Amir Halfon

I’ve met David Schaeffer, the founder and CEO of Cogent Communications, a few times over the years. In those interactions, Dave always came across as brilliant - measured in tone, armed with precise statistics, and deeply analytical. He could confidently explain a complex capital structure, including the securitization of IP address assets, and offer thoughtful prognostications about long-term internet usage and declining unit costs. For years, Dave expressed skepticism about owning fiber and data center infrastructure, arguing that such assets were commodities and easily sourced. Then Cogent acquired from T-Mobile Sprint’s legacy assets for $1, including aging fiber routes and obsolete telco data centers. That’s why the news this summer that Dave had purchased Washington, D.C. real estate, pledging his Cogent shares as collateral to lever office buildings and that the shares were ultimately seized by lenders following a margin call was so jarring. It was a remarkable, and frankly sad, story to hear relayed in real time. With Cogent stock down roughly 70% over the past 18 months, it is hard not to conclude that the combination of a collapsing share price and a deteriorating office real estate market proved catastrophic. Since Dave’s mea culpa, the stock is down further, and the company slashed its quarterly dividend by about 98%, reducing it from roughly $1.02 per share to just $0.02 per share. This week we learned that Cogent’s Board amended Dave’s employment agreement to grant substantial new equity awards, extend his term through 2028, and introduce a $1,000,000 base salary, despite his long-standing practice of being compensated primarily through dividends tied to his equity ownership. After cutting the dividend, Dave is now paid a salary and bonus; after losing a significant portion of his stock in a margin call tied to bad real estate decisions, the company is now issuing him new equity.   Only in America.  https://lnkd.in/gpGc_Jxe

57

17 Comments

At its last earnings conference, Coherent brought up “200-gig VCSEL-based 1.6T transceivers ramping in the second half of this calendar year” (please see: https://lnkd.in/eJETqkmi), albeit putting an emphasis on CPO and MPO apps (please see: https://lnkd.in/eQ6uVt2p). At least right now, the opportunities for those devices at that rate, especially in the west, appear to be substantially lower than what occurred at 100G per lane. As the major hypersclalers moved away from reliance on multimode fibers, the potential for significant demand was mainly at the mercy of Nvidia (please see: https://lnkd.in/dbN9zdF6). It turns out that the supplier, like those large users, decided against planning to employ 200G VCSELs for 1.6T MM optics modules. In fact, it is fibeReality’s understanding that Nvidia even designated the 100G VCSEL-based solution for scale-out as “end of life” because of the problem with link flapping. Although there is still at least theoretically a market for VCSELs, involving short reach with active optical cables, the bad news is that with scale-up, the requirement is even more demanding, with the bit error rate higher than for silicon photonics (please see: https://lnkd.in/eXCmm5g5). Therefore, it seems that the main, current opportunity for sales of 200G VCSELs, presumably also effecting Broadcom, which excels at making bare die for those products, could seemingly be for cloud usage in China, probably not large in size (please see: https://lnkd.in/eSvw4eJa). Furthermore, the presently most optimistically new play for VCSELs would be for slow and wide, 50G-NRZ deployment (please see: https://lnkd.in/eSVdfH-b). The way it has been characterized to fibeReality by a large end-user, is in possibly utilizing a “micro-VCSEL.” In general, though, just as the case with NRZ, one should still be careful about quickly throwing in the towel on the future of what was a short time ago, considered to be a critically important, low-cost network element. There has been movement (including standards work) in the direction of developing VCSELs up to 50 meters at the 200G rate and at the 1060nm wavelength for scale-up, requiring a new fiber with sufficiently high EMB to reach at least 30 meters.

53

3 Comments

A few weeks ago at the NHAI summit in NYC, Rom Carmel 🚀(aside from Francis Odum) was among the only people talking about the role intent will play in access control for AI agents. Everyone I spoke to at the event told me it was the first time they'd heard an approach to agentic security that felt both concrete and designed for the dynamics that make agents fundamentally different from anything that's come before. I've been at Apono for 3 years. I've seen a lot of impressive things from this team. Today's release of Agent Privilege Guard is the first one to leave my jaw on the floor (Ofir Stein and our product/engineering org are mavericks every one of them). The state of AI adoption today is moving fast, breaking things, and patching access gaps after the damage is done. Amazon just showed us what happens when AI-assisted actions run without adequate guardrails. Last week an autonomous agent hacked McKinsey's AI platform in two hours. The privileges agents need to deliver on the productivity gains everyone is chasing are the same privileges that put organizations at risk. The market needs a model for secure AI adoption that actually works at runtime. Evaluating declared intent and enforcing continuous authorization is that model. See how Apono is defining the future of privileged access: https://lnkd.in/gPqtrbst

29

1 Comment

Every release starts with a problem we can't stop thinking about. For this one it was simple: too many personal injury cases fall apart not because of bad lawyering, but because of gaps. A document that didn't make it in. A lien that got misclassified. A follow-up that never happened. This week we shipped our March release at Supio to close those gaps. Automatic document intake that routes files to the right case. Instant lien classification across every document in the file. A complete economic picture for every case in under 30 minutes. Not features for features' sake. Solutions to problems we heard directly from the firms we work with every day. Proud of the team that built this one. Read more about the release here → https://lnkd.in/gi3BEzGM Jerry, our CEO, and @Niki, our CMO, also wrote great breakdowns on what this means for customers. Take a look here: https://lnkd.in/gw6XqR5u https://lnkd.in/gsN6h9hE #LegalTech #AI #ProductManagement #PersonalInjury #Supio

18

1 Comment

Stoked to announce that TracPlus is partnering with CAL FIRE, one of the world's leading wildland firefighting agencies! We're a Kiwi company that builds mission-critical intelligence for aerial firefighting operations, and we're honoured that CAL FIRE has chosen our FireFlyte platform to power their next-generation aviation management system (CATIS). CAL FIRE protects California's 31 million acres with 9,000 firefighters and operates one of the world's largest aerial firefighting fleets - and they're trusting TracPlus to provide the real-time tracking and operational insights that help them deploy resources more effectively. This partnership sets a new global standard for aerial firefighting and directly benefits communities by enabling faster response times, more effective resource deployment, and operational intelligence when wildfires threaten homes and lives. Pretty special for a small company from New Zealand!

32

4 Comments

Just returned from CES and one thing is clear - AI is everywhere and it's here to stay! All of which simply validated what we at Invysta have thought about for a long time - How to Neutralize AI-Driven Identity Theft? 1. The Threat: The Death of Biometrics & Passwords By 2026, generative AI has turned traditional Multi-Factor Authentication (MFA) into a liability: • Deepfake Bypass: AI can now spoof facial recognition and voice patterns in real-time during video calls or "liveness" checks. • Automated Social Engineering: AI agents can trick employees into revealing SMS codes or "push" notifications with 99% success rates. • Session Hijacking: High-speed bots steal "browser cookies," allowing hackers to bypass the login process entirely. 2. The Solution: Invysta’s AI-Immune "Hardware Anchor" Invysta’s Anonymous Access Key (AAK) removes the "Human" from the vulnerability equation. It replaces spoofed signals with a cryptographic "Hardware DNA" that AI cannot simulate or steal. • Invisible to AI Interrogation: AAK is generated at the system level, not the application level. It is shielded from the "data-gathering interrogation" techniques used by AI scrapers and botnets. • Device-Defined Defense: Even if a deepfake convinces an admin to reset a password, the attacker still cannot log in unless they possess the exact physical device tied to the user’s account. • Zero-Knowledge Handshake: The system validates the endpoint device without ever transmitting a password or biometric signature over the network, leaving no "digital footprint" for AI to harvest.

16

2 Comments

To my Boston network: If AI is on your radar (and let’s be honest, whose isn’t these days?), join me on April 1st for a live in-person event featuring real humans talking about AI. (A sequel event hosted by AI talking about real humans may or may not be in the works - stay tuned for updates.) In all seriousness, big thanks to Pipeline360 for organizing this - I’m thrilled to be part of the panel alongside some incredible B2B marketing leaders. We’ll be diving into how marketers are moving from AI experiments to real execution, with plenty of actionable insights to take back to your teams. Hope to see you there! https://lnkd.in/eE6V_69v

33

1 Comment

Most teams don’t struggle with AI agents because the technology is immature. They struggle because prompting is not treated with design discipline. An agent prompt is not a single instruction. It’s a system of assumptions, constraints, and decision rules. When those aren’t explicit, reliability breaks down fast. That’s why two teams can use the same models and get completely different outcomes. This guide breaks down how to design agent prompts the way operators design systems. Intentionally, repeatably, and with scale in mind. Full breakdown here: https://lnkd.in/gzhuQgRW #Agents #Prompting #AI

294

12 Comments