Pakistan’s $507 Million Spectrum Auction, Seven Months On: What 1,782 5G Sites Mean for AI Workloads at the Edge

Pakistan’s $507 Million Spectrum Auction, Seven Months On: What 1,782 5G Sites Mean for AI Workloads at the Edge now has measurable answers rather than projections. The Pakistan Telecommunication Authority raised approximately $507 million in its March 10–12, 2026 auction, selling 480 MHz of the 597.2 MHz offered and lifting national holdings from 274 MHz to 754 MHz. As of late October 2026, operators run 1,782 live 5G sites across 24 cities, backed by $178.6 million in site-upgrade capital expenditure to June 2026. The 3500 MHz band is what makes edge inference technically viable here for the first time — and electricity tariffs, not spectrum, are now the binding constraint.

Introduction

The answer comes before the backstory. Seven months after the PTA concluded its Next Generation Mobile Services auction, Pakistan’s $507 Million Spectrum Auction, Seven Months On: What 1,782 5G Sites Mean for AI Workloads at the Edge resolves into hard figures — 1,782 live sites, 24 cities, three operators — because licences signed on March 19, 2026 carried phased rollout obligations that forced construction rather than inviting it.

Spectrum auctions are usually reported as revenue events, and that framing obscures the engineering consequence. The 3500 MHz holdings acquired by Jazz, Ufone and Zong are the reason AI inference can move from centralised data centres to the tower edge, because that band supplies both the capacity per cell and the latency floor real-time workloads budget against. This analysis tests three constraints against the published data — coverage density, electricity tariffs and device affordability — and ends with a verdict rather than a recap.

About this analysis: Written by The Verge PK’s telecom and digital policy desk. Every figure below is attributed to the PTA, the Ministry of IT and Telecommunication, or operator disclosures, and dated. Where data is operator-reported rather than independently measured, the Limitations section says so explicitly.

Transparency note: The Verge PK received no compensation, access or review rights from Jazz, Zong, Ufone, PTCL or the PTA in connection with this article. No operator was given advance sight of the text.

What Pakistan’s $507 Million Spectrum Auction Actually Sold

The PTA conducted its NGMS auction on March 10–12, 2026, selling 480 MHz of the 597.2 MHz offered and raising approximately $507 million for the national exchequer. Licences were signed on March 19, 2026. National mobile spectrum holdings rose from 274 MHz to 754 MHz as a result. The 117.2 MHz that went unsold matters commercially, because unsold inventory gives the regulator a reserve it can reprice — and operators a reason to expect softer terms next time.

Operator	Spectrum won (MHz)	Amount paid (USD)	Bands acquired
Jazz (VEON)	190	$239.37m	700 / 2300 / 2600 / 3500 MHz
Ufone (PTCL Group)	180	Not stated	Includes 3500 MHz
Zong (China Mobile Pakistan)	110	Not stated	Includes 3500 MHz
Auction total	480 of 597.2 offered	~$507m	700–3500 MHz range

The band mix determines workload capability more than the headline price does. Low-frequency 700 MHz propagates further per site, which means it extends reach into rural Punjab and remote districts at lower capex per square kilometre. The 3500 MHz band delivers the capacity density and latency profile enterprise workloads require, which is why it is treated globally as the core 5G band. Consequently Jazz’s 190 MHz across four bands buys coverage and capacity simultaneously, while Ufone’s 180 MHz and Zong’s 110 MHz produce materially different build economics — and, as the next section shows, different deployment curves.

The competitive consequence is the one worth holding onto. Because all three operators bought into 3500 MHz rather than one winner taking the band, Pakistan has contested low-latency capacity rather than a single-supplier bottleneck. That drives enterprise edge connectivity pricing down faster than a single-winner outcome would have, since a buyer can run a procurement with three credible bidders.

Seven Months On: Where the 1,782 5G Sites Actually Are

As of late October 2026, Federal Minister for IT and Telecommunication Shaza Fatima Khawaja confirmed 1,782 live 5G sites across 24 cities. The split is uneven: Jazz operates 1,015 sites, Zong 640 and Ufone 127. Those figures are ministry-confirmed and operator-reported, dated to the end of October 2026, and they describe installations rather than measured coverage.

From auction to live sites: the 2026 timeline

1. March 10–12, 2026 — PTA conducts the NGMS auction; 480 MHz sold for approximately $507 million.
2. March 19, 2026 — Licences signed by Jazz, Ufone and Zong, starting the phased rollout obligations.
3. To June 2026 — Operators report $178.6 million in 5G site-upgrade capital expenditure.
4. January–August 2026 — Over 607,000 5G-capable handsets locally assembled.
5. Late October 2026 — Ministry of IT confirms 1,782 live 5G sites across 24 cities.
6. October 2026 — Jazz, Zong, Ufone and PTCL CEOs jointly petition the Prime Minister for industrial power tariffs.

City	5G sites	Share of national total
Karachi	385	~22%
Lahore	373	~21%
Islamabad	276	~15%
Faisalabad	144	~8%
Peshawar	122	~7%

The concentration is not accidental. Karachi, Lahore and Islamabad together hold roughly 58 per cent of all sites because those three cities contain the enterprise customers, the data centre capacity and the power reliability that justify the capital outlay. Faisalabad and Peshawar form a second tier, and the remaining 19 cities share a thin residual between them. The engineering conclusion follows directly: an edge AI service-level agreement with a latency guarantee is deliverable in three cities today, not nationally, and any vendor promising otherwise is selling a roadmap rather than a network.

Operator	Live 5G sites	5G site capex to June 2026 (USD)	Spectrum held (MHz)
Jazz	1,015	$59.67m	190
Zong	640	$98.9m	110
Ufone	127	~$20m	180
Total	1,782	$178.6m	480 sold

The capex ranking inverts the site-count ranking, and that inversion is informative. Zong spent $98.9 million against 640 sites while Jazz spent $59.67 million against 1,015, which means Zong is committing more per site to core and transport upgrades while Jazz converts existing 4G assets faster and cheaper. Jazz has separately committed $1 billion over three years to modernise its core network. Core modernisation is the layer that decides how close to the radio an inference workload can sit, because it governs where user-plane functions are placed — and a workload cannot run nearer the antenna than the user plane allows.

Why 3500 MHz Changes the Economics of AI Inference

Pakistan’s $507 Million Spectrum Auction, Seven Months On: What 1,782 5G Sites Mean for AI Workloads at the Edge is, stripped of policy framing, a question about transit cost. Before 2026, Pakistani deployments routed inference to cloud regions outside the country, which meant every video frame and sensor reading paid an international transit penalty in both latency and bandwidth charges. As a result, any workload with a sub-100-millisecond control loop was uneconomic regardless of how good the model was.

Three things change with 3500 MHz. Available capacity per cell rises sharply, the latency floor drops, and operators gain the practical ability to host user-plane functions and inference accelerators at or near the site. Edge inference becomes economically viable when round-trip latency to the compute node falls below the application’s control-loop budget, which the 3500 MHz band delivers and 4G backhaul routed offshore did not. The cost consequence is structural: because inference runs on local data, the buyer stops paying upstream bandwidth on raw streams and pays only for extracted results, which cuts transport cost per camera or per sensor substantially. That reallocates spend from bandwidth to electricity and accelerator hardware — precisely the pressure point the tariff section addresses. Moving computation to the data rather than data to the computation is a long-established principle in large-scale scientific computing, as decades of work at Oak Ridge National Laboratory on data-movement cost in high-performance systems demonstrates; it is not a 5G marketing claim.

Dimension	Centralised cloud inference	5G edge inference
Round-trip latency	High; includes international transit	Low; local to the metro or site
Upstream bandwidth consumed	Full raw stream	Extracted results only
Where raw data resides	Foreign cloud region	Within Pakistan
Behaviour during link loss	Inference stops	Local inference continues
Dominant recurring cost	Transit bandwidth	Electricity and accelerator hardware

theverge.pk – AI Governance and Data Standards

The decisive position: for computer vision, industrial control and anomaly detection in financial transaction streams, edge inference is now the correct default architecture in Karachi, Lahore and Islamabad. For model training, centralised compute remains correct because training is latency-insensitive and throughput-bound. Teams deploying across all three operators’ edge platforms should insist on open standards in AI at the orchestration layer, because three proprietary platforms otherwise produce three incompatible deployment pipelines.

Edge AI Workloads Already Running on the New Network

Most current deployments are pilots rather than commercial services, and they cluster for identifiable reasons. Vision pilots sit inside the three cities with dense 3500 MHz coverage, while sensing deployments use 700 MHz reach where bandwidth demand per node is low. The four active sectors — urban traffic analytics, agricultural sensing, remote healthcare diagnostics and financial or industrial control — divide cleanly along that band dependency.

Where edge AI is being deployed on Pakistan’s 5G network

• Urban traffic analytics — AI camera pilots in Lahore processing video at the edge rather than backhauling it to central data centres.
• Smart agriculture sensing — field telemetry across Punjab, carried on long-range 700 MHz capacity.
• Remote healthcare diagnostics — AI-assisted screening in underserved districts where 700 MHz provides reach.
• Financial and industrial control — banks and plant operators keeping sensitive data localised to cut transit cost and shorten transaction processing times.
• AI-ready Tier-III data centre build-out — operator investment, including Zong drawing on China Mobile’s network-optimisation research capability.

5 Critical AI Governance Challenges in Multi-Institution Research Labs – theverge.pk

The pattern explains itself once the physics is stated. Vision workloads concentrate in dense urban cells because video is bandwidth-hungry and latency-sensitive; sensing workloads spread into agriculture and rural health because telemetry payloads are small and tolerate lower-capacity 700 MHz links. Therefore the spectrum portfolio an operator bought in March 2026 directly determines which class of edge AI customer it can serve in October 2026 — a causal link between auction strategy and addressable market that rarely gets stated aloud. Public-space camera deployments also carry oversight obligations that academic work on vision systems, including AI research at the University of Michigan College of Engineering, treats as inseparable from the technical design. Sensing deployments, meanwhile, succeed or fail on ingestion discipline, which is where automated data analysis pipelines earn their cost.

The Power Tariff Fight That Decides Whether Edge AI Scales

In October 2026 the CEOs of Jazz, Zong, Ufone and PTCL jointly wrote to Prime Minister Shehbaz Sharif warning that commercial electricity tariffs are undermining $2 billion of planned investment in 5G networks and AI-ready data centres, and asking that telecom load be reclassified as industrial. The mechanism behind the petition is straightforward: edge inference moves energy consumption away from a handful of large centralised facilities and distributes it across thousands of sites, while accelerator hardware raises draw per site. Which means the tariff class applied to telecom load now materially changes the cost of every inference performed at the tower.

Take the position squarely. Because Pakistan classifies telecom sites as commercial rather than industrial consumers, the marginal cost of hosting compute at the tower is higher here than in jurisdictions that treat data infrastructure as industrial load. Consequently operators rationally defer accelerator deployment and keep sites as pass-through radio, monetising connectivity instead of compute. The effect is that spectrum capacity now sits ahead of compute capacity by a widening margin, and the gap closes on a meaningful timeline only if reclassification is granted.

Governance Obligations That Arrive With Edge Inference

Distributed inference creates four governance problems that barely exist in a single cloud region: model version sprawl, the absence of a central inference log, data residency ambiguity, and unmonitored drift across thousands of nodes. Moving inference to 1,782 physical locations multiplies the places a model version can diverge, which turns provenance and audit trails from a configuration question into a distributed-systems problem. The NIST AI Risk Management Framework is the voluntary baseline most multi-institution teams map against, and its requirements to document model characteristics and monitor deployed behaviour are exactly the controls that strain at the edge — which is one reason how AI governance differs from traditional data governance matters operationally rather than semantically.

AI – theverge.pk

Teams that have already worked through model provenance challenges in multi-institution AI labs will recognise the following requirement set.

Four governance requirements for distributed inference

• Model provenance — a signed, queryable record of which model version, training dataset and configuration runs on each edge node.
• Inference logging and retention — upstream shipping of decision records before local storage rotates them out.
• Data residency controls — documented proof that raw sensitive data never left the jurisdiction, which is the stated reason many financial deployments chose the edge.
• Drift and performance monitoring — per-node accuracy tracking, because conditions differ between a Karachi intersection and a Punjab field sensor.

Retention is where edge architectures fail quietly. Because edge nodes have finite local storage, inference logs are either shipped upstream or lost, and losing them destroys the evidentiary chain needed for post-incident review — a principle the National Archives and Records Administration applies to electronic records generally, including the authenticity and retention obligations that attach to automated decision records. Documented lineage of the kind catalogued on Data.gov shows what a defensible residency record looks like in practice, while data management plan expectations set by the National Science Foundation establish the reproducibility standard collaborative deployments are measured against. One further dimension deserves naming: pushing proprietary model weights onto operator-owned nodes raises trade secret and ownership questions, and U.S. Patent and Trademark Office guidance on AI-related inventions is the relevant starting point. For implementation detail this section deliberately does not duplicate, see The Verge PK’s AI data governance framework guide.

Limitations of This Assessment and the Alternative Reading

The data has real boundaries. Site counts are operator-reported and ministry-confirmed, not independently measured coverage, and no public disclosure states how many of the 1,782 sites carry 3500 MHz versus lower bands. Independent analysts note that many sites are 5G running on upgraded 4G towers, which means a site count overstates usable low-latency capacity. Capex figures run only to June 2026, so the October position is partly inferred. Several edge AI deployments described publicly are pilots without published performance data, and reproducibility expectations of the kind NSF applies to funded research are not met by pilot announcements.

The sceptical reading deserves a fair hearing. A critic argues the auction was primarily a fiscal exercise, that 1,782 sites in seven months represents a tower conversion programme rather than a greenfield build, and that with the cheapest 5G handset at roughly Rs 36,000 the consumer revenue needed to fund accelerator hardware will not arrive quickly — despite over 607,000 5G-capable handsets being locally assembled between January and August 2026. On that reading, enterprise edge AI stays a three-city niche through 2027 unless tariff reclassification lands.

Conclusion

Pakistan’s $507 Million Spectrum Auction, Seven Months On: What 1,782 5G Sites Mean for AI Workloads at the Edge resolves decisively: the auction bought Pakistan a usable low-latency radio layer, and as a result edge inference is now the correct architecture for computer vision, industrial control and transaction-monitoring workloads in Karachi, Lahore and Islamabad. That is a narrower claim than national 5G, and a far stronger one than the pre-2026 position, where offshore transit made those workloads uneconomic.

One variable decides the next phase: whether telecom and data centre load is reclassified from commercial to industrial power tariffs. Granted, the $2 billion in planned investment has a path to accelerator deployment at the tower. Refused, spectrum capacity stays ahead of compute capacity. Read more in technology coverage from The Verge PK.

Frequently Asked Questions

What did Pakistan’s 5G spectrum auction raise and what has it delivered so far?

Approximately $507 million, and 1,782 live 5G sites. Pakistan’s $507 Million Spectrum Auction, Seven Months On: What 1,782 5G Sites Mean for AI Workloads at the Edge now has concrete figures: 480 MHz of the 597.2 MHz offered sold in the March 10–12, 2026 PTA auction, national holdings up from 274 MHz to 754 MHz, and 1,782 sites across 24 cities as of late October 2026.

Which operator has the most 5G sites in Pakistan?

Jazz, with 1,015 live 5G sites. Zong follows at 640 and Ufone at 127. Jazz’s lead reflects faster conversion of existing 4G towers plus its 190 MHz multi-band holding, while Zong spent more per site — $98.9 million to June 2026 against 640 sites — on core and transport upgrades rather than site count.

Why does the 3500 MHz band matter for AI workloads?

Because it supplies the capacity and low latency floor that real-time inference requires. Sub-100-millisecond control loops were uneconomic on 4G backhaul routed to foreign cloud regions, so moving inference to the tower edge cuts both latency and upstream bandwidth cost. All three operators bought 3500 MHz, which means competitive enterprise pricing rather than one supplier setting terms.

Does 1,782 sites mean Pakistan has national 5G coverage?

No. A site is a physical installation, not a coverage guarantee, and independent analysts note many of the 1,782 sites are 5G running on upgraded 4G towers. Karachi, Lahore and Islamabad hold roughly 58 per cent of all sites, which means dependable edge AI service levels are currently restricted to specific urban zones in those three cities.

What is blocking wider edge AI deployment in Pakistan?

Electricity tariffs and device affordability, not spectrum. In October 2026 the CEOs of Jazz, Zong, Ufone and PTCL warned Prime Minister Shehbaz Sharif that commercial power tariffs are undermining $2 billion in planned 5G and AI data centre investment. Separately, the cheapest 5G handset still costs around Rs 36,000, which caps consumer-side demand.

How much have operators invested in 5G upgrades so far?

$178.6 million in site upgrades to June 2026 — Zong $98.9 million, Jazz $59.67 million, Ufone roughly $20 million — on top of auction payments. Jazz has separately committed $1 billion over three years to core network modernisation, which is the layer determining how close to the radio an inference workload can be placed.

What governance controls should an edge AI deployment have?

Signed model provenance records per node, upstream inference logging, documented data residency and per-node drift monitoring. Distributing inference across thousands of sites multiplies the places a model version diverges and where decision logs are lost to local storage rotation. The NIST AI Risk Management Framework is the voluntary baseline; The Verge PK’s governance framework guide covers implementation.

References

• Pakistan Telecommunication Authority — NGMS spectrum auction results, March 10–12, 2026; licences signed March 19, 2026.
• Ministry of IT and Telecommunication, Government of Pakistan — statement by Federal Minister Shaza Fatima Khawaja confirming 1,782 live 5G sites across 24 cities, late October 2026.
• Operator disclosures — Jazz (VEON), Zong (China Mobile Pakistan), Ufone (PTCL Group): spectrum acquisitions, site counts and 5G capex to June 2026.
• National Institute of Standards and Technology — AI Risk Management Framework.
• National Archives and Records Administration — electronic records retention and authenticity standards.
• Oak Ridge National Laboratory — large-scale scientific computing and data-movement cost research.
• University of Michigan College of Engineering — artificial intelligence research, computer vision and sensing.
• National Science Foundation — data management plan and reproducibility expectations.
• Data.gov — open data cataloguing and provenance documentation practice.
• U.S. Patent and Trademark Office — guidance on AI-related inventions and proprietary models.

Related Reading

• Home – theverge.pk
• What Are Self-Driving Labs? – theverge.pk