Pulumi Blog: Gcp

Kubernetes Agent Sandbox: What It Is and How to Deploy It with Pulumi

Adam Gordon Bell — Tue, 21 Jul 2026 00:00:00 +0000

When you use a coding agent, it can seem like there’s a trade-off between autonomy and permissions. If you approve every command, it’s safe but slow. Let it do whatever it likes and it works more autonomously, but as the nx supply-chain attack showed, that can go badly.

The fix is to give the agent a sandbox: a box it’s allowed to wreck, with limited permissions and scoped network access. The only files are the checkout you handed it, the only credentials are the task’s own, and trashing the machine just means a disposable pod gets garbage-collected early. Pulumi Neo works this way, and if you want to scale that pattern up inside your own organization, the Kubernetes project Agent Sandbox is a great path to building your own. This post is what it is and how to deploy it on GKE with Pulumi.

What is Agent Sandbox?

Agent Sandbox is a Kubernetes SIGs project that gives AI agents isolated, disposable environments as Kubernetes resources: a Sandbox custom resource, backed by gVisor or Kata Containers for kernel-level isolation.

You could build Agent Sandbox yourself. You’d need gVisor support, a userspace kernel that sits between the agent’s code and your host. With that in place, you could approximate a sandbox for every agent by stringing together a StatefulSet of size one, a headless Service, and a PersistentVolumeClaim, plus some lifecycle machinery to keep warm pools of nodes around. Agent Sandbox wraps all of that up as a CRD, so you can run a Kubernetes cluster where each sandbox is a disposable, kernel-isolated environment a coding agent runs in.

apiVersion: agents.x-k8s.io/v1beta1
kind: Sandbox
metadata:
 name: demo-sandbox
spec:
 podTemplate:
 spec:
 runtimeClassName: gvisor
 containers:
 - name: agent
 image: ubuntu:24.04  # swap in your coding-agent image
 command: ["sleep", "infinity"]

The Sandbox CRD, the whole idea in one manifest.

There are two common patterns for using it. In the first, every coding agent session in the organization maps to its own pod with a persistent volume, basically Claude Code running across a whole fleet of pods, each one an individual user’s session.

In the second, you’re building your own Neo or Devin: a productized agent harness that needs to spin up execution environments. The harness lives outside the sandboxes and hands each task a disposable box, and Agent Sandbox is what spins those boxes up and down, suspending and resuming them.

Why not just run agents in Docker?

So why do we need gVisor or Kata Containers at all? The agent already runs inside a container. Isn’t that the box?

Not really. A container isn’t much of a security barrier. Every container on a host shares one kernel, and the Linux kernel exposes a huge surface area, 450+ syscalls, to every one of them. CVE-2019-5736 is the canonical example: a malicious container tricks runc into overwriting the host’s own runc binary, and after that every docker run on the host runs attacker code as root. Put a prompt-injectable AI agent in that container, and the risk is obvious.

gVisor, the runtime underneath Agent Sandbox’s default path, puts a userspace kernel written in Go between your agent and the host, which limits the possible security surface area. A kernel bug the agent can reach becomes, mostly, a crash in a userspace process rather than a root shell on your node. That is what makes an Agent Sandbox secure.

What is GKE Sandbox?

GKE Sandbox is GKE’s gVisor feature: a RuntimeClass named gvisor that runs a pod under the userspace kernel described above. It predates Agent Sandbox by years and is a per-pod isolation primitive, not an agent tool. Agent Sandbox is the lifecycle layer that sits on top of it: the CRD’s runtime selection points down at gVisor (or Kata) for the actual kernel isolation. They compose. On GKE, Agent Sandbox with runtimeClassName: gvisor is literally using GKE Sandbox underneath.

Google engineers in Kubernetes SIG Apps maintain Agent Sandbox, Janet Kuo and Justin Santa Barbara (of kOps) among them, and it launched at KubeCon NA in November 2025. It’s also one layer of a broader cloud-native agent stack taking shape: agent-sandbox for isolation, kagenti (IBM) for identity, kagent (Solo.io, a CNCF Sandbox project) for agent logic, agent-substrate for density. As Solo.io’s Lin Sun puts it, “Sandboxing your agents is necessary, but not sufficient.”

The rung above gVisor: hardware microVMs

gVisor filters syscalls in a userspace kernel. The heavier option gives the workload its own guest kernel behind hardware virtualization (KVM), so an escape has to cross a CPU-enforced boundary. Firecracker, the VMM AWS built for Lambda (and the isolation layer under Bedrock AgentCore), keeps that boundary cheap: roughly 50,000 lines of Rust against QEMU’s ~1.4 million lines of C. Google’s kvmCTF pays $250,000 for a KVM escape, which is how the market rates that boundary. Agent Sandbox reaches this rung through Kata Containers via the same runtime selection, and the tradeoff isn’t linear: Kata can beat gVisor on I/O-heavy work, because its guest runs a real kernel servicing syscalls natively.

The one-second problem

Isolation is the hard part, but it isn’t the only one. If you’re in scenario 2, using Agent Sandbox pods as the backing instances for your own agent harness, startup time is a challenge too. People abandon chat sessions, and when they come back they expect the agent to pick up quickly, not wait on a machine to boot.

Booting a fresh Kubernetes pod costs about a second of overhead. That’s nothing for a rolling deployment, but enough that the maintainers say it “breaks the continuity” of an interaction. So Agent Sandbox avoids paying that cost twice: suspend a sandbox and its pod goes away while its workspace and identity persist, and warm pools keep pre-booted pods ready for new sandboxes to adopt. The managed GKE version goes further and restores from memory snapshots, the same snapshot-restore move E2B (~150ms) and Fly’s Sprites (~300ms) use to pull the effective cold start well under a second¹.

Deploying it on GKE with Pulumi

If you already have a cluster and want to try Agent Sandbox, the kubectl apply from the project’s quickstart will probably get you going. But productionized use takes a bit more. Let’s walk through setting up a cluster with a sane access policy, named individuals, and a dedicated environment per user. The full program is at pulumi/examples/gcp-ts-agent-sandbox.

Move 1: the cluster. First we stand up a GKE cluster with a gVisor node pool. The one line that matters is the sandboxConfig, with sandboxType: "gvisor". As far as I’m aware, this can only be done on GKE:

const gvisorPool = new gcp.container.NodePool("gvisor-pool", {
 cluster: cluster.name,
 nodeConfig: {
 machineType: "n2-standard-4",
 imageType: "COS_CONTAINERD", // gVisor requires COS + containerd
 sandboxConfig: { sandboxType: "gvisor" },
 },
});

GKE installs the gvisor RuntimeClass, labels the nodes, and taints them so only pods that opt in land there.

That auto-install is the GKE convenience, not a requirement. Agent Sandbox itself is just CRDs and a controller, so on any other cluster you can install gVisor yourself (or Kata), register a RuntimeClass, and everything below works the same.

Move 2: the install, pinned. Next we install the controller and the CRDs, using the all-in-one sandbox-with-extensions.yaml manifest that shipped in last week’s v0.5.2²:

const base = `https://github.com/kubernetes-sigs/agent-sandbox/releases/download/v0.5.2`;

const agentSandbox = new k8s.yaml.v2.ConfigGroup("agent-sandbox", {
 files: [`${base}/sandbox-with-extensions.yaml`],
});

Pinning matters here: releases land every week or two, and the API group already graduated once (v1alpha to v1beta1).

Move 3: sandboxes are a loop, not a manifest. This is where Pulumi really shows its advantages. In practice you don’t create sandboxes by hand, because they’re per-user or per-task. So the demo reads a list and maps over it:

const developers = config.requireObject<Developer[]>("developers");

// Who may open which box, derived from that same list.
const acl = sandboxAcl(developers);

export const sandboxes = developers.map(dev => new AgentSandbox(`sbx-${dev.name}`, {
 owner: dev.name,
 credentials, // this owner's Claude Code creds, per box, not per cluster
 prompt: BOOT_PROMPT,
 // ...
}, { dependsOn: [agentSandbox, operator, acl] }));

That developers array is read at runtime, so it could just as easily be a GitHub team or whoever currently has a session open, none of which kubectl apply -k can loop over. (The developers list and the per-box credentials are stack config; the example’s README has the exact pulumi config set commands.)

Nothing in the sandbox is Claude-specific, either. The agent is just what the image installs: swap in Codex CLI, or point one of the Claude Code orchestration frameworks at it, and the isolation story doesn’t change.

Move 4: the egress policy. Create a Sandbox directly, as we do here, and no NetworkPolicy applies³: kernel isolation, but wide-open egress. (The SandboxTemplate extension manages a default policy for template-created sandboxes; a raw Sandbox gets nothing.) We set a policy that lets the agent reach api.anthropic.com and npm but not any private IPs:

egress: [{
 to: [{ ipBlock: {
 cidr: "0.0.0.0/0",
 except: ["10.0.0.0/8", "172.16.0.0/12", "192.168.0.0/16", "169.254.169.254/32"],
 }}],
}, /* + a DNS rule to kube-system:53 */]

We also exclude 169.254.169.254 to keep the agent off the node metadata server.

Move 5: private access with Tailscale. The sandbox runs an IDE (code-server) on port 13337. Instead of a public load balancer with a password, we put each sandbox on a tailnet with the Tailscale Kubernetes operator. Three annotations on the Service do it:

"tailscale.com/expose": "true",
"tailscale.com/hostname": name, // sbx-adam
"tailscale.com/tags": `tag:${name}`, // tag:sbx-adam

Move 6: lock down each sandbox. Finally, we generate the tailnet access policy from the same developer list that creates the boxes, so each person can reach only their own box:

acls: developers.map(d => ({
 action: "accept",
 src: [d.email], // this person
 dst: [`tag:sbx-${d.name}:13337`], // this box, this port. Nothing else.
})),

Because the boxes and the grants come from one array, a developer’s sandbox and their permission to open it can’t drift apart.

Move 7: run it. With all of that declared, pulumi up builds the whole environment in one command: cluster, gVisor pool, controller, per-developer sandboxes, egress policy, and private access. Get the stack’s URL and open it:

$ pulumi stack output sandboxUrls
["http://sbx-adam:13337"]

Open it on any device signed into your tailnet and you’re in a full VS Code (code-server), running inside the sandbox pod, on the gVisor node pool, behind the egress policy set above.

Each box also boots with a task. The prompt in the demo asks the agent to figure out where it is: read /proc/version, decide whether it’s in a container or a VM, and write up the evidence in a REPORT.md. An agent doing forensics on its own jail is a decent smoke test that the isolation is real.

Here’s the finished result:

Recorded a little demo: Here's how every developer gets their own kernel-isolated dev box on Kubernetes: VS Code + Claude Code in a pod behind Tailscale with gVisor underneath for security isolation. write up coming
— Adam Gordon Bell (@adamgordonbell) July 20, 2026

Wrapping up

Agent Sandbox makes a kernel-isolated, disposable agent environment a first-class Kubernetes object, and Pulumi is how you stand one up as one program instead of a runbook.

The full program, everything in this post, deploy to teardown, is at pulumi/examples/gcp-ts-agent-sandbox. Clone it, point it at a GCP project, and you have your own dedicated agent cluster in about twelve minutes. If you’re new to Pulumi, get started here.

github.com/pulumi/examples

Warm pools plus snapshot restore are how the managed version keeps this fast at scale: Google’s GKE Agent Sandbox launch cites 300 sandboxes per second at sub-second latency. See Bringing you Agent Sandbox on GKE and Agent Substrate. ↩︎
The all-in-one manifest is new in v0.5.2 (July 17): earlier releases split the install into manifest.yaml plus extensions.yaml, and v0.5.2 renamed the core file — one more reason to pin the version you deploy. ↩︎
Verified against the v0.5.2 release manifests: sandbox-with-extensions.yaml installs no NetworkPolicy objects, and networkPolicyManagement defaults to Managed only for SandboxTemplate-created sandboxes. A directly-created Sandbox gets neither, so restricting egress is left to you. See the v0.5.2 release. ↩︎

Fully Automated AI Inference on AWS, Azure, and Google Cloud with Pulumi

Engin Diri — Tue, 30 Jun 2026 00:00:00 +0000

Putting Ollama on a cloud GPU is something I keep coming back to. A while ago I wrote up running open-source LLMs on an AWS EC2 box with Ollama and Pulumi, and the shape never really changes: a GPU instance, a model server, and a firewall rule in front. Infrastructure as code earned its place by making that kind of setup predictable and repeatable, and AI infrastructure is no exception. A GPU box serving a model is still a VM, a disk, and a firewall rule, and it should be declared like one.

Thorsten Hans made exactly that case in his Akamai post, Fully Automated AI Infrastructures with Terraform and Akamai Cloud, which stands up a single GPU instance on Linode, installs the drivers, runs Ollama, and pulls a model, with no manual steps after terraform apply.

I liked the shape of it, so this post ports the same idea to Pulumi and runs it across AWS, Azure, and Google Cloud instead of one. The result is one program shape per cloud: a single pulumi up brings up a GPU box that installs its own driver, runs Ollama, and pulls a model with no manual steps, and a single pulumi destroy takes it back down. Along the way it drops the two imperative bits the Terraform version leans on: a static access token sitting in an environment variable, and a null_resource running a shell loop to wait for the model. The first becomes an OIDC login from a Pulumi ESC environment, so no long-lived key lives anywhere. The second turns out not to be a resource at all.

What you are building

Strip away the per-cloud naming and every version of this is the same three things: a GPU virtual machine, a firewall in front of it, and a cloud-init script that turns a bare Ubuntu box into a running inference server. The model serving runs on Ollama, which exposes an HTTP API on port 11434 and keeps the model resident in GPU memory between requests.

flowchart LR
Dev([Your machine / curl]) -->|"HTTP :11434"| FW["Firewall / security group<br/>(allow 11434, optional 22)"]
FW --> VM["GPU VM (Ubuntu 24.04)<br/>NVIDIA driver + Ollama"]
VM -->|"resident in GPU memory"| Model["qwen2.5:14b"]
ESC["Pulumi ESC<br/>pulumi-idp/auth"] -.->|"OIDC login (short-lived creds)"| Pulumi["pulumi up"]
Pulumi -.->|"AWS / Azure / GCP"| VM

Component	Port	Description
GPU VM	-	Ubuntu 24.04 with an NVIDIA T4-class GPU. Installs the driver and Ollama from cloud-init, then pulls the model.
Ollama	11434	Serves the model over an HTTP API and keeps it in GPU memory between calls.
Firewall / security group	-	Allows inbound `11434` (and `22` when you ask for it). Open to the world for a demo; lock the CIDR down for anything real.
`pulumi-idp/auth` (ESC)	-	One environment that brokers an OIDC login into AWS, Azure, and GCP. No static keys in code or CI.

One detail here is worth pausing on, and it explains why the Pulumi version comes out shorter than the Terraform one. The Akamai project ends with a null_resource that runs a curl loop in local-exec to block terraform apply until the model finishes downloading. That is not infrastructure but a runtime check wearing a resource costume. Pulumi has a Command provider that would let you reproduce it line for line, and this post deliberately does not. The program provisions the box, the box pulls the model on its own, and the program prints the endpoint. Whether the model has finished downloading yet is a question you answer with a curl, not a question your IaC tool should be holding a deployment open to ask.

Prerequisites

Before getting started, ensure you have:

Pulumi CLI installed and configured
A Pulumi Cloud account (ESC and OIDC live here)
An account on at least one of AWS, Azure, or Google Cloud, with an OIDC trust set up for Pulumi (covered below)
Enough GPU quota in your target region for one T4-class instance; a fresh account often starts at zero, which is the most common reason a first deploy fails, so check it before you run pulumi up
Node.js 18+ for the TypeScript program
curl to talk to the inference endpoint once it comes up

This post serves the qwen2.5:14b model, the same one the Akamai article uses, because the 4-bit quant fits comfortably on a single 16 GB T4. The model is one config value (model), so swap in anything Ollama supports; match the GPU to the model’s memory footprint, because a model that overflows VRAM still runs but spills onto the CPU and crawls.

Credentials without the copy-paste

The Terraform version authenticates the only way a single-cloud demo can: you mint a personal access token, export it as LINODE_TOKEN, and the provider reads it from the environment. It works, but that token is long-lived, it sits in your shell history and your CI secrets, and you mint one per cloud. Across three clouds that is three static keys to rotate and worry about.

Pulumi ESC (Environments, Secrets, and Configuration) replaces it all with an OIDC login. The idea: instead of storing a cloud key, you store a trust relationship. At deploy time, ESC presents a short-lived OIDC token to AWS, Azure, or Google Cloud, and each one hands back temporary credentials scoped to a role you control. Nothing long-lived is ever written down.

I keep that wiring in one environment, pulumi-idp/auth, and every stack imports it. Here is the whole thing:

# pulumi-idp/auth: one ESC environment that brokers an OIDC login into all three
# clouds. Every stack imports it. No static cloud key lives here or anywhere else.
values:
 aws:
 login:
 fn::open::aws-login:
 oidc:
 roleArn: arn:aws:iam::123456789012:role/pulumi-esc
 sessionName: pulumi-esc
 azure:
 login:
 fn::open::azure-login:
 clientId: aaaaaaaa-bbbb-cccc-dddd-eeeeeeeeeeee
 tenantId: aaaaaaaa-bbbb-cccc-dddd-eeeeeeeeeeee
 subscriptionId: /subscriptions/00000000-0000-0000-0000-000000000000
 oidc: true
 gcp:
 login:
 fn::open::gcp-login:
 project: 123456789012 # numeric project number, not the project ID
 oidc:
 workloadPoolId: pulumi-esc
 providerId: pulumi-esc
 serviceAccount: pulumi-esc@my-project.iam.gserviceaccount.com
 environmentVariables:
 # AWS: read by the Pulumi AWS provider, the AWS SDKs, and the aws CLI
 AWS_ACCESS_KEY_ID: ${aws.login.accessKeyId}
 AWS_SECRET_ACCESS_KEY: ${aws.login.secretAccessKey}
 AWS_SESSION_TOKEN: ${aws.login.sessionToken}
 # Azure: read by the azure-native provider
 ARM_USE_OIDC: "true"
 ARM_CLIENT_ID: ${azure.login.clientId}
 ARM_TENANT_ID: ${azure.login.tenantId}
 ARM_SUBSCRIPTION_ID: ${azure.login.subscriptionId}
 ARM_OIDC_TOKEN: ${azure.login.oidc.token}
 # Google Cloud: read by the Pulumi Google Cloud provider
 GOOGLE_CLOUD_PROJECT: ${gcp.login.project}
 GOOGLE_OAUTH_ACCESS_TOKEN: ${gcp.login.accessToken}

You only need the blocks for the clouds you actually deploy to. Opening this environment makes ESC perform a live OIDC login for each cloud listed, so trim it to the one you use, or keep all three if, like me, you bounce between them.

Each cloud needs a one-time trust setup so this login is allowed at all: an IAM OIDC identity provider and role on AWS, a federated credential on an Azure app registration, and a Workload Identity Pool on Google Cloud. You do that once; after that, pulumi-idp/auth is the only thing any stack references for credentials.

A stack opts into it with one block in its stack config. The AWS stack’s Pulumi.aws.yaml:

# Pulumi.aws.yaml
environment:
 - pulumi-idp/auth

That is the entire credential story. No keys in the program, no keys in CI, nothing to rotate. The program below never mentions a secret; it only creates resources, and the ambient credentials from pulumi-idp/auth carry the request.

One cloud-init, three clouds

The part that turns a bare Ubuntu box into an inference server is identical on every cloud, so it lives in one file, cloud-init.yaml, that all three programs read. It follows the Akamai cloud-config closely, with a couple of robustness tweaks for a multi-cloud run, and it runs in a deliberate order because the GPU driver needs a reboot before Ollama can see the card:

Update packages and install the kernel headers and ubuntu-drivers.
Install the NVIDIA driver, then reboot so the kernel loads it.
On the next boot, a one-shot systemd service installs Ollama, binds it to 0.0.0.0:11434, and pulls the model.
The service disables itself, so it never runs again.

#cloud-config
# Zero-touch Ollama GPU box. Cloud-agnostic: no provider metadata, no static creds.
# The Pulumi program substitutes the model name on the `ollama pull` line below
# before this is passed as user-data (AWS/GCP) or custom-data (Azure).

write_files:
 # Make Ollama listen on every interface and never unload the model from VRAM.
 - path: /etc/systemd/system/ollama.service.d/override.conf
 content: |
 [Service]
 Environment="OLLAMA_HOST=0.0.0.0:11434"
 Environment="OLLAMA_KEEP_ALIVE=-1"

 # Runs once on the post-reboot boot, after the GPU driver is loaded.
 - path: /usr/local/bin/ollama-setup.sh
 permissions: '0755'
 content: |
 #!/usr/bin/env bash
 set -euxo pipefail

 # Install Ollama; the installer creates and starts the ollama systemd unit.
 curl -fsSL https://ollama.com/install.sh | sh

 # Pick up the OLLAMA_HOST / OLLAMA_KEEP_ALIVE override written above.
 systemctl daemon-reload
 systemctl enable ollama.service
 systemctl restart ollama.service

 # Wait for the daemon to bind its socket before pulling. This is the box
 # waiting on its OWN local daemon, not an external readiness gate.
 until curl -fsS http://127.0.0.1:11434/api/tags >/dev/null 2>&1; do
 sleep 2
 done

 # Pre-pull the model so the endpoint answers on the very first request.
 ollama pull __MODEL__

 # One-shot: never run again on future boots.
 systemctl disable ollama-setup.service

 - path: /etc/systemd/system/ollama-setup.service
 content: |
 [Unit]
 Description=One-time Ollama install and model pull
 After=network-online.target
 Wants=network-online.target

 [Service]
 Type=oneshot
 ExecStart=/usr/local/bin/ollama-setup.sh
 RemainAfterExit=true
 # The model pull runs for several minutes; without this, systemd's default
 # 90s start timeout kills the service mid-download and the model never lands.
 TimeoutStartSec=0

 [Install]
 WantedBy=multi-user.target

runcmd:
 - apt-get update
 - DEBIAN_FRONTEND=noninteractive apt-get -y upgrade
 # linux-headers-generic also covers the kernel the upgrade above may have pulled
 # in, so DKMS builds the NVIDIA module for the kernel that boots next.
 - DEBIAN_FRONTEND=noninteractive apt-get install -y "linux-headers-$(uname -r)" linux-headers-generic ubuntu-drivers-common
 # --gpgpu selects the headless server driver branch, the right one for a compute
 # GPU like the T4; bare `ubuntu-drivers install` would pull the desktop stack.
 - ubuntu-drivers install --gpgpu
 - systemctl enable ollama-setup.service

power_state:
 mode: reboot
 message: Rebooting to load the NVIDIA driver before Ollama setup
 condition: true

The model name is the literal token __MODEL__. The Pulumi program reads this file at deploy time and substitutes your model config value before handing it to the instance. Two environment settings are doing quiet but important work: OLLAMA_HOST=0.0.0.0:11434 makes Ollama listen on every interface instead of localhost alone, and OLLAMA_KEEP_ALIVE=-1 keeps the model pinned in GPU memory so only the first request pays the load cost.

The programs

Every program has the same five beats: read the config, template the cloud-init.yaml, open the inbound ports, launch the GPU instance with that cloud-init as its user data, and export the endpoint. What differs is only the dialect each cloud speaks for “GPU instance” and “firewall rule.”

The shared contract keeps the three listings legible side by side: the same model and allowSsh config keys, the same cloud-init.yaml, and the same four exports (publicIp, ollamaEndpoint, generateEndpoint, tagsEndpoint). Pick your cloud:

import * as pulumi from "@pulumi/pulumi";
import * as aws from "@pulumi/aws";
import * as fs from "fs";

const cfg = new pulumi.Config();
const model = cfg.get("model") ?? "qwen2.5:14b";
const allowSsh = cfg.getBoolean("allowSsh") ?? false;
const region = (cfg.get("region") ?? "us-east-1") as aws.Region;

// Inject the model name into cloud-init at deploy time (a file read, not a shell-out).
const userData = fs.readFileSync("cloud-init.yaml", "utf8").replace(/__MODEL__/g, model);

// Region comes from config; credentials arrive ambiently from pulumi-idp/auth.
const provider = new aws.Provider("aws", { region });

// Most recent Ubuntu 24.04 LTS (amd64, hvm, gp3) published by Canonical.
const ubuntu = aws.ec2.getAmi({
 mostRecent: true,
 owners: ["099720109477"],
 filters: [
 { name: "name", values: ["ubuntu/images/hvm-ssd-gp3/ubuntu-noble-24.04-amd64-server-*"] },
 { name: "virtualization-type", values: ["hvm"] },
 ],
}, { provider });

const ingress: aws.types.input.ec2.SecurityGroupIngress[] = [{
 description: "Ollama HTTP API. NOTE: restrict cidrBlocks to your own range in production.",
 fromPort: 11434,
 toPort: 11434,
 protocol: "tcp",
 cidrBlocks: ["0.0.0.0/0"],
}];

if (allowSsh) {
 ingress.push({
 description: "SSH",
 fromPort: 22,
 toPort: 22,
 protocol: "tcp",
 cidrBlocks: ["0.0.0.0/0"],
 });
}

const sg = new aws.ec2.SecurityGroup("ollama", {
 description: "Ollama inferencing access",
 ingress,
 egress: [{
 description: "Allow all outbound",
 fromPort: 0,
 toPort: 0,
 protocol: "-1",
 cidrBlocks: ["0.0.0.0/0"],
 }],
}, {
 provider,
});

const server = new aws.ec2.Instance("ollama", {
 ami: ubuntu.then(a => a.id),
 instanceType: "g4dn.xlarge", // 1x NVIDIA T4
 vpcSecurityGroupIds: [sg.id],
 associatePublicIpAddress: true,
 userData, // plain text; the AWS provider base64-encodes it for you
 rootBlockDevice: {
 volumeSize: 40,
 volumeType: "gp3",
 },
 tags: { Name: "ollama" },
}, {
 provider,
});

export const publicIp = server.publicIp;
export const ollamaEndpoint = pulumi.interpolate`http://${publicIp}:11434`;
export const generateEndpoint = pulumi.interpolate`http://${publicIp}:11434/api/generate`;
export const tagsEndpoint = pulumi.interpolate`http://${publicIp}:11434/api/tags`;

import * as pulumi from "@pulumi/pulumi";
import * as gcp from "@pulumi/gcp";
import * as fs from "fs";

const cfg = new pulumi.Config();
const model = cfg.get("model") ?? "qwen2.5:14b";
const allowSsh = cfg.getBoolean("allowSsh") ?? false;
const project = cfg.get("project");
const zone = cfg.get("zone") ?? "us-central1-a";

// Inject the model name into cloud-init at deploy time (a file read, not a shell-out).
const userData = fs.readFileSync("cloud-init.yaml", "utf8").replace(/__MODEL__/g, model);

// Latest Ubuntu 24.04 LTS amd64 image.
const ubuntu = gcp.compute.getImage({
 family: "ubuntu-2404-lts-amd64",
 project: "ubuntu-os-cloud",
});

// Network tag binds the firewall rule to this instance.
const networkTag = "ollama";

const firewall = new gcp.compute.Firewall("ollama-fw", {
 network: "default",
 project: project,
 // Inbound to Ollama from anywhere; restrict this CIDR for production.
 allows: allowSsh
 ? [{ protocol: "tcp", ports: ["11434"] }, { protocol: "tcp", ports: ["22"] }]
 : [{ protocol: "tcp", ports: ["11434"] }],
 sourceRanges: ["0.0.0.0/0"],
 targetTags: [networkTag],
});

const instance = new gcp.compute.Instance("ollama", {
 machineType: "n1-standard-4",
 zone: zone,
 project: project,
 tags: [networkTag],
 bootDisk: {
 initializeParams: {
 image: ubuntu.then(i => i.selfLink),
 size: 40,
 },
 },
 guestAccelerators: [{
 type: "nvidia-tesla-t4",
 count: 1,
 }],
 // GPUs cannot live-migrate, so host maintenance must terminate (and restart) the VM.
 scheduling: {
 onHostMaintenance: "TERMINATE",
 automaticRestart: true,
 },
 networkInterfaces: [{
 network: "default",
 accessConfigs: [{}], // an empty config requests an ephemeral public IP
 }],
 metadata: {
 "user-data": userData, // Ubuntu cloud-init reads this key, not startup-script
 },
}, {
 dependsOn: firewall,
});

export const publicIp = instance.networkInterfaces.apply(nics => nics[0].accessConfigs![0].natIp);
export const ollamaEndpoint = pulumi.interpolate`http://${publicIp}:11434`;
export const generateEndpoint = pulumi.interpolate`http://${publicIp}:11434/api/generate`;
export const tagsEndpoint = pulumi.interpolate`http://${publicIp}:11434/api/tags`;

import * as pulumi from "@pulumi/pulumi";
import * as resources from "@pulumi/azure-native/resources";
import * as network from "@pulumi/azure-native/network";
import * as compute from "@pulumi/azure-native/compute";
import * as random from "@pulumi/random";
import * as fs from "fs";

const cfg = new pulumi.Config();
const model = cfg.get("model") ?? "qwen2.5:14b";
const allowSsh = cfg.getBoolean("allowSsh") ?? false;
const location = cfg.get("location") ?? "eastus";

// Inject the model name into cloud-init at deploy time (a file read, not a shell-out).
const userData = fs.readFileSync("cloud-init.yaml", "utf8").replace(/__MODEL__/g, model);

const resourceGroup = new resources.ResourceGroup("ollama-rg", {
 location,
});

const vnet = new network.VirtualNetwork("ollama-vnet", {
 resourceGroupName: resourceGroup.name,
 addressSpace: { addressPrefixes: ["10.0.0.0/16"] },
});

const subnet = new network.Subnet("ollama-subnet", {
 resourceGroupName: resourceGroup.name,
 virtualNetworkName: vnet.name,
 addressPrefix: "10.0.1.0/24",
});

// Standard SKU public IPs use static allocation.
const publicIpAddress = new network.PublicIPAddress("ollama-pip", {
 resourceGroupName: resourceGroup.name,
 sku: { name: network.PublicIPAddressSkuName.Standard },
 publicIPAllocationMethod: network.IPAllocationMethod.Static,
});

// Azure permits all outbound by default, so only inbound rules are needed.
const nsg = new network.NetworkSecurityGroup("ollama-nsg", {
 resourceGroupName: resourceGroup.name,
 securityRules: [
 {
 name: "allow-ollama",
 priority: 1000,
 direction: network.SecurityRuleDirection.Inbound,
 access: network.SecurityRuleAccess.Allow,
 protocol: network.SecurityRuleProtocol.Tcp,
 sourcePortRange: "*",
 destinationPortRange: "11434",
 sourceAddressPrefix: "0.0.0.0/0", // prod: restrict to your client CIDR
 destinationAddressPrefix: "*",
 },
 ...(allowSsh ? [{
 name: "allow-ssh",
 priority: 1001,
 direction: network.SecurityRuleDirection.Inbound,
 access: network.SecurityRuleAccess.Allow,
 protocol: network.SecurityRuleProtocol.Tcp,
 sourcePortRange: "*",
 destinationPortRange: "22",
 sourceAddressPrefix: "0.0.0.0/0",
 destinationAddressPrefix: "*",
 }] : []),
 ],
});

const nic = new network.NetworkInterface("ollama-nic", {
 resourceGroupName: resourceGroup.name,
 networkSecurityGroup: { id: nsg.id },
 ipConfigurations: [{
 name: "ipconfig1",
 subnet: { id: subnet.id },
 publicIPAddress: { id: publicIpAddress.id },
 privateIPAllocationMethod: network.IPAllocationMethod.Dynamic,
 primary: true,
 }],
});

// Azure requires an admin credential even though we never log in. Generate one
// instead of hard-coding it; it stays a Pulumi secret and is never exported.
const adminPassword = new random.RandomPassword("ollama-admin-password", {
 length: 24,
 special: true,
 overrideSpecial: "!#$%*",
 minLower: 1,
 minUpper: 1,
 minNumeric: 1,
 minSpecial: 1, // Azure requires 3 of 4 character classes
});

const vm = new compute.VirtualMachine("ollama-vm", {
 resourceGroupName: resourceGroup.name,
 hardwareProfile: { vmSize: "Standard_NC4as_T4_v3" }, // 1x NVIDIA T4
 networkProfile: {
 networkInterfaces: [{ id: nic.id, primary: true }],
 },
 osProfile: {
 computerName: "ollama",
 adminUsername: "azureuser",
 adminPassword: adminPassword.result,
 customData: Buffer.from(userData).toString("base64"), // Azure wants base64
 linuxConfiguration: {
 disablePasswordAuthentication: false,
 },
 },
 storageProfile: {
 imageReference: {
 publisher: "Canonical",
 offer: "ubuntu-24_04-lts",
 sku: "server", // Gen2; NC4as_T4_v3 is a Gen2 size
 version: "latest",
 },
 osDisk: {
 name: "ollama-osdisk",
 createOption: "FromImage",
 diskSizeGB: 40,
 managedDisk: { storageAccountType: "StandardSSD_LRS" },
 },
 },
});

// A Standard/Static public IP is reserved at creation, so its address is known
// once the resource exists; no separate lookup needed.
export const publicIp = publicIpAddress.ipAddress.apply(ip => ip!);
export const ollamaEndpoint = pulumi.interpolate`http://${publicIp}:11434`;
export const generateEndpoint = pulumi.interpolate`http://${publicIp}:11434/api/generate`;
export const tagsEndpoint = pulumi.interpolate`http://${publicIp}:11434/api/tags`;

A few per-cloud details are worth calling out, since they are the places the “same program” abstraction leaks:

AWS is the shortest program, because the GPU comes with the instance shape: a g4dn.xlarge is a T4 box, so there is no separate accelerator to attach. The one thing to do before you deploy is raise the Running On-Demand G and VT instances vCPU quota in your region; a fresh account starts at zero, and pulumi up fails with VcpuLimitExceeded until you do.
Google Cloud attaches the GPU explicitly with guestAccelerators, and that brings the rule that trips people up: a GPU instance cannot live-migrate, so scheduling.onHostMaintenance must be "TERMINATE" or the apply is rejected. The cloud-init also has to ride on the user-data metadata key, not startup-script, and the empty accessConfigs: [{}] is what hands the box a public IP. T4 quota is also zero on a new project.
Azure is the longest listing, because the network is à la carte: the resource group, virtual network, subnet, public IP, security group, and NIC are each their own resource before you reach the VM. One detail surprises people: a Linux VM requires an admin credential even when you never log in, so the program generates a throwaway password with random.RandomPassword rather than committing one. The NC4as_T4_v3 is a compute GPU, so the standard server (CUDA) driver from the cloud-init is the right choice here; the GRID driver is only needed for GPU-accelerated visualization workloads.

The full programs, all three Pulumi.<cloud>.yaml files, and the shared cloud-init.yaml are in the companion repo:

github.com/dirien/fully-automated-ai-inference-pulumi

Run GPU inference across three clouds

Stand up an Ollama GPU box on AWS, Azure, or Google Cloud from one Pulumi program, with short-lived credentials brokered by Pulumi ESC and no static keys to rotate.

Get started

Deploying

Create a project, install the provider for your cloud, point the stack at pulumi-idp/auth, and deploy. For AWS:

mkdir ai-inference && cd ai-inference
pulumi new typescript
npm install @pulumi/aws

Drop the AWS listing into index.ts, copy the cloud-init.yaml shown earlier into the same directory (or grab it from the companion repo), point the stack at the auth environment, and set your region:

pulumi stack init aws
# add `environment: [pulumi-idp/auth]` to Pulumi.aws.yaml (shown above)
pulumi config set region us-east-1
pulumi config set model qwen2.5:14b
pulumi up

The other two clouds are the same flow with a different provider package and a couple of config keys: npm install @pulumi/gcp with pulumi config set project <your-project-id> and pulumi config set zone us-central1-a, or npm install @pulumi/azure-native @pulumi/random with pulumi config set location eastus.

The slow part is the GPU box: it boots, installs the driver, reboots, installs Ollama, and pulls the model, all without you. Depending on the instance, the region, and the model size, plan on roughly 10 to 15 minutes before the endpoint answers. When pulumi up finishes you get the endpoints straight back as stack outputs:

Outputs:
 generateEndpoint: "http://<public-ip>:11434/api/generate"
 ollamaEndpoint : "http://<public-ip>:11434"
 publicIp : "<public-ip>"
 tagsEndpoint : "http://<public-ip>:11434/api/tags"

Testing the inference server

pulumi up returns as soon as the infrastructure exists, which is before the model has finished downloading. This is exactly where Terraform reached for that null_resource loop, and where Pulumi hands you a URL instead. To check whether the model is ready, ask Ollama what it has loaded:

curl -s $(pulumi stack output tagsEndpoint) | grep -q "qwen2.5:14b" && echo "ready" || echo "still pulling"

Once it reports ready, send it a prompt. This is the same request the Akamai post makes, pointed at your generateEndpoint output:

curl -s $(pulumi stack output generateEndpoint) -d '{
 "model": "qwen2.5:14b",
 "system": "Answer every question with a three-line poem.",
 "prompt": "Why is the sky blue?",
 "stream": false
}'

The first call is slower, because Ollama loads the model into GPU memory before it answers. Every call after that is fast, since OLLAMA_KEEP_ALIVE=-1 keeps it resident:

{
 "model": "qwen2.5:14b",
 "response": "Sunlight scatters, short waves fly,\nBlue light paints the open sky,\nViolet fades as day drifts by.",
 "done": true
}

When you are done, take it all down with one command, so the GPU stops billing:

pulumi destroy

Cost

A GPU instance is the whole bill, and you pay by the hour whether the model is busy or idle. The numbers below are rough on-demand rates for a single T4-class box left running 24/7; in practice you spin it up, use it, and pulumi destroy it, so what you actually pay tracks the hours it stays up.

Cloud	Instance	GPU	~Hourly	~Monthly (24/7)
AWS	`g4dn.xlarge`	1× T4 (16 GB)	~$0.53	~$384
Azure	`Standard_NC4as_T4_v3`	1× T4 (16 GB)	~$0.53	~$384
Google Cloud	`n1-standard-4` + 1× T4	1× T4 (16 GB)	~$0.54	~$394

These are on-demand Linux rates in a cheap US region (us-east-1, eastus, us-central1) as of mid-2026, and the GPU dominates the bill. Spot or preemptible capacity cuts it sharply if your workload tolerates interruption (often 60 to 70 percent off), Google Cloud’s sustained-use discount trims an always-on instance on its own, and the 40 GB disk adds only a few dollars a month.

A GPU box left running overnight is the expensive mistake here. Because the whole stack is declarative, the safe habit is cheap: pulumi destroy when you stop using it, pulumi up when you need it back. The state and config are version-controlled, so standing it up again is one command, not a rebuild.

Security considerations

The architecture in this post matches the Akamai original on purpose, which means it carries the same caveat: Ollama is exposed over plain HTTP on a port open to the entire internet. That is fine for a demo on a box you tear down the same day. It is not fine for anything that outlives the afternoon, and self-hosted AI endpoints get found fast: scanners like Shodan enumerate a freshly exposed one within hours, and an open Ollama instance is free compute for whoever finds it first.

The credential side, though, is genuinely better than a static-token setup, and that is the part worth keeping. There is no long-lived cloud key in the program, in your shell, or in CI; every deploy gets short-lived credentials minted through pulumi-idp/auth and thrown away when it finishes.

Concern	Akamai/Terraform demo	This deployment
Cloud credentials	Static PAT in an env var	OIDC login, short-lived, via `pulumi-idp/auth`
Inbound exposure	Port 11434 open to `0.0.0.0/0`	Same by default; one config flag scopes the CIDR
Transport	Plain HTTP	Plain HTTP (front with TLS for real use)
Readiness wait	`null_resource` + `local-exec` shell loop	A runtime `curl`, no resource

My recommendations for taking this past a demo:

Scope the inbound rule to your own CIDR instead of 0.0.0.0/0; the allowSsh flag in the program is the pattern to copy for the Ollama port
Put a reverse proxy with TLS in front of Ollama, or keep the box off the public internet entirely and reach it over a private network or a Tailscale tailnet, the way I did for the Hermes agent
Keep credentials on OIDC through ESC; never fall back to a static key to “just get it working”
Treat the box as disposable, and pulumi destroy it when you are not using it, which shrinks both the bill and the attack window

What’s next?

The program is a foundation, and the interesting work is what you layer on once a model is a pulumi up away:

Lock it down. Scope the firewall, add TLS, or move the box behind a private network so the endpoint is not on the public internet at all.
Scale the model to the GPU. qwen2.5:14b on a T4 is the entry point. Bigger models want more memory, which means an A10G, an L4, or an A100, and the only change in the program is the instance type and the model value.
Swap the serving engine. Ollama is the simplest thing that works. For higher throughput, the same shape holds with vLLM in place of Ollama in the cloud-init.
Generate the next one. Pulumi Neo can take a target like “a GPU inference box on Azure behind a private endpoint” and produce a first draft of the program, which you then review in a PR like any other change.

Conclusion

Declaring an AI inference server as code buys you the same thing it buys for any other infrastructure: you can reproduce it on any of the three clouds, version-control it, and tear it down with a single pulumi destroy. Porting the Akamai project made the contrast sharp. The credentials moved from a static token to an OIDC login that leaves nothing behind, and the readiness wait disappeared entirely once you stop treating a runtime check as a resource.

That is the line worth carrying to the next thing you deploy. Not every step in a runbook is a resource. A GPU box and a firewall rule are; waiting for a download to finish is not. Draw that line first, and the program gets shorter on its own.

If you run into issues or have questions, drop by the Pulumi Community Slack or GitHub Discussions. New to Pulumi? Get started here.