The modern world runs on networks. Progress in fields as diverse as self-driving cars, large LLM training runs, and e-commerce is desperately bottlenecked on designing and debugging large complex networks. Traditionally, top-tier network engineers earn their expertise through countless hours of study, rigorous practice, and internalizing principles that become second nature. They rely on real-time dashboards and command-line interfaces to detect sluggish connections, reconfigure hardware, and alleviate traffic bottlenecks, all while keeping systems online. We want to give network engineers a 100x multiplier by building a foundation model for networking from the ground up—fast, informed, and always learning.

We don't just manage racks like this—we build them. Every component in this diagram, from PDUs and security appliances to switches, access points, and cellular hardware, is developed in-house from the ground up. We write the firmware, operating systems, distributed systems, and APIs that power them. Meter's different because we make our own product decisions across the entire stack: hardware, firmware, software, and operations. This full-stack approach lets us optimize performance at every layer ensuring a seamless experience for our customers.
Manually troubleshooting a rack like this often involves repetitive tasks: verifying firmware versions, tracing cables, adjusting firewall rules, or rerouting traffic. Each misstep can lead to downtime, financial loss, and frustrated customers.
Consider this snippet from a live network:
timestamp | interface | rx_packets | tx_packets | errors | latency (ms) | packet_drops |
2025-01-26T10:00 | eth0 | 105,000 | 95,000 | 0 | 15 | 0 |
2025-01-26T10:01 | eth0 | 112,000 | 102,000 | 3 | 20 | 0 |
2025-01-26T10:02 | eth0 | 118,000 | 108,000 | 7 | 35 | 1 |
2025-01-26T10:03 | eth0 | 124,000 | 115,000 | 12 | 50 | 2 |
In this scenario, not only do error counts increase, but latency and packet drops begin to climb—a red flag that something deeper is amiss. Typically, these anomalies require a seasoned network engineer to embark on an exhausting diagnostic journey. Imagine the following multi-layered decision tree that outlines this process:
Anomaly Detected: [↑]Errors [↑]Latency [↑]Drops ──────────────────────────────┬─────────────────────────────── │ ┌────────────────┴───────────────┐ │ │ ▽ ▽ Physical Layer Logical/Configuration Examination Analysis │ │ │ ┌───────┴───────┐ │ │ │ ▽ ▽ ▽ Inspect Cables & Examine Review Analyze Routing Hardware Sensors Firmware & Policies │ │ │ │ Connectors Temperature, Version, Load │ Voltage, etc Logs Balancing, │ │ │ ACLs │ │ │ │ ▽ ▽ ▽ ▽ ┌Connectors┐ ┌Overheating?┐ ┌Outdated?┐ ┌Misconfigured?┐ │ │ │ │ │ │ │ │ │ │ │ │ │ ▽ │ │ ▽ │ │ ▽ │ │ ▽ │ │[ ] Yes │ │ [ ] Yes │ │[ ] Yes │ │ [ ] Yes │ │ │ │ │ │ │ │ │ │ │ │ │ │ ▽ │ │ ▽ │ │ ▽ │ │ ▽ │ └────Fix───┘ └───Replace──┘ └──Update─┘ └────Correct───┘ │ │ │ │ └─────────────┴─┬───────────────┘ │ │ │ ▽ ▽ ┌──────────If No Issues, Evaluate ─────────────Audit Firewall┐ │ Environmental Interference & Rules & QoS │ │ Power Stability Policies │ │ │ └───────Cross-Reference with Real-Time Traffic Analytics─────┘
In practice, each branch of this decision tree can further subdivide. For example, "Analyze Routing & Policies" might involve verifying if recent configuration changes correlate with the observed anomalies, or if a transient spike in traffic has overwhelmed load-balancers. The process is iterative, often requiring the engineer to switch between physical diagnostics and logical analysis as new data emerges.
Meter's end-to-end autonomy speeds everything up. Our model ingests time-series data, packet headers, and hundreds of JSON events in real time, continuously calling Meter's internal APIs for milliseconds-fast diagnostics and resolutions. We train on the largest corpus of networking text—and validate against every support ticket Meter has ever logged—to ensure the system not only matches human results but resolves issues before they are flagged. Beyond real-world examples, we expose the model to a vast array of synthetic data, covering even the rarest edge cases. So whether it's a routine glitch or an obscure anomaly, our model has already seen a version of the problem and can fix it before it escalates. It doesn't stop at reactive troubleshooting—it is engineered to drive network design, configuration, and management, continuously optimizing the network to adapt to evolving demands. Our model transforms traditional network management from a reactive process into a proactive, self-optimizing system.
We recently released our custom-model product: Command that acts as an interface to our backend — a generative UI that lets users manage networks and build custom dashboards in plain English.
Compute
We just signed a long-term compute partnership with Microsoft, giving us access to tens of thousands of GPUs. With the team we're building, each person will effectively have 2,000+ H100s at their disposal.
People
At Meter, we believe the best products are built at the intersection of applied research and obsessive product development. We operate networks in the real world, at scale—thousands of locations, millions of devices—where mistakes carry real costs, and the stakes are high.
We're assembling a team of people who live at the center of a three-circle Venn diagram: the top 0.1% in research, the application of that research, and product thinking. This work demands a blend of ambition and velocity that stretches comfort zones—and we're looking for the builders who thrive on it. If you're excited by real-world impact and the prospect of shaping how networks evolve, join us.