Your Snowflake Cortex Training Bill Depends on GPU Throughput

Cortex Training is worth a pilot if your training data already lives in Snowflake, but its "up to 61% lower cost" comes from one training method on one GPU type. You pay by the GPU-second, at a rate that only beats token pricing if your job sustains high throughput. Snowflake's list price works out to $9.60 per H200 GPU-hour, and the press release does not state the throughput behind its claim. Run your own workload through the formula below before you move a fine-tuning budget.

Cortex Training is Snowflake's serverless service for customizing open-weight LLMs, announced in public preview on November 3, 2026 at Expedition 2026. The release note is dated October 30 and limits it to commercial AWS regions, with enablement through your account team. If your Snowflake Cortex account runs on Azure or Google Cloud, this decision waits.

---

What Snowflake's 61% Claim Measures

The 61% figure is a calculated price per million tokens for LoRA supervised fine-tuning on H200 GPUs, set against competitors' published list prices. LoRA trains small adapter layers on frozen base weights, which Snowflake says uses less GPU memory and runs faster than full-parameter training. SFT (supervised fine-tuning) is the other half of the label, and reinforcement learning and full-parameter runs are outside the footnote.

The press release footnote spells out the method. Fireworks and Tinker pricing came from their web pages on October 22, 2026. Snowflake says, "We ran LoRA SFT training workloads on Cortex Training, and measured the throughput per H200 GPU." It then computed (price_per_second / throughput) x 1M.

The footnote gives no throughput number, model or context length. Figure 5 of the launch blog, which repeats the claim against "Tinker and others," supplies the models. Measured against "the cheapest available alternative," it shows $2.92 vs $7.46 per million tokens on Qwen3.8-27B at 256K (61%), $2.29 vs $4.10 on the same model at 128K (44%), $2.82 vs $5.58 on GLM-5.3-Flash at 256K (49%) and $1.26 vs $2.14 on Qwen3.6-35B-A3B at 256K (41%). So the 61% is one bar of four. At $9.60 an hour, those Snowflake prices imply about 913 to 2,116 tokens per second per GPU, a range Snowflake does not state itself. No independent benchmark turned up in our search on November 4, 2026.

The comparison set is narrower than two vendors suggests. On the models both list at 256K context, Tinker and Fireworks Serverless Training charge the same training rate: $2.14 per million tokens for Qwen3.6 35B-A3B, $2.32 for Nemotron 3.5 Super VL 120B, $4.98 for DeepSeek V4.1 Flash 1017 and $14.58 for GLM 5.3. Beating one on the train meter means beating both by the same margin.

---

What Cortex Training Costs Per GPU-Hour

Cortex Training costs 4.80 AI Credits per H200 GPU-hour on demand, or $9.60 at Snowflake's $2.00 global credit price. Both numbers are in the Service Consumption Table effective October 29, 2026, which also lists the B300 at 9.00 credits ($18.00). The table's regional credit price is $2.20, which would make an H200 hour $10.56. Snowflake's AI pricing page ties the regional rate to cross-region routing settings without addressing training jobs, so ask which applies to yours if residency is the reason you are here.

For scale, Fireworks' Dedicated Training API lists the same H200 at $8.00 an hour and the B300 at $15.00, checked November 4, 2026, with a 1.5x charge for region-restricted deployments. Snowflake's raw GPU hour costs more than Fireworks' standard rate and less than its region-restricted one. Snowflake's case therefore depends on doing more work per GPU-second.

The docs describe billing per second with no minimum, starting when GPU capacity is assigned (model loading included) and ending when capacity is released. The models page lists eight GPUs for every trainable model, and Snowflake's quickstart job uses eight, four for training and four for sampling. If a job holds all eight, that is $76.80 an hour. The quickstart's training job sets an 1,800-second idle timeout, so a forgotten notebook could burn $38.40 before that timeout fires, and the guide does not say what happens then.

Snowflake's docs leave idle time unclear. They say you pay for "the GPU hours you use instead of holding idle capacity while non-GPU parts of your pipeline run," and that scheduling several jobs on the same GPUs raises utilization. They also say billing ends when Snowflake releases the job's capacity, on completion, failure, cancellation or a dropped deployment, and that you should cancel a job you no longer use. Your loop runs in your own workspace and sends batches over REST, so if a job keeps its capacity between batches, slow data loading and pauses for evals land on your bill. Token pricing leaves that risk with the vendor.

How Much Throughput You Need to Beat Tinker

To tie Tinker's list price on Qwen3 8B, Cortex Training has to sustain about 6,061 tokens per second per H200 GPU. A 61% saving on that model would take about 15,540, though Snowflake's own 61% is on Qwen3.8-27B at 256K, not this model. Applying Snowflake's own formula to the four trainable Qwen models that have a Tinker rate gives these targets, assuming the $9.60 global rate and a fully busy GPU:

Model (context)	Tinker train, per 1M tokens	Fireworks Serverless train, per 1M	Tokens/sec per GPU to tie Tinker	Tokens/sec per GPU for 61% lower
Qwen3 8B (32K)	$0.44	Not listed	6,061	15,540
Qwen3.5 4B (256K)	$1.34	Not listed	1,990	5,103
Qwen3.6 35B-A3B (256K)	$2.14	$2.14	1,246	3,195
Qwen3.8 27B (256K)	$7.46	$4.103 (at 128K)	357	917

Tinker's rates come from its pricing page, stamped October 19, 2026, and context length moves them a lot: Qwen3.8 27B drops to $4.103 at 64K, which raises the tie point to about 650 tokens per second. If your sequences are short, compare against the short-context rate.

Published per-token rates need no measurement at all. Fireworks' Managed Training charges $0.50 per million training tokens for LoRA SFT on models up to 16B parameters, doubling for full-parameter work. That simplicity has a value that belongs in your comparison, especially for a team running a few jobs a month.

Where Snowflake's Case Is Strongest

The strongest cost argument for Cortex Training sits outside the benchmark Snowflake published. The open source Arctic Platform behind the service describes prompt deduplication for PPO and GRPO-style reinforcement learning, and says that on long sequence tasks 80% to 95% of the tokens in a batch are redundant prompt tokens. The same page lists SFT among the trainers still to come. Token-metered vendors bill prefill and sampling separately, so an RL team with long shared prompts could plausibly save more on GPU time than an SFT team would. Snowflake has published no such number, and the page does not say the hosted service turns deduplication on, so treat it as a test worth running.

The backend's Apache 2.0 license also lets your engineers read what runs on the GPUs. And if your legal team would take six months to approve sending labeled customer records to a new vendor, a higher per-token cost inside an existing Snowflake agreement may still be the cheaper project.

The evidence from customers is thin on cost. Ian Fischer, co-CEO of Poetiq, praises flexibility, "from LoRA, to SFT, to full-parameter post-training with our custom GRPO implementation," and gives no savings figure. The other named users in the launch blog are ResolveAI, OpenHands and Snowflake AI Research. All four build AI systems for a living, and none is a regulated enterprise reporting a bill.

---

How Far the Governance Boundary Really Extends

Governance is the better reason to look at Cortex Training, though the boundary is smaller than the press release implies. The release says, "Every step, from preparing training data to deploying the finished model, happens inside Snowflake's governed environment." The docs add that your workspace can be a laptop or a CPU-only VM, and the quickstart shows batches tokenized locally before being sent to the training job. Run that loop on a laptop and those batches pass through the laptop. The quickstart's route, a Snowflake Notebook on Container Runtime, keeps the loop on the platform, and that is the setup to mandate.

Four other limits belong in your review:

1. Snowflake's preview policy says preview features "should not be used in production systems or with production data." That collides directly with fine-tuning on real regulated records, so get your account team's position in writing.
2. The API reference says, "Only EVALUATION deployments are supported, and a deployment holds its GPUs until you drop it," although the product page describes production endpoints. The quickstart's four-GPU endpoint would cost $38.40 an hour at the H200 rate, or about $28,000 for a 730-hour month, if deployments bill at that table rate. Our fine-tuning cost breakdown found a $326 to $28,908 monthly swing in serving costs for the same weights, against about $30 to train.
3. Weight portability is undocumented. The docs say you "don't download model weights to your machine," the API reference lists export_checkpoint(job_id, checkpoint_id) with no documented return value or destination, and the quickstart warns that a saved checkpoint "is not recoverable after teardown." Find out whether you can take your adapter elsewhere before you train something you depend on.
4. The trainable catalog is six Qwen models. The launch blog names Qwen, GLM, Nemotron, Gemma and DeepSeek, but the models page on November 4 listed training support for Qwen3 0.6B, 1.7B and 8B, Qwen3.5 4B, Qwen3.6 35B-A3B and Qwen3.8 27B, with DeepSeek V4 Flash, GPT-OSS 120B and GLM 5.2 as inference only. GLM-5.3-Flash, which Figure 5 prices, is not on that page. Qwen3.6's LoRA currently covers only the attention projections.

Expect your security review to ask about multi-tenant scheduling. The docs say the service "dynamically schedules multiple training or inference jobs on the same GPUs," and the pages we read do not describe tenant isolation or how a shared GPU is billed. Ask for both.

Full-parameter runs and continued pretraining also change your regulatory position. Our EU AI Act analysis found that a typical LoRA run sits far below the one-third compute test, while continued pretraining on an undisclosed base can cross it. Cortex Training supports both, so log which method each job used.

---

What to Do Before You Move a Fine-Tuning Budget

This Week:

1. Ask your account team five questions and get written answers: does preview allow production data under your agreement, does training bill at the global or regional credit rate, does a job bill while your loop sits idle between batches, how is a shared GPU billed, and what does export_checkpoint produce?
2. Check whether you need fine-tuning at all. DSPy GEPA and similar prompt optimizers are worth trying first, and Snowflake's own product page says to tune when prompting and retrieval stop improving accuracy.

This Month:

3. Run one real dataset on non-production or synthetic records through Cortex Training and through Fireworks or Tinker, on the same Qwen model, with the same context length and epochs. Record total tokens, wall-clock GPU seconds and the final invoice line, and compare cost per million tokens using your measured throughput. Score the output with your own evals as well, because cost per token says nothing about quality.
4. Log how long your loop sits idle during that run, then compare it with the invoice line. A throughput benchmark leaves that time out, and the comparison shows whether you paid for it.

Before Any Reserved Commitment:

5. Hold off on the 1-3 year reserved capacity terms until the service is GA and production deployments are documented. Snowflake's table says previews may carry different pricing, and the docs warn that models and APIs may change before general availability.
6. Price the serving line for 12 months before you approve the training job, and compare it with your other open-weight inference options.

The Bottom Line

GPU-second billing moves at least some of the utilization risk of fine-tuning from the vendor to you, and how much depends on the idle-time answer you get from Snowflake. Snowflake's older Cortex Fine-tuning function lists 0.64 Platform Credits per million training tokens on llama3.1-8b, or $1.92 at the $3.00 Enterprise credit price in US East, so you could estimate a job from its token count. Cortex Training adds methods and models, and its price per token now depends on the throughput your job gets. That trade suits an RL team running long jobs on data it cannot move. A team tuning a small classifier a few times a quarter is probably better served by a token meter.

If you are already reviewing the rest of your Snowflake AI spend, our alternatives analysis covers moving AI compute without moving tables. For this decision, divide 2,667 by your measured tokens per second per GPU to get your cost per million tokens at $9.60, then put that number next to Tinker's rate for the same model.

Continue Reading

• Fine-Tuning vs RAG Cost: The Training Bill Isn't the Bill
• DSPy GEPA Lets Teams Test Prompts Before Fine-Tuning
• EU AI Act's one-third test can make fine-tuners model providers
• Snowflake Cortex vs Databricks Mosaic AI: Pick on Exit Cost
• Snowflake Takes Cortex AI Gateway GA, Previews Business Ontology
• Together vs Fireworks vs Groq: Fireworks Leads Moonshot's Audit
• Microsoft Starts Billing for Copilot Harness on December 1