Partho Mishra

Co-Packaged Optics (CPO) is now confirmed as the technology of choice for scaling AI Compute Clusters - RANOVUS Inc. XPU CPO, Broadcom and NVIDIA Ethernet/InfiniBand Switch CPO. RANOVUS Inc. has been focused on replacing copper with ODIN CPO for scaling GPU interconnects while Broadcom and NVIDIA are tackling replacement of optical modules with CPO to power efficiently scale AI networking. This OFC Conference is going to be remarkable. Come see us in San Francisco next week to discuss your requirements and see how RANOVUS Inc. can help you step into CPO for your AI platforms. 📍 Find us at Booth #6273 📅 March 31st – April 3rd 2025 #OFC #IndustryTrend #AI #Compute #CPO #SiliconPhotonics

348

7 Comments

We just published a post on how we went from FreeSWITCH to SWML. This expands on an earlier post about the origin of SignalWire and the challenges we faced scaling real-world communications infrastructure. FreeSWITCH gave us deep control over media and signaling. The next frontier was a way to express workflows in a way that horizontally scaled across physical regions, teams, products, and use cases and exposed powerful IVR features like Conversational AI SWML is our answer. It’s a programmable, composable framework that combines real-time media control with declarative logic and AI. You can define entire voice, video, and messaging applications in a single document that can be static or generated on demand. It’s built for people who want more than basic IVRs or cookie-cutter CPaaS functions. This is the evolution of the FreeSWITCH technology into something a web programmer can really embrace. You can build the client into a single-page browser app with our customizable widget and define the server as you wish using basic tools to generate SWML on demand to create powerful IVRs, Conversational AI, and Contact Center functionality. Read the blog here: https://lnkd.in/ex-T_JAA #SignalWire #SWML #FreeSWITCH #ProgrammableVoice #AIinTelecom #VoiceAI #TelecomInfrastructure #CPaaS #PUC #DeveloperTools #ProgrammableUnifiedCommunications

37

1 Comment

I’ve never been so happy to eat my own words! When I last joined Wi-Fi Alliance’s The Signal podcast, I predicted a quicker-than-usual ramp up speed for #WiFi 7. The reality has been a much faster deployment than even we anticipated, across all market segments. Qualcomm as the innovation leader in the #WiFi7, we are truly encouraged to see widespread adoption across #IoT, industrial IoT, access infrastructure, #smartphone, #PC, and #automotive. Best-in-class Wi-Fi is the key to keeping up with speeds and feeds and latency requirements of pretty much every environment—including the modern home, which has about 20 connected devices on average. Our tech can handle high-speed demand, ensuring seamless video streaming, gaming and, increasingly, augmented reality, virtual reality, mixed reality, and extended reality experiences. Tune in below to learn more about why Wi-Fi 7 is the perfect complement to advanced Fixed Wireless Access and Fiber to the Home and how it’s unlocking the benefits of edge #AI in home networks. Link: https://lnkd.in/gQZsiex9

146

3 Comments

Big news at #MWC25: the Qualcomm X85 5G Modem-RF solution leads Android connectivity, increasing global adoption of our Dragonwing branded 5G Open RAN cellular infrastructure platform, plus new 5G Advanced FWA and industrial IoT solutions. Visit us to see agentic AI on phones, PCs, smart glasses, and our latest industrial innovations. Video recap below!

812

14 Comments

. Eben Albertyn, EVP and CTO of EchoStar Corporation and its Boost Mobile 5G network, said the operator's open radio access network (RAN) design will make it easier for Boost to embrace AI RAN technology. "The ... thing that's important to remember about open RAN is that open RAN and what people are calling AI RAN are not competitive to each other. They're orthogonal to each other," Albertyn said during a keynote session at the Competitive Carriers Association (CCA) event in Denver Wednesday. "And because of how open RAN works and how it's been put together, it is far easier to introduce AI RAN components into open RAN than it is any other technology," Albertyn said. Albertyn did not say whether EchoStar's Boost Mobile network would deploy AI RAN technology, but he suggested that the company has been experimenting with various AI applications in its 5G network. "AI has been really kind to us. Because our core network lives in the cloud, and we're adjacent to all of these AI technologies, we're able to ingress that immediately into our network," he said. MORE: https://lnkd.in/ggFqNnhR

77

2 Comments

Several days after this writer posted about the lopsided market narrative related to co-packaged optics (please see: https://lnkd.in/eReYuVp8), there were a couple of references to the CEO of Nvidia, Jensen Huang, stating that such a solution is not inevitable, starting with a news article (https://lnkd.in/egyNwa92). The piece quotes a supposedly contemporaneous video on a social media site in which Jensen says the following to a press outlet about working with TSMC on silicon photonics: “It’s still several years away. We should stay with copper for as long as we can, and then after that, if we must, we’ll use SiPh.” (https://lnkd.in/eRPv9RDQ). It should be noted that there are couple of comments on the post, despite the August 2025 date, questioning the timing of the statement, and it does appear that the tape was made in January of 2025. It looks like he is in Taiwan for the New Year's event. Yet, the comments are still instructive. While signing autographs, Jensen is somewhat distracted, but he is obviously providing his frank view of the technology. In stark contrast, only a few months later, at GTC 2025, he has effusive praise for SiPh and CPO, with the targeted audience also involving investment analysts. To be fair, it could be asserted that in the first case, Jensen’s comment was only for the NVLink scale-up fabric. Nevertheless, it still provides insight into his actual expectations for CPO. There is the view that was notably expressed very well by Loi Nguyen, EVP Emeritus, Optics, at Marvell Technology at the Optica Executive Forum at OFC 2025 on scale-up that Nvidia will do everything possible to cramp more GPUs in a single rack with liquid cooling and other techniques to allow them to stay with copper. He also believes that the present CPO effort is on a noncritical path in scaling out on a switch platform in which the pluggable is the primary option -- and so, essentially agrees that it will take several years more for CPO to be ready. Such a timeframe would obviously conflict with Nvidia’s quite optimistic 2026 schedule for shipment. Thus, the business narrative for CPO itself will inexorably remain on a more solid foundation compared with the actual, past realities of engineers struggling over the years to get the solution to work. There will continue to be the basic challenges with production, testing, and dependability. 

37

9 Comments

Telco GPU-as-a-Service doesn't work at the cell site In the wake of the NVIDIA investment in Nokia, there seems to be renewed expectation for GPUs deployed at cell sites for non-RAN applications. Telcos betting on running Gen-AI workloads at every cell-site are playing a high-cost, low-return game. The Math Doesn't Add Up Cell-site GPU deployment ($15K + $650/month OPEX) at 15% utilization costs $33 per 1M tokens. Cloud-scale GPU ($35K + $1,120/month OPEX) at 75% utilization costs $2.85 per 1M tokens. That's 10x more expensive per token at the edge. Adjust any variable, the gap only widens. Utilization Is the Killer A typical cell-site serves ~1,000 users in a small area. Demand peaks together (morning commute, evening), then flatlines. Size for peak? You're idle 85% of the time. Cloud pools thousands of users across time zones. Higher utilization equals lower unit cost. You can't optimize away fundamental demand patterns. When Does Edge AI Actually Make Sense: Autonomous vehicles (<10ms): In-vehicle compute. Wireless adds risk Factory robotics (<50ms): On-premises. You already have infrastructure and staff Medical imaging (<100ms): Hospital data centers. On-prem for compliance + existing infra Consumer AI (<20ms): Cloud or select metro nodes. This is where scale wins Cell sites? Nowhere. They're built for radio, not general-purpose AI workloads. The Brutal Truth: Cell-site inference costs 10-15x more than cloud Utilization stuck at 10-20% Every use case has a better alternative Bottom line: The real edge AI market isn't GPU racks at every cell tower. It's: On-premises enterprise gear (where latency/control matter) Hyperscale metro nodes (cloud pushed closer) Cell towers will remain for network functions, not broad customer AI workloads. #EdgeComputing #AI #5G #TelcoCloud #InfrastructureEconomics #Telecommunications #CloudComputing

362

98 Comments

Dear Friends, For over 16 years, Mobile Future Forward has been where the industry's most consequential conversations happen first. Before 5G was deployed anywhere, it was debated here. Before AI infrastructure became a boardroom priority, the architectural questions were framed in the keynotes and panel discussions. Before FWA, before edge computing, before the current race for sovereign AI, the leaders who would shape these movements were testing ideas, challenging assumptions, and building relationships at MFF. This has never been a conference of scripted keynotes and sponsor pitches. It is a working session among peers: CxOs, chief architects, and strategy leaders from operators, hyperscalers, semiconductor companies, infrastructure vendors, and enterprises, who come not to be talked at, but to think together as a community. The questions facing our ecosystem in 2026 are as thorny as any we have tackled: - The wireless industry is being asked to become AI infrastructure. What does that actually require and who captures the value? - Billions are flowing into AI compute, but the bottlenecks are shifting from training to inference, from cloud to edge, from software to silicon. Where should the next bets be placed? - 6G standardization is accelerating while 5G monetization and diffusion remains incomplete. How do leaders plan across two horizons simultaneously? - Robotics, autonomous systems, and spatial computing are moving from demos to deployment. What will connectivity need to deliver that it cannot deliver today? - What are key deep-tech areas that will shape the global conversations? These are not questions with easy answers. They are questions that demand the kind of direct, unfiltered exchange that only happens when the right people are in the room - without cameras, without agendas, without the theatrics. That platform convenes again on September 10th in Seattle. The 2026 theme will be announced soon. For now, if you could please mark your calendar. If past years are any indication, the conversations that begin here will shape decisions for the year ahead.   If you would like to partner with us, please get in touch at https://lnkd.in/gzeB2Uw More details to follow. Warm regards, Chetan

45