What a cloud AI agent is, and why hosting region matters

Direct answer: A cloud AI agent is a real-time Human AI Agent, face, voice and presence, hosted on infrastructure you do not own, so it scales to thousands of live conversations without you provisioning GPUs. Where it runs is a compliance decision: an agent processing EU customers' faces and voices has to meet GDPR residency requirements.

A cloud AI agent runs on infrastructure you do not own, and the part most teams underestimate is where that infrastructure sits.

What a cloud AI agent is

A cloud AI agent is a real-time Human AI Agent hosted on cloud infrastructure rather than on hardware you manage. You call it over the network, it runs on GPUs someone else keeps warm, and it scales up and down with demand. You will also see this called a cloud AI avatar, which is the same idea in a different word.

At Ojin the agent itself is the flagship Human Agents product: an orchestration layer that bundles speech-to-text, an LLM, a face model, and a voice layer into one pipeline. You pick Oris Portrait or Oris Presence as the developer-tier face model, at two fidelity tiers, each working to a sub-200ms face-model latency budget. The voice layer produces the multilingual speech, and the orchestration keeps the pipeline reading as live rather than played. Hosting it in the cloud is what lets one agent greet ten people or ten thousand without you forecasting GPU purchases a quarter ahead. You trade control of the box for elasticity, which most teams should be glad to do. Teams typically wire the agent into their product through a real-time agent API, which is part of why hosting region and network one-way transmission time matter as much as raw GPU capacity does. Region choice can be measured: Google Cloud's Performance Dashboard publishes median latency and average packet loss between region and zone pairs across its estate.

Voice-only agent platforms and contact-centre-as-a-service (CCaaS) incumbents run on cloud infrastructure too, and none of them add a visual layer to the hosted agent. Where you host a voice agent and where you host a human AI agent raise the same infrastructure questions, latency, region, GDPR, but only one of them can put a face on the call. The conversational side of that comparison is set out in our guide to the conversational AI agent.

Why the GPU capacity sits on someone else's balance sheet

Cloud providers are the ones absorbing that capital cost. Gartner forecasts that worldwide AI infrastructure spending, the GPU capacity, network fabric, and hosted compute that live agents run on, will rise from USD 975.5 billion in 2025 to USD 1.43 trillion in 2026, as part of a total AI spend of USD 2.59 trillion. That buildout sits on cloud providers' balance sheets rather than on yours. The physical side lands there too: the International Energy Agency's Energy and AI report of April 2025 sets out how much electricity data centres will draw to meet AI demand, and where that power comes from.

Cloud versus on-prem, and the assumption worth checking

The usual case for on-prem is control: your hardware, your building, your data never leaves. It sounds safer. For a real-time agent it often is not, because of GPUs.

A live agent needs warm, high-end GPUs standing by to answer in real time. Google's Cloud Run documentation describes the same trick for ordinary services, letting you specify a minimum number of container instances to be "kept warm and ready to serve requests". Running your own means buying for peak and paying for idle, plus the staff to keep it healthy, a pattern measured in a two-month trace from a multi-tenant GPU cluster at Microsoft, published at USENIX ATC 2019.

An EU-region cloud gives you the thing on-prem was supposed to give you, data that stays in your jurisdiction, without the capital and the ops burden. The large providers commit to that contractually. AWS's data privacy FAQ states: "You choose the AWS Region(s) in which your content is stored," and "We will not move or replicate your content outside of your chosen AWS Region(s) without your agreement." The useful question is narrower than cloud-or-on-prem on principle. Do you need the box itself for a specific reason, or do you need your data to stay in a specific place? Those are different problems, and the second one has a cleaner answer.

Where Ojin sits, and what it is not

Ojin is not a hyperscaler, one of the global cloud operators running data centres at planetary scale, and we do not compete with AWS, Azure or GCP. Ojin is the real-time agent layer that runs on top of cloud infrastructure, the part that turns raw GPUs into a face that holds a conversation. When you compare clouds for an agent workload, you are choosing where Ojin runs, not choosing between Ojin and a cloud. Ojin, the Human AI Company, builds the agent layer to sit inside whichever cloud and region a customer has already chosen. Some buyers call this shape a BYOC AI agent, bring your own cloud, since the provider and region stay the customer's choice and Ojin supplies the face, voice and orchestration on top of it.

None of the hyperscalers hand you the agent itself. Ojin adds the layer on top: a cloud AI solution that holds a live conversation on your own site. Recorded avatar tools and voice-only platforms stop short of that.

AI agent data residency decides the architecture

A live agent processes faces, voices, and whatever a customer says out loud. Under EU law that is sensitive data, and where it is processed is not optional once you have EU users. How closely European regulators look at facial data is a matter of record: the European Data Protection Board's Guidelines 05/2022, finalised in May 2023, work through facial recognition in law enforcement in detail, and that is the level of scrutiny any face-processing pipeline should expect. Data residency is therefore an architectural constraint, decided before the first line of integration code.

Ojin is a Berlin company built around the EU AI Act, GDPR and Saudi Arabia's PDPL from the start. Most of the voice-AI startup field this category gets compared against is US-based, and its compliance posture reflects that starting point: EU data residency is either an enterprise add-on or absent from the conversation entirely. The EU AI Act adds transparency duties for AI that interacts with people, and the GDPR governs how that biometric-adjacent data is handled and where it can travel. For a regulated buyer, a vendor built EU-first from incorporation can commit to more in a contract, so ask directly where processing happens for your region and what residency options exist before you sign.

Security and compliance are their own decision, separate from where the agent runs. Ask a vendor about SOC 2, ISO 27001, GDPR, and encryption in transit and at rest, and about how a live face-and-voice pipeline handles biometric-adjacent data.

Scaling a cloud AI agent without breaking the budget or the latency

Scaling a live agent is not like scaling a web request. A request is stateless: it arrives, it returns, and the box can sleep afterwards. A conversation is a held-open session with several models running the whole time, so the usual scale-to-zero playbook does not apply.

What works is warm capacity plus region-aware concurrency-based autoscaling: enough GPUs kept hot to answer the first word on time, scaling against live-session count rather than request count, with region placement treated as a compliance decision from the start. Knative's Serving documentation scales on simultaneous requests per replica, with a soft concurrency limit of 100 and a target utilisation of 70% by default, which is the same shape of control a session-based workload needs. Done right, a cloud real-time AI platform built around a sub-200ms face model holds up across thousands of simultaneous conversations.

How to decide where to host your agent

The decision comes down to four questions. Do your users sit in the EU, and does that fix your region for you? Do you need the agent embedded in your own product surface? Are you rolling out across several markets at once? And can the provider commit to a processing region in writing? Answer those four before you compare vendors. The adjacent decisions in this cluster are covered across the Ojin insights library.

Cloud AI solutions and cloud AI avatars are the same thing

Procurement teams often evaluate this category as cloud AI solutions or a cloud real-time AI platform. The architecture under those labels is the same one described above. If your shortlist also includes a face-and-voice deployment, the same hosted architecture is sometimes marketed as a cloud AI avatar, though the underlying stack, and the human AI agent running on top of it, is identical to what this page describes. Larger organisations run this evaluation across several departments at once, and some buyers weigh a specialist GPU host against a hyperscaler for the workload.

Frequently asked questions

What is a cloud AI agent?

A real-time Human AI Agent hosted on cloud infrastructure instead of your own hardware, so it can scale to many live conversations on demand without you running GPUs.

Cloud or on-prem for an AI agent?

Cloud usually wins for a real-time agent because of GPU elasticity. An EU-region cloud also gives you data residency without the capital and staffing of running your own hardware.

Is a cloud AI agent the same as a cloud AI avatar?

The terms get used interchangeably. "Avatar" points at the face; the agent is the live system, face, voice, and presence, that holds the conversation. Ojin builds the live kind.

Where is the data processed?

Ask directly about the processing region and the residency options for your specific deployment before signing, since a live agent handles biometric-adjacent data and the answer varies by vendor and by plan.

Is Ojin a cloud provider like AWS?

No. Ojin runs on cloud infrastructure as the real-time agent layer. It is not a hyperscaler, and you choose where it runs.

Can a cloud AI agent handle thousands of concurrent users?

Yes, with warm capacity and region-aware autoscaling. The real constraint is keeping the face model's latency budget intact under load, not raw request volume.

Is EU-aligned hosting available from voice-only agent platforms?

Sometimes as a custom enterprise arrangement, not as the default. Most are US-headquartered companies. Ojin is a Berlin company built around GDPR and the EU AI Act from incorporation, which matters specifically for regulated buyers in healthcare, finance, and the public sector who cannot treat data residency as an afterthought.

How to build a cloud AI agent with Ojin

Spin up a real-time Human AI Agent, Oris Portrait for a lighter embed or Oris Presence for the higher-fidelity face, without touching a GPU. Build your first agent at ojin.ai/signin and use docs.ojin.ai to guide you, on Ojin's self-serve AI platform. Current plans and what you can try for free are listed on ojin.ai/pricing, and residency options are discussed as part of onboarding.

Read next