Plan and Budget a Platform Project

The Loop Platform Connects#

Every production vision system runs the same loop, and the expensive part is the arrow between two boxes, not the box itself:

Platform's value is that each stage records what the next one needs. Use this table as your map; the last column is the lesson that teaches the stage:

Two tools span the loop. Agents builds visual workflows, such as collecting uncertain detections into a review dataset, and Ask AI can train, compare, annotate, export, and deploy from a conversation using the same ul cloud commands you can run yourself.

Decide What You Still Own#

Platform stops at a few deliberate boundaries. Give each one a named owner in your plan:

• Getting frames out of proprietary systems (DICOM archives, OEM cameras, plant historians): your code lands them in a bucket or an upload.
• Paging. Platform shows endpoint metrics and posts job events to Slack, but it doesn't alert on latency or error thresholds. You'll build that check in lesson 8.
• Accuracy over time. A labeled holdout and its history are yours (lesson 8).
• Application logic. Counting, tracking, and zone rules live in your service; Ultralytics Solutions and the Ultralytics YOLO in Production course cover them.
• Traffic routing. Endpoints have generated URLs and no custom domains, so canaries, multi-region routing, and branded hostnames need your own gateway.

Platform is not a separate framework. A Platform model downloads as a normal best.pt that YOLO("best.pt") loads, validates, and exports, and a local model.train(..., project="username/my-project") streams into Platform with an API key set. Moving between the two never means converting anything.

Worked Example: Budget the Project Before You Start#

Three things cost credits: cloud training (GPU rate × time used), Batch Annotation and whole-dataset face blurring ($1.00 per 1,000 images), and custom-size endpoints (billed per hour of uptime). Everything else in this course is free. This calculator uses Platform's real rates for a small detection project:

# Monthly Platform budget for a small detection project (rates from the billing and deployment docs)
TRAIN_RUNS = 8  # experiments per month
COST_PER_RUN = 0.23  # dialog estimate: ~1,000 images, YOLO26n, 100 epochs, RTX PRO 6000 ($2.09/h)
BATCH_ANNOTATION_IMAGES = 3_000  # $1.00 per 1,000 images
HOURS_PER_MONTH = 730

# Custom endpoint sizes keep one warm instance; us-central1 rates are $0.000024/vCPU-s and $0.0000025/GiB-s
warm_endpoint_hourly = (2 * 0.000024 + 4 * 0.0000025) * 3600  # 2 vCPU, 4 GiB

lines = {
    "cloud training": TRAIN_RUNS * COST_PER_RUN,
    "batch annotation": BATCH_ANNOTATION_IMAGES / 1000 * 1.00,
    "default endpoint (1 vCPU, 2 GiB, scales to zero)": 0.0,
    "warm endpoint (2 vCPU, 4 GiB, always on)": warm_endpoint_hourly * HOURS_PER_MONTH,
}
for item, usd in lines.items():
    print(f"{item:<50} ${usd:>7.2f}")
print(f"{'total with a warm endpoint':<50} ${sum(lines.values()):>7.2f}")

cloud training                                     $   1.84
batch annotation                                   $   3.00
default endpoint (1 vCPU, 2 GiB, scales to zero)   $   0.00
warm endpoint (2 vCPU, 4 GiB, always on)           $ 152.42
total with a warm endpoint                         $ 157.26

How to read it:

• COST_PER_RUN comes from the training dialog's estimate, which accounts for dataset size, model, epochs, image size, and GPU. Replace it with the estimate for your own dataset (lesson 5); 5,000 images × YOLO26s × 100 epochs on an H100 SXM is about $1.56.
• The endpoint line dominates. Training and labeling a small project costs a few dollars; keeping an instance warm 730 hours a month costs about $150 in Iowa (rates vary by region, and the deployment dialog shows the exact hourly price).
• Decision rule: start every endpoint at the free default size. Move to a custom size only when users can't tolerate a cold start of up to a minute and traffic is steady enough to justify the uptime bill.

Plans#

Decision rule: stay on Free while you learn; move to Pro when you need connected cloud storage, a team, more than three endpoints, or more than three concurrent trainings. Pro's $30 monthly credit covers the training and labeling lines of the budget above. The billing docs list every rate.

Pitfalls#