AI Agent Platform Deployment: Cloud vs Self-Hosted vs Hybrid (2026)

The security review meeting was supposed to take 30 minutes.

Four hours later, the engineering lead at a fintech startup was still explaining to their compliance officer why an AI agent platform needed access to production code. The vendor's answer to "where does our code go?" was a hand-wavy diagram with arrows pointing to "secure cloud infrastructure" and a promise that "enterprise customers trust us."

That's not an answer. And that team wasted three months evaluating a platform that was never going to pass their security review.

Deployment model isn't a checkbox at the end of your evaluation. It's the first filter. If you're in a regulated industry — fintech, healthcare, defense, anything touching PII — the wrong deployment model disqualifies a platform before you ever see a demo. (For a deeper look at how DevOS approaches AI agent platform deployment, keep reading.)

Here's the matrix we wish that team had used.

Quick Verdict

TL;DR: Cloud-hosted is fastest to adopt but least control. Self-hosted gives you full control at the cost of ops burden. Hybrid splits the difference — vendor-managed control plane, your infrastructure for the data plane. Air-gapped is for government, defense, and regulated industries where zero external connectivity is non-negotiable.

If you're a startup without compliance requirements, go cloud. If your auditor says source code can't leave your network, go self-hosted or air-gapped. If you want control without the full ops burden, evaluate hybrid options.

The Deployment Matrix

Factor	Cloud-Hosted	Hybrid	Self-Hosted	Air-Gapped
Control over code/data	Low — vendor infra	Medium — data plane yours	High — everything yours	Maximum — no external calls
Compliance flexibility	Vendor-dependent	Good for most regs	Full control	Required for some industries
Ops burden	Zero	Low-Medium	High	Very High
Time to production	Days	Weeks	Months	Months
Cost model	SaaS pricing	SaaS + infra	Infra only	Infra + headcount
Scaling responsibility	Vendor	Shared	You	You
Model access	Vendor-selected	Configurable	Any	Self-served
Network requirements	Internet	Hybrid	Your cloud	None
Typical adopter	Startups, SMBs	Mid-market, regulated	Enterprises, fintech	Government, defense

Deep Dive: Cloud-Hosted Deployment

Cloud-hosted is what most people try first. You sign up, connect your GitHub repo, and agents start working. The vendor handles everything — infrastructure, scaling, model orchestration, security patches, monitoring.

Where it shines:

Speed. Nothing beats "create account → connect repo → agents running" in under an hour. For teams validating whether AI agents fit their workflow at all, cloud is the right starting point. You can always migrate later. Maybe. (Famous last words.)

Zero ops. No Kubernetes clusters. No database backups at 2 AM. No certificate rotations that ruin your weekend. Your team stays on product work instead of platform work. I've seen startups burn entire quarters on self-hosted infrastructure they didn't need — don't be that team. (For teams handling telephony compliance, VeloCalls has written about TCPA-compliant call tracking — similar ops-vs-compliance tradeoffs.)

Predictable pricing. Most cloud platforms charge per-seat or per-usage with clear tiers. DevOS's planned tiers (all pre-launch waitlist) are $0 for Free, $25/user/month for Pro, $49/user/month for Team — unlimited agents and tasks on paid plans. You know your bill before the month ends.

Where it breaks:

Data residency. Your code, prompts, and agent outputs live on the vendor's infrastructure. For some teams, that's fine. For fintech processing payment data, healthcare handling PHI, or any company whose customers require specific data geography — it's a non-starter.

Vendor trust. "SOC 2 certified" means they've passed an audit. It doesn't mean your auditor will accept it. If your compliance framework requires that source code never touch third-party infrastructure, certification doesn't matter.

Lock-in risk. Workflows built on a specific platform's agent orchestration may not port cleanly. If the vendor raises prices, pivots, or gets acquired, your options narrow. (For context on tracking agent ROI as platform costs shift, the team at ClickzProtect wrote about measuring real spend versus fake clicks — different domain, same "where is my money actually going?" problem.)

Who should pick this:

Startups without compliance requirements. Teams evaluating AI agents for the first time. Anyone who'd rather ship features than operate infrastructure.

Deep Dive: Hybrid Deployment

Hybrid splits responsibilities: the vendor operates the control plane (auth, UI, orchestration, billing), you operate the data plane (where agent code actually runs, where your source code lives).

Where it shines:

Compliance without the full burden. Your code and agent execution happen on infrastructure you control — your AWS account, your GCP project, your Azure subscription. The vendor never sees your source code. But you don't have to build the orchestration layer, the agent marketplace, or the monitoring stack from scratch. That stuff takes months. Ask me how I know.

Data sovereignty. For companies with European customers, GDPR data residency isn't optional. Hybrid lets you run the data plane in EU regions on your infrastructure while still using the vendor's control plane (assuming it doesn't process PII, which it shouldn't in a well-architected split).

Model flexibility. Some hybrid deployments let you bring your own model endpoints — your Anthropic keys, your OpenAI organization, your self-hosted open-source models. The control plane routes tasks; your infrastructure runs inference.

Where it breaks:

Complexity. You're now operating two things: the vendor relationship and your own infrastructure. Debugging issues requires understanding both. "Is this the control plane's fault or ours?" becomes a common question.

Partial trust. You're still trusting the vendor's control plane with metadata — which tickets exist, which agents ran, when, for how long. For some threat models, that's acceptable. For others, it's not.

Cost unpredictability. You pay the SaaS fee plus your own infrastructure costs. Agent workloads can be bursty; scaling your data plane to handle spikes means either over-provisioning or accepting latency during peaks. Budgeting becomes guesswork. Not fun.

Who should pick this:

Mid-market companies with compliance requirements but without dedicated platform teams. Enterprises who want control over code residency without building everything themselves. Teams transitioning from cloud to self-hosted who want a stepping stone.

Deep Dive: Self-Hosted Deployment

Self-hosted means you run the entire platform — control plane, data plane, model inference (or API calls from your network), database, observability, auth. The vendor delivers software; you operate it.

Where it shines:

Maximum control. Nothing leaves your network that you don't explicitly allow. Your code, your prompts, your agent outputs, your logs — all on infrastructure you own and audit. For threat models where "third-party vendor access to source code" is unacceptable, this is the only option. Period.

Custom everything. You can modify the agent runtime, add custom observability, integrate with internal systems that don't have public APIs, run any model you want (including fine-tuned models on internal GPUs). No feature request queue — you just build what you need.

Audit clarity. Your auditor can inspect the actual infrastructure, not a vendor's certification letter. "Where does code go?" has a concrete answer: "This Kubernetes cluster, these subnets, these S3 buckets — here's the Terraform."

Where it breaks:

Ops burden. Expect to dedicate at least a part-time engineer (more likely a full platform team for larger deployments) to running the platform. Kubernetes upgrades, database migrations, model serving infrastructure, certificate rotation, security patches, incident response — all yours. If you've never run a self-hosted platform before, budget for a painful first three months. Possibly six. I'm not exaggerating.

Slower upgrades. Cloud platforms ship features continuously. Self-hosted means planning upgrade windows, testing in staging, handling migration scripts. You'll be behind on features compared to cloud users.

Cost model changes. No more predictable SaaS bill. Now you're paying for EC2 instances, managed databases (or worse, operating them yourself), model API calls at your own rate limits, and engineering time. For smaller teams, self-hosted is often more expensive than cloud despite "owning the infrastructure."

DevOS's planned Enterprise tier (custom pricing, contact sales) includes self-hosted deployment with BYOK encryption and SOC 2 / HIPAA compliance posture. But we're clear-eyed: self-hosted is a serious commitment. If you're not already running Kubernetes in production, don't start just for an AI agent platform. Honestly? That's terrible advice we've seen too many teams take.

Who should pick this:

Large enterprises with existing platform teams. Companies where compliance explicitly requires no third-party infrastructure access. Teams with specialized model requirements (fine-tuned models, on-premise GPUs, air-gapped inference). If you're evaluating browser automation agents with similar self-hosted requirements, JustBrowser covers headless browser deployment options in their technical guides.

Deep Dive: Air-Gapped Deployment

Air-gapped is self-hosted with an additional constraint: zero external network connectivity. The platform runs entirely within a secured network boundary that cannot reach the internet. This matters for government contractors, defense, classified environments, and some healthcare / critical infrastructure settings.

Where it shines:

Compliance requirements that mandate no external connectivity. If your environment literally cannot make outbound API calls — not "we'd prefer not to," but "the network physically cannot" — air-gapped is your only option. ITAR, FedRAMP High, some CMMC levels, certain healthcare facilities with isolated networks.

Maximum threat model. Supply chain attacks, DNS hijacking, API credential theft — none of these attack surfaces exist if the platform never talks to the outside world. It's paranoid, sure. But sometimes paranoia is warranted.

Where it breaks:

Model access. No API calls to Anthropic, OpenAI, or Google. You're running open-source models on-premise or operating self-hosted inference infrastructure for proprietary models (which requires licensing negotiations we won't pretend to understand). Model capability in air-gapped environments is often 6-12 months behind frontier.

Updates. No pulling container images from Docker Hub. No automated security patches. Every update is a manual artifact transfer through your security boundary, tested in a staging environment, then deployed. It's slow.

Cost. Air-gapped typically means running your own GPU clusters for model inference. A100s or H100s aren't cheap. Neither are the engineers who know how to run them. (And good luck hiring them right now.)

Who should pick this:

Government contractors. Defense. Critical infrastructure. Classified environments. If you're reading this section and nodding, you probably already know whether air-gapped applies to you.

Pricing Across Deployment Models

This gets messy because the cost structures differ:

Model	Direct Platform Cost	Infra Cost	Hidden Costs
Cloud	SaaS tiers	$0	Model tokens (if own keys)
Hybrid	SaaS tiers (often higher)	Data plane compute + storage	Engineering time for integration
Self-hosted	License (often per-seat)	Full infra stack	Platform team salaries
Air-gapped	License + support contract	Full infra + model inference	Specialized ops + clearances

For DevOS specifically: Free tier ($0) and Pro ($25/user/month) are cloud-only. Team ($49/user/month) adds SSO/SAML, audit logs, and Linear/Jira sync — still cloud. Enterprise (custom) unlocks self-hosted deployment with white-label, BYOK, and compliance features. All tiers are pre-launch and waitlist-only.

The honest math: for teams under 20 engineers without compliance requirements, cloud is almost always cheaper. Self-hosted only makes economic sense when compliance requirements mandate it OR when your scale is large enough that per-seat pricing exceeds infrastructure costs. Run the numbers before committing. Most teams don't — and regret it.

Compliance Considerations by Deployment Model

Compliance Framework	Cloud	Hybrid	Self-hosted	Air-gapped
SOC 2 Type II	Vendor-dependent	Usually achievable	Full control	Full control
HIPAA	Vendor must sign BAA	Achievable with proper split	Full control	Full control
GDPR data residency	Some vendors, some regions	Achievable	Full control	Full control
FedRAMP (Moderate)	Few platforms qualified	Some offer this	Common path	Required for High
CMMC Level 2+	Unlikely	Maybe with right vendor	Typical	Often required
PCI-DSS	Vendor-dependent	Possible	Common	Possible

The pattern: as compliance requirements tighten, deployment model shifts from cloud → hybrid → self-hosted → air-gapped. Talk to your compliance officer before your vendor selection process, not after. For teams running payment processing alongside AI agents, VeloCards has documented PCI-compliant card handling workflows worth reviewing.

(If you're also evaluating privacy-focused analytics for tracking agent activity without third-party data leakage, JustAnalytics covers cookie-free event logging — same "keep data on your infrastructure" mindset.)

The Decision Tree

Start here:

Does your compliance framework require that source code never touch third-party infrastructure?

• Yes → Self-hosted or air-gapped
• No → Continue

Does your environment have outbound internet access?

• No → Air-gapped
• Yes → Continue

Do you have a platform team with Kubernetes experience?

• Yes → Self-hosted is viable, evaluate hybrid as simpler alternative
• No → Hybrid or cloud

Is data residency (geographic location of data) a hard requirement?

• Yes with specific regions → Hybrid or self-hosted
• Yes but vendor regions work → Cloud with data residency options
• No → Cloud

Is time-to-production critical (under 1 month)?

• Yes → Cloud or hybrid
• No → Any model is viable

Who Should Pick What

Cloud-hosted: Startups, SMBs, and teams without compliance requirements. Speed matters more than control. You'd rather ship features than operate infrastructure. Budget is predictable and you want it to stay that way. This is most teams.

Hybrid: Mid-market companies with compliance requirements who want data sovereignty without full ops burden. You have some infrastructure experience but not a dedicated platform team. Your auditor accepts vendor-managed control planes as long as your code stays on your infra.

Self-hosted: Large enterprises with platform teams. Compliance frameworks that require no third-party access to source code. Organizations already running Kubernetes who can absorb another workload. Teams with model requirements that cloud platforms don't support.

Air-gapped: Government contractors. Defense. Classified environments. Critical infrastructure with mandated network isolation. If you need air-gapped, you already know — and you probably have a procurement process that takes longer than this evaluation anyway.

Our Honest Take

Most teams evaluating AI agent platforms should start with cloud. Seriously. The ops burden of self-hosted is real, and most companies underestimate it until they're three months in and their platform engineer is burned out. I've watched it happen. It's not pretty.

But "start with cloud" doesn't mean "ignore deployment model." If you're in a regulated industry, confirm the platform supports your eventual deployment model before you invest in it. Migrating from cloud to self-hosted is painful. Migrating from a platform with no self-hosted option to a completely different platform is worse.

DevOS is pre-launch — every plan CTA is "Join Waitlist." We're building Enterprise tier with self-hosted, BYOK, and SOC 2 / HIPAA compliance because we know certain teams can't adopt cloud-only platforms. But we're not going to pretend self-hosted is right for everyone. If you don't have compliance requirements driving you there, cloud is faster, cheaper, and less painful.

The deployment model conversation should happen in week one of your evaluation, not week twelve. Save yourself a four-hour security review meeting. Or don't. Your calendar.

Frequently Asked Questions

Which deployment model offers the most control over AI agent data?

Air-gapped (on-premise) deployment offers the most control — your code and prompts never leave your network. Self-hosted is next, running on your cloud account with full infrastructure ownership. Cloud-hosted platforms retain some data (prompts, outputs, logs) on vendor infrastructure, though most offer data residency options. For teams with strict data sovereignty requirements, self-hosted or air-gapped is the only real option.

What's the ops burden difference between cloud and self-hosted AI agent platforms?

Cloud-hosted platforms handle upgrades, scaling, monitoring, and security patches — you manage zero infrastructure. Self-hosted shifts all of that to your team: Kubernetes clusters, database maintenance, model serving, certificate rotation, and incident response. Expect to dedicate at least a part-time engineer to self-hosted infrastructure. Hybrid sits in the middle — control plane managed by the vendor, data plane on your infra.

Can cloud AI agent platforms meet SOC 2 or HIPAA requirements?

Some can. Look for platforms with published SOC 2 Type II reports, BAA signing for HIPAA, and data residency options (EU, US, specific regions). But cloud compliance means trusting the vendor's controls — if your auditor requires that PHI or source code never touch third-party infrastructure, you'll need self-hosted or air-gapped regardless of vendor certifications.

When does hybrid deployment make sense for AI agents?

Hybrid works well when you need data sovereignty (agent execution and code stay on your infrastructure) but don't want to operate the full platform stack (auth, UI, orchestration, billing). The vendor manages the control plane; you manage the data plane. This is common for mid-market companies with compliance requirements but without dedicated platform teams.

---

Join the DevOS Waitlist

AI agents that work as employees inside your sprints, standups, and tickets — not single-task copilots. Planner / Developer / QA / DevOps agents pick up work from the backlog, ship in branches, request review. Linear-shaped backlog UI with AI underneath. Pre-launch.

Join the waitlist → · How agents-as-employees works