Qualcomm struck a deal to acquire AI software company Modular for about $3.9 billion, in all stock-deal aimed at strengthening its software foundation for generative and agentic AI across data centre and edge environments.

The deal will pair Qualcomm’s silicon expertise with Modular’s AI-native software platform, which is designed to run AI models efficiently across CPUs, GPUs, NPUs and custom ASICs without requiring separate rewrites for each type of accelerator.

The chipmaker stated developers and enterprises will only need to build once ahead of deploying across multiple environments at a lower total cost of ownership.

Qualcomm framed the acquisition as a response to a shift in what is constraining AI deployment at scale. As AI use grows, the company argued, efficiency rather than raw capability becomes the limiting factor, since performance-per-watt drives inference costs and those costs determine what can scale.

Closing this gap, according to Qualcomm, requires software that connects system-level optimisation with heterogeneous, disaggregated compute, turning silicon performance into AI services which work reliably across different accelerators and use cases.

The deal is intended to expand Qualcomm’s data centre ambitions, supporting more efficient inference, orchestration and deployment in distributed AI systems while deepening relationships with model creators, developers, hyperscalers, and enterprise customers.

It is also expected to help the company achieve strong day-zero performance on new Qualcomm AI hardware as it rolls out.

Qualcomm President and CEO Cristiano Amon called the acquisition a pivotal moment for the AI industry. He stated agentic AI’s growth across data centres and edge environments is pushing the sector toward disaggregated, multi-vendor architectures which need a more open, modern software foundation.

He said Qualcomm believes the future belongs to developer-friendly platforms which can run across diverse compute environments, giving customers real choice in how and where they deploy AI.

“With Modular, we’re accelerating that shift, combining our scale and energy-efficient data center technologies with an open ecosystem approach to help drive the next chapter of AI,” Amon stated.

Modular co-founder and CEO Chris Lattner stated the company was built on the belief AI needs a more open and efficient software foundation spanning diverse hardware and deployment environments.

The transaction is expected to close in the second half of 2026, subject to customary closing conditions and regulatory approvals. Qualcomm plans to issue 19.2 million shares to Modular equity holders.

The post Qualcomm inks $4B deal for software specialist Modular appeared first on Mobile World Live.

Verizon’s road to Level 4 network autonomy was boosted by turning 33,000 employees into software developers earlier this year using Anthropic’s Claude model.

Verizon CTO Yago Tenorio (pictured) told Mobile World Live (MWL) the company gave about 500 application developers on his team access to Claude Code in December 2025.

“We started using it like there is no tomorrow,” he said. “In a few weeks we made everyone in technology proficient as software developers and then people started coding.

“There is a mantra now that says: ‘If you can imagine it, Claude can code it for you.’ It is amazing the massive amount of enablement, simplification and efficiency we achieved in a few months.”

He explained Verizon’s legacy as a long-established operator means its path to network autonomy is far more complex compared to a greenfield player. Years of M&A, overlapping legacy systems, merged networks and coexisting OSS platforms have created operational complexity which must be abstracted away before AI-driven automation can work at scale.

The AI autonomy revolution is well underway. In 2025, Verizon’s closed-loop automation platforms executed over 70 million network configuration changes autonomously.

“I feel like it’s making the network better. It is reacting in real time and it is freeing up hours of labour that can be then thrown on something that requires a human.”

But Tenorio draws a line between that kind of automation and what’s coming next. Those 70 million loops were largely programmatic, following decision trees and fixed logic which engineers wrote in advance, the network equivalent of a Python script running the same routine repeatedly.

The shift now underway replaces fixed logic with AI agents which can reason through a situation they were never explicitly programmed for because they are trained on the skills and judgment which make Verizon’s network teams effective, rather than coded with rules for every scenario.

Verizon is moving away from traditional hardware constraints to turn its infrastructure into a software-defined, self-organising reasoning engine which optimises itself in real time. The company describes this not as simply adopting tools, but as engineering an AI-native enterprise.

Level 4 network autonomy
In the TM Forum’s six-stage taxonomy for autonomous networks, Level 4 is a major shift from traditional automation to true intelligence. For this level, the network moves away from rigid, rule-based scripting and task-execution, transitioning instead to automating decisions.

CCS Insight research director Ian Fogg told MWL a part of an operator’s network may be at one level of autonomy, but other parts will likely be at a different level depending on how far along each operator’s suppliers and network teams are in modernisation.

“The barriers are that this is simply a very complicated area,” he said. “AI and especially generative AI and agentic AI are increasingly at the heart of autonomous network innovation as operators move to level 3 and beyond. 

“AI is an essential technology to enable the creation of autonomous networks,” he added.

Tenorio is careful not to overclaim on autonomy levels. He said Tesla represents something closer to Level 3, a Waymo represents Level 4, and Level 5 would be a car with no steering wheel at all.

He is not convinced the network needs to go Level 5. In his view, building a network with literally no human element anywhere in it is not necessarily the goal, even if individual domains or use cases can get there.

Across the board, in every domain and every use case, he considers Level 4 a reasonable target, while cautioning operators in the industry tend to generalise their autonomy successes across specific areas.

Rather than chase a number, Tenorio stated Verizon’s focus is on making autonomy matter operationally, whatever level it ends up being called.

He also pushes back on the idea that full autonomy means removing humans from the picture entirely. Even with what he calls a “totally dark” network operations center, people must still be dispatched to base stations and customer premises for hardware replacement or tuning which cannot be done remotely.

He argues agents are trained, not programmed, meaning the human expertise which has shaped a network for years does not disappear, it just gets transferred into the skills the agents learn from.

The loop which runs the network day-to-day can become fully closed, but the loop of deciding how those agents should behave still runs through human judgment.

AI models
To get there, Verizon has doubled down on generative and agentic AI. Tenorio describes Verizon’s approach as model-agnostic. For agents now running parts of its network, it is primarily using Anthropic’s Claude and Google’s Gemini.

Verizon is also optimising across model size and cost: simpler problems get routed to small language or cheaper models, while more complex tasks get the most capable ones available, including Claude Opus 4.7

Verizon had access to Anthropic’s Claude Fable 5 model prior to the US government shutting it down.

Tenorio used Fable 5 on a side project built with Claude Cowork and came away impressed, particularly by what felt like a broader contextual awareness and more lateral reasoning than he was used to. However, he is cautious about claiming a measurable gap versus Opus 4.8, since the same project has continued just as well on Opus 4.8 since Fable 5 access ended.

“I thought it was fantastic,” he said of Fable 5. “I was blown away by it.”

Starting in mid-January, the company expanded Claude access to its entire 33,000-person technology organisation over about six weeks, turning employees across modelling, capacity planning, capital deployment, and operations into capable software developers regardless of their formal coding background.

It now counts between 7,000 and 10,000 weekly active users of the tool. Tenorio frames this as a deliberate first step on the road to autonomy, not a side project.

Before trying to make the network autonomous, Verizon wanted to find out what happens when everyone in the organisation can write software.

Architectural blueprint
Verizon is currently conducting an autonomous network programme with intelligent agents running full-time, although Tenorio resists calling the effort a pilot.

He instead describes it as a deliberate, already-in-production evolution. The same 70 million automated loops from last year are progressively having reasoning and token-based decision-making layered on top of them, with the foundational pieces of true network autonomy now live rather than experimental.

He identifies two areas for genuine autonomous network operations.

The first is a common data layer. Since an autonomous network is only as good as the data feeding it, the telemetry needs to be shared across the organisation in near real time, and it needs to reflect actual customer experience rather than what the network itself reports.

Tenorio stated Verizon began rolling this out at the start of June and it is now in production.

The second is a common repository of skills, tools and agents, the framework by which years of collective engineering expertise and pattern recognition gets encoded into a form AI agents can draw on.

Once both pieces are in place, Tenorio stated deploying domain-specific agents becomes comparatively straightforward.

He said this approach has no off-the-shelf component and is not organised around a specific network domain or generation. It is not built around radio, transport, core, or optical separately, and it is not tied to 5G or 6G.

Instead, Verizon designed it to run across the network as a whole, starting with making every employee a developer rather than buying a vendor’s packaged autonomy product.

A master agent can immediately detect a localised anomaly, take control and instantly spin up a network of specialised sub-agents focused on the exact root cause, pinpointing it down to the precise network domain and specific element.

Work which used to take engineering teams hours to manually diagnose and correct is now identified and resolved in less than two minutes, long before the end user ever experiences a dropped call or a lag in their data connection.

An autonomous system, however, is only as good as the live data feeding it. Traditional network KPIs tend to skew toward favourable conditions and do not always fully capture the true, localised end-user experience.

Verizon’s new AI-driven models aim to close the telemetry gap. By using aggregated, anonymised network performance signals, including floor-by-floor indoor environments, the company builds an accurate, hyper-granular picture of service quality, feeding the common data layer described above.

The rich physical telemetry, fused with existing network data, feeds a continuous optimisation loop which detects and resolves coverage and capacity issues proactively, often before a customer is affected.

Tenorio said the end-result is near real-time visibility into in-building performance environments, enabling automated alerting and network tuning at a level of granularity the company did not previously possess.

“We started by making everyone a software developer,” he said. “I don’t know if other companies are doing it.”

The post Feature: How Verizon is crafting its network for the agentic AI era appeared first on Mobile World Live.

The US Federal Communications Commission (FCC) closed an AWS-3 re-auction with gross winning bids topping $3.5 billion in what was the agency’s first spectrum sale after a four-year lapse in activity.

Bidding concluded for Auction 113 yesterday (23 June) after 72 rounds for the 200 licences across the 1695-1710MHz, 1755-1780MHz and 2155-2180MHz bands.

All the licences received bids in the process, which had 17 qualified participants and began on 2 June.

The frequencies were originally auctioned to Dish Network, which is now part of EchoStar, in 2014 but never made it into service after a series of defaults and bid withdrawals left them sitting unused in the FCC’s inventory for more than a decade.

In 2015, Dish Network affiliates Northstar Wireless and SNR Wireless surrendered a number of spectrum licences worth $3.5 billion after a dispute with the FCC over discounts.

In April, the FCC and EchoStar reached an agreement which included the latter dropping a lawsuit it filed in a US Court of Appeals over the defaults by Northstar and SNR.

The financial total carries significant regulatory implications. Up to $3.3 billion of the proceeds will be legally directed to cover funds borrowed for the FCC’s Secure and Trusted Communications Networks Reimbursement Program, commonly known as the “rip and replace” initiative to swap out Huawei and ZTE gear from US networks.

Surpassing the $3.5 billion mark clears the $2.9 billion threshold required for EchoStar to avoid paying a potential shortfall payment to the regulator under a prior settlement framework.

While specific winning bidders remain anonymous under strict anti-collusion quiet period rules, AT&T, Verizon, T-Mobile US and SpaceX are widely expected to have driven the demand.

The event marks a milestone for FCC chair Brendan Carr, representing the return of the agency’s auction power.

“After years on the sidelines, FCC auctions are finally back,” Carr stated. “Today’s successful auction generated billions of dollars in competitive bids to put spectrum to effective commercial use, and it bolsters competition in the wireless marketplace.”

The FCC plans to carry the momentum into its upcoming Upper C-Band auction, tracking on schedule to meet a congressional deadline to auction at least 100MHz of the spectrum by July 2027.

The post FCC wraps AWS-3 re-auction with $3.5B in bids appeared first on Mobile World Live.

SpaceX reportedly signed a significant computing power deal with Reflection AI, making the open-source start-up the latest outside company to tap Elon Musk’s Colossus infrastructure.

CNBC reported Reflection will get access to Nvidia’s GB300 chips, the company’s top-tier hardware for training and running advanced AI models, housed at SpaceX’s Colossus 2 facility in Memphis, Tennessee.

Reflection agreed to pay SpaceX $150 million per month starting 1 July through 2029, with total payments reaching roughly $6.3 billion if the agreement runs its full term.

Either party can terminate the contract with 90 days’ notice after the first three months. Bloomberg reported Reflection will pay SpaceX’s AI division, SpaceXAI, under the arrangement.

Colossus was originally built in part to power Grok, Musk’s AI chatbot as a rival to ChatGPT, but SpaceX has increasingly opened the infrastructure to outside AI companies seeking compute capacity.

The deal is smaller than two of SpaceX’s other compute agreements. According to TechCrunch, SpaceX’s existing contracts with Anthropic and Google cost those companies $1.25 billion and $920 million per month, respectively, with both agreements also running through July 2029.

CNBC reported SpaceX has also struck a computing power-related deal with Cursor prior to buying the company for $60 billion earlier this month.

In October 2025, Nvidia led a $2 billion funding round for the open-source start-up, directly contributing $800 million of the total to value Reflection AI at $8 billion.

Reflection AI was founded in 2024 by Misha Laskin and Ioannis Antonoglou, both alumni of Google’s DeepMind lab.

The post SpaceX inks computing deal with Reflection AI appeared first on Mobile World Live.

Groq raised $650 million in funding as the company revamps itself from a chip startup into a data centre operator focused on AI inference.

The round was led by growth-focused investment company Disruptive and hedge fund Infinitum, both of which already hold board seats at Groq.

The funding follows a non-exclusive licensing agreement Groq struck with Nvidia in December 2025. At Nvidia’s GTC event earlier this year, the chipmaker unveiled its next-generation LPX platform, which incorporates Groq’s inference technology.

US-based Groq stated it plans to use the new cash to expand its network of 13 data centres in North America, Europe, the Middle East and Asia-Pacific, and quadruple total capacity to 200 megawatts by the end of 2027.

The new capital will also fund the fit-out of its existing footprint with Groq’s latest inference technology, including Nvidia’s new LPX system.

In September 2025 Groq raised $750 million, which valued it at $6.9 billion. Its backers include KDDI, Samsung Electronics, Cisco Systems and BlackRock.

Groq chair Alex Davis stated the company spent years building the technology, infrastructure and operational expertise needed for AI’s next phase, adding it is positioned to become a foundational layer of the AI economy.

Groq also announced several executive additions alongside the funding news. Alan Rice, who previously worked on xAI’s Colossus data centre project and Meta data centres, is now COO.

Sinclair Schuller will start as CTO in July, with Rakesh Malhotra joining as chief product officer the same month. Schuller and Malhotra previously worked together at enterprise cloud platform Apprenda and later co-founded a separate software engineering firm.

The post Groq lands $650M to aid shift from chips to AI Inference appeared first on Mobile World Live.

Nvidia and telecom partners including SoftBank Corp and NTT Data plan to demonstrate a stack of data, models, simulation tools and secure runtimes at DTW Ignite 2026, aimed at helping operators move from task-based automation to fully autonomous network operations.

The company argues while generative AI has delivered strong returns in network management, customer care and back-office automation, most of the impact has been limited to speeding up predetermined steps while humans still correlate insights and direct next actions.

Nvidia’s claims automation is now “the launchpad to autonomy” rather than the end goal, with AI agents expected to proactively monitor for problems and coordinate changes across network, IT and business systems while keeping humans in control of policy.

Nvidia cites figures showing 54% of telecom operators view data-related issues as their biggest obstacle to building telecom-specific AI models, largely because the most valuable network and customer data is too sensitive to use directly for training.

SoftBank Corp is tackling the problem by using Nvidia’s NeMo Safe Synthesizer and NeMo Anonymizer tools to generate privacy-preserving synthetic datasets which mirror the structure of real network performance and configuration datasets.

Those datasets are being used to fine-tune its Large Telecom Model (LTM) and build specialised network agents.

On the runtime side, Nvidia stated its NemoClaw blueprints and OpenShell secure runtime are designed to give long-running agents policy-based guardrails and sandboxed access to telecom systems, allowing operators to expand agent responsibilities while keeping behaviour auditable and governed.

NTT Data is applying Nvidia’s open Nemotron models alongside NemoClaw to build agents which track long-term network performance trends and escalate anomalies to specialised research agents for deeper telemetry analysis.

Other partners building on the same framework include: AdaptKey, which is piloting self-healing 5G agents with operators; ServiceNow, which is bringing its Project Arc incident-response tooling to telecom network operations centres; and Tata Consultancy Services, which is developing a multi-fidelity AI sensor architecture for faster anomaly detection.

Simulation is the third pillar of the effort, intended to let agents test recommendations in near-real-time digital environments before acting on live systems.

Software company Forsk stated its GPU-accelerated radio propagation model now achieves ray-tracing-level accuracy up to 200 times faster than CPU-only systems, while Viavi Solutions reports order-of-magnitude gains in RAN simulation throughput after shifting workloads to Nvidia’s RTX Pro 6000 Blackwell GPUs.

KDDI and KDDI Research, working with Nvidia, Keysight and Samsung Research America, are building a high-fidelity RAN digital twin aimed at the 6G era, allowing multiple autonomous agents to simulate scenarios for future radio conditions and traffic shifts.

The TM Forum’s DTW Ignite conference runs 23–25 June 2026 in Copenhagen.

The post Nvidia rallies telcos around AI agents at DTW Ignite appeared first on Mobile World Live.

Mavenir unveiled an Agentic Service Assurance Framework designed to automate complex network operations across multiple domains while allowing operators to retain existing infrastructure.

The vendor stated the TM Forum-compliant, multi-agent system combines an intent orchestrator with a layered agentic ecosystem to help detect, diagnose, recommend and resolve network faults faster.

Mavenir positioned the framework as a step toward higher operator productivity and autonomous network operations.

The vendor said network operations across 5G, cloud-native and IP domains are becoming increasingly fragmented and expensive to manage at scale. Traditional assurance systems can identify alarms but often struggle to determine root causes, while the retirement of experienced engineers risks taking years of institutional knowledge out of operations teams.

Multi-vendor environments add another layer of complexity by creating data silos which individual tools cannot easily correlate, according to Mavenir.

At the heart of the framework is Mavenir’s Intent Ops engine, which observes how network operations centre teams work, learns patterns behind successful issue resolution and converts those workflows into repeatable automation templates with explainability and guardrails.

The company explained the workflows are validated against a human baseline before running autonomously, with vetted automations published to an expanding orchestration catalogue.

The framework supports cross-domain monitoring, correlation and root-cause analysis, with closed-loop remediation built on Mavenir products.

It also connects with existing OSS AI agents through the agent-to-agent protocol defined in TM Forum’s IG1453, which Mavenir stated enables operators to preserve prior technology investments while extending automation across the network stack.

Mavenir CEO Pardeep Kohli told Mobile World Live the foundation AI agents are designed to automate 60% to 75% of routine network operations work, to help operators lower costs, improve service quality and reduce the need to train staff on repetitive tasks.

“In a way, you have a kind of personal assistant always watching you, but this personal assistant is for the network,” he said.

Mavenir did not say when it will be available but a representative stated the initial release is being trialled in customer deployments and future feature releases will be delivered through software upgrades.

The post Mavenir debuts agentic automation framework appeared first on Mobile World Live.