diff --git a/README.md b/README.md
index f9e62b6..724f4da 100644
--- a/README.md
+++ b/README.md
@@ -113,6 +113,36 @@ find the following examples of 64 GPU (8 node) runs:
 - [H100](./mpi-operator/nccl-test-distributed-h100-64-mpijob.yaml)
 - [H100 with SHARP](./mpi-operator/nccl-test-distributed-h100-64-sharp-mpijob.yaml)
 - [GB200 NVL72](./mpi-operator/nccl-test-distributed-gb200-nvl72-mpijob.yaml)
+- [GB200 NVL72 Multi-rack](./mpi-operator/nccl-test-distributed-gb200-128-multirack-mpijob.yaml)
+
+#### Generating Custom Manifests
+
+For custom configurations, use `nccl.nu` to generate manifests dynamically.
+Requires [Nushell](https://www.nushell.sh/).
+
+```bash
+# Generate manifest for 64 H100 GPUs
+./nccl.nu generate --gpu h100 --scale 64
+
+# Generate with SHARP and multi-rack topology
+./nccl.nu generate --gpu h100 --scale 128 --sharp --multirack
+
+# Generate and apply directly to cluster
+./nccl.nu generate --gpu a100 --scale 64 --apply
+
+# Use custom container image
+./nccl.nu generate --gpu h100 --scale 64 --image ghcr.io/custom/image:tag
+```
+
+Available options:
+- `--gpu` - GPU type (a100, h100, a40, gb200, rtxp6000_8x)
+- `--scale` - Total GPU count (must be divisible by GPUs per node)
+- `--multirack` - Enable multi-rack topology spread constraints
+- `--noib` - Disable InfiniBand
+- `--gdrcopy` - Enable GDRCopy
+- `--sharp` - Enable SHARP/COLLNET
+- `--apply` - Apply manifest directly to cluster
+- `--image` - Override container image
 
 #### Running Jobs
 
diff --git a/cue.mod/module.cue b/cue.mod/module.cue
new file mode 100644
index 0000000..f243535
--- /dev/null
+++ b/cue.mod/module.cue
@@ -0,0 +1,9 @@
+module: "github.com/coreweave/nccl-tests"
+language: {
+	version: "v0.15.3"
+}
+deps: {
+	"cue.dev/x/k8s.io@v0": {
+		v: "v0.6.0"
+	}
+}
diff --git a/generate.cue b/generate.cue
new file mode 100644
index 0000000..b6331e8
--- /dev/null
+++ b/generate.cue
@@ -0,0 +1,223 @@
+package main
+
+import (
+	"strings"
+
+	corev1 "cue.dev/x/k8s.io/api/core/v1"
+	metav1 "cue.dev/x/k8s.io/apimachinery/pkg/apis/meta/v1"
+	"github.com/coreweave/nccl-tests/gpus"
+)
+
+// Input parameters (set via -t flag)
+gpu:          *"a100" | string @tag(gpu)
+scale:        *64 | int        @tag(scale,type=int)
+name:         *"" | string     @tag(name)
+multirack:    *false | bool    @tag(multirack,type=bool)
+noib:         *false | bool    @tag(noib,type=bool)
+gdrcopy:      *false | bool    @tag(gdrcopy,type=bool)
+sharp:        *false | bool    @tag(sharp,type=bool)
+image:        *"" | string     @tag(image)
+defaultImage: string           @tag(defaultImage)
+
+// Lookup the GPU config
+_gpu: gpus.gpus[gpu]
+
+// Calculate workers (integer division)
+_workers: scale div _gpu.gpusPerNode
+
+// Resolve image (override takes precedence, then GPU-specific, then default)
+_image: defaultImage
+if image != "" {
+	_image: image
+}
+
+// Generate job name (lowercase)
+_name: "nccl-test-\(scale)-\(strings.ToLower(_gpu.name))"
+if name != "" {
+	_name: name
+}
+
+// Build env var flags string from mpiEnv list
+_gpuEnvFlags: strings.Join([for k, v in _gpu.mpiEnv {"-x \(k)=\(v)"}], " \\\n  ")
+
+// Build gdrcopy env flags
+_gdrcopyEnvFlags: string | *""
+if gdrcopy {
+	_gdrcopyEnvFlags: "-x NCCL_GDRCOPY_ENABLE=1 \\\n  -x NCCL_DEBUG=INFO \\\n  -x NCCL_DEBUG_SUBSYS=INIT"
+}
+
+// Build sharp env flags (COLLNET + ALGO for H100)
+_sharpEnvFlags: string | *""
+if sharp {
+	if gpu == "h100" {
+		_sharpEnvFlags: "-x NCCL_COLLNET_ENABLE=1 \\\n  -x NCCL_ALGO=COLLNETCHAIN"
+	}
+	if gpu != "h100" {
+		_sharpEnvFlags: "-x NCCL_COLLNET_ENABLE=1"
+	}
+}
+
+// Combine all extra env flags
+_allExtraEnv: [for s in [_gpuEnvFlags, _gdrcopyEnvFlags, _sharpEnvFlags] if s != "" {s}]
+_extraEnvStr: strings.Join(_allExtraEnv, " \\\n  ")
+_extraEnv:    string | *""
+if _extraEnvStr != "" {
+	_extraEnv: " \\\n  \(_extraEnvStr)"
+}
+
+// Determine IB HCA setting (eth0 when noib, else ibp)
+_ibHca: string | *"ibp"
+if noib {
+	_ibHca: "eth0"
+}
+
+// Build UCX_NET_DEVICES part (only when has ibDevices and not noib and not gdrcopy)
+_ucxPart: string
+if _gpu.ibDevices != "" && !noib && !gdrcopy {
+	_ucxPart: " \\\n  -x UCX_NET_DEVICES=\(_gpu.ibDevices)"
+}
+if _gpu.ibDevices == "" || noib || gdrcopy {
+	_ucxPart: ""
+}
+
+// Build bind-to part (only when not gdrcopy)
+_bindPart: string | *" \\\n  -bind-to none"
+if gdrcopy {
+	_bindPart: ""
+}
+
+// Build the mpirun command
+_mpirunArgs: "mpirun \\\n  -np \(scale)\(_bindPart) \\\n  -x LD_LIBRARY_PATH \\\n  -x NCCL_SOCKET_IFNAME=eth0 \\\n  -x NCCL_IB_HCA=\(_ibHca)\(_ucxPart)\(_extraEnv) \\\n  /opt/nccl_tests/build/all_reduce_perf -b 512M -e 8G -f 2 -g 1\n"
+
+// Launcher container spec (typed)
+_launcherContainer: corev1.#Container & {
+	name:  "nccl"
+	image: _image
+	env: [
+		{name: "OMPI_ALLOW_RUN_AS_ROOT", value: "1"},
+		{name: "OMPI_ALLOW_RUN_AS_ROOT_CONFIRM", value: "1"},
+	]
+	command: ["/bin/bash", "-c"]
+	args: [_mpirunArgs]
+	resources: {
+		requests: {cpu: "2", memory: "128Mi"}
+		limits:   {cpu: "2", memory: "128Mi"}
+	}
+}
+
+// Determine if we should request IB resources (GPU supports IB and noib is not set)
+_useIB: _gpu.useIB && !noib
+
+// Worker volume mounts (always dshm, optionally gdrdrv)
+_workerVolumeMounts: [...corev1.#VolumeMount]
+if !gdrcopy {
+	_workerVolumeMounts: [{mountPath: "/dev/shm", name: "dshm"}]
+}
+if gdrcopy {
+	_workerVolumeMounts: [
+		{mountPath: "/dev/shm", name: "dshm"},
+		{mountPath: "/dev/gdrdrv", name: "gdrdrv"},
+	]
+}
+
+// Worker container spec (typed)
+_workerContainer: corev1.#Container & {
+	name:  "nccl"
+	image: _image
+	resources: {
+		requests: {
+			cpu:              "\(_gpu.cpu)"
+			memory:           _gpu.memory
+			"nvidia.com/gpu": "\(_gpu.gpusPerNode)"
+			if _useIB {"rdma/ib": "1"}
+		}
+		limits: {
+			"nvidia.com/gpu": "\(_gpu.gpusPerNode)"
+			if _useIB {"rdma/ib": "1"}
+			memory: _gpu.memory
+			cpu:    "\(_gpu.cpu)"
+		}
+	}
+	volumeMounts: _workerVolumeMounts
+}
+
+// Worker pod affinity (typed)
+_workerAffinity: corev1.#Affinity & {
+	if _gpu.podAffinity {
+		podAffinity: preferredDuringSchedulingIgnoredDuringExecution: [{
+			weight: 100
+			podAffinityTerm: {
+				labelSelector: matchLabels: "metadata.coreweave.cloud/job": "nccl-test"
+				topologyKey: "topology.kubernetes.io/zone"
+			}
+		}]
+	}
+	nodeAffinity: requiredDuringSchedulingIgnoredDuringExecution: nodeSelectorTerms: [{
+		matchExpressions: [{
+			key:      _gpu.nodeKey
+			operator: "In"
+			values: [_gpu.nodeType]
+		}]
+	}]
+}
+
+// Worker volumes (always dshm, optionally gdrdrv)
+_workerVolumes: [...corev1.#Volume]
+if !gdrcopy {
+	_workerVolumes: [{name: "dshm", emptyDir: medium: "Memory"}]
+}
+if gdrcopy {
+	_workerVolumes: [
+		{name: "dshm", emptyDir: medium: "Memory"},
+		{name: "gdrdrv", hostPath: path: "/dev/gdrdrv"},
+	]
+}
+
+// Topology spread constraints (typed)
+_topologySpreadConstraints: [...corev1.#TopologySpreadConstraint]
+if multirack {
+	_topologySpreadConstraints: [{
+		maxSkew:           1
+		minDomains:        2
+		topologyKey:       "topology.kubernetes.io/zone"
+		whenUnsatisfiable: "DoNotSchedule"
+		labelSelector: matchLabels: "metadata.coreweave.cloud/job": "nccl-test"
+	}]
+}
+
+// The MPIJob manifest
+manifest: {
+	apiVersion: "kubeflow.org/v2beta1"
+	kind:       "MPIJob"
+	metadata: metav1.#ObjectMeta & {
+		name: _name
+	}
+	spec: {
+		slotsPerWorker: _gpu.gpusPerNode
+		runPolicy: cleanPodPolicy: "Running"
+		mpiReplicaSpecs: {
+			Launcher: {
+				replicas: 1
+				template: {
+					spec: corev1.#PodSpec & {
+						containers: [_launcherContainer]
+					}
+				}
+			}
+			Worker: {
+				replicas: _workers
+				template: {
+					metadata: metav1.#ObjectMeta & {
+						labels: "metadata.coreweave.cloud/job": "nccl-test"
+					}
+					spec: corev1.#PodSpec & {
+						topologySpreadConstraints: _topologySpreadConstraints
+						containers: [_workerContainer]
+						affinity: _workerAffinity
+						volumes:  _workerVolumes
+					}
+				}
+			}
+		}
+	}
+}
diff --git a/gpus/a100.cue b/gpus/a100.cue
new file mode 100644
index 0000000..a96fbe4
--- /dev/null
+++ b/gpus/a100.cue
@@ -0,0 +1,10 @@
+package gpus
+
+gpus: a100: #GPU & {
+	name:        "A100"
+	gpusPerNode: 8
+	nodeType:    "gd-8xa100-i128"
+	cpu:         112
+	memory:      "1000Gi"
+	ibDevices:   "ibp0:1,ibp1:1,ibp2:1,ibp3:1"
+}
diff --git a/gpus/a40.cue b/gpus/a40.cue
new file mode 100644
index 0000000..8322e9d
--- /dev/null
+++ b/gpus/a40.cue
@@ -0,0 +1,11 @@
+package gpus
+
+gpus: a40: #GPU & {
+	name:        "A40"
+	gpusPerNode: 8
+	nodeType:    "A40"
+	nodeKey:     "gpu.nvidia.com/model"
+	cpu:         90
+	memory:      "400Gi"
+	ibDevices:   "ibp0:1,ibp1:1,ibp2:1,ibp3:1"
+}
diff --git a/gpus/gb200.cue b/gpus/gb200.cue
new file mode 100644
index 0000000..f8b88f9
--- /dev/null
+++ b/gpus/gb200.cue
@@ -0,0 +1,21 @@
+package gpus
+
+gpus: gb200: #GPU & {
+	name:        "GB200"
+	gpusPerNode: 4
+	nodeType:    "gb200-4x"
+	cpu:         140
+	memory:      "900Gi"
+	useIB:       false
+	podAffinity: true
+	mpiEnv: {
+		NVIDIA_IMEX_CHANNELS:       "0"
+		NCCL_NET_GDR_C2C:           "1"
+		NCCL_MNNVL_ENABLE:          "1"
+		NCCL_CUMEM_ENABLE:          "1"
+		NCCL_SHM_DISABLE:           "0"
+		UCX_TLS:                    "tcp"
+		UCX_NET_DEVICES:            "eth0"
+		OMPI_MCA_coll_hcoll_enable: "0"
+	}
+}
diff --git a/gpus/gpus.cue b/gpus/gpus.cue
new file mode 100644
index 0000000..09a4d83
--- /dev/null
+++ b/gpus/gpus.cue
@@ -0,0 +1,14 @@
+package gpus
+
+import "github.com/coreweave/nccl-tests/schema"
+
+// Default image passed via tag from nccl.nu (which reads workflow YAML and git SHA)
+defaultImage: string @tag(defaultImage)
+
+// Alias for schema.#GPU with our default image injected.
+// Individual GPU configs can optionally override the default image.
+#GPU: schema.#GPU & {
+	image: string | *defaultImage
+}
+
+gpus: [string]: #GPU
diff --git a/gpus/h100.cue b/gpus/h100.cue
new file mode 100644
index 0000000..e8b4c89
--- /dev/null
+++ b/gpus/h100.cue
@@ -0,0 +1,11 @@
+package gpus
+
+gpus: h100: #GPU & {
+	name:        "H100"
+	gpusPerNode: 8
+	nodeType:    "gd-8xh100ib-i128"
+	cpu:         110
+	memory:      "960Gi"
+	ibDevices:   "ibp0:1,ibp1:1,ibp2:1,ibp3:1,ibp4:1,ibp5:1,ibp6:1,ibp7:1"
+	mpiEnv: SHARP_COLL_ENABLE_PCI_RELAXED_ORDERING: "1"
+}
diff --git a/gpus/rtxp6000_8x.cue b/gpus/rtxp6000_8x.cue
new file mode 100644
index 0000000..07bbbe9
--- /dev/null
+++ b/gpus/rtxp6000_8x.cue
@@ -0,0 +1,10 @@
+package gpus
+
+gpus: rtxp6000_8x: #GPU & {
+	name:        "RTXP6000-8X"
+	gpusPerNode: 8
+	nodeType:    "rtxp6000-8x"
+	cpu:         112
+	memory:      "960Gi"
+	useIB:       false
+}
diff --git a/nccl.nu b/nccl.nu
new file mode 100755
index 0000000..05cf1b5
--- /dev/null
+++ b/nccl.nu
@@ -0,0 +1,223 @@
+#!/usr/bin/env nu
+
+# NCCL Test CLI - Sync cluster GPU types and generate manifests
+
+# Get default image from workflow file and git SHA
+def get-default-image [] {
+    let workflow = (open .github/workflows/ubuntu-22.yml)
+    let job = $workflow.jobs.cu129.with
+    let sha = (git rev-parse --short HEAD)
+    $"ghcr.io/coreweave/nccl-tests:($job.base-tag)-nccl($job.nccl-version)-($sha)"
+}
+
+# Get image tag arguments for CUE commands
+def get-image-tags [] {
+    ["-t", $"defaultImage=(get-default-image)"]
+}
+
+# Generate CUE config file content from node data using cue eval
+def generate_cue_file [node: record] {
+    let labels = $node.metadata.labels
+    let cap = $node.status.capacity
+
+    let gpu_count = $labels | get "gpu.nvidia.com/count" | into int
+    let node_type = $labels | get "node.coreweave.cloud/type"
+    let cpu = $labels | get "cpu.coreweave.cloud/cores" | into int
+    let mem_ki = $cap.memory | str replace "Ki" "" | into int
+    let mem_gi = ($mem_ki / 1048576 | math floor)
+    let ib_speed = $labels | get -o "ib.coreweave.cloud/speed" | default "0G"
+    let has_ib = $ib_speed != "0G"
+    let config_name = $node_type | str replace "-" "_" --all
+    # Detect IB devices from node labels (pattern: ib.coreweave.cloud/neighbors.current.ibp<N>.device)
+    let ib_device_labels = ($labels | columns | where { |k| $k =~ 'ib.coreweave.cloud/neighbors.current.ibp\d+\.device' })
+    let ib_devices = if ($ib_device_labels | length) > 0 {
+        $ib_device_labels
+        | each { |k| $k | parse --regex 'ibp(?P<num>\d+)' | get num.0 | into int }
+        | sort
+        | each { |n| $"ibp($n):1" }
+        | str join ","
+    } else {
+        ""
+    }
+
+    let img_tags = (get-image-tags)
+    let content = (cue eval ./utils/... -e newGpu
+        ...$img_tags
+        -t $"nodeType=($node_type)"
+        -t $"gpuCount=($gpu_count)"
+        -t $"cpu=($cpu)"
+        -t $"memGi=($mem_gi)"
+        -t $"hasIB=($has_ib)"
+        -t $"ibDevices=($ib_devices)")
+
+    {
+        name: $config_name
+        content: $"package gpus\n\n($content)\n"
+    }
+}
+
+# Sync cluster GPU types with CUE config
+def "main sync" [
+    --kubeconfig: string = "",  # Path to kubeconfig file
+] {
+    let kubeconfig = if $kubeconfig != "" { $kubeconfig | path expand } else { "" }
+    let kc = if $kubeconfig != "" { ["--kubeconfig", $kubeconfig] } else { [] }
+
+    # Get GPU types from cluster
+    print -e "Fetching GPU nodes from cluster..."
+    let cluster_nodes = (kubectl ...$kc get nodes -l node.coreweave.cloud/class=gpu -o json
+        | from json
+        | get items)
+
+    let cluster_gpus = ($cluster_nodes
+        | each { $in.metadata.labels | get "node.coreweave.cloud/type" }
+        | uniq)
+
+    # Get GPU configs from CUE
+    print -e "Reading GPU configs from CUE..."
+    let img_tags = (get-image-tags)
+    let cue_gpus_raw = (cue export ./gpus/... -e gpus ...$img_tags | from json)
+    let cue_gpu_names = ($cue_gpus_raw | columns)
+
+    let cue_gpus = ($cue_gpu_names | each { |name|
+        let cfg = ($cue_gpus_raw | get $name)
+        {name: $name, nodeType: $cfg.nodeType}
+    })
+
+    let cue_node_types = ($cue_gpus | get nodeType)
+
+    # Build status table
+    let all_types = ($cluster_gpus | append $cue_node_types | uniq | sort)
+
+    let status = ($all_types | each { |t|
+        let in_cluster = $t in $cluster_gpus
+        let in_config = $t in $cue_node_types
+        let config_name = if $in_config {
+            ($cue_gpus | where nodeType == $t | get name.0)
+        } else { "" }
+
+        {
+            "GPU Type": $t
+            "Cluster": (if $in_cluster { "✓" } else { "✗" })
+            "Config": (if $in_config { $"✓ \(($config_name)\)" } else { "✗" })
+        }
+    })
+
+    print ""
+    print ($status | table)
+
+    # Find missing configs
+    let missing = ($cluster_gpus | where { |t| not ($t in $cue_node_types) })
+
+    if ($missing | length) > 0 {
+        print $"\n($missing | length) GPU type\(s\) in cluster without config:\n"
+
+        for gpu_type in $missing {
+            let create = (input $"Create config for ($gpu_type)? [Y/n] ")
+            if ($create | str downcase) in ["", "y"] {
+                # Get node details
+                let node = (kubectl ...$kc get nodes -l $"node.coreweave.cloud/type=($gpu_type)" -o json
+                    | from json
+                    | get items.0)
+
+                let cue_file = (generate_cue_file $node)
+                let file_path = $"gpus/($cue_file.name).cue"
+                $cue_file.content | save -f $file_path
+                print $"Created ($file_path)"
+            }
+        }
+    } else {
+        print "\nAll cluster GPU types have configs ✓"
+    }
+}
+
+# Generate NCCL test manifest
+def "main generate" [
+    --kubeconfig: string = "",  # Path to kubeconfig file
+    --gpu: string = "",         # GPU type (skip prompt if provided)
+    --scale: int = 0,           # Number of GPUs (skip prompt if provided)
+    --multirack,                # Enable multi-rack topology spread
+    --noib,                     # Disable InfiniBand
+    --gdrcopy,                  # Enable GDRCopy
+    --sharp,                    # Enable SHARP/COLLNET
+    --apply,                    # Apply manifest to cluster
+    --image: string = "",       # Override container image
+] {
+    let kubeconfig = if $kubeconfig != "" { $kubeconfig | path expand } else { "" }
+
+    # Get image tags from workflow
+    let img_tags = (get-image-tags)
+
+    # Get available GPUs from CUE
+    let gpus_json = (cue export ./gpus/... -e gpus ...$img_tags | from json)
+    let gpu_names = ($gpus_json | columns)
+
+    # Select GPU
+    let selected_gpu = if $gpu != "" { $gpu } else {
+        $gpu_names | input list "Select GPU type:"
+    }
+
+    if $selected_gpu not-in $gpu_names {
+        print -e $"Error: GPU '($selected_gpu)' not found. Available: ($gpu_names | str join ', ')"
+        exit 1
+    }
+
+    let gpu_info = ($gpus_json | get $selected_gpu)
+    let gpus_per_node = $gpu_info.gpusPerNode
+
+    # Select scale
+    let selected_scale = if $scale > 0 { $scale } else {
+        let default = $gpus_per_node * 8
+        let scale_input = (input $"Number of GPUs [default: ($default), ($gpus_per_node)/node]: ")
+        if $scale_input == "" { $default } else { $scale_input | into int }
+    }
+
+    # Validate scale is divisible by gpusPerNode
+    if ($selected_scale mod $gpus_per_node) != 0 {
+        print -e $"Error: Scale ($selected_scale) must be divisible by gpusPerNode ($gpus_per_node)"
+        exit 1
+    }
+
+    let workers = ($selected_scale / $gpus_per_node | into int)
+    print -e $"Generating: ($selected_gpu) × ($selected_scale) GPUs \(($workers) workers\)\n"
+
+    # Generate manifest using CUE
+    let mr = if $multirack { ["-t", "multirack=true"] } else { [] }
+    let noib_flag = if $noib { ["-t", "noib=true"] } else { [] }
+    let gdrcopy_flag = if $gdrcopy { ["-t", "gdrcopy=true"] } else { [] }
+    let sharp_flag = if $sharp { ["-t", "sharp=true"] } else { [] }
+    let img_override = if $image != "" { ["-t", $"image=($image)"] } else { [] }
+    let manifest = (cue export -e manifest --out yaml -t $"gpu=($selected_gpu)" -t $"scale=($selected_scale)" ...$mr ...$noib_flag ...$gdrcopy_flag ...$sharp_flag ...$img_tags ...$img_override .)
+
+    if $apply {
+        let kc = if $kubeconfig != "" { ["--kubeconfig", $kubeconfig] } else { [] }
+        $manifest | kubectl ...$kc apply -f -
+    } else {
+        print $manifest
+    }
+}
+
+def main [] {
+    print "NCCL Test CLI
+
+Commands:
+  sync      - Compare cluster GPU types vs CUE configs, create missing
+  generate  - Interactive manifest generation
+
+Options:
+  --kubeconfig <path>  - Path to kubeconfig file
+  --noib               - Disable InfiniBand
+  --gdrcopy            - Enable GDRCopy
+  --sharp              - Enable SHARP/COLLNET
+
+Examples:
+  ./nccl.nu sync --kubeconfig ~/Downloads/kubeconfig
+  ./nccl.nu generate
+  ./nccl.nu generate --gpu a100 --scale 64
+  ./nccl.nu generate --gpu a100 --scale 64 --noib
+  ./nccl.nu generate --gpu a100 --scale 64 --gdrcopy
+  ./nccl.nu generate --gpu a100 --scale 64 --sharp
+  ./nccl.nu generate --gpu h100 --scale 64 --sharp
+  ./nccl.nu generate --gpu h100 --scale 128 --apply --kubeconfig ~/kube.conf
+  ./nccl.nu generate --gpu h100 --scale 64 --image ghcr.io/coreweave/nccl-tests:custom-tag"
+}
diff --git a/schema/gpu.cue b/schema/gpu.cue
new file mode 100644
index 0000000..54e9bb1
--- /dev/null
+++ b/schema/gpu.cue
@@ -0,0 +1,15 @@
+package schema
+
+#GPU: {
+	name:        string
+	image:       string
+	gpusPerNode: int
+	nodeType:    string
+	nodeKey:     string | *"node.coreweave.cloud/type"
+	cpu:         int
+	memory:    string
+	ibDevices: string | *""
+	useIB:       bool | *true
+	mpiEnv: [string]: string
+	podAffinity: bool | *false
+}
diff --git a/utils/new_gpu.cue b/utils/new_gpu.cue
new file mode 100644
index 0000000..e100ea0
--- /dev/null
+++ b/utils/new_gpu.cue
@@ -0,0 +1,36 @@
+package utils
+
+import (
+	"strings"
+	"github.com/coreweave/nccl-tests/schema"
+)
+
+// Input parameters via tags
+_nodeType:     string @tag(nodeType)
+_gpuCount:     int    @tag(gpuCount,type=int)
+_cpu:          int    @tag(cpu,type=int)
+_memGi:        int    @tag(memGi,type=int)
+_hasIB:        bool   @tag(hasIB,type=bool)
+_ibDevices:    string @tag(ibDevices)
+_defaultImage: string @tag(defaultImage)
+
+_configName: strings.Replace(_nodeType, "-", "_", -1)
+
+// The generated config - validated against #GPU schema
+_newConfig: schema.#GPU & {
+	name:        strings.ToUpper(_nodeType)
+	image:       _defaultImage
+	gpusPerNode: _gpuCount
+	nodeType:    _nodeType
+	cpu:         _cpu
+	memory:      "\(_memGi)Gi"
+	if _hasIB {
+		ibDevices: _ibDevices
+	}
+	if !_hasIB {
+		useIB: false
+	}
+}
+
+// Output structure for file generation
+newGpu: gpus: (_configName): _newConfig