On July 11, 2026, five researchers from Nanyang Technological University published the results of an experiment that should get the attention of the Telecom community.

The researchers built a multi-agent system called iFinder and pointed it at seven open-source 4G and 5G core implementations. The agents found 84 previously unknown vulnerabilities. Developers confirmed 83; 81 received CVE numbers, and 58 were patched; still, 23 confirmed vulnerabilities lacked a code fix.

What makes the research important is how the vulnerabilities were found. One agent searched the code for suspicious trust assumptions, another checked the behavior against 3GPP specifications, and a third generated and tested exploits until it could prove whether the weakness was real. A process that traditionally requires scarce security expertise can now be repeated by tireless Agent Swarms across thousands of components, versions,sions and interfaces.

As mentioned many times before, Telecom security is entering a completely new era. Vulnerability discovery is moving toward machine speed, while remediation still runs through engineering teams, vendor releases, regression testing, staging, and maintenance windows measured in days or weeks. If attackers can search continuously, defenders cannot afford to keep humans in every first response loop.

What the AI Agent Actually Found Inside the 5G Core

The most serious finding was a session hijacking flaw in the 5G core. It affects the relationship between the Session Management Function and the User Plane Function, which are connected via the N4 interface using PFCP. The SMF tells the UPF how to identify a subscriber’s packets and where to forward them. Those instructions include Packet Detection Rules, or PDRs, that become part of the subscriber session.

The researchers found that one implementation would accept a PFCP Session Modification Request containing a PDR identifier already being used by another subscriber session. The malicious rule could then be assigned a lower PFCP Precedence value, which counterintuitively means higher priority. The UPF would evaluate the attacker’s rule first and associate the traffic with a forwarding action controlled by the attacker. Instead of simply crashing a network function, the vulnerability could redirect a subscriber’s uplink traffic.

There is one important condition for a session hijack: the attacker must first gain access to the internal PFCP interface. Historically, N4 was protected largely because it sat inside a trusted core domain. But the same research showed how that assumption can fail. In five of the seven open-source cores tested, a device with a valid SIM could tunnel control messages via GTP-U and reach interfaces that were supposed to remain internal.

This assumption becomes much weaker in a cloud-native core. Network functions now operate across virtual infrastructure, containers, Kubernetes clusters, and increasingly complex network policies. The paper specifically considers internal interfaces becoming reachable through cloud configuration errors, but it also demonstrates a more interesting route from the subscriber side. A malicious device with a valid, ordinary SIM can place crafted PFCP or GTP-C messages inside the GTP-U tunnel carrying its own uplink traffic. If the boundary between the user plane and internal control interfaces is insufficiently enforced, those messages can cross into parts of the core they were never intended to reach. The researchers successfully demonstrated this protocol tunneling and boundary bridging technique against five of the seven open-source implementations.

Read more

Three operators on three continents moved on AI Voice within six weeks. Ooredoo Qatar launched AI-powered outbound sales calling using Microsoft Azure OpenAI. Deutsche Telekom presented the Magenta AI Call Assistant at MWC Barcelona, built with ElevenLabs and embedded in the network rather than in an app. Orange Business followed with branded calling, deepfake detection and agentic telephony, while disclosing that it had already handled 80 million AI conversations across 7,000 enterprise customers in 2025. The products were different, but all three were making the same bet on Voice.

Generative AI changes what a phone call can do. The old automated experience was built around menus and routing. The new one can understand what the customer wants, access the systems behind the business, and execute part of the task. Voice can become an interface for sales, service, and enterprise workflows rather than simply another transport service, and operators are starting to build around that opportunity.

Elon Musk is moving into the same space.

When xAI launched its Voice Agent Builder in July, it advertised $0.05 per audio minute, plus $0.01 per minute for telephony on a provisioned number. Grok Voice Think Fast 2.0 is now listed at $0.08 per audio minute. Those figures are not the full cost of running an enterprise contact center because integration, compliance and the underlying systems still have to be paid for. But they set a public price, in cents per minute, for a capability that Telcos are preparing to sell as a premium enterprise service.

Musk also has a live telecom business in which to prove it. In August, SpaceXAI said Grok Voice was resolving more than 15,000 inbound Starlink support and sales calls a day and fulfilling more than 3,000 orders a week across Voice and chat. The Starlink agent converted around 20% of sales inquiries and resolved 70% of support inquiries without a human in the loop, while using 28 tools across hundreds of workflows including hardware diagnosis, replacement shipments and service credits.

In those 28 tools released by these Starlink Agent there are interesting aspects to understand.

A traditional contact center agent listens to the customer and then moves through CRM, billing, order management, logistics and support systems until something gets resolved. Grok is beginning to automate that navigation. A customer explains that a Starlink terminal is not working; the agent diagnoses the issue, calls the relevant tools, and can initiate a replacement or credit. Voice gets the customer into the process, but the value comes from what happens behind the conversation.

I do not think Musk is viewing Voice primarily as another telecom revenue stream. Voice is a practical spearhead for the “28 Trillion TAM” AI-native enterprise. Every call contains intent that must eventually become an action within the business. If Grok becomes the layer that understands that intent and connects into CRM, payments, orders, scheduling, logistics, or support, the opportunity is no longer measured by how many Voice minutes xAI sells. It is measured by the amount of enterprise work that passes through Grok.

Read more

Last week I sat in a meeting between a Tier One telco and one of the frontier AI companies to discuss partnership opportunities.

The AI model company quickly moved the conversation to enterprise distribution.

Sadly, the telco’s immediate proposition was the old playbook. Bundle the model, add it to a plan, and add another logo to the catalog. I stopped the executive there. AI is not another WhatsApp bundle. The opportunity is much bigger.

Frontier AI companies have created an enormous economic problem for themselves.

Anthropic is preparing to tell investors that AI can address a market worth more than $30 trillion, with roughly $22.7 trillion coming from enterprise applications alone. OpenAI, meanwhile, is planning to spend around $600 billion on compute through 2030, after generating $13.1 billion in revenue in 2025. These companies are no longer being valued against the software market. They are being valued against a meaningful share of the global economy.

But there is a significant gap between what model companies promise and what is actually happening within enterprises. PwC found that 56% of CEOs still report no significant financial benefit from AI, while other studies show enterprises spending millions and struggling to turn pilots into measurable P&L impact.

That changes the distribution question completely. Frontier models do not simply need more users. They need someone capable of taking intelligence deep into enterprises across regulation, sovereignty, data, performance, and implementation. That is where telcos suddenly become interesting.

The Models Are Not the Bottleneck Anymore

The economics behind frontier AI explain why enterprise adoption has become economically critical for companies like OpenAI, Anthropic, Mistral, and Cohere.

Those fancy AI data centers do not run on $20 subscriptions.

OpenAI’s audited 2025 figures, reported by the Financial Times in June 2026, showed $13.07 billion in revenue, $34 billion in total costs, and a $20.9 billion operating loss. The detail that matters most is R&D, which reached $19.18 billion. That single line was roughly 147% of everything the company earned that year. You can build extraordinary technology that way, but eventually the revenue curve has to catch the cost curve. Consumer subscriptions, API calls, and a few thousand large enterprise accounts will not carry the valuation story forever.

Read more

Every telco wants AI agents. Marketing wants one to reduce churn. Care wants one to resolve complaints. Finance wants one to optimize investment. The NOC wants one to find root causes. RAN wants one to optimize performance. There is only one problem. None of them necessarily agree on what a customer, service, network problem, or even revenue actually means.

Giving them more data will not solve it. An LLM can read millions of records and an agent can call thousands of APIs, but neither inherently knows that Customer A owns Product B, which contains Service C, which depends on Resource D, which is currently affected by Alarm E, while Policy F prevents the obvious corrective action. Data contains the values. Ontology gives those values meaning, relationships, and constraints.

This is no longer semantic web bullcrap. In June 2026, TM Forum published TR326, proposing a layered semantic architecture for AI-native autonomous networks to enable runtime contextual reasoning, semantic interoperability, and explainable operations. Telstra has already built a Knowledge Plane around a telco ontology derived from TM Forum SID, while TM Forum’s 2026 A2A T work combines agents with knowledge graphs because agents from different vendors still struggle to share operational meaning.

This points to a much larger architectural shift driven by AI. The model is not the brain; the agent is not the brain; APIs are not the brain. The brain starts with the semantic representation of the Telco itself.

If your AI strategy includes models, agents, RAG, APIs, and data lakes, but no one can show you your ontology layer, you may not have an AI architecture yet. You may simply have a collection of AI projects.

Data Is Not Context

Telcos have spent years centralizing data, building lakes, exposing APIs, and creating increasingly sophisticated analytics platforms. None of that creates a shared understanding of the business.

A data platform can tell an AI system that CUST_98211, SRV_4482 and CELL_319 exist. An ontology indicates that one is an enterprise customer, another is a service contracted by that customer, and a third is a network resource supporting that service. It can also specify that the service has an SLA, that certain faults constitute degradation, that specific policies restrict the actions that can be taken, and that an outage affecting this resource has commercial consequences.

That difference becomes critical when AI moves from prediction to decision-making.

Read more

Marcus Weldon ran Bell Labs and served as Nokia’s chief technology officer until 2021, so when he loses his temper about Bell Labs in public, the industry should pay attention.

In a LinkedIn post this week, he estimated that Nokia has cut roughly half of the research jobs at Bell Labs since he left, from more than 1,200 people to perhaps 600, and called the contraction shocking and unprecedented. Nokia replied that Bell Labs remains a deeply important part of the company and is entering a new chapter under Guru Parulkar, with its research teams unified around breakthrough science for the AI era. Both statements can be true at the same time. Bell Labs is smaller, and Nokia is reorganizing what remains.

I share Weldon’s sadness about what has been lost. But I don’t think Nokia is the problem. Nokia picked up Bell Labs a decade ago in its €15.6 billion takeover of Alcatel-Lucent, and it inherited an institution whose economic engine had disappeared decades earlier. Expecting a company with €19.9 billion in annual sales to recreate the Bell Labs of the AT&T monopoly ignores the financial logic of today´s telecom, but it surely exposes the biggest weakness in our industry.

It is worth remembering our glorious past. Bell Labs invented the transistor in 1947 because vacuum tubes were becoming a maintenance problem for the telephone network, and in trying to fix telephony it accidentally launched the semiconductor industry. Claude Shannon wrote information theory there. Its scientists made foundational contributions to the laser, built Telstar for its 1962 launch, developed the CCD image sensor, and wrote Unix and the C programming language.

Decades later, Yann LeCun’s work at AT&T Bell Labs showed that convolutional neural networks could read handwritten digits on real cheques, a direct ancestor of modern AI. Work associated with Bell Labs has earned ten Nobel Prizes and five Turing Awards. In the 1960s, NASA created Bellcomm to borrow systems engineering talent from AT&T, Bell Labs, and Western Electric for the Apollo program.

Please don’t get me wrong here. AT&T’s engineers were no smarter than today’s, but the difference was the economic machine underneath them.

The Bell System controlled most of the American communications value chain, from research and manufacturing to the network and the customer, and a small slice of every telephone dollar flowed continuously into research. In 1974, AT&T booked $26 billion in revenue, about 1.4% of US GDP, and Bell Labs alone spent more than $500 million on nonmilitary R&D, roughly 2% of that revenue, while Western Electric spent even more on engineering and development. More than four cents of every AT&T dollar went into R&D that year.

The stability of the money and its vast amount meant that a researcher knew the budget would still exist in fifteen years and could work on problems that would not produce a product for fifteen years.

Telcos Outsourced the R&D of a $1.3 Trillion Industry

Telecom deliberately dismantled that structure, for the most part, for good reasons.

The 1984 breakup of the Bell System, liberalization in Europe and Asia, and the arrival of competition pushed operators to concentrate on customers, spectrum, deployment and operations, while Ericsson, Nokia, Huawei, Qualcomm, ZTE, Samsung and a long tail of smaller suppliers took over most of the underlying technology development.

Read more

The Nordic ceasefire is over

For years Ericsson and Nokia fought over different versions of essentially the same machine. Radios got lighter, basebands got faster, Massive MIMO got better, and each generation arrived with a fresh argument about watts, capacity and spectral efficiency, conducted between two companies that shared every underlying assumption about what a base station was. Those days are gone. Ericsson still believes radio is peculiar enough to justify purpose-built silicon and continues to fund its own. Nokia has said the Marvell-based chips in its current products are the last custom silicon it will ever develop, and it is tying a large part of its 5G and 6G roadmap to Nvidia accelerated computing and CUDA.

The stakes are symmetrical even if the evidence, for now, is not. If Nokia can demonstrate anything close to a 100% improvement in spectral efficiency at commercial site power and cost, Ericsson’s purpose-built strategy is finished, along with the Networks margin structure that sits on top of it, and Ericsson becomes a software tenant on a platform its rival helped design. If Nokia cannot, it has already dismantled its only alternative: cutting the laboratory that was supposed to produce the algorithms, and it runs a radio business with very little margin to absorb a failed bet. Both companies have wagered the RAN business on a physics question that will be answered on live networks between now and 2028.

That is key to operators because RAN equipment outlives the people who bought it.

An operator can trial Ericsson AI, Nokia GPU RAN, and Nvidia in parallel, and most large ones will, but trials are cheap, and architecture is not. Once thousands of sites are built around one path, the power design, the software, the skills, the upgrade cycles, and the economics lock in, and equipment bought in 2027 will still be running when 6G arrives. For a decade, technology agnostic was a useful procurement posture. This time it risks delaying a choice that the installed base will eventually make for you.

Both sides have strong arguments

Read more

The endgame of the SaaSpocalypse

On August 26, in the middle of its Q2 FY27 earnings call, Salesforce and Anthropic announced Claudeforce. The headline product is Salesforce in Claude, a plugin with 37 prebuilt sales skills that lets a seller review pipeline, check deal health, and update records without opening Salesforce; live with pilot customers now and in open beta from September. Underneath sits what the joint press release calls AIforce, the enterprise harness that exposes Salesforce data, workflows, business logic, and governance to agents through MCP servers, APIs, and CLI tools. The pipes were laid months earlier, with Headless 360 announced in April 2026 and Hosted MCP Servers in beta since July, so Claudeforce is mostly the commercial packaging of an architecture Salesforce had already built.

The timing is extremely interesting to understand what is going on here.

Investors spent 2026 trading a thesis the press named the SaaSpocalypse. The selloff started in late January after Anthropic launched Claude Cowork, and tallies cited by Forbes put the damage to the software industry at more than $2 trillion by April. By early August, CNBC reported that Airtable had sold itself for about one-tenth of its peak valuation, while HubSpot, Datadog, and Figma fell on earnings. The idea itself is not new either, since Satya Nadella argued back in December 2024 that business applications collapse into the agent tier. What changed on August 26 is that the loudest denier conceded the architecture.

And Salesforce did deny it, loudly. Benioff told the Wall Street Journal in April that the opportunity had never been bigger, and he opened the Claudeforce launch by saying, “This nonsense of the SaaSpocalypse, I think it’s time for it to stop.” He then presented a product whose main selling point is that a seller will never have to open a Salesforce tab again, and the stock jumped in after-hours trading.

Ironically, both moves are rational at once. The application layer is losing value, and Salesforce intends to survive by keeping the data, business rules, and permission checks on its side of the MCP connection. Telcos should study this very carefully, because OSS/BSS is the most extreme version of the same condition.

The employee was always part of the integration layer

For fifty years, most enterprise software has been a translation layer, built because machines could not understand what humans wanted.

Read more

Telcos Put the Edge on a Map

Telcos have spent years defining edge computing by where it sits in the network. Central offices, metro data centers, RAN sites, and enterprise premises all became potential “edge locations.” The assumption was that, because operators owned infrastructure closer to customers than the large cloud regions did, applications would eventually pay to run there.

That business never really developed at scale. Operators launched MEC platforms, partnered with AWS, Microsoft, and Google, and tested dozens of use cases, but edge revenues remained small relative to the attention and investment behind them.

Recently, the head of enterprise at a Tier One operator told me that something had finally changed. Customers were starting to ask for local AI inference. Some wanted models closer because of latency. Others wanted enterprise data to remain on-site or within the country. After years of pushing edge to customers, customers were now coming to them with actual workloads.

The problem was that they still did not have a clear product to sell.

AI may finally create real demand for distributed compute. But it also exposes a basic problem with how telcos have approached the edge. We started with our locations, but the customer starts with the workload.

AI Does Not Care Where the Edge Is

AI changes the edge discussion because inference does not have a preferred location; It has requirements. The model can run on the device, inside the enterprise, in a telco facility, in a regional data center, or in a hyperscale cloud. The right answer depends on what the workload needs at that moment.

A small model on a smartphone may run entirely on the device because there is no need to send the request elsewhere. A bank may keep inference inside its own premises because customer data cannot leave. A voice application may use regional compute because response time matters. A large reasoning task can run hundreds of kilometers away if the user is willing to wait a few seconds and the economics are better.

Read more

In my day, when a startup reached a $1 billion valuation, it was a pretty big thing. We even invented a mythical animal for it. Unicorns were rare, founders dreamed about becoming one, and investors treated the billion-dollar mark as evidence that somebody had built something extraordinary.

Apparently, $1 billion is for babies now. Anthropic, founded in 2021, is reportedly preparing for an IPO that could value the company at around $2 trillion and raise as much as $100 billion. To support that valuation, Anthropic is preparing to tell investors that its potential market opportunity exceeds $30 trillion a year. Its quarterly revenue also reportedly more than doubled to $11.6 billion, which helps explain why investors are at least willing to listen to the story.

That scale of addresable market (TAM) is ridiculous. The IMF puts the 2026 world economy at roughly $126 trillion, while the U.S. economy is around $32 trillion. Anthropic is effectively putting a theoretical opportunity in front of investors almost as large as the annual economic output of the United States and equivalent to roughly one quarter of the global economy. SpaceX used similar logic in its IPO, presenting a $28.5 trillion TAM, with $26.5 trillion attributed to AI.

You can reasonably argue that these numbers contain industrial quantities of imagination. TAM slides have never been famous for modesty. Still, laughing at the $30 trillion figure misses the important part of this story. Anthropic reached that number by changing what it considers its market.

The previous generation of technology startups mostly targeted technology spending.

Salesforce competed for CRM budgets. ServiceNow attacked workflow software. CrowdStrike went after cybersecurity. Snowflake targeted data infrastructure. Thousands of SaaS companies built businesses around selling better tools to companies and their employees. Gartner expects worldwide software spending to reach about $1.468 trillion in 2026. Enormous market, but there is an obvious ceiling to the story. If you sell software, eventually you are competing for some part of what companies spend on software.

AI changes that denominator.

Read more

The Consumer Robot Revolution Has Already Started

A few weeks ago, a Tier One telecom operator asked me to analyze what humanoid robots could mean for telcos over the next five years. I am still working through the answer, but the more I research the market, the more convinced I become that everyone waiting for the “next smartphone” may be looking in the wrong direction. Physical AI could become one of the biggest new consumer technology categories of the next decade.

The numbers already look pretty wild for something supposedly futuristic. Nori Robotics is advertising a $1,688 household robot that can help fold clothes, load dishes, fetch items from the refrigerator, and tidy rooms. Unitree already sells a fully bipedal humanoid for under $5,000, while 1X offers its NEO home robot for $20,000 or $499 per month.

China is now testing robots in real apartments, hotels, and libraries, doing laundry, making beds, receiving packages, and shelving books. Bosch, meanwhile, plans to start series production of Humanoid robots in Germany in 2027.

The hardware is arriving surprisingly fast. The real race now is to make these machines intelligent, reliable, and autonomous enough to be genuinely useful at home.

The Physical AI “ChatGPT Moment”

Cheap hardware does not automatically give us a useful household robot. The real bottleneck is generalization or getting a machine to handle situations it has never seen before without someone retraining it every time the environment changes. Unitree Chairman Wang Xingxing says robots trained for fixed environments can already achieve near-100 % task success, but small changes in the object, its position, or the surroundings can cause performance to deteriorate quickly. He defines the real breakthrough as a robot entering an unfamiliar home, receiving normal voice instructions, and autonomously completing roughly 80% of everyday tasks. His estimate for reaching that point is an aggressive two to three years at the earliest.

Read more

In 2016, I was discussing the early concept of Open RAN with a Group CTO. The O-RAN Alliance did not yet formally exist, but the industry was already circling the same promise. Break the RAN apart, open the interfaces, separate radios from basebands, move processing onto standard x86 hardware, and let operators combine components from different suppliers. The vision sounded like what had already happened in IT. Instead of buying a vertically integrated system from Ericsson, Nokia, Huawei, or ZTE, operators would build the RAN like Lego, picking the best component from each supplier, pushing hardware costs down and making room for a new generation of vendors.

I told him I thought the idea was stupid.

Some people assumed I was defending Ericsson, where I had spent many years. That was never the point. I had no problem with open interfaces or with more competition. My problem was the assumption that because Radio A could technically work with Baseband B, a large operator would want to make that combination its standard operating model across tens of thousands of sites.

Anyone who has spent time around real mobile networks knows that building the box is the small part. The hard part is planning, designing, deploying, integrating, optimizing, and maintaining a national network while keeping it running every minute of every day, with spectrum layers working together, mobility tuned, software releases tested, faults isolated quickly, and thousands of configuration changes under control. Energy consumption and performance must be managed continuously, and Massive MIMO has only tightened these dependencies because radio performance, beamforming, scheduling, software, silicon, and baseband processing are now deeply interconnected.

Which is why I never understood the obsession with proving that different vendors could interoperate. Of course, they could. Give competent engineers a specification, enough time, and enough money, and they will make Radio A talk to Baseband B. The problem was that doing it at scale would deliver better performance, lower total cost and energy consumption, faster upgrades, and simpler operations than buying an optimized radio and baseband system from a single supplier.

Ten years later, the market has answered. Dell’Oro now explains Open RAN as a story about interface standardization rather than supplier diversification, with Open Fronthaul expected to become the preferred interface across next-generation platforms, while operators continue to buy radio and baseband predominantly from the same vendor. In February 2026, the firm cut its forecast for multivendor RAN to under 5% of total deployments by 2030.

That is a long way from the original ambition. The lasting achievement of Open RAN looks like the standardization of the fronthaul interface. Useful, probably important, and considerably smaller than the promised transformation.

The operating model never made any sense

The early Open RAN narrative treated the RAN like an IT architecture. Separate each component behind a standard interface, competition improves, operators mix suppliers, and everybody wins. That view underestimated system-level optimization and, more importantly, operational accountability.

Read more

Unfortunately for Telcos, they are standing right in the middle of Musk's Galactic ambitions

Telecom has a strange insecurity. We spend years complaining that our industry is low-growth, capital-intensive, heavily regulated, and chronically under-monetized, and then, the moment Elon Musk buys spectrum, we assume the richest technology entrepreneur on Earth has developed an irresistible urge to become a mobile operator.

I don’t buy it.

SpaceX clearly intends to compete for mobile customers. On its Q2 call, Gwynne Shotwell said the company will build terrestrial infrastructure, use the 65 MHz of spectrum acquired from EchoStar for $19.6 billion, and expects to take “quite a few” customers from AT&T, Verizon, and T-Mobile. Its next Starlink Mobile generation is set to begin launching in 2027, with roughly 10 times as many satellites and 65 MHz instead of the roughly 5 MHz used by today’s partner-based service. SpaceX describes the combination as potentially 100 times better.

So yes, SpaceX is coming into mobile. But coming into communications and wanting to become a Telco are not the same thing.

The industry keeps looking at Musk through its own tiny window: How many macro sites will he need? Can his rooftop cells hand over properly? Will he sign an MVNO? Can Starlink really compete with Verizon coverage?

Sure, these are important questions, but for a Telco. They may also be completely missing the scale of what he is trying to build.

Stop Looking for the Fourth Carrier

The latest SpaceX filing basically tells us how Musk thinks.

SpaceX estimates a $740 billion addressable market for Starlink Mobile. Nice business. But the same prospectus claims a $26.5 trillion AI opportunity, including $22.7 trillion from enterprise applications.

Yes, my friend, you read well; a 22.7-trillion opportunity. Just to give you an idea, the global economy, measured by global Gross Domestic Product, is approximately $126 trillion, and the top contributor is the United States, at roughly $32 trillion.

SpaceX calls its overall $28.5 trillion opportunity the largest actionable TAM in human history. Whether that number is a brilliant strategy or TAM abuse requiring medical supervision is another discussion. The important point is scale. Inside Musk’s own model, mobile communications are a small component of the much larger AI economy he wants to participate in.

That changes how you should look at Starlink Mobile.

Most analysts are currently asking whether millions of Starlink rooftop radios can reproduce an American cellular network. MoffettNathanson and others have good reasons to be skeptical. Random customer rooftops are a lousy substitute for an engineered macro network, and a serious nationwide mobile product probably still requires some combination of MVNO access, towers, cable infrastructure, and more spectrum.

Read more

Hackers are patient people, very patient. The popular image is someone in a hoodie breaking into a network at three in the morning, grabbing what he can, and disappearing before breakfast. Serious state groups play a very different game. They get in, stay quiet, learn how the network works, map credentials and dependencies, and keep that access for years before doing anything visible.

Telecom has already lived through one of these. Salt Typhoon, the Chinese state-linked group that surfaced publicly in October 2024, compromised at least nine US carriers, including AT&T, Verizon, and Lumen, according to Senator Maria Cantwell, who spoke at a Senate Commerce Committee hearing in April 2026. The FBI assessed in August 2025 that the campaign reached more than 80 countries and that roughly 600 organizations were notified of potential compromise. Dwell time at major Telcos ranged from 12 to 24 months before anyone noticed. In the UK, parallel access to the communications of aides across three successive Prime Ministers ran from 2021 to 2024. By May 2026, Singapore had confirmed that all four of its national carriers were compromised.

Where they went inside those networks is a very interesting aspect. They reached the lawful intercept systems mandated under CALEA; infrastructure carriers are legally required to build them so law enforcement can wiretap, thereby giving a foreign intelligence service the same visibility the host country had built for itself. Cisco disclosed in February 2025 that the entry was mostly via stolen credentials and, in one case, via a router vulnerability that had been in the NIST database for seven years.

But there is another detail to this story that should worry any telco security team reading this. Cantwell stated publicly in February 2026 that Salt Typhoon may still be inside US networks, and the FBI’s own description of the threat as ongoing did not contradict her. As of early 2026, full remediation across the affected carriers has not been confirmed.

Which means nobody has to imagine an attacker quietly copying encrypted traffic out of a carrier core. One did it at the exact interfaces where key-exchange and subscriber-identity traffic converge, and we cannot currently prove he left.

Quantum attackers can start today

Security people have a name for the strategy this enables. Harvest Now, Decrypt Later. An attacker does not need a quantum computer today. He needs access, the ability to collect the right material, and enough storage to keep it. TLS and IPsec handshakes, certificates, public keys, subscriber records, roaming data, encrypted backups, and management traffic could all still be worth reading in 2033.

Harvest Now, Decrypt Later only works if the cryptography protecting that archive can eventually be broken, so it is worth spending a minute on what a quantum computer actually changes. Less than most headlines suggest, but in a very specific place.

It is not simply a faster machine. It works differently enough that a handful of mathematical problems become tractable, and cryptography happens to sit on exactly those problems.

Classical machines work with bits that are either zero or one. Quantum machines work with qubits that can exist in a superposition of states, and the mathematical space doubles with each qubit added. That does not mean the machine tries every answer at once and hands you the right one. Physics is not that generous. When you measure, you still get one result. The trick happens before measurement because quantum states behave like waves, and computational paths can interfere constructively or destructively. A good quantum algorithm arranges for cancellation so that useless possibilities fade and the pattern you want becomes likely to appear.

In 1994, Peter Shor ruined public-key cryptography. His algorithm solves integer factorization and the discrete logarithm problem, the two hard problems underlying RSA, Diffie-Hellman, and elliptic-curve cryptography. Multiplying two large primes is easy, and reversing it is not, unless you have a quantum shortcut through the mathematics.

For telcos, the important distinction is between encrypting the data and establishing the trust that allows two systems to exchange it securely.

Picture what happens when your phone opens a secure connection to a network function. Two things take place, one after the other. First, the two sides have to agree on a secret key while strangers are listening, and each has to prove they are who they claim to be. Once that is settled, the data itself gets scrambled with the shared key and sent.

Quantum does very different things to those two steps.

The scrambling step is much less exposed. That work is done by AES and similar symmetric algorithms, and the best-known quantum attack against them is Grover’s algorithm. In a simplified model, AES-256 has roughly the brute-force security margin of a 128-bit classical key, which is still an absurdly large search space. This is why AES is not where most quantum-security teams are losing sleep.

But the first step is where everything breaks. Agreeing on a secret key in the open and proving identity is exactly the job of RSA, Diffie-Hellman, and elliptic curve cryptography, which is exactly what Shor’s algorithm dismantles. And once you start listing where telcos rely on that, the list becomes really, really long.

TLS on every interface; IPsec and IKE between base stations and security gateways; VPN tunnels; the certificates and the entire PKI behind them; Network function authentication within the 5G core; Software and firmware signing; roaming agreements between operators; and eSIM provisioning. Even SUCI, the 5G mechanism that hides a subscriber’s permanent identity during registration, relies on elliptic-curve cryptography today.

So an attacker with a working quantum machine never has to decrypt traffic packet by packet. He can forge a certificate, impersonate a network function, or recover the secret behind a key exchange captured years earlier, which can be far more useful than trying to decrypt traffic packet by packet.

That difference should decide the migration sequence. Bulk traffic encryption can wait a few years, but the trust machinery cannot, and plenty of coverage still has this backward.

How close is the Q-Day

Read more

Over the past few weeks, I have been doing some soul-searching regarding telecom network automation. If you have read my recent LinkedIn posts, you know I am constantly looking for the line between actual operational reality and the endless stream of industry marketing.

The market is flooded with promises of “zero-touch operations” and fully autonomous networks. Yet, anyone who has worked with real-world telecom infrastructure knows that dropping an AI model on top of a fragmented legacy stack does not suddenly make the network autonomous.

Recently, my former Google colleague Gabriele Di Piazza, who leads Product Management, Alliances and Architectures at Blue Planet, shared some deep technical blueprints and operational data from their field deployments. Reviewing those materials gave me a chance to step back, compile my key learnings, and synthesize what actually works versus where the industry is fooling itself.

When you go into the rabbit hole of automating a multi-vendor, multi-domain network, it is clear to me that most operators are doing network automation wrong.

Sure, they are not failing because of a lack of effort, investment, or technical skill. They are struggling because they are automating in isolated silos. Essentially, they are applying next-generation intelligence to fragmented, legacy workflows.

If you are evaluating or building an automation roadmap today, here are the 10 basic non-negotiables that determine whether you build a genuinely autonomous network or just a faster set of problems.

The 10 Basics You Need to Get Right

If you want to move from isolated scripts to a functioning autonomous network, these are the ten structural requirements you cannot skip.

1. You Need Air Traffic Control, Not Just Smart Planes

Automating isolated network domains, whether radio access, optical, or IP routing, creates intelligent silos. If every domain executes local optimizations without coordination, cross-domain policies can conflict.

TM Forum architectures emphasize layer decoupling, where resource domains execute local actions while a cross-domain intelligence layer coordinates overall operational outcomes. Without that coordination to reconcile business intent across access, transport, and core, domain-level automation simply accelerates policy clashes.

2. You Cannot AI What You Cannot See

Data fragmentation is one of the biggest bottlenecks in telecom operations. You cannot deploy AI agents if your inventory is static, out of sync, or fragmented across legacy databases. AI needs a continuously reconciled view of what actually exists and its current state. Field deployments report up to a 95% reduction in capacity reporting time through automated, federated inventory.

Using model-driven discovery and standardized catalog and inventory interfaces such as TMF634, TMF638, and TMF639, operators can map physical, virtual, and cloud assets so AI agents act on live network state rather than outdated records.

3. Stop Scripting, Start Declaring

Imperative scripting, with rigid playbooks defining step-by-step commands, becomes brittle as firmware, configurations, and topologies change. Autonomous operations require declarative, intent-based automation.

The operator defines the target outcome, while orchestration translates that intent into resource actions across access, core, and transport and continuously reconciles the network toward the desired state. Field deployments report reductions of up to 80% in service activation times.

4. Standard LLMs Need an OSS Knowledge Graph

Off-the-shelf Large Language Models parse text well, but they lack native understanding of network topology, service dependencies, and radio context. To give AI real operational reasoning, operators need an OSS knowledge graph that represents how network resources and services relate.

This graph maps both “North-South” service-to-resource relationships and “East-West” multi-vendor network adjacencies. Advanced implementations can use Graph Neural Networks to analyze IP topology, link utilization, and traffic-engineering constraints such as Segment Routing and IGP metrics. This topological context allows agents to perform more accurate root-cause analysis rather than guessing.

5. Agentic AI is a Three-Layer Stack, Not a Feature

Adding a conversational AI widget to a legacy OSS dashboard is not agentic automation. Operationalizing AI agents across a complex, multi-vendor environment requires a dedicated three-layer architecture. At the foundation, agentic tooling exposes OSS knowledge, telemetry, and network capabilities through governed APIs and open interfaces such as the Model Context Protocol.

Above it, an agentic core manages the agent lifecycle, model access through secure LLM gateways, orchestration, policy, and Agent-to-Agent communication. Finally, agentic channels embed that intelligence into inventory, orchestration, assurance, and other operational workflows, making reasoning part of how the network is run rather than another isolated portal.

6. If You Cannot Govern Change, You Cannot Trust AI

The skepticism surrounding AI-driven network control is justified by real-world outage data. Industry reports indicate that 80% of serious outages are preventable, stemming directly from process and configuration errors, while faulty software changes have been linked to hundreds of millions of lost user-hours.

Telcos cannot scale automation without Configuration and Change Management. This governance layer enforces real-time drift detection, validates live configurations against compliance baselines, and maintains a strict audit trail of every change, whether executed by a human engineer, an imperative script, or an AI agent.

7. Never Act Without a Digital Twin

In modern software development, code is tested before it hits production; high-impact network changes need the same guardrail. Before an automated system changes routing, capacity, or critical configuration, it should validate the action against a network digital twin.

Using route optimization and analysis tools, operators can run predictive “what-if” simulations. If an AI agent proposes changing routing costs or shifting traffic off an optical path, the twin first simulates the impact to verify that latency, packet loss, and enterprise SLAs remain within acceptable limits.

8. Reactive Monitoring is Obsolete

Traditional operations wait for an alarm before an engineer opens a ticket and investigates. In dynamic multi-domain networks, that is too slow. Assurance must evolve into a proactive, closed-loop engine that uses real-time telemetry to detect anomalies, predict degradation, and understand service impact before customers are affected.

When a risk is identified, assurance can trigger orchestration workflows to reroute traffic or adjust resources. In 5G slicing deployments, field data show a 75% reduction in root-cause identification time and an 85% reduction in mean time to resolve slice disruptions.

This is stronger because the number is now specific and defensible, rather than making 85% sound universal.

9. Automation is a Commercial Weapon, Not an OPEX Trim

Viewing network automation merely as a back-office cost-cutting exercise misses its commercial value: service velocity and Network-as-a-Service monetization. Lumen now reports more than 3,000 NaaS customers, with new customer adoption growing 22% quarter-on-quarter as enterprises shift toward programmable connectivity for cloud and AI workloads.

Automated, intent-based orchestration enables this commercial model by compressing service delivery from days or months to hours. In one European Tier 1 5G slicing deployment, AI-driven order management reduced activation from months to hours while delivering 80% cost savings, transforming infrastructure into a programmable commercial asset.

10. Level 4 Autonomy is an Evolution, Not a Product

“Level 4 Autonomous Networks” cannot be purchased off a vendor price sheet. Under TM Forum’s Autonomous Network Levels, Level 4 marks the shift from human-defined automation toward intent-driven, predictive, closed-loop decision-making. Operators are increasingly reaching this level in specific domains and operational scenarios rather than across the entire network at once.

Operators must progress systematically: starting with data unification, moving to domain automation and closed-loop assurance, and ultimately extending autonomous reasoning across the service lifecycle. Skipping the foundational data, orchestration, and governance layers simply automates the problems underneath.

The Long story short of Network automation

Autonomous networking is not a fixed destination or a product you buy off a shelf. And adding more AI does not make a fragmented network more autonomous.

If your network strategy skips the hard work of unifying data, governing change, and moving from scripts to intent-based orchestration, you are not modernizing the network. You are digitizing your legacy debt. Put AI on top of a fragmented OSS architecture, and you simply automate chaos at a higher speed.

The operators that win the next decade will not be those that bought the most AI. They will be the ones who built the operational foundations to make AI useful: live inventory, cross-domain orchestration, governed change, digital twins, and closed-loop assurance working as one coherent system. If your automation roadmap starts with an LLM, you are probably starting in the wrong place.

Where are you seeing operators get stuck? I would be interested to hear which of these ten foundations is proving hardest to get right.

Read more

Telecom has spent the last decade proving that Open RAN, private 5G, edge computing, network APIs, 5G standalone, and network slicing can all work. The problem is that working and making money are not the same thing.

The pattern is always the same. A new technology appears, standards get written, vendors publish enormous TAMs, consultants draw a hockey stick, operators spend billions, and a few years later the technology is still alive while the original revenue forecast has quietly disappeared.

That is key these days because telecom is standing next to an even larger pile of capital labeled AI. Telcos are again being shown huge opportunities around inference, sovereign AI, GPU-as-a-service, AI-RAN, and edge infrastructure. Some of them may be real. But before spending another few billion proving they were technologically right, operators should ask a much simpler question:

If we are right, who actually gets paid?

Telecom has proven every tech, but the revenue was less cooperative.

Private 5G is a good place to start because, technically, it worked. By the end of 2025, roughly 6,500 private LTE and 5G networks were deployed worldwide, up from 4,700 a year earlier. The problem is that the entire market was worth only about $2.4 billion, while forecasts had spent years pointing toward opportunities of $80 billion, $100 billion, and even $150 billion. Nobody lied about the technology. The factories got connected, and the networks got deployed. The money simply forgot to show up at the scale promised.

Network APIs are following a similar, concerning track. Telcos can now expose identity, fraud, location, and network capabilities through standardized APIs, which is genuinely useful. Actual operator API revenue in 2025, however, was roughly $284 million, while forecasts continue to point toward a few billion by 2030, although earlier industry estimates talked about hundreds of billions in broader value creation. Apparently, the easiest part of creating a $100 billion telecom market remains putting $100 billion in the title of the consulting report.

Read more

Like it or not, you are already distributing billions of AI tokens every day. To enterprises, to consumers, and increasingly to your best new customers, which are not people at all. AI agents are already on your network. Humanoids, delivery robots, drones, and every other intelligent machine anyone can imagine will follow, and they will all move their atomic units of intelligence across your infrastructure.

You can be in denial about this and call them bits. Nokia has already measured AI at roughly a fifth of network traffic, with global production exceeding 100 trillion tokens a day and more than half of generative AI usage occurring on mobile devices. Bell Labs expects AI to account for about 30% of wide-area traffic by 2034. So you can keep spending 20% of revenue on capex, watch traffic compound in the mid-twenties, and tell investors this is a growth story.

But we know it is not. None of these customers will pay you more for connectivity. Moving bits will continue to erode in price, as they have every year for thirty years. Sorry, I don´t write those rules for this industry.

Think like this: you are in the tap water business, crucial for life, but nobody pays more than 50 cents for a bottle of water.

But now, you have a golden opportunity in this gold rush, and it is not the one most operators are chasing.

Read more

AI Agents Are Not Applications

For most of the software era, the client was a stable application. A company built it, registered it, gave it credentials, connected it to an API, and kept that integration running for months or years. Authentication, billing, quotas, versioning, documentation, and support were all designed around software that stayed around.

AI agents work differently. A persistent agent can take one objective and create thousands of smaller agents to work on different parts of it at the same time. One may search suppliers while another reads contracts; others compare prices, analyze documents, test scenarios, or verify results. Kimi Moonshot already runs swarms with up to 300 concurrent subagents, more than 4,000 tool calls in a single task, and reports execution up to 4.5 times faster than sequential execution.

In the future, thousands will become tens of thousands. As inference gets cheaper and swarm infrastructure improves, a complex enterprise task may create several generations of temporary agents during only a few minutes of work. Some may live for several minutes, while others may exist for only a few seconds, return one answer, and disappear. One durable agent may represent the person or company, keep the important memory and objectives, while thousands of temporary agents are created underneath it whenever more reasoning, search, comparison, or verification is required.

Calling all of them digital employees gives them too much dignity. Many will be closer to processes with opinions, temporary pieces of cognition created to solve one small part of a larger problem. Temporary software can still call APIs perfectly well, so lifespan alone does not make APIs obsolete. But when one objective can create thousands of agents that live for seconds, each responsible for external integrations, credentials, retries, and state, the API logic starts to make little sense.

To understand why, we need to look at how these swarms will operate once they scale from hundreds to thousands of agents.

How AI Swarms Will Work

A swarm with ten thousand agents cannot behave like ten thousand employees sitting in the same Slack channel. If every agent tried to communicate with every other agent, the number of possible communication links would approach 50 million.

Real systems will never allow that pattern because most of those exchanges would add cost without adding intelligence. Every model message has to be generated, transmitted, placed into another context, interpreted, and sometimes summarized again before the next piece of work can begin.

Large swarms will therefore look more like distributed computing than digital organizations.

Read more

If Space Does Not Replace Some Infrastructure, What Is the Point?

Satellites can rescue stranded hikers, keep phones alive after hurricanes, and fill the ugly holes on a carrier’s coverage map. That is useful, but it does not change Telco economics at all.

If orbital D2D is really going to move the needle for AT&T, Verizon, or T-Mobile, it eventually needs to do something much more important with their networks. It needs to carry enough real traffic that the carrier can avoid building, upgrading, or maintaining some terrestrial infrastructure. Otherwise, what exactly is the capex benefit?

AST already says this openly, as its pitch to mobile operators is that they can extend coverage into places where terrestrial economics do not work without building additional towers or other land-based infrastructure. SpaceX is pursuing a different commercial model, but the infrastructure logic is the same. If meaningful rural and semi-rural traffic moves into orbit, some radios on the ground should eventually become unnecessary.

As of the end of 2025, the U.S. had roughly 158,500 purpose-built cellular towers supporting a broader footprint of almost 255,000 macrocell sites. So let’s stretch the case and assume Starlink, AST, and others become ridiculously successful over the next decade, eventually carrying 20% of rural and semi-rural mobile traffic. Does that put 10% or even 20% of American macro locations under economic pressure?

At least in theory, yes.

But that is only half the story. The same orbital networks attacking marginal rural towers may desperately need the best urban and suburban towers once they start carrying serious traffic. So is space a wrecking ball for towercos, or the beginning of their next major tenant cycle?

If Orbital D2D Really Works, What Is Actually at Risk?

Let’s assume a very successful outcome for orbital D2D. Over the next ten years, Starlink, AST, and other satellite networks become capable of carrying meaningful mobile traffic across rural and semi-rural America. In our stress case, 20% of traffic in those areas eventually moves to space.

The U.S. has roughly 255,000 macrocell locations, but no reliable public breakdown shows exactly how many are rural or semi-rural. We estimate roughly 25% to 35%, equivalent to about 64,000 to 89,000 sites.

Most of those sites will still be needed. A tower serving a small town, highway interchange, or growing suburb may be rural geographically but still carry significant traffic. The locations with the greatest exposure are low-utilization sites whose main purpose is maintaining continuous geographic coverage.

Read more

I am working with a Tier One telecom operator on its agentic AI strategy. Much of the work has nothing to do with choosing a model or buying more GPUs. We are trying to understand what the customer becomes when software starts making decisions, buying services, using networks, and acting with authority on somebody else’s behalf. Telecom has spent decades designing around humans holding phones and machines sending traffic.

The next customer may be an agent that sits between both.

Several pieces are already real. Agents can operate computers, retain memory, write software, use payment rails, and run across distributed inference infrastructure. Those are no longer research demos dressed up as strategy slides. They are shipping capabilities.

But the hot mess begins when you combine them. Give an agent memory, money, credentials, and enough autonomy to work for days, and suddenly reliability, identity, liability, verification, and security matter more than another benchmark point. A buggy chatbot is annoying. A buggy agent with a corporate card and production access is a financial disaster nobody wants to clean up.

I cannot share the details of the work, but I can share the ten shifts shaping it.

They cover agents that work longer, learn after deployment, build software, reason in new ways, delegate authority, trade with other agents, simulate outcomes, verify each other’s work, create new security risks, and eventually operate machines in the physical world. For telecom, the real issue is not whether agentic AI arrives. The issue is what kind of customer it creates, what that customer expects from the network, and who gets paid when machines start buying from machines.

10 Agentic AI Trends That Change Everything

1. Endless Work Agents

AI agents can already work for hours, but give them a long enough job and they start losing the plot. METR measures how long frontier agents can complete professional tasks at 50% reliability, and that horizon has been rising quickly. Yet research on web agents shows the problem clearly: success rates around 40% to 50% on short tasks can fall below 10% when the same work stretches across a longer interaction history. METR is also seeing early evidence of agents attempting coding work that would take humans weeks.

The good news is that the pieces needed to close that gap are already being built, such as external memory, checkpoints, independent evaluators, better context management, and systems that can detect when an agent has gone off course. Put those together and an agent stops being something you ask to finish a task before lunch. It becomes something you can leave running on a problem for several days. That is when AI starts competing with jobs, not prompts.

2. Agents That Learn on the Job

Today’s agents can remember what happened without necessarily becoming better because it happened.

Systems such as Mem0 and Letta already give agents persistent memory across sessions, but most deployed models still leave the factory with roughly the same underlying intelligence they will have months later.

Researchers are now trying to change that. The NeurIPS 2026 work on continual learning agents, Evo Memory, Titans, and related approaches explore how agents can absorb experience while they operate, rather than waiting for somebody to retrain them.

Read more

Satellites are hot again, and SpaceX has decided the next prize is mobile. On its Q2 call, Gwynne Shotwell pointed at AT&T, Verizon, and T-Mobile, framed the opportunity at roughly $600 billion, and said SpaceX expects to take customers from them. That number isn't an accurate U.S. mobile TAM, but the level of ambition is clear. SpaceX needs Starlink to become much bigger.

Starlink Connectivity generated about $4.3 billion in Q2 revenue and remains the company’s most proven cash engine, while Musk is simultaneously pouring tens of billions into rockets, satellites, and AI.

So let’s stretch Musk Telecom's ambitions. Assume Musk eventually wants just $100 billion of that $600 billion pie. At $75 a month, he needs roughly 111 million customer-equivalents. Surely, this is not about rescuing hikers or filling a few rural dead zones. It means building a serious mobile business.

During the Q2 call and likely under pressure from investors, Musk’s proposed answer was wonderfully simple: satellites overhead, 65 MHz of spectrum, and millions of small cellular radios attached to Starlink dishes on homes and businesses.

Anyone who lived through the femtocell era should immediately ask the obvious question: Haven’t we tried this shit before?

SpaceX engineers know the limits of femtocells. So the real question is whether Musk actually believes Femtocells 2.0 plus rockets can build a national mobile network, or whether it is a smarter commercial weapon for cheap offload, negotiating leverage, and forcing carriers to sell him the network pieces he does not want to build.

Femtocells 2.0: We Have Seen This Movie Before

The first problem with Musk’s “Small cells” idea is that a large number of radios does not automatically create a good mobile network. Telcos learned this the hard way with femtocells. The original idea was attractive for the same reasons SpaceX’s pitch sounds attractive today: put a cheap cellular radio inside a home or business, use somebody else’s power and broadband connection, and offload traffic that would otherwise hit the macro network. It worked reasonably well as an edge solution. It was never a serious substitute for an engineered nationwide mobility layer.

Read more