Nectar
01
The Spark trial
Prove fleet compute fits your stack: a clean Kubernetes join and telemetry that reconciles with your ground truth. Integration and trust, not performance claims.
live in days, not quarters.
A new category · fleet compute
Physical AI runs on fleet compute.
Shared GPUs on the factory floor for a whole robot fleet - run for you as a service. One pool makes the whole fleet smarter - and takes you from POC to production.
Backed by & building with
The opportunity
Robots are finally getting a brain.
A new frontier of models - robotics foundation models - lets general-purpose robots perceive, reason, and act. That makes factory deployment finally practical.
01
Perceive
Make sense of complex, changing environments.
02
Reason
Plan across long-horizon, multi-step tasks.
03
Act
Execute precise actions in the real world.
a continuous loop, many times a second
The dead end
One model, one robot, one GPU doesn't scale.
- Idle by design. The GPU waits on the robot's control loop; one robot can't batch.
- Bought for peak. Every robot sized for its own worst case.
- Outgrown by the model. Robot-grade silicon trails server-grade by 29× memory bandwidth and 3.5× compute - the planning models don't fit.
Provision the whole fleet this way and per-robot and naively-shared serving both collapse at scale: in factory-scale benchmarks, SLO-met falls to 0% as the fleet grows.
The category
The missing layer has a name.
Fleet compute / noun /
Shared, on-site AI compute that serves a whole fleet's real-time inference from one pool - close enough for the control loop, big enough for the model - delivered and managed as a service.
The published pattern
One pool. The whole pipeline. Every agent.
NVIDIA Research + Stanford published ROSA (2026) - a fleet of robots sharing one server-class GPU pool, an SLO-aware scheduler routing every agent's full RFM pipeline, every action qualified against its latency budget.
Abstraction of ROSA, Figure 2 - system overview
~7×
less hardware
than a GPU sized per robot
than a GPU sized per robot
12×
factory productivity
SLO-qualified actions per hour, pooled vs dedicated
SLO-qualified actions per hour, pooled vs dedicated
99.96%
SLOs held at fleet scale
SLO-met at 32 robots in benchmarks
SLO-met at 32 robots in benchmarks
Why Nectar
The pattern is public. The operating layer is the product.
As Physical AI scales, the value moves up-stack - off the box and onto the loop that keeps it dependable. Nectar is that managed operating layer: the substrate, the fused reliability loop, and 24/7 ops, run for you.
Who runs the compute where Physical AI works?Procure
Zero capex. Opex only - capital stays in your product, not depreciating GPUs.
Reliability
24/7 ops, run for you. No NOC to staff, no alert fatigue to own.
Capacity
Scale in a day. Compute headroom provisioned Day 1 - grow at will.
Staffing
Your team builds robots. We run the infrastructure.
Your compute, your data, your models - our loop to keep them production-grade.
The system
Nectar's Box + Brain = fleet compute as-a-service.
One node on your floor: the hardware that serves the pool, and the loop that keeps it dependable - delivered and operated for you.
Inside the node
on your floor · on your LAN
Hardware
The Box
The immersion-cooled enclosure and the GPUs inside it - the shared pool that serves your fleet's inference, milliseconds away. Sized in robots, not chips: your whole model set resident, every loop in budget, with headroom for the fleet you're growing into.
Software · reliability loop
The Brain
The on-box control loop that keeps the pool dependable - it fuses GPU, power, cooling, network, and workload signals, autoscales within reserved headroom, and holds tail latency in band.
$0Upfront capexOpex · monthly fee
~30%Less power drawimmersion vs air cooling
~50%Less floor footprintimmersion vs air · DIY
1 wkSpark trial proves integrationone-week POC
Data sovereignty
Your data stays on your floor.
Workload data - payloads, prompts, logs - never leaves your network.
Only operational telemetry goes out; only managed updates come in. Never your data, never models learned from your fleet.
Your floor · LAN
Your fleet
workload
Nectar Box
workload data stays inside
boundary · workload data never crosses
OUTHealth & performance telemetry
INManaged software & policy updates
Nectar · managed NOC · off-site
How it starts
Two steps. Trust first, scale next.
A free one-week trial that proves fleet compute fits your stack - then a production Box pilot that proves the numbers at site scale.
02
The Box pilot
A production node that proves the operating KPIs at site scale: tail-latency stability, time-to-capacity, intervention reduction.
then managed capacity - more Boxes as the fleet grows; refresh risk stays ours.
Your workload data never leaves your network.
FAQ
Questions fleet operators ask
Shared, on-site AI compute that serves a whole fleet's real-time inference from one pool - close enough for the control loop, big enough for the model. Instead of a GPU sized for each robot's peak, one pooled Box on your floor serves every agent and trains between shifts. Nectar delivers and operates it as a service.
Tight control loops run 30–200 Hz - a 5–33 ms budget. Practical cloud round-trips measure 50–150 ms: outside the loop, before egress cost. The Box keeps inference and data on-site.
Neither. The Box serves the inference your stack calls; your agent control and orchestration stay yours, with a safe fallback if the Box is unavailable.
Two paths. K8s Join - the Box joins your existing control plane as a worker node via kubeadm or k3s-agent - a join designed for minutes, not days; the one-week Spark trial proves it on your cluster before anything scales. Standalone - it runs alongside your stack with no cluster join. Either way, your workloads aren't re-platformed and your CI/CD runs as it stands.
You choose Brain's access, too: a scoped-RBAC joined-cluster mode, or a read-only shadow mode with no cluster access at all.
Sized in robots, not chips: a Box serves your whole fleet's model set with every loop in budget and headroom to grow. Under the hood that's high-VRAM, H200/B200-class GPUs - up to 24 per Box, with reserved headroom so you can scale utilization instantly. We describe the tier, not exact SKUs.
No. Your workload data never leaves your LAN. What crosses the boundary is operational, both ways: health and performance telemetry goes out to Nectar; managed software and policy updates and remote remediation come in from the NOC. Never your data, never models learned from your fleet.
A two-step on-ramp. First, a one-week POC on a Brain-managed NVIDIA DGX Spark - it proves Nectar joins your stack cleanly and that its telemetry reconciles with your ground truth. Then a production Box pilot that proves the operating KPIs at site scale over 60–90 days. No upfront capex - you pay opex.
Still weighing it? The trial is non-binding - start with the week.
Request a trial
Kick off the Spark trial.
About a week on a pre-installed NVIDIA DGX Spark. The week proves integration - that Nectar slots into your stack and that its telemetry reconciles with your ground truth. Not the operating KPIs; those come with the Box pilot.
- Proves integration and trust. Telemetry reconciles with your ground truth - not the operating KPIs.
- Self-serve K8s join. No disruption to running workloads.
- Your workload data never leaves your network.
- Non-binding. A clean, reversible test.
Trial request received.
We'll reach out within one business day to schedule your Spark unit and walk through the K8s join. Nothing changes on your floor until you say so - and your workload data never leaves your network.