Ajit Singh (@ajitgaur) header
Ajit Singh (@ajitgaur) profile picture

AVP Calix Cloud ☁️ @CalixHQ, My opinions are my own

Dallas, TX
Joined June 2008
For twenty years, Wi-Fi has been the part of the network nobody respected. The last hop. The thing you blame when the video call freezes and the thing nobody budgets for until it breaks. That era is ending, and most operators haven't noticed yet. Wi-Fi 7 is not an incremental upgrade. Multi-link operation across 2.4, 5 and 6 GHz, 320 MHz channels, theoretical throughput past 46 Gbps — the numbers matter less than what they enable. AI inference is moving to the edge, onto factory floors, into offices and campuses. The last mile of AI runs over Wi-Fi, and it carries traffic that looks nothing like a video call: latency-sensitive, bursty, unforgiving. Quality of service, an afterthought in enterprise WLAN for a decade, is suddenly mission-critical. Meanwhile the old operator reflex — treat Wi-Fi as a threat to cellular revenue, tolerate it only as offload — is running out of road. Passpoint is making Wi-Fi an extension of the cellular core, not a parallel system. Private 5G and Wi-Fi are settling into coexistence, not a fight to the death: Wi-Fi 7 for dense, device-heavy spaces; private 5G where mobility and latency are non-negotiable. And look at what's next: Wi-Fi 8 isn't chasing a bigger peak-speed number. It's chasing deterministic performance — the same latency and reliability guarantees 6G is being built around. The pattern is the same one we've seen across this whole AI cycle. The unglamorous layer becomes the strategic layer. Wi-Fi spent two decades as a commodity. In the AI era, it's infrastructure — and operators who still treat it as a cost center will find the enterprise AI business built on someone else's network.
19
Most of the coverage of Nvidia's leadership change treated it as a succession story. The more consequential story is financial. For the last few years Nvidia has not just sold GPUs — it has underwritten the AI buildout, backstopping customer purchases and financing guarantees that let mid-tier AI cloud providers scale far beyond their balance sheets. CoreWeave's debt, secured against its GPU inventory, is the clearest example of how deep that arrangement ran. If that support gets less predictable, the squeeze lands first on the smaller providers, and the compute market consolidates further into the hyperscalers — Microsoft, Google, Amazon — who can buy at scale without anyone's help. That is not a neutral outcome. It decides who controls the price of inference. For telecom, this is where it gets interesting. Operators have been building AI into the network itself — AI-RAN, edge inference, autonomous operations — and the carriers with real balance sheets and real estate at the edge are in a stronger position than the cloud-only players. The smart move now is to stop betting the network's AI future on a single vendor's credit line and a single GPU architecture. AMD, Intel, custom silicon — diversifying the silicon underneath AI-RAN is no longer hedging. It is engineering discipline. The demand for AI compute is not going away. The question is who pays for it, and on whose terms.
26
6G won't get a clean band. It will have to live on 5G's carriers, and 3GPP is designing that sharing in from the start. The requirement is blunt: the 6G RAN "shall support efficient Multi-RAT Spectrum Sharing between 6GR and NR." For any band 5G already occupies, sharing the carrier is the main way in. The design makes it possible. 6GR is built on NR's grid: same waveforms, subcarrier spacing, cyclic prefix and symbol boundaries, 14-symbol slots, 12-subcarrier resource blocks, boundaries aligned with NR's. A resource block is NR in one slot and 6GR in the next, and nothing is wasted. The expensive part was always the overhead, not the data. In October, RAN1#126bis agreed dynamic sharing of 5G and 6G control resources — one region, drawn on slot by slot, invisible to the device. RAN1's assessment: adding 6GR costs a median ~1.9% of downlink resources on FDD carriers with shared control, against ~11.9% with separate regions. On TDD carriers it's ~1.8% shared against ~10.1% separate. DOCOMO's worked example on a 10 MHz FDD carrier: 17.83% overhead for NR alone, 19.09% with 6GR sharing control — 1.26 points for a second generation. Contrast 5G's DSS against LTE, a retrofit: LTE's always-on reference signal eats 9.5% of the carrier before either generation schedules anything. All of this is Release 20 study text — nothing specified yet, first 6G specs freeze December 2028. But the direction is set: 6G's migration will be won or lost on spectrum sharing, not peak speeds.
2
44
The hardest problem in agentic AI is not how much the agent knows. It is whether it knows which ten things matter right now. A latency spike in a metro 5G sector could be RF interference, a bad slice, backhaul congestion, a bad rollout, or an attack. Each one needs a different investigation, different data, different expertise. An agent that reaches for its most-used tools risks missing the root cause — or making it worse. In telecom, a wrong automated fix pushed to thousands of base stations is catastrophic, with SLA penalties attached. The real differentiator is not the size of the skill library. It is orchestration intelligence: the layer that decides which agents, tools, and data to activate for this fault, scores its own confidence before acting, and escalates when it is not sure. As one senior network architect at a Tier 1 European operator put it: "I don't need an AI that knows a thousand things. I need one that knows exactly which ten things matter right now, and why." Yesterday, broadband operators had the AI capability and not yet the product. Inside the network, it is the same gap: plenty of capability, not enough judgment about where to aim it. The operators who win will not be the ones with the longest list of agent skills. They will be the ones whose AI knows when to act — and when to ask a human.
1
39
Every broadband operator knows AI is coming to their network. Almost none of them know what to sell it yet. A Fiber Broadband Association survey out yesterday is unusually honest about that. 90% report growing interest in AI, and 73% have moved to pilots or beyond. But 68% are unsure which AI transport services they'd offer over their own networks, and 41% can't predict AI traffic growth through 2027. That's the gap that matters: a network gap, not a marketing gap. The broadband network is where AI traffic lands: the fiber between data centers, the middle mile, the last mile to the home. Operators own the pipes AI runs on. What they don't yet have is the product: an AI-ready service with real capacity planning, real SLAs, real pricing. The pressure lands first between facilities: 42% name data center interconnect the segment most likely affected. And 44% are unsure when they'll invest in the supporting infrastructure. The barrier isn't money. Only 8% cite limited capital. 42% cite limited organizational bandwidth and competing priorities. The industry has the capital and the fiber. It doesn't yet have the plan, the people, or the product. Yesterday's wavelengths story said AI traffic is the revenue line. This one says broadband hasn't worked out how to sell it yet. Operators who turn their networks into AI-ready products will set the terms. The rest will be carrying someone else's AI traffic on yesterday's terms.
2
41
The AI buildout just lost the person who was the platform. Jensen Huang resigned as Nvidia's president, CEO and director on Thursday, effective immediately, after 33 years. CFO Colette Kress is interim CEO while the board searches for a permanent successor, and Harvey Jones becomes its first board chair. The filing gives no reason. For telecom, this is not a stock story. Nvidia sits underneath the industry's AI plans. It put $1 billion into Nokia last year to build AI-RAN. Just days ago it signed with SK Telecom and Rebellions to build sovereign AI infrastructure in Korea, opening NVLink Fusion to a Korean NPU. Those bets were made on a roadmap — and a person — operators trusted to hold. The near term is probably fine. Demand for AI compute still exceeds supply, and Nvidia says its November 17 earnings will go ahead as scheduled. The question is what slows while the top job is interim: the dealmaking, the financing, the platform openness that turned a chip company into the centre of the AI ecosystem. If that slows, the knock-on is physical. As one telecom analyst put it today, a slower AI build means slower data-centre builds — and slower fibre, inside and outside those centres. Watch November 17. That is the first test of whether the platform outlasts its founder.
1
61
The most interesting AI revenue line in telecom isn't a model, a GPU, or an app. It's a wavelength. Omdia just surveyed 122 operators, and 61% named high-bandwidth wavelength services one of their top two revenue opportunities from AI over the next three years — ahead of IP services, dark fiber, and private networks. The reason is the shift nobody talks about enough: AI is moving from hyperscaler training to mass enterprise inference, and inference traffic has to move between a lot of places, fast. The fiber between the clusters is where the money is. The numbers tell you where the race stands. Nearly half of operators already sell 400G wavelengths or will by the end of this year. But only 3% have 800G deployed today — while 57% expect to launch it commercially in 2027 or 2028. Everyone sees the same finish line; almost nobody is there yet. And the buyers are not patient. Hyperscalers now treat delivery intervals as seriously as price — one Tier 1 exec told Omdia they're "less and less negotiable than anything else." KT is already rolling out 800G AI-dedicated lines in Korea. The operators with fiber in the ground and crews that can light it in weeks, not quarters, will take this market. The rest will be explaining why their 800G roadmap slipped.
1
25
Germany just showed the rest of the industry how fiber gets built at scale. Telefónica's O2 has signed a long-term deal to sell its broadband over Deutsche Glasfaser's open-access fiber network — about 3 million more households, taking O2's fiber footprint to roughly 19 million, a 20% jump. No duplicate network to build, no second set of trenches to the same street. Just access to fiber that is already there, or already being built — up to 1 Gbit/s to the home. Three of Germany's four national providers now run on the same fiber. Same terms, same service, same fault resolution. The competition moves up the stack, to the product and the experience, instead of being fought in the ground. That is the model worth watching. Fiber's biggest problem was never demand. It was the economics of every operator building its own parallel network to the same street. Open access splits the two jobs apart: one party builds the infrastructure once, properly, and everyone competes on top of it. Investment and competition stop being enemies. The US is still arguing about this in municipal broadband fights and wholesale debates. Germany just did the math: shared fiber gets built faster, used harder, and reaches further. The network you compete on beats the network you compete with.
1
1
41
5G bolted AI on. 6G is being built with it inside. That is the real story of the next generation, and it is already written into the standards. The ITU's IMT-2030 framework — the document that defines what 6G is — lists "AI and communication" as one of its six core usage scenarios. Not a feature. A design principle. ITU-T's own framework calls AI-native networks the core architectural principle of the next generation. 3GPP is already doing the work. Release 20, the last 5G-Advanced release, standardizes AI and machine learning in the radio — for mobility, for XR, for sensing combined with communications. Release 21, the first 6G release, starts normative work in March 2027 and freezes in March 2029. The network that learns is being specified now, not imagined later. And this is not theory. On a live 5G network in Egypt, Ericsson's AI-native scheduler delivered 50% higher downlink throughput and 14% better spectrum efficiency — the radio itself learning to use spectrum better, in real time, on commercial spectrum. That is the shift operators need to understand. Every generation before this one was designed first and optimized later. 6G is the first generation where intelligence is part of the design — in the radio, in the core, in how the network plans itself. The question is not whether your network will have AI. It is whether AI will be something you buy and bolt on, or something your network is.
37
1/6: AI in broadband was never going to be a chatbot. Treating it as a feature to bolt on — a smarter script, a nicer dashboard — misses the point.
1
53
5/6: Then the agentic layer: Agent Workforce puts AI agents inside real operator workflows, while people keep the judgment and an AI Control Center sets the rules.
1
8
6/6: Your network is already a distributed data center. The operators who win will put that compute to work. The rest will be explaining the gap.
7
The brave new world isn't a slogan. In telecom, AI is quietly dissolving the lines between connectivity, cloud, and intelligence. Operators who see it as a feature will lose to those who build on it.
34
Ajit Singh retweeted
What an incredible #ConneXions26! We’re leaving with one message loud and clear from Michael Weening: There is no future that does not include AI. 🎥 ConneXions OnDemand arrives November 4 on MyCalix. 🔗calix.com/events/connexions/…
2
8
434
Ajit Singh retweeted
It’s game day and Coach Gerry Dee’s all-star roster is ready to take the field: 📣 Marketing Agent ⚙️ Operations Agent 🎧 Service Agent 🏠 Subscriber Agent One team. One platform. One goal: victory. #ConneXions26
1
1
1
122
Ajit Singh retweeted
There’s plenty to look forward to today at #ConneXions26! Check out today’s highlights and get ready for a full day of learning, innovation, and connection! Can’t join us in Las Vegas? Watch the Opening General Session live at 8 a.m. PT. 🔗 linkedin.com/events/75162207…
2
5
128