Snowflake launches Cortex Training public preview | Magica
Snowflake previews Cortex Training with a qualified cost claim
Editorial Team
••📖6 min read
Cortex Training gives developers custom training APIs on managed GPUs. Snowflake is extending its existing model-customization tools, with enterprise-data integration and a cost claim limited to LoRA fine-tuning.
Cortex Training puts GPU operations behind managed APIs while developers control model-customization logic.
The public preview requires account enablement and is restricted to commercial Amazon Web Services (AWS) regions.
Snowflake’s claim of up to 61% lower costs covers its own LoRA fine-tuning benchmark; deployment and total application costs require separate assessment.
Snowflake, the enterprise data-platform provider, announced the public preview of Cortex Training on November 3, giving developers APIs to customize open-weight language models while the company runs the GPU infrastructure. Its October 30 release notes had already recorded preview availability.
The company says more than 14,500 customers use its AI Data Cloud to build, use and share data, applications and AI. Cortex Training adds another computing workload to that platform: preparing enterprise data, adapting a model and deploying it under Snowflake’s governance controls. The commercial opportunity is to sell model customization alongside the data businesses already keep there.
But Snowflake already offered in-platform fine-tuning, and competing managed APIs already separate customer-controlled algorithms from provider-operated GPUs. The question is whether Snowflake’s combination delivers better results and economics for its customers. Its launch announcement claims costs up to 61% below existing serverless training options—a comparison of a specific fine-tuning workload, rather than the full expense of building and running an AI application.
Snowflake’s Cortex Training architecture connects training scripts and frameworks to APIs backed by managed GPU clusters. Source: Snowflake. Source: Snowflake.
What developers control, and what Snowflake operates
Cortex Training adapts open-weight models—models whose parameters are available for customization—to domain-specific tasks. Supervised fine-tuning, or SFT, learns from examples of desired behavior. Reinforcement learning, or RL, uses feedback on outcomes.
The service documentation describes a split architecture. A laptop or CPU-only virtual machine can run orchestration, data loading and batch preparation. Input batches go over REST to Snowflake for GPU-intensive operations, including forward and backward passes, optimizer updates, weight synchronization and generation. Snowflake holds the model weights and optimizer state.
Xiaomi says its humanoid robot sorted car parts for a full shift and corrected errors autonomously. The result extends its factory trials, but published evidence does not yet establish deployment costs or reliability across repeated shifts.
Editorial Team
Developers can use low-level APIs for custom loops or training-framework integrations. The preview supports full-parameter updates and low-rank adaptation, or LoRA, which freezes the base model and trains smaller additions. Customers can save and resume checkpoints, evaluate them, and deploy base models or customized checkpoints through OpenAI-compatible inference endpoints with streaming responses.
Snowflake says its open-source Arctic Platform backend underpins the service. It uses GPU multiplexing to schedule multiple training or inference jobs on the same GPUs, seeking higher utilization. Its technical launch post also says inference endpoints use dedicated GPUs, a deployment detail buyers need to include in their cost estimates.
The preview has boundaries. Its model catalog is curated, and support differs by model, context length and training mode. Framework integrations are at different stages of upstream acceptance; Snowflake directs customers to their account teams for current status. Models, APIs and specifications can change before general availability.
Snowflake’s earlier ArcticTraining workflow runs training code inside ML Jobs with access to internal data and stages. Source: Snowflake. Source: Snowflake engineering blog.
Integration is the competitive argument
This is an extension of Snowflake’s customization work. In a February 2, 2026 engineering post, the company described fine-tuning directly inside Snowflake with ArcticTraining, its open-source post-training framework, and ML Jobs. That approach runs containerized Python workloads on managed GPU compute pools, with direct access to Snowflake tables and stages. Cortex Training instead exposes GPU operations to a training loop that can run in the customer’s workspace.
There is external precedent, too. Thinking Machines Lab, the developer of the competing Tinker fine-tuning service, introduced its API on October 1, 2025, initially in private beta. It described low-level post-training operations, customer control over algorithms and data, provider-managed scheduling and failure recovery, and LoRA-based sharing of compute across training runs. Those design principles predate Snowflake’s preview.
Snowflake’s pitch is to connect customization to enterprise data and production feedback. Its launch post describes role-based access controls across training data, training, deployment and inference. Teams can combine agent traces, logs, application outcomes and human feedback into datasets for further model adaptation.
The open-source backend also qualifies how tightly workflows must depend on the managed service. The client repository describes an ArcticRLClient interface whose backend can be either Cortex Training or an on-premises Arctic Platform deployment, supporting portable RL recipes. That does not establish that every managed workflow is interchangeable across providers.
Customization is only one route Snowflake is offering to reduce agent costs. The same Expedition announcement describes dynamic model routing for its CoCo coding agent, assigning lighter models to high-volume background tasks and reserving stronger reasoning models for complex queries. Those routing features are in private preview. Buyers therefore have a separate decision about whether a task needs model adaptation at all.
Snowflake’s published on-demand rates are 4.80 AI Credits per GPU-hour for H200 and 9.00 for B300. These rates alone do not establish dollar savings against competitors. Chart: Magica; source: Snowflake documentation. Source: Snowflake documentation.
What the 61% comparison measures
Snowflake’s stated methodology runs LoRA SFT workloads on Cortex Training at specified context lengths, measures throughput per NVIDIA H200 GPU, and converts its GPU price into a price per million tokens. The calculation divides price per second by tokens processed per second, then multiplies by one million. The company used published Fireworks and Tinker prices from October 22, 2026.
That is a company-run training-cost comparison. The maximum saving does not establish the result for every model or context length, or equivalent savings for full-parameter updates, RL, evaluation or production serving. It also does not demonstrate equivalent task quality after customization.
Usage is metered per second with no minimum, but billing begins when GPU capacity is assigned, including initialization while the model loads. A queued job is not billed before assignment. Billing ends when capacity is released, including when a job finishes, fails or is canceled, or a model deployment is dropped. Customers must cancel jobs they no longer use to stop charges.
The documentation’s example—two H200 GPUs held for 90 minutes—consumes 2 × 1.5 × 4.80, or 14.40 AI Credits. These rates explain the unit of consumption; they do not by themselves supply a dollar comparison with token-priced competitors. Reserved-capacity pricing requires an account-team discussion.
The next decision is whether adaptation pays for the task
Customers first need account enablement and a supported commercial Amazon Web Services region. They must confirm the required model, training mode and framework integration before treating the preview as a deployment plan. The cited release notes give no general-availability date.
Snowflake recommends benchmarking base models on quality, latency and cost before investing in customization. That baseline matters: a lower training bill is useful only alongside the quality achieved and the subsequent cost of serving the model.
An independent analysis of the launch makes the broader economic test explicit: an agent request can involve multiple agents, retries and tools, so operators need to measure cost per successfully completed task and connect execution traces to business outcomes. Evidence on that basis—against an unmodified model and other available approaches—would show whether Cortex Training’s claimed training savings translate into a better enterprise application.
Railway’s CEO apologized after an engineer’s AI agents targeted rival Primcloud during testing. Primcloud says attempts to obtain customer secrets failed, leaving questions about the agents’ instructions and Railway’s controls.
Apple’s developer preview extends Siri AI beyond predefined app categories and adds guidance for complex tasks. Testing is limited to development builds and internal testers, with App Store submissions planned for spring 2027.
Jill Hazelbaker will join OpenAI as chief corporate affairs officer after a decade at Uber, bringing experience in technology communications and political campaigns to a role focused on reputation, positioning and public trust.
The watchOS 27.3 developer beta brings Live Rewind and Siri Recap to Apple Watch Series 12 and Ultra 4. The opt-in tools offer brief transcripts and conversation summaries, but require a compatible iPhone and carry Apple warnings about accuracy and sensitive information.
Traders reportedly extract Samsung and SK hynix memory from processor packages shipped into China, selling it mainly to Huawei customers. The account highlights the memory constraint behind Huawei’s AI chip rollout, but leaves trade volumes, recovery yields and transaction legality unresolved.
Creative-industry groups are challenging OpenAI and Microsoft’s arguments against licensing AI training material, as a New York court weighs fair use. Their briefs share an emphasis on licensing but differ over what the judge should decide.
Traversal's Series B backs its expansion from diagnosing software outages toward autonomous production operations. Company-reported adoption, cloud competition and unfinished access controls frame the next tests.
The startup formerly called TesterArmy is replacing its testing agent with an open-source framework, betting that adoption will generate demand for paid infrastructure and enterprise services.
Fal has announced Sequoia-led and Bessemer-led funding rounds for its 2027 expansion, adding post-trained media models to an established inference business. The announcements leave the new rounds’ size and comparative serving economics undisclosed.
SK hynix bought existing Penguin Solutions preferred shares from an SK Telecom investment vehicle. The deal brings director nomination rights and a convertible stake in an existing AI infrastructure partner, while sending the purchase price to the seller.
Google plans to switch off Gemini and Search for school-managed K–12 student accounts unless administrators enable them. Schools will gain control over access, but Google says keeping Search without its AI features will not be an option.
Anthropic is expected to exceed 3.5 GW of computing capacity in 2026 and aims to double it in 2027. Its filing shows $895.4 billion in commitments, substantial financial resources and an immediate cost from scarce capacity.
Independent tests reproduce MIT researchers’ repair of Qwen’s repetitive reasoning. The edited model still beats a stronger unedited configuration, but by much less; broader claims about creating capabilities remain unproven.
Gartner forecasts that open-weight models will execute 90% of enterprise operational agent steps by 2030. Its cost estimates favor routine, high-volume work, but infrastructure costs and the choice of comparison model leave enterprise savings to be proved.
FTAI Aviation is financing a planned 100-turbine production run in 2027, while ProEnergy has a 650 MW equipment order from Crusoe. Converted aircraft engines offer another supply route for data-center power, but delivery targets, affiliated financing and unproven operating economics qualify the opportunity.
Meta is marketing dollar bonds in a plan to borrow at least $20 billion, including an expected euro offering. Capital spending exceeded operating cash flow last quarter, but the debt is part of a wider financing strategy that also includes equity offerings and a proposed non-voting share-sale program.
A nonbinding opinion targets real-time retrieval and recognisable reuse of journalism, while rejecting a monopoly over news facts and recommending that questions about model training be dismissed as hypothetical.
A study of Lovable projects found passport details, customer records and private messages accessible without login, raising questions about database permissions and the limits of automated security checks. Its sample does not establish a platform-wide leak rate.
CrowdStrike is extending Falcon Guardian into Claude Cowork’s virtual machine and AWS’s AI guardrail layer. The Cowork beta addresses a visibility gap, while existing AWS safeguards and unfinished acquisitions limit what buyers can infer about the broader platform’s capabilities.
