Gradient AI provides cloud GPU instances for AI model training and inference, supports fast notebooks, preconfigured images, and elastic clusters, reducing infrastructure management costs.

GradientAI

Core parameters and statistics of Gradient AI

Gradient AI (formerly Paperspace Gradient) is DigitalOcean's cloud GPU infrastructure platform for AI training and inference. Its core value is "obtaining GPU computing power on demand, without requiring upfront hardware investment, and enabling training to be started in minutes." It is on the same track as AWS SageMaker, Lambda Labs, RunPod and other competing products, but it forms a differentiated positioning with "simplified developer experience + DigitalOcean's ecological integration".

Projects	Public Information
Official positioning	AI training and inference cloud platform (Paperspace by DigitalOcean)
Core capabilities	GPU instance Jupyter Notebook, model deployment, elastic cluster
GPU Type	NVIDIA H100, A100, L40S, RTX 5000 Ada, etc.
Deployment form	Cloud SaaS
Billing granularity	Billed by second, billed by hour
Main customer groups	AI developers, startups, small and medium-sized enterprises
Parent company	DigitalOcean (acquisition completed in 2024)
Latest version	platform-2026-Q2 (continuous iteration of cloud services)

Differences from AWS SageMaker: Gradient is aimed at the developer group who "need GPU but don't want to learn AWS". Its Notebook startup speed is typically 30-60 seconds, much faster than SageMaker's 3-5 minutes; the billing model is also more intuitive - a flat rate per hour by GPU instance type, rather than SageMaker's multiple overlays of instance + storage + network + API calls. The trade-off is significantly less GPU model selection, regional coverage, and network customization capabilities than hyperscale clouds.

The significance of DigitalOcean ecological integration: After Gradient was acquired by DigitalOcean, it can be used in combination with its Droplets (virtual machines), Managed Databases, App Platform and other products to form a one-stop architecture of "database + backend + AI reasoning". For teams already running operations on DigitalOcean, this means adding layers of AI computing power without having to switch between multiple cloud vendors.

Users and market recognition of Gradient AI

Specific user numbers and revenue data have not been officially disclosed. Its market position can be assessed from the following dimensions:

Acquisition signal: DigitalOcean announced the acquisition of Paperspace in late 2023, with integration completed in 2024. As a Nasdaq-listed company (NYSE: DOCN), DigitalOcean's acquisition decision has undergone public market compliance review, which has verified Paperspace's technology and customer base. Post-acquisition Gradient will remain operating as the core brand for DigitalOcean's AI/ML products.

Target customer group: Gradient's main user profile is "AI startups and independent developers" - that is, groups that require GPU training and reasoning capabilities, but have small teams, no dedicated DevOps, and difficulty supporting the complex billing and permissions systems of AWS/Azure. This highly overlaps with DigitalOcean’s SME customer base.

Competitive position: In the "on-demand GPU cloud" market segment, Gradient is differentiated from Lambda Labs (for large model training), RunPod (for inference and low cost), and Vast.ai (for crowdsourced GPUs). The core difference of Gradient is "product completion" - it provides a complete link from Notebook (development) → Machines (training) → Deployments (deployment), rather than a pure computing power market or a single inference endpoint.

Cost Advantages of Gradient AI

The cost advantage of Gradient AI does not come from the absolute low price, but from the three-layer structure of "transparent per-second billing + no hidden consumption traps + collaborative discounts with the DigitalOcean ecosystem".

C client/individual: Billed by GPU instance per second, starting from RTX 5000 Ada (approximately $0.3-0.5/hour), no minimum consumption, no long-term contract. Storage and network traffic are billed separately, with storage space billed per GB/month and outbound traffic billed per GB. Cost Trap: Continuously running instances without shutting down is the largest hidden cost - billing per second means that "forgetting to shut down" will directly incur ongoing charges. It is recommended to set an automatic shutdown policy for non-production environments.

Developer/API Calls: Gradient provides a CLI and Python SDK to manage instances and deployments through APIs. There is no additional billing for the API calls themselves, but GPU instances launched through the API are still billed per instance type. In contrast, AWS SageMaker charges API call fees and storage fees in addition to instance fees, making the total cost structure more complex. Taking a medium training task (single card A100, running for 100 hours) as an example, the price range comparison of the three platforms is as follows:

Expenses	Gradient (DigitalOcean)	AWS SageMaker	Lambda Labs
GPU instance (A100 80G)	About $1.5-2.5/hour	About $3-5/hour	About $1.1-1.8/hour
Storage fee	$0.1/GB/month	$0.2/GB/month	Including instance storage
API call fee	None	$0.01/call	None
Outbound traffic	$0.01/GB	$0.05-0.12/GB	$0.01/GB
Estimated total cost for 100 hours	~$200	~$400+	~$150

The above is an estimate of the public price range, which is subject to the real-time pricing page of each platform.

Enterprise/Contract: Gradient itself is a pure SaaS platform and does not provide private deployment solutions. But for large-scale needs, Reserved Instance discounts and dedicated support can be negotiated through DigitalOcean's Enterprise Contract. Hidden costs: Enterprises need to verify the data residency policy before adopting it - Gradient's data centers are mainly located in North America and Europe, with limited coverage in the Asia-Pacific region, which may affect compliance choices. Additionally, Gradient does not offer an on-premises deployment option, a limitation that data sovereignty-sensitive industries need to evaluate.

Main features of Gradient AI

Gradient's product design follows the concept of "the same pipeline from development to deployment", and the functional links cover the three stages of the model life cycle:

• Jupyter Notebook (Development): Web-based ML development context, pre-configured with mainstream framework images such as PyTorch, TensorFlow, JAX, etc., ready to use at startup. Supports code completion, version control integration and real-time collaboration. Floor Tips: Notebook will automatically disconnect after being idle for a period of time. Please save checkpoints regularly to avoid unexpected interruptions during long-term training.
• GPU Machines (training): Single machine and single card to multi-machine and multi-card clusters, supporting on-demand creation and release. Covering NVIDIA H100, A100, L40S, RTX 5000 Ada and other models. Core Value: From boot to SSH uptime is typically 30-60 seconds, much faster than the 3-10 minutes of traditional cloud vendors. This significantly improves the efficiency of "trial and error" in the research phase.
• Model Deployments (Inference): The trained model can be deployed as a REST API endpoint with one click, automatically handling request routing, expansion and contraction, and health checks. Supports both GPU inference and CPU inference modes. Acceptance concerns: The trigger conditions and cold start delay of automatic expansion and contraction need to be tested before production, especially in scenarios with large fluctuations in inference load.
• Preconfigured container images: Provides support for standard images and custom images of the deep learning framework, avoiding repeated installation of dependencies each time you create a container. Synergy: Linked with Notebook and Machines - the code debugged in the Notebook can be directly submitted as a training task, and the model weights of the training output can be directly pointed to the Deployment, forming an isomorphic environment of "development-training-deployment" and reducing cross-section configuration deviations.
• Team Management and Usage Monitoring: Organize members and resources by Project, support RBAC permission control, usage quotas, and bill separation. Ideal for multi-project management for small teams or startups. Note: The team management granularity is project level rather than resource level, which may not be enough for large teams that require fine-grained permission isolation (such as restricting users by GPU type).

Version evolution of Gradient AI

As a cloud service product, Gradient does not have a traditional software version number system. The following is a minimal context based on publicly verifiable milestones:

Time nodes	Milestones	Description
~2014	Paperspace was established	Co-founded by Dillon Erb, Daniel Kobran, etc., initially focused on cloud GPU game streaming
~2017	Transforming ML infrastructure	Shifting from game streaming to GPU cloud computing platform for AI/ML, launching Gradient brand
~2020	Gradient platform officially released	Core product line Notebook + Machines + Deployments taking shape
2023-11	DigitalOcean announces acquisition	DigitalOcean announces acquisition of Paperspace for an undisclosed amount
2024	Acquisition completed, brand integration	Gradient operates as DigitalOcean AI/ML product line
2025	Continuous iteration	Add H100 instances, expand data center area, and improve Notebook experience
2026-Q2	Current platform version	Continuous iteration, no fixed version number, please refer to the official status page and Changelog

Version Features: Gradient’s server is continuously updated, and users do not need to manually upgrade. Feature updates are released via the DigitalOcean Changelog and Paperspace Blog. For teams with critical business dependencies, it is recommended to follow the DigitalOcean Status Page for service change notifications.

Technical advantages of Gradient AI

The technical value of Gradient does not lie in the innovation of AI models (it does not develop large models), but in "making the out-of-box experience of GPU computing as close as possible to the local development environment."

Containerized Bounded Isolation: Each Notebook and Machines instance runs in an independent container, with NVIDIA CUDA, cuDNN and major ML frameworks pre-installed. Users do not need to manually configure GPU driver and CUDA version compatibility. Mechanism → Effect: Compress context configuration time from hours to minutes, allowing developers to focus on model experiments rather than context debugging.

Snapshots and Automatic Versioning: Notebooks and Machines support automatic snapshots, saving disk state every time you stop. Training scripts, data, and context configurations are versioned for easy traceability and reproducibility. Applicable scenarios: When parameters are frequently adjusted during the experimental phase, the snapshot mechanism allows "rolling back to the previous version and starting again after failure" to reduce the cost of trial and error.

Deep integration with DigitalOcean infrastructure: Gradient instances run directly on DigitalOcean's global data center network and belong to the same intranet as Droplets, Spaces (object storage), and Managed Database. Synergy: Training data can be loaded from Spaces through the intranet, and inference results can be written directly to the Managed Database, avoiding the delay and traffic costs of public network transmission. This is the core difference between Gradient and pure GPU computing power markets (such as Vast.ai, RunPod) - the latter usually only provide computing power and do not provide supporting storage and database services.

Per-second billing engine: Instances are measured by seconds and billed by the hour. The billing granularity accurate to the second means that short-term experiments (such as inference tests of several minutes) will not be wasted by the "minimum billing unit". For high-frequency experiments and low-load inference scenarios, this billing model has a cost advantage over cloud vendors that charge by the hour.

How to use Gradient AI

Gradient provides two access paths: Web console and API/CLI, covering different levels of needs from visual operations to automated management:

How to use	Suitable scenarios	Features	Cost
Web console	Personal development, small team collaboration	Visit console.paperspace.com with a browser to visually create and manage instances	Billed by GPU instance
CLI (gradient CLI)	Automated script CI/CD integration	Command line to create/stop/manage machines and deployment	Same as above
Python SDK	Embedded control of training scripts	Call the API in the training code to complete instance management and status query	Same as above
REST API	Deep integration, platform embedding	Complete RESTful API, suitable for embedding Gradient into your own platform	Same as above

Quick Start Steps:

1. Visit console.paperspace.com to register a DigitalOcean account (supports Google/GitHub login).
2. After creating the Project, select Notebook or Machines, select the GPU type and preconfigured image.
3. After the instance is started, access the Notebook interface through the browser or connect to the terminal through SSH.
4. The model weight after training can be used to create an inference endpoint through Deployments and specify the model path and number of copies.

Implementation Tip: It is recommended to enable the Auto-shutdown policy when creating an instance and set an idle timeout (such as automatically stopping if there is no activity for 30 minutes) to avoid continuous billing caused by forgetting to shut down. For long-running production inference endpoints, you can configure Health Checks and Auto-scaling to ensure availability.

Product Pricing for Gradient AI

Gradient's pricing continues DigitalOcean's "transparent, no hidden fees" pricing philosophy. There is no subscription fee, and pricing is entirely based on GPU instance type and usage time.

• C Client/Individual: Billed by second per GPU instance and billed by hour. Entry-level instances (RTX 5000 Ada) are about $0.3-0.5 per hour, mid-range instances (L40S) are about $0.8-1.2, high-end instances (A100) are about $1.5-2.5, and flagship instances (H100) are about $3-5. Storage is billed separately per GB/month, and egress traffic is billed per GB. Free Credit: Gradient does not provide permanent free GPU credit, but new DigitalOcean users usually receive a $200 trial credit (subject to official real-time activity).
• Developer/API Calls: There is no additional billing for the API calls themselves. Developers are mainly responsible for GPU instance fees + storage fees + outbound traffic. This means that for short-term, high-frequency API-triggered training tasks, the total cost is determined by the "running time of the instance launched per trigger" rather than the number of calls.
• Enterprise/Contract: Negotiate Reserved Instance discounts and dedicated support through the DigitalOcean Enterprise sales team. Enterprise contracts typically include: discounted GPU hourly rate, priority support SLA, bill consolidation (Gradient + Droplet + other DigitalOcean services). Purchasing Prerequisites: Enterprises need to first evaluate whether the annualized GPU usage reaches the discount threshold of reserved instances. Usually, the average monthly usage of $1000+ is worth negotiating.

Pricing Transparency: DigitalOcean is known for its transparent pricing. The hourly rates for all GPU models are publicly listed on the official website and can be viewed without registration. This is in contrast to AWS/Azure's "price calculator" model, which requires entering multiple variables to estimate monthly costs. However, it is still recommended to create an instance through the console to confirm the real-time rate before official use, because the cloud service price may be adjusted depending on the region and supply and demand.

Application scenarios of Gradient AI

The following four scenarios have been ecologically verified by Gradient and DigitalOcean, covering the complete link from personal experimentation to production deployment:

• Quick Model Experimentation and Prototyping Validation: Data scientists need to quickly validate new model architectures or hyperparameter configurations. Launch a pre-configured environment in 30 seconds via Notebook, write and debug code directly in the browser, and close as soon as the experiment is complete. Compared with traditional methods: Local GPU limited hardware configuration, IT approval purchase takes 1-2 weeks; Gradient compresses the time "from idea to first training" from days to 30 minutes. Implementation Tips: It is recommended to turn on automatic snapshots during the experimental phase to facilitate backtracking of different versions of experimental configurations.
• Short-term computing power expansion and overflow calculation: When the local GPU cluster is full, some training tasks will be flexibly expanded to Gradient. The per-second billing feature makes this "hybrid computing" model economically feasible - cloud computing power is only used when needed and released immediately after the task is completed, without incurring idle costs. Human-machine collaboration boundary: The scheduling and monitoring of expansion tasks can be 100% automated (through API/CLI scripts), but the evaluation of training results and model selection still require manual intervention - automation cannot replace the judgment of domain experts on model quality.
• Model inference endpoint deployment: Package the trained model into a REST API endpoint for online inference. Gradient's Deployments service automatically handles request routing and expansion and contraction, which is suitable for scenarios where the number of users fluctuates greatly. Note: The cold start latency of the inference endpoint (the time between starting a new replica and accepting requests) is usually 1-3 minutes in GPU mode. It is recommended to set Minimum Replicas to ensure that instances are still resident during low periods and avoid request thrashing. For latency-sensitive real-time inference (<100ms), verify that end-to-end latency meets the SLA in a test environment.
• Education Training and Workshop: Create notebooks with the same configuration in batches for students or workshop participants to ensure that everyone starts experiments on the same starting line. Gradient’s project-level permission management enables

Instructors can uniformly manage instance quotas and cost caps. Quantified cost reduction and efficiency improvement: Taking a deep learning workshop with 20 people as an example, the traditional method requires preparing a pre-installed physical machine or local VM for each person, and the preparation time is about 2-3 days; after using Gradient, students can scan the QR code to register and obtain a uniformly configured cloud environment within 5 minutes, and the preparation time is compressed to less than 1 hour. The instance is released immediately after the course ends, with no ongoing costs.

Applicable groups of Gradient AI

Gradient's positioning determines that it is most suitable for groups that "need GPU computing power but don't want to be slowed down by the complex systems of cloud vendors." Its users are clearly stratified:

• AI Independent Developers and Freelancers: Personal projects or freelance contracts require GPU training and inference capabilities. The per-second billing + no monthly fee model allows low-frequency users to only pay for actual usage. Misfit Boundary: If GPU usage is extremely low (<10 hours) throughout the year, the total cost of on-demand billing may still be higher than a "buyout" on-premises GPU solution (such as a personal workstation with an RTX 4090). It is recommended to estimate the annualized GPU demand before making a decision.
• AI startups (1-20 people): This is Gradient’s core customer group. The team needs to quickly iterate the model and flexibly expand capacity, but cannot support full-time DevOps or large cloud contracts. The combination of Gradient + DigitalOcean provides a single-vendor solution for "AI training + backend deployment". Implementation Tips: Startups should establish instance cost monitoring habits in the early stages - Although Gradient's billing is transparent, when multiple people share accounts, unexpected overspending may occur due to a lack of budget awareness. It is recommended to use project-level quotas to cap monthly budgets for each project.
• Small and medium-sized enterprises and technical teams: Teams that are already running business on DigitalOcean can add an AI computing power layer without switching cloud vendors. Purchasing Prerequisites: The enterprise should have at least one clear AI application scenario (such as customer intent classification, intelligent document extraction, image quality inspection), and have PoC verification results. It is not recommended to purchase GPU computing power first to "explore AI possibilities" - you should first complete the PoC through Gradient's on-demand instances, and then decide whether to commit to reserved instances based on usage and effects.

Not suitable for groups: Enterprises that require large-scale distributed training (100+ GPUs), financial institutions that have strict compliance requirements for data sovereignty and must be deployed locally, and research institutions that require extremely rich GPU instance types (such as TPU, AMD MI300X, etc.). These scenarios are more suitable for AWS/Azure or specialized computing power markets (Vast.ai, RunPod).

Summary and Outlook

Gradient AI has found a precise niche in the "on-demand GPU cloud" market - it is not as complex and expensive as hyperscale clouds, nor does it lack product experience like the computing power market. Through the product matrix of Notebook + Machines + Deployments and the ecological support of DigitalOcean, it provides tens of thousands of AI developers and startups with a complete link from development to deployment.

Core advantages: Fast startup (30-60 seconds), transparent billing (per-second/public pricing), integration with DigitalOcean intranet (reduces data transmission costs and delays), team management and permission control suitable for small teams.

Current limitations: GPU model types and regional coverage are far less than AWS/Azure; privatized deployment is not supported, and industries sensitive to data sovereignty need to assess compliance risks; the cluster capabilities of large-scale distributed training (multiple machines and multiple cards) have not been widely verified; data center coverage in the Asia-Pacific region is limited, and user latency in China is relatively high.

Follow-up observation points: DigitalOcean's continued investment in Gradient (frequency of new product feature releases, data center expansion plans); whether to launch Serverless GPU or more fine-grained inference billing options; depth of integration with the DigitalOcean App Platform - if the Gradient inference endpoint can be directly referenced in the App Platform in the future, it will further lower the developer threshold.

Procurement and Adoption Risk Assessment: For individual developers and startups, the flexible pay-as-you-go model has almost no upfront sunk cost risk. It is recommended to start with the Notebook free experience to verify the workflow. For enterprises already running operations on DigitalOcean, incorporating Gradient into their existing technology stack is a low-risk option—with intranet integration providing immediate latency and cost benefits. For enterprises considering large-scale migration, it is recommended to conduct 2-4 weeks of grayscale testing on 1-2 non-critical models, focusing on verifying: GPU instance stability (whether there are frequent restarts or resource contention), P99 latency of the inference endpoint under target concurrency, and whether the price is lower than AWS/Azure Reserved Instances with equivalent configurations. In terms of data compliance, it is necessary to confirm whether the area where Gradient's data center is located meets industry regulatory requirements; if local deployment is required, Gradient is currently not a suitable option.

Related tools: , replicate

Gradient AI model and version evolution

Continuous iterative updates, the latest version introduces performance optimization and new features. Historical version information can be viewed through the official release page. There is currently no complete public version evolution timeline.

How to use Gradient AI

• Web client: You can use it by visiting the official website and registering an account. Most functions do not require installation.
• API access: Provides RESTful API, developers can obtain the API Key and integrate it into their own applications.

Version Info

• Gradient June 2026 Update ：A continuously iterative cloud service platform with no fixed version number. （2026-06）
• Gradient platform release ：There is no official precise date yet. （~2023-01）