Hosted-Model Tier

What It Is

A hosted-model tier is the cloud arrangement where a third-party AI model, built by a frontier lab like Anthropic or OpenAI, runs entirely inside a hyperscaler’s infrastructure and behaves, to the customer, like a native cloud service. The customer hits the model through the cloud’s API surface, pays through the cloud’s billing fabric, authenticates through the cloud’s identity service, and inherits the cloud’s data-residency guarantees. The model vendor still wrote the weights and ships updates, but the cloud is the one running the inference, holding the data, and signing the contract.

A useful contrast: the older arrangement, which is still common, is a hosted-on-vendor tier. There the customer’s request leaves the cloud, hits the model vendor’s API at a different domain under a different contract, and returns. The traffic still moves, but the auth, the billing, the governance, and the audit logs sit in two different places. A hosted-model tier collapses that into one place. It is the difference between “the model is available on this cloud” and “the model is native to this cloud.”

How It Actually Works

The cloud provider stands up the model in its own data centers under a license from the model vendor. Customers route requests through the cloud’s standard model-serving API, often the same API the cloud already uses for its own first-party models, and the request never leaves the cloud’s network. Inference is processed in cloud regions the customer selects, billing rolls into the customer’s existing cloud invoice, and identity is handled by the cloud’s IAM. Logs land in the cloud’s logging service. Microsoft’s “Hosted on Azure” deployment for Claude inside Microsoft Foundry is a current example; Amazon Bedrock’s hosted Claude is another. The hard work happens at the contract layer: how the model is shipped to the cloud, how updates land, who owns the customer relationship, and how the vendor and the cloud split responsibility when something breaks.

Why It Matters Right Now

Procurement teams have one foot in the cloud’s enterprise agreement and the other in a tangle of point contracts with model vendors. A hosted-model tier collapses that tangle. A Microsoft customer can now call Claude under the same Microsoft Customer Agreement that already governs their Azure billing, their Defender alerts, and their data-residency policy. That changes who can buy and how fast they can sign. It also changes the model vendor’s leverage. The lab is no longer a vendor reaching for the buyer over the cloud’s shoulder; it is a tenant that depends on the cloud’s pipes, the cloud’s invoice, and the cloud’s account team to reach the customer.

The Cost and Tradeoff

Pricing inside a hosted-model tier sometimes runs at parity with the model vendor’s own API, sometimes at a small premium for the cloud’s governance overlay. The bigger cost is feature lag. The hosted tier may be a release or two behind whatever the model vendor ships on its own API, because the cloud has to certify, document, and run the new endpoint through its own SLA process before promoting it. Customers who need the absolute latest model the morning it ships sometimes keep a fallback to the vendor-hosted path; customers who need data-residency or single-pane procurement usually do not.

How TWO Uses It

For the operator readers TWO writes for, the practical question is which buying motion you can actually run on a given Tuesday. If your team already pays for Azure or AWS and your CFO has signed off on data-residency policy, the hosted-model tier is the path of least friction: one contract, one auth header to rotate, one log to audit. If you are running on a sovereign-ai posture, or you need the freshest model the moment a vendor ships it, the vendor-hosted path may still win. The decision is rarely about the model; it is about whose hand is on the bill and whose pipes the prompt sits in. TWO’s rule of thumb: pick the tier whose audit log your security team will actually open, not the one whose marketing page reads best.

What to Watch Next

Watch two signals. First, when a frontier-model launches and the hosted-model tier ships within the same week, the lab and the cloud are coordinating release calendars; that means the lab is closer to native than just available. Second, watch the pricing column. If hosted-tier pricing routinely runs below the vendor’s own API, the cloud is subsidizing the model to keep tenants inside; if it runs above, the cloud is taxing the convenience. Either tells you who is holding power in that particular partnership, and how much room the model vendor has to negotiate the next contract.