does not fit neatly into any single existing bucket. Is it a telecommunication service (DoT's domain), or is it an IT / data-centre / cloud service regulated through the Ministry of Electronics and Information Technology (MeitY)? The answer determines licensing, taxation, security obligations, and who you answer to. As of now, this boundary has not been definitively drawn for AI-RAN specifically, and that ambiguity is itself a planning risk.
Jurisdiction is split, and convergence is still a work in progress
Today, telecom networks sit with DoT, while data centres, content-delivery networks, internet exchange points, and over-the-top services are handled by MeitY and the IT Act, 2000. TRAI has explicitly recommended that data centres and related digital infrastructure be treated as integral digital communication infrastructure under a converged policy, arguing that it is not in the sector's interest to have one converged technology governed by different ministries. Until that convergence is realised, an AI-RAN business may straddle two regulators with different rulebooks. The new Cloud-Hosted Telecom Network (CTN) provider authorisation, which brings virtualised network functions into the telecom framework, is one piece of this puzzle worth tracking.
Network elements are Critical Information Infrastructure
Under Section 70 of the IT Act, the computer resources of telecom network elements have been designated Critical Information Infrastructure, subject to heightened security practices monitored by the national protection centre. Putting third-party AI workloads onto that same hardware raises real questions about isolation, auditability, and lawful-interception obligations that any developer will have to engineer around from the start, not bolt on later.
Data protection now reaches directly to AI workloads
The Digital Personal Data Protection Act, 2023 was operationalised by the DPDP Rules, notified on 13 November 2025, with a phased timeline and substantive compliance obligations landing by 13 May 2027. For anyone hosting AI inference at the edge, the salient points are: India has not imposed blanket data localisation, but the central government retains broad discretion to restrict cross-border transfers to specified countries and to mandate localisation of certain categories of data. Entities designated Significant Data Fiduciaries face enhanced obligations, including a India-based data protection officer, independent audits, data protection impact assessments, and, notably for AI, the appointment of an independent algorithmic auditor to assess the models used in processing. Sectoral localisation mandates (the Reserve Bank's rules on payments data, for example) continue to apply on top. An edge AI platform hosting enterprise inference will need to know whose data is being processed where, and design for the possibility that the rules tighten.
A DCIP can build and share the compute-bearing active infrastructure, but the spectrum-using radio service remains the authorised operator's domain. Structuring a business cleanly across that line is the central design problem.
Why the New Regime Makes Investing More Convenient
The complexity above is real, but it sits on top of a reform trajectory that, taken as a whole, makes India a markedly easier place to put money into infrastructure than it was a few years ago. For an investor weighing a compute-bearing infrastructure play, several of these changes lower the barrier directly.
- From a licensing labyrinth to a lighter authorisation regime. India's telecom sector ran for years on a multi-tiered, overlapping set of licences and registrations that was slow and expensive to navigate. The Telecommunications Act, 2023 replaces much of that with an authorisation framework: general authorisations are quicker to obtain, are not tied to a fixed licence duration, and lower the entry barrier. The Act folds a cluster of niche services, machine-to-machine connectivity, public Wi-Fi, and others, into a single, light-touch "miscellaneous" authorisation, signalling a clear preference for market entry over gatekeeping. For a new entrant building edge or shared-compute infrastructure, that means a shorter, more predictable path from decision to deployment.
- Lower cost of entry for the people who build the infrastructure. Both the existing IP-1 category and the new DCIP authorisation are exempt from authorisation-fee obligations. A developer building and sharing the active and passive infrastructure that AI-RAN runs on is not handing over a slice of revenue as a licence fee, which materially improves the economics of a capital-intensive, multi-tenant build.
- Foreign capital can come in cleanly. India now permits 100% foreign direct investment in the telecom sector, with up to 49% on the automatic route and higher stakes via approval, and 100% under the automatic route for infrastructure providers such as towers and optical-fibre networks (subject to standard security clearances when a country shares a land border with India). The overwhelming majority of FDI into India now flows through the automatic route, meaning a foreign investor in compute-ready infrastructure can, in most structures, enter without a separate government approval step.
- The levy base has been narrowed. Part of the 2021 reform package rationalised the definition of Adjusted Gross Revenue to exclude operators' non-telecom revenue from the base on which statutory levies are charged. The principle is directly relevant to shared compute: revenue earned from selling AI capacity is not obviously "telecom" revenue, which strengthens the case that it should sit outside the traditional telecom levy net, though the precise treatment of compute revenue under the new authorisation regime is not yet settled, so this is a point to confirm rather than assume.
Set against the genuine open questions on jurisdiction and spectrum, the direction of policy is to stop taxing adjacent, non-core revenue as if it were connectivity. For the investor, certainty and a predictable direction of travel matter as much as any single incentive.
What Infrastructure Developers Should Be Doing Now
The opportunity is real and the regulatory direction is broadly favourable, but the execution window rewards preparation over haste. A few practical moves stand out:
- Build compute-ready, not compute-now. New sites and fibre routes should be specified for power, cooling, and space at densities that AI inference needs, even where the immediate driver is connectivity. Retrofitting later is far more expensive than provisioning for headroom up front.
- Treat the DCIP authorisation as a real option. For neutral developers, the new category is the cleanest current route to building and sharing the active, compute-bearing infrastructure AI-RAN needs. Engineering the business so it sits cleanly on the DCIP side of the spectrum line is worth doing early.
- Design isolation, auditability, and lawful-interception obligations in from the start. Hosting third-party AI on Critical Information Infrastructure raises real security and compliance questions. These belong in the architecture, not bolted on after a regulator asks.
- Map data flows against a tightening DPDP regime, and design for the possibility that localisation rules harden. Knowing whose data is processed where is both a compliance requirement and a commercial selling point to enterprise tenants.
- Watch the convergence and AGR questions closely: they will move, and the compute bits sit on them. The treatment of compute revenue and the DoT/MeitY boundary are the two variables most likely to change the economics. Track them as live risks, not settled facts.
The shift underway is genuine: the network is becoming a distributed AI factory, and the value is migrating toward whoever owns the distributed footprint and the compute that sits on it.
In India, the opportunity is unusually large and the regulatory path unusually specific. The developers who read both correctly (and act now) are the ones likely to own the corridor.
Sources & Further Reading
This analysis draws on industry reporting and primary policy material, including:
- IoT Business News / IoT Analytics, and MWC 2026 analysis on AI-RAN and shared compute
- NVIDIA, "AI-RAN Goes Live" and AI Aerial / AI-RAN platform documentation
- Nokia, AI-RAN platform and MWC 2026 partnership announcements
- Ericsson and Intel, AI for the RAN collaboration (MWC 2026)
- Telecom Tech News, NVIDIA–Marvell AI-RAN alliance (on OSS/BSS and radio-stack integration)
- IEEE Xplore / ScienceDirect, technical papers on AI-RAN architecture and compute economics
- Lexology and Khaitan & Co, analyses of the draft Authorisation for Telecommunication Network Rules, 2025 and DCIP authorisation
- TRAI Recommendations on digital communication infrastructure and convergence
- AZB & Partners and Cyril Amarchand Mangaldas, on the Telecommunications Act, 2023 (authorisation regime, migration, AGR)
- Invest India, UNCTAD Investment Policy Monitor, on 100% FDI in telecom via the automatic route
- EY, PIB and secureprivacy.ai, on the DPDP Act, 2023 and DPDP Rules, 2025
- US ITA Country Commercial Guide, on data localisation and Budget 2026 data-centre incentives
