Fluidstack

Our network fleet grows faster than we can hire network engineers, so we built an agent to absorb the investigation work. Agent-deploy runs in the terminal. An operator types a question in plain English, "overall fleet status?", and the agent fans out across the fleet, gathers device state through typed tools, and lands on a conclusion the evidence supports. It walks the whole troubleshooting path without stopping to ask what to check next, and it cannot touch the network without operator approval. An independent analysis of a popular agent loop put ~1.6% of the code in AI decision logic and ~98.4% in permission gates, context management, tool routing, and recovery. Our engineering went into the 98.4%, and a network agent needs different things there than a coding agent does. Determinism: the same question about the same fleet state takes the same path to the same answer. Eval-ability: every run leaves a recorded artifact we can replay and grade, and a prompt tweak or model upgrade that regresses a graded trajectory does not ship. Impact foresight: before this earns write access, it has to predict blast radius deterministically. A general coding agent keeps its safety in the system prompt as a polite request. Ours lives in code, and a test covers it. Every device query returns structured results tagged with the device they came from, and before the app writes an export or a saved config, it checks that every line traces back to tool output from the device it claims to describe. The payoff we did not plan for: new engineers use it to learn the network. What does this alarm mean, why did we design it this way, what does healthy look like. They ask the fleet instead of waiting on a senior engineer. Kavan Smith (Kavan S.), Senior Network Automation Engineer, wrote up how we built it: https://lnkd.in/e2HNTfqZ Open roles on the team: fluidstack.com/jobs

On Tuesday we hosted The Decade Ahead in San Francisco, a dinner on the future of AI infrastructure. Fluidstack builds supercomputers for the world's leading AI labs, and progress at this pace will not be carried by any one industry. We were joined by construction leads and electricians, software engineers and operators, dealmakers and schedulers. Everyone has a part to play. Thank you to everyone who attended. This was the first of many, and there's more coming soon. We're hiring across every department. Open roles: fluidstack.io/jobs

Another one this week: NYSE celebrated Fluidstack's $830M Series A with a Trading Floor Takeover. Thank you NYSE. We're hiring across all departments: fluidstack.io/jobs

A nice moment for the team this week: Nasdaq featured Fluidstack’s $830M Series A on their Times Square tower. We’re hiring across all departments: fluidstack.io/jobs

The Decade Ahead: a dinner on the future of AI infrastructure. We just confirmed our $830M Series A. Fluidstack engineers the skeleton of artificial intelligence: the steel, the silicon, and the software. We were chosen to lead one of the largest infrastructure projects in US history, and our job is to deliver it at speed and scale: deploying data centers the size of Central Park in months, and standing up the supercomputers within them in days. Progress at this scope and pace requires experts from across industries. It needs construction leads and electricians, software engineers and datacenter operations techs, dealmakers and schedulers. Everyone has a part to play in building the future of technology. If that sounds interesting to you, come join us for dinner tomorrow in SF: good food and a real conversation about where AI infrastructure is heading, with the people building it. We're hiring across every department, and an evening at the table is the best way to meet us. Open roles: fluidstack.io/jobs Space is limited. Apply here: luma.com/osffvtq0

In January, Fluidstack raised an $830M Series A at a $7.5B valuation, led by Situational Awareness, with participation from world-leading investors. Fluidstack builds infrastructure for the leading AI labs - our goal is to be the fastest on the planet at deploying hundreds of gigawatts of compute. As we stand at the event horizon of the singularity, humanity’s best chance is if democracies, with error correcting institutions, free speech, and checks on power, imbue those same principles into superintelligence. Whoever deploys frontier compute infrastructure fastest will decide whether Al expands human freedom or shrinks it. We are hiring for hundreds of roles - come join us. https://lnkd.in/gzBQ3MBY

I’ve joined Fluidstack as Director of Design Engineering, Capacity Delivery — designing and building gigawatt-scale AI compute infrastructure at speed and scale, while staying lean and efficient. FluidStack builds and operates the data centers behind frontier AI, and every function owns its outcomes end to end. We’re hiring broadly across the company, especially in Engineering. If you’re the kind of person who wants to see your work energized and serving racks, my DMs are open!

We are hiring across the security program: detection engineering, incident response, infrastructure security, and physical security for live sites. Incident Response: https://lnkd.in/gWtnwmqe Infra Security: https://lnkd.in/gh7CDiTT Detection: https://lnkd.in/gnDYsqEC

The uncomfortable truth about securing frontier AI infrastructure is that the adversaries most capable of reaching it have effectively unlimited time, budget, and technical depth to spend compromising you and your systems. It's not a question of if, it's a question of when, and your only option is to out-adapt them. Most organizations start this off by attempting to build the perfect defensive perimeter. Unfortunately, capable adversaries treat every defensive architecture as a starting position, not an insurmountable barrier. They get a vote in how the game is played, and they'll take the path of least resistance to gain access: a phishing email, a supply-chain attack on a vendor, a misconfigured API, an interdicted truck, or a guard shack at 3am. The list goes on and on. History bears this out: states have buried massive nuclear enrichment facilities inside mountains to defeat overhead surveillance, and state-nexus intrusion campaigns against critical infrastructure have dwelled inside networks for years before anyone noticed. Forecasters thinking about the next decade of AI competition have extended this logic in published thought experiments, imagining data centers disguised as heavy industry or built underground with deep water cooling to hide their thermal signature from satellites. Whether or not those specific scenarios ever materialize, the underlying assumption is the right one: actors with infinite resources and patience should be presumed to be probing frontier AI infrastructure today. Real detection often comes down to something vanishingly small: one anomalous log line, one badge swipe that doesn't fit a pattern, a single alert that a tired analyst almost closes. Generating the signals is not hard, making sense of them across a complicated socio-technical landscape is. We're not chasing a mythical "complete" security architecture: we're building for agility against adversaries with high offensive optionality, which is a different kind of program than what most enterprises run. This looks like unified detection spanning IT, OT/ICS, and physical security, correlating signals across domains that traditionally never talked to each other. We're building towards something more ambitious: a security program sophisticated enough to defend the infrastructure underpinning the most consequential technology of this century, with defensive depth that holds up at every layer of the pyramid of pain, from the trivial indicators an adversary discards, to the tools, techniques, and procedures that force them to retool entirely. If you're interested in these problems - bare metal, corporate, or physical security; detection and response; data pipelines - browse open roles: https://lnkd.in/gDwr-FGv

Our latest hackathon winner: RackScout, an AR copilot for the data hall. Put on the headset and the hall annotates itself - a glowing path to the rack you need, checklists pinned to the hardware, and racks turning green as they go production-ready. It runs on real placement, inventory, and handover data: power and thermal heatmaps across the hall, fiber uplink paths drawn in place, NOC tickets routed to the rack they belong to. The usual version of this job is heads-down on a laptop, tabbing between systems, stitching the picture together yourself. RackScout also carries a voice copilot built on Claude that knows which rack you are standing at and which ticket you are working - so guidance stays hands-free while your hands stay on the hardware. Built in a company hackathon by Alexander Coppess, a senior staff network engineer at Fluidstack. Engineers here build things like this: https://lnkd.in/gXQ7yrki