Cosmic Labs

GPU builders, inference engineers, and training teams are meeting in SF on July 23. Join engineers ai&, NVIDIA, ByteDanceOSS, RadixArk, Liquid AI, ZML, Graphsignal, Cosmic Labs, Kernelize, and more for an evening of lightning talks and networking. Hosted by Crusoe and SGLang. RSVP: https://luma.com/rxsn0u0h

Nearly US$1 trillion in announced sovereign compute commitments are now being made globally as governments treat AI infrastructure as a strategic asset. Meg McNulty, Co-founder of Cosmic Labs, argues that sovereign compute isn't about building a domestic Nvidia or achieving full-stack AI independence. It's about retaining agency over critical layers of the AI stack while managing dependencies elsewhere. With frontier AI concentrated among a small number of labs, GPU clusters, and cloud providers, countries are increasingly viewing compute alongside energy, communications, and defense as foundational infrastructure. The real differentiator won't be capital commitments alone. It will be operational competence, utilization, and the ability to run production AI workloads effectively. Read the full piece: https://lnkd.in/d6aQT35Y Christine Tsai | Tony Wang | Vishal Harnal | Atika French | Ally Boffa-Taylor | Iris Sun | Leonard Lee

Anthropic is down because their recovery systems are too slow. High demand just exposes the gray failures their monitoring and response abilities miss. If you know someone on Anthropic's platform team, tell them there's a better way to do this. Tell them we need to chat. GPUs and racks and networks break at scale. It is predictable. The only thing that matters is how fast you predict, detect, reroute, and recover. We operate at the hardware layer to stop failures before they cascade. That shrinks the blast radius and keeps the platform running. If they were running Cosmic Labs, the scale of this outage would not be happening.

Three things happened in the last 24 hours that tell you everything about where AI infrastructure is headed. Meta signed a $60 billion, five-year deal with AMD for 6 gigawatts of AI compute, starting with AMD's MI450 GPUs in the second half of this year. Meta gets an option to buy up to 10% of AMD. This is AMD's second anchor deal after OpenAI, and it's the clearest signal yet that hyperscalers are done being single-vendor. MatX, a startup founded by former Google TPU engineers, raised $500 million in a Series B led by Jane Street and Leopold Aschenbrenner's Situational Awareness fund. Their chip, the MatX One, claims 10x better throughput for inference than NVIDIA GPUs. It doesn't exist in silicon yet. They're fabbing at TSMC with shipments targeted for 2027. The round valued the company at several billion dollars for a pre-production chip. Then NVIDIA reported Q4 earnings: $68.1 billion in revenue, beating estimates by nearly $2 billion. Data center revenue grew 75% year over year to $62.3 billion. They guided $78 billion for next quarter, five billion above what Wall Street expected. NVIDIA said Vera Rubin samples shipped to customers this week. Read these three stories together. NVIDIA is still growing at 73% YoY and printing $97 billion in free cash flow. But Meta is spending $60 billion to make sure it doesn't depend entirely on one supplier. And VCs are writing $500 million checks to startups whose chips won't ship for another year, betting the demand will be there when they arrive. The GPU market isn't shrinking. It's getting company. ASIC share in AI servers is projected to grow from 20.9% in 2025 to 27.8% this year, while GPU share drops from 75.9% to 69.7%. Not because GPUs are failing, but because the workloads are diversifying faster than any single architecture can cover.

Attending #PCT26 Sunday through Tuesday 🎉 If you are building in telecom, data centers, or edge compute and want to talk about what we are doing at Cosmic Labs, shoot me a message. We are building the control plane for complex infra: real-time awareness and control. For telecom and data center operators, that means: ➡️ Everything after rack-and-stack: config, flashing, validation, lifecycle ➡️ Automated debugging that cuts MTTR from days to minutes ➡️ Hardware-aware orchestration and #GPU scheduling across any vendor ➡️ Visibility & control into across physical layer (GPUs, CPUs, FPGAs, and more) that other tools cannot see The teams running satellite networks, terrestrial backhaul, and clouds are dealing with the same fundamental problem: hardware complexity is scaling faster than engineering headcount. 📩 If any of this resonates, or you know someone at PCT I should meet, drop me a DM or comment below.

Happy New Year from Cosmic Labs 🚀 2025 was a year of building—our team, our product, and relationships with engineers who are tired of the same broken workflows. The pain point we're obsessed with: systems integration. It's fragmented, manual, and eats months of engineering time. There's no unified way to flash, configure, validate, and debug across hardware, vendors, and environments. We're building the engine that changes that. One platform, one intelligence layer, universal control. This quarter alone, we launched our SF office, kicked off some insane contracts, and grew our engineering team. (If you want in, check the comments for our open roles!) We also laid serious groundwork across industries we care deeply about. Lots more to share soon. To our team: thank you for betting on hard problems. To the data center/infra engineers, embedded teams, and systems integrators pushing aerospace, automotive, AI infrastructure, and silicon forward: we're building this for you. 2026 is going to be fun. Let's build 💪 💻 🚀

Hi, LinkedIn! I'll be attending #NeurIPS2025 this year in San Diego. DM me if you'll be around and want to grab coffee! We're especially excited about the frontiers of agentic system design and the intersection of ML and hardware. (P.S. we're hiring... details listed in the comments.)

We're #hiring a new Founding Engineer, Embedded in San Francisco Bay Area. Apply today or share this post with your network.