Denvr

What is Denvr?

Denvr is an AI cloud and private AI infrastructure platform operated by Denvr Dataworks Corp., with its headquarters located in Calgary, Canada. It addresses the needs related to GPU computing power, networking, storage, and operational management required for model training, fine-tuning, inference, and private enterprise deployments.

The platform offers both on-demand cloud resources and dedicated infrastructure, which can be used in the Canadian and U.S. regions; it also enables the creation of local environments for organizations that require data sovereignty, low latency, or isolation. It is not a general-purpose chatbot – users typically create computing resources through the console, APIs, or infrastructure tools, and then run their own models and workloads.

Product composition

Main functions

GPU and CPU computing

Users can create nodes with either a single GPU or multiple GPUs, and they can choose between bare metal, virtual machines, or containers depending on their workload. The available hardware includes NVIDIA H200, GH200, H100, A100, A40, and Intel Gaudi2; the models available vary depending on the region, inventory levels, and purchasing method.

• Training and fine-tuning: High-memory GPUs, multi-GPU NVLink connections, and high-speed interconnects are used to handle large models, vision models, and scientific computing tasks.
• Elasticity testing: Run Notebook tasks, evaluations, data processing operations, or model conversion tasks on-demand using instance instances with short duration.
• Reserved clusters: Capacity is locked in through multi-month or annual commitments, making them suitable for continuous training, production inference, and deterministic scheduling.
• CPU workload: Use Intel or AMD instances for data preprocessing, service orchestration, general computing, and components that do not require a GPU.
• Machine learning images: Images with pre-installed drivers, Docker, and common frameworks are used to reduce the time required to set up the underlying environment.

High-performance networking and storage

Multi-GPU nodes can be interconnected using technologies such as InfiniBand, RoCE, and NVLink; the specific bandwidth depends on the model of the instance. Users can also configure local NVMe drives, persistent storage, or network-shared file systems, allowing them to balance speed, accessibility, and persistence.

• Local NVMe: Suitable for caching, temporary training data, and high-throughput intermediate files; however, it is necessary to consider the lifecycle of the data once the instance is released.
• Persistent disk: Stores data independently of compute instances, and is suitable for model weights, datasets, and files related to long-term projects.
• Shared file system: It allows cluster nodes to access the same data space, facilitating distributed training and teamwork.
• Network isolation: Access boundaries are established through VPCs, VPNs, as well as public or private IPs, so it is not necessary to expose all services to the public internet.

Dedicated AI inference endpoint

AI Inference deploys models on dedicated, single-tenant GPUs; it is not charged based on the number of input or output tokens, but rather on the time the GPU is utilized. Endpoints can be configured with public or private IPs, and the compatibility with the OpenAI API reduces the costs associated with application migration.

• Model Directory: Select the models and hardware listed on the platform; once launched, inference endpoints that can be used by applications will be available.
• Custom models: Deploy your own, open-source, or fine-tuned models using the vLLM Server or Ollama Server.
• Stable capacity: Dedicated GPUs do not share inference computing power with other tenants, making them suitable for services that have specific requirements regarding latency and throughput.
• Cost characteristics: The absence of a token limit does not mean there are no costs; GPU usage fees continue to be incurred during the operation of the endpoint.
• Compatibility boundaries: Just because the interface format is compatible does not mean that every parameter, model capability, or response field is identical to those of other services; testing should be carried out before migration.

Notebook and interactive applications

The platform offers packaged applications such as Jupyter Notebooks and OpenWebUI, allowing developers to access an interactive environment via a browser. Notebooks are suitable for working with code, data, and conducting model experiments, while OpenWebUI facilitates teams in testing deployed models through a chat interface.

These applications continue to run on the computing and storage resources chosen by the user; costs, access controls, and data storage methods are determined by the underlying configurations. Closing a browser page does not necessarily result in the termination of the instance, so it is necessary to check the status of the resources after use.

The workflow from registration to operation

1. Create an organizational account and access the Web console to complete the settings for accounts, teams, and billing.
2. Select the region, project, and VPC, and verify the requirements regarding data storage, network access, and capacity.
3. Select GPU or CPU instances based on model memory, training scale, latency requirements, and budget.
4. Select a machine learning image, Notebook, OpenWebUI, or a custom image, and mount a persistent disk or shared storage.
5. Configure SSH, keys, user permissions, public or private IPs, as well as any required firewall and VPN connections.
6. Upload data and code, or retrieve model weights from controlled storage, and then carry out training, fine-tuning, or deployment tasks.
7. Monitor performance, capacity, and anomalies through logging, and automate repetitive tasks using APIs, CLI, or Terraform.
8. Once the task is completed, stop or delete the computing resources that are no longer in use, while retaining the disks, snapshots, and logs that are indeed necessary.

The process of deploying inference services

1. Confirm the model license, weight format, context length, memory requirements, and expected concurrency level.
2. Select a version from the model directory, or prepare a custom model that can be loaded by the vLLM Server or Ollama Server.
3. Select the GPU model, quantity, region, and whether to use a public or private endpoint to assess the ongoing costs.
4. Set up authentication, network policies, and keys; do not include access credentials in the frontend code or in public repositories.
5. Create an endpoint and conduct low-volume tests using a client that is compatible with the OpenAI API, in order to verify the parameters, responses, and error handling.
6. First, conduct stress testing on the token latency, throughput, context length, and maximum concurrency levels, before deciding whether to increase the number of instances or deploy multiple endpoints.
7. Monitor runtime and GPU utilization continuously; services with low utilization can have their capacity adjusted or can be shut down based on business requirements.

Prices and billing

Compute on demand is billed on a per-minute basis; the page displays reference prices per GPU hour or per instance hour. Reserving resources for several months or a whole year can save up to 20%, but the available capacity, regions, contract duration, and final price are subject to what is shown in the console or in the quotation.

On the page showing the total price, the H100 is listed at $2.30, while the AI Compute product page indicates a price of $2.45; therefore, the figure shown on a single page cannot be considered the fixed price for all accounts. Factors such as region, inventory levels, settlement periods, reservation conditions, and the timing of page updates can all affect the final price, so it is necessary to check the estimated actual cost before creating a resource.

CPU, storage, and networking

The current pricing page lists internet inflows and outflows, public IPv4 addresses, VPN gateways, and VPCs as being free of charge; however, this does not mean that all custom networks, inter-regional connections, or dedicated links are cost-free. The total cost for large-scale projects should be determined in the proposal and contract.

Startup credit limit and payments

The AI Ascend startup program offers an initial cloud quota of $1,000; eligible teams can apply for a production quota of up to $500,000. These quotas require an application and a qualification assessment, and they are not permanent free packages that are available automatically to all new accounts.

• For on-demand resources, payment can be made with a prepaid credit card; companies can also request invoices via bank transfer.
• Reserving resources usually requires a time commitment; early termination, changes to capacity, and the handling of unused amounts are subject to the terms of the order.
• The public page does not provide a unified refund policy applicable to all cloud services; it is necessary to check the terms regarding cancellation, refunds, SLAs, and service disruptions before making a purchase.
• Inference endpoints are charged based on GPU time; even if the volume of requests is low, charges may still apply as long as the endpoint remains running.

API, SDK, and open-source status

Denvr offers REST APIs, command-line tools, Python SDKs, Go SDKs, and a Terraform Provider, which can be used to create clusters, virtual servers, networks, and application resources. Teams can integrate these interfaces into continuous delivery processes, testing environments, and Infrastructure as Code workflows.

When using SDKs or examples, it is necessary to check the licenses and dependencies of the specific version repository; not all projects within an organization should be assumed to have the same licensing terms. When deploying third-party models, it is also required to comply with the licensing conditions related to the model weights, the base models, the datasets, and the content generated.

Supported platforms

Suitable for users and scenarios

• AI team: There is a need to obtain GPUs such as H100, A100, and GH200 as quickly as possible for training, fine-tuning, and evaluation.
• Enterprise developers: wish to deploy custom model endpoints in a way that is compatible with the OpenAI API, while also having control over the hardware and network.
• Regulated organizations: Require plans for the Canadian or U.S. region, virtual private networks, dedicated environments, and data residency.
• Platform engineering team: Aims to manage development and production resources in a unified manner using APIs, CLI, SDKs, or Terraform.
• Research and engineering teams: They require high-speed interconnections, shared storage, and multi-node clusters to run distributed workloads.
• Startups: Those that meet the AI Ascend criteria and require early-stage cloud capacity, technical support, or resources for scaling up production.

Advantages

• There are paths available ranging from a single GPU to multi-node clusters, covering tasks such as experimentation, training, and production inference.
• The inference endpoints utilize dedicated, single-tenant GPUs; the cost is calculated based on the amount of hardware time used, which facilitates capacity planning for tasks with high and stable workloads.
• It also offers a web console, API, CLI, SDK, and Terraform, making it suitable for teams at various levels of maturity.
• VPCs, private IPs, VPNs, and regional data centers help to create solutions for isolation and data residency.
• Private AI solutions can be deployed on dedicated clouds or on-premises infrastructure, and are not limited to public cloud instances.

Restrictions and Precautions

• GPU availability is affected by region, inventory, and reservation conditions; the hardware listed on the page is not guaranteed to be available for purchase at any time.
• The unit price of the H100 varies across different product pages; the actual cost must be confirmed in the console, order, or sales quote.
• Dedicated inference is charged based on runtime, and services with low request volumes may not be as cost-effective as those based on tokens.
• Public information does not provide unified rules regarding refunds, cancellations, SLAs, and data deletion timelines that apply to all products.
• The website terms and the cloud service agreements are not the same document; companies purchasing such services should review the actual service contract as well as the data processing terms.
• SOC 2 controls, security capabilities, and data residency requirements need to be determined based on specific regions, products, configurations, and the scope of the contract.
• The quality of the output provided by the platform depends on the models, versions, prompts, data, and inference settings used by the user; the infrastructure does not guarantee the accuracy of the model’s responses.

Privacy, Security, and Commercial Use

The platform handles account and company contact information, login credentials, communication tools, resource utilization rates, system metrics, as well as device and browsing data. Some service providers may process data in Canada, the United States, or other jurisdictions; companies should verify the applicable data processing agreements and cross-border requirements.

• The customer is responsible for obtaining the necessary legal permissions to process the data they upload; sensitive personal information should not be submitted in the absence of appropriate agreements and control measures.
• The platform employs management, technical, and physical security measures, and restricts access based on operational requirements; however, any cloud environment requires users to properly configure identities, keys, and networks.
• There is no fixed deadline set for the deletion of private content, and public activities may continue even after an account is deleted.
• The storage, backup, deletion of training data, fine-tuning weights, and logs, as well as incident response, should be specified in the enterprise contract.
• The license for using webpage content does not replace the commercial authorization for customer data, third-party models, or the generated results.
• Whether a model can be used commercially depends on the licenses for the model itself, its dependencies, and the training data; the infrastructure services provided by Denvr do not automatically grant such rights.

Frequently Asked Questions

Is Denvr a chatbot?

No. It is a platform for AI computing, reasoning, and private infrastructure; OpenWebUI is merely an interactive interface used to interact with the models that have been deployed.

Is there a free version of Denvr?

There is no permanently free computing tier available to all users. AI Ascend offers allocated resources on an application-based basis to eligible startup teams; it cannot be considered a universal free package that is provided automatically upon registration.

Is the inference service charged on a per-token basis?

The dedicated AI inference endpoints are not charged on a per-token basis; instead, the cost is determined by the amount of time the GPU is in use. The absence of a token limit does not mean that the service is free – ongoing usage of the endpoint can result in additional charges.

What is the current price of H100?

The total price page shows $2.30 per GPU hour, while another product page indicates $2.45 per hour. The price applicable before creation should be based on the selected region, account console, or official quote.

Can I deploy my own models?

Custom, open-source, or fine-tuned models can be deployed using the vLLM Server or Ollama Server. Users still need to verify format compatibility, memory requirements, and model licenses.

Is the interface exactly the same as the OpenAI API?

The inference endpoint uses a compatible calling format, which facilitates the reuse of common clients. The specific models, parameters, return fields, rates, and error behaviors still need to be confirmed through testing.

Is Denvr an open-source platform?

It has not been verified whether the cloud platform itself is open source. The Python SDK, Go SDK, Terraform Provider, as well as some example repositories have their code made available publicly, under the MIT or Apache-2.0 licenses; these licenses apply only to those specific projects.

Are there any mobile apps or browser extensions?

No official iOS or Android apps or browser extensions have been confirmed yet. The available access methods that have been verified are the Web console, API, CLI, SDK, and Terraform.

Will the company’s data be stored in Canada?

The platform offers options for deployment in Canada and the United States, as well as for private deployment; however, the actual location of the data depends on the selected region, services, backup options, and arrangements with the supplier. Organizations with specific requirements regarding where the data must be stored should verify each detail in the architecture and contract.

Summary

Denvr is suitable for teams that need a dedicated GPU, private inference capabilities, access to the Canadian region, or full control over their infrastructure. When making a choice, it is necessary to consider factors such as the availability of hardware, ongoing operational costs, network and storage expenses, model licenses, as well as the security, deletion, and refund provisions outlined in the contract.