Le Perspective

Armadin, an AI-native cybersecurity company, has raised $255.5 million in Series B funding at a valuation of over $2.5 billion. The round was co-led by Andreessen Horowitz and Accel, with new investors Bain Capital Ventures (BCV) and Redpoint joining, alongside returning backers including 8VC, Ballistic Ventures, GV (Google Ventures), In-Q-Tel, Kleiner Perkins and Menlo Ventures. That takes Armadin's total funding to $445 million. The problem it targets: AI is shrinking the gap between a vulnerability being disclosed and a working exploit existing. Periodic penetration tests can't keep up, and scanners rate each flaw in isolation, missing how attackers actually operate. Armadin's approach is to attack first. It deploys a swarm of AI agents that reason like a skilled adversary across a company's systems chaining together individually minor weaknesses into complete, validated attack paths, from an initial break-in at the perimeter through to full cloud compromise. Security teams see the exact routes a real attacker would take, and how much damage each would cause, so they can close them first. It's already running these campaigns in production for Fortune 500 and government customers. "AI lets an attacker find and chain weaknesses faster than any human team can respond," said CEO Kevin Mandia. "The only way to build a defense that keeps pace is to train it against the best offense available, every day." Armadin was founded by Kevin Mandia (CEO) who previously founded Mandiant, sold to FireEye for $1 billion and later to Google for $5.4 billion along with Travis Lanham (CTO), Evan Pena (Chief Offensive Security Officer) and David Slater (Chief Architect). Barbara Brian Frank Michael David Trevor Ping David Steph Nick Anna Varun Christian Andrew Harry Scott Kevin Meg

Halluminate, a data research lab, has raised $30 million in Series A funding led by Oak HC/FT, with participation from Y Combinator, Orange Collective, FT Partners and Heavybit. That takes its total capital raised to $38.5 million. The problem it's working on: AI can now solve some of the world's hardest math problems and write most of our software domains where answers are easily verifiable. But those same systems still struggle with everyday professional work, like building a realistic financial model, producing a client-ready presentation, or writing a balanced strategic analysis. Halluminate builds the benchmarks and reinforcement learning environments essentially realistic simulated workplaces where frontier AI models can practise that kind of work and be measured on it. Its first focus is financial services and consulting, with plans to expand into adjacent verticals. In August it released Westworld Finance Diligence Bench, a benchmark of 88 problems testing AI agents across a complete company acquisition due-diligence process, drawn from anonymised real transactions and reviewed by practising deal professionals. The company says it has gone from zero to a mid-eight-figure revenue run rate in 10 months while remaining profitable, and has built RL environments with four of the five leading closed-source US AI labs. Halluminate was founded in 2024 by Jerry Wu (CEO), previously product and research lead at Capital One Labs, and Wyatt Marshall (CTO), whom he met at Cornell. Alinа Victoria Jeremy Ronok Aravind Matt Dave Jon Leah Colleen Lexi Christopher Jackie

doxx.net has launched the open beta of its Agentic Defined Networking platform, alongside a $38 million Series A led by Andreessen Horowitz, with participation from Animo Ventures and Focal.vc. a16z partner Joel De La Garza joins the board. The idea: a private, parallel network that people and their AI agents can build and run on. Users can browse privately, transfer files directly, connect devices, and make calls, video and chat peer to peer with no server in the middle, and no email, phone number or account required to sign up. The agent angle is central. As AI agents browse and act on a user's behalf carrying their data, accounts and authority they can stumble into fake login pages or malware-hosting sites. doxx.net gives agents their own network identity and a policy-defined environment with DNS-level threat protection that blocks malicious destinations before a connection is made. The company says it has blocked over 38 million threats since its closed beta began in December 2025. Its pitch is that privacy comes from architecture rather than policy: the platform runs end to end on infrastructure the company owns its own IP space, DNS root, certificate authority and bare-metal servers rather than renting capacity from public clouds. doxx.net was founded by Barrett Lyon, a serial founder who pioneered the DDoS mitigation industry with Prolexic (acquired by Akamai), and later founded BitGravity (acquired by Tata Communications), XDN (acquired by Fortinet) and Defense.net (acquired by F5). Nico Pascal Daniel Zenobia Martin Zane Yoko Jason Owen Antonio

Volantis Semiconductor, a semiconductor startup using light to move data inside computers, has raised an $88 million Series A led by Lachy Groom and Abstract, taking its total funding to $97 million. Backers include Sam Altman, Jeff Dean, Dylan Patel and John Doerr. The problem it's attacking is AI's real bottleneck: memory. Today's chips force a trade-off on-chip memory is fast but tiny, while the high-capacity memory used in GPUs is large but slow to access. That ceiling is what limits how big a model can be and how quickly it can respond. Volantis says its optical interconnect delivers both, boosting memory capacity and bandwidth by over an order of magnitude by replacing electrical links with light. To get there, the team built its own optical stack eliminating external lasers in favour of integrated micro-VCSELs, and doing away with traditional optical fibre. What that could unlock, per the company: running models with over 10 trillion parameters at speeds up to 10,000 tokens per second per user. Its own example a coding agent finishing a task in 30 seconds instead of 30 minutes. Volantis was founded in 2022 by Tapa Ghosh (Founder & CEO), a Thiel Fellow and serial chip entrepreneur, and Roy Meade (Co-Founder & CTO), previously the first employee and VP of Engineering at Ayar Labs, who led high-bandwidth memory development at Micron. The wider team has built industry first products at NVIDIA, AMD, Broadcom and Ayar Labs. Beth Rishabh Ramyad Ashutosh Dachuan Sreeni Horst John David John Naveen Dwarkesh Sholto Cameron Steven Dan

Satlyt, a startup building software to run AI directly on satellites, has raised an $8 million seed round led by Houston-based non sibi ventures. The problem it's solving: satellites generate huge amounts of data, but sending it back to Earth ("downlink") is slow and expensive, and spacecraft typically rely on ground controllers to fix anything that goes wrong. Satlyt's software lets AI models run onboard the satellite itself resolving anomalies and processing sensor data in orbit. Earlier this year it deployed Google DeepMind's Gemma model on a spacecraft operated by Momentus, cutting the size of a transmission about onboard software errors by more than 60%. Unlike SpaceX, Google or startups building their own orbital data centres, Satlyt isn't building spacecraft. It's building software that works across many companies' satellites — an approach its founder sums up neatly: "The folks like SpaceX who are doing orbital data centres — if they are the iPhone, we'll build Android as a horizontally integrated, open ecosystem." Satlyt was co-founded by Rama Afullo, a Kenyan American engineer who previously worked in Google's cloud business and at SpaceX's Starlink. He says he pitched the idea internally at both companies and was turned down — so he left to build it himself. The company is headquartered in Sunnyvale, California, with an office in Nairobi. Notably, this week's mission puts Satlyt's software onboard a spacecraft built by TakeMe2Space, an Indian startup making computing hardware for satellites, with NASA among the customers testing cloud-computing protocols in space. Nelson Junn Leina Sharon Isaac Bernard Kent Ronak Uwem Okendo Zachariah

Hewlett Packard Enterprise has announced a $1.2 billion order from Vultr, the world's largest privately-held cloud infrastructure company, to deploy AMD Helios AI systems across Vultr's US data centres. It's HPE's first order for the new AMD Helios system. The systems are built for the heaviest AI workloads training very large models and running them at high volume. Each rack packs 72 AMD Instinct MI455X GPUs alongside AMD EPYC "Venice" CPUs, connected by HPE Juniper Networking switches and cooled with direct liquid cooling. The design uses open, standards-based Ethernet rather than proprietary interconnects. "AI is driving the largest infrastructure buildout in history, and together with AMD and Vultr, we are building an open foundation to accelerate the next phase of AI," said Antonio Neri, President and CEO of HPE. J.J. Kardwell, CEO of Vultr, framed the demand side: "Demand for high-performance AI infrastructure continues to outpace available capacity… Together, we're delivering the memory capacity, interconnect performance, and energy efficiency customers need to run the largest and most demanding AI training and inference workloads." The deal builds on nearly three years of work between Vultr and Juniper Networks, which is now part of HPE. Bhawna Sajan Aditya Mark Nilotpal Mayank Radhakrishnan Jaya Vinay Arvind Jerome Vivek Rami Mark Kevin Prashanth Rahul Pratik Ajay Amaresh

Unveilr AI, a Mumbai-based startup, has raised a pre-seed round from ajvc, the pre-seed fund founded by Aviral Bhatnagar. Founded by Sanditya Srivastava and launched in February 2026, Unveilr AI helps B2B and consumer brands get discovered not just on Google, but inside AI search tools like ChatGPT, Perplexity, Gemini and Claude, where a growing share of buyers now start their research. Its AI agents work to improve how often and how prominently a brand shows up in those AI-generated answers, turning that visibility into leads. The company says results show up fast: for one Indian B2B legal services platform, it increased AI-sourced traffic around 4.7x and leads 5x, along with an 80% rise in ChatGPT citations. Its team includes IIT alumni alongside SEO and AEO specialists, and it currently serves brands across India, the UAE and the US. The funds will go toward expanding its engineering team, building more capabilities into its AI agents, and growing its sales team across those three markets. Aviral Subajit Mohit Puja Kartik Debanirmalya

Google has announced Gemini 4 Argon, its new frontier AI model built for complex, long-running professional work software engineering, enterprise knowledge work like legal and finance, and cybersecurity defence. A few notable points from the announcement: It expands the output limit to 1 million tokens (up from 64K), giving the model room to reason through a problem in one go. Google says Argon is already running internally helping optimise quantum computing subroutines, freeing over 300 TiB of memory across its data centres, and migrating large C/C++ codebases to Rust. On cybersecurity, it can autonomously find, validate and patch software vulnerabilities. Security firm Wiz says Argon uncovered a critical flaw exposing personal data in healthcare software used by hospitals worldwide one earlier models had missed. Rollout is phased: Argon is going first to trusted cyber defenders through Google's Fairwind Program, while Google strengthens safeguards around misuse, prompt-injection attacks and misalignment, before a wider release to developers, enterprises and consumers. Introductory pricing is $2 per million input tokens and $10 per million output tokens. Sundar koray Hugo Antonio Marcelo Manish Prateek Gautam Rahul Abhinav Rahul Shraddha Mustafa Tulsee Logan Vivek Naren Roma Ragini

President Donald Trump hosted some of the biggest names in technology and AI at the White House, including Elon Musk, Mark Zuckerberg, Sundar Pichai, Jensen Huang, Dario Amodei, Greg Brockman, Satya Nadella, Jeff Bezos and Lisa Su. Trump and leaders from Google, Meta, Nvidia, Anthropic, OpenAI and xAI signed the voluntary “Joint Commitment on Frontier Responsibilities,” outlining internal controls, independent audits and board-level oversight for frontier AI systems. But one detail quickly drew attention: in the publicly shared signed document, “United States” was misspelled as “Unites States” directly beneath Trump’s signature.

Restate, a durable execution company founded by the creators of Apache Flink, has raised $20 million in a Series A round led by Singular, with participation from Redpoint Ventures and Capital One Ventures. That brings its total funding to $27 million. Software has always had to survive interruptions servers restart, connections drop, APIs go down mid-task. "Durable execution" is what lets software remember what it has already done and pick up correctly after a failure, instead of developers hand-coding recovery into every step. AI agents raise the stakes. They run for long stretches, call models and APIs, use tools, and interact with other agents and humans so when something breaks, there's far more to preserve than just the task itself. Restate's answer is a purpose-built distributed log and event-driven orchestrator that holds execution, state and communication together, without needing an external database. It's already in production: Replit uses Restate for all durable orchestration behind its Replit Agent, alongside customers including DOSS, Bolt.new, Coralogix, BILT Rewards and Fortune 500 firms. Restate was co-founded by Stephan Ewen (CEO), Till Rohrmann, Igal Shilman and Ahmed Farghal the team behind Apache Flink, the stream-processing framework used at Apple, Netflix, Uber, LinkedIn and Stripe. With teams in Berlin and San Francisco, the funding will support product development and a new US commercial hub. Shane Hollie Jamie Francesco Jan Giselle Daniela Nikrooz Zack Henri Erica Michele Kelsey Naomi