Deploy and Manage Foundation Models

Bundled into the existing Operationalizing Machine Learning and Generative AI Solutions premium course — no separate purchase.

14-day money-back guarantee — no questions asked.

Included in this chapter:

Decision axis	Standard	Provisioned (PTU)	Global Batch
Billing basis	Pay per token	Per PTU per hour (reservation-discountable)	Pay per token, 50% off Global Standard
Capacity	Best-effort, shared	Reserved, dedicated	Separate async batch quota
Latency	Variable at high volume	Low and predictable	Async, up to 24-hour turnaround
Throughput guarantee	None (TPM/RPM rate limit)	Guaranteed by allocated PTUs	No real-time SLA
Best for	Variable or bursty real-time traffic	High, steady real-time volume	Large offline, cost-minimizing jobs
Cost lever	Rebalance TPM or request an increase	Provisioned Throughput Reservation	The batch discount itself

Cheat sheet

Serverless API deployment bills per token and needs no compute quota
Managed compute uses dedicated VMs, needs compute quota, and bills hourly while running
Standard deployments are best-effort per-token; Provisioned reserve capacity
Global routes anywhere; Data Zone keeps processing within a geography
A deployment content filter tunes four harm categories and severity thresholds plus blocklists and Prompt Shields against direct jailbreak and indirect (XPIA) prompt injection
Standard deployments draw a per-deployment TPM quota from a region-subscription pool
Global Batch is the cost-optimal deployment type for high-volume asynchronous workloads
Protected-material text and code are output-only content filters, on by default, and separate from the ProtectedMaterial evaluator
Use the model leaderboard to compare quality, safety, cost, and throughput
The catalog spans many model providers
Match the model class to the workload before comparing individual models
Model router auto-selects the best underlying model per prompt
Three deployment version-upgrade options control model updates
Automatic updates apply to Standard only; Provisioned migrates manually
GA models get at least 60 days' retirement notice
PTUs reserve fixed capacity billed hourly regardless of tokens
Spillover routes overflowing PTU requests to a paired standard deployment
Provisioned Throughput Reservations discount committed PTUs for a fixed term

Unlock with Premium — includes all practice exams and the complete study guide.

Cheat sheet