Salesforce shared numbers this week: they used new AWS tools to compress AI inference (running large models to compute answers) GPU costs to 1/8 of the original. What's worth paying attention to: saving money is not the whole story — the truly hard part is what comes after the savings.
What this is
Salesforce's flagship product, Agentforce, is their AI Agent platform, running on AWS underneath. Originally one GPU per model, they changed it so multiple models share the same batch of GPUs, dropping costs directly to 12.5%. The problem: AWS's default scheduling logic only optimizes for cost, not for multi-AZ disaster recovery — that is, whether replicas of the same model are distributed evenly across different data centers. If one data center fails, the entire model becomes unavailable. Enterprise-grade AI cannot accept this risk.
AWS added a new parameter, SchedulingConfig, for this use case — letting enterprises specify how many replicas go into each availability zone and by what placement strategy. Essentially, they decoupled "saving money" from "being reliable," letting enterprises configure to their compliance requirements.
Industry view
Supporters see this as a sign of cloud vendor maturity — no longer just selling compute, but selling "configurable engineering capability." The concrete 8x cost reduction also shows that AI infrastructure's optimization headroom is larger than people imagined.
But we notice the other side: this solution is itself a patch on AWS's existing scheduling mechanism. Salesforce can write code to solve it because they have top-tier engineering teams; most enterprises do not. Put differently, the "enterprise-grade" standard for AI infrastructure is still being forced to mature by big customers — it's nowhere near plug-and-play. There's also an overlooked cost: every additional layer of scheduling configuration adds another layer of operational burden. The GPU money saved may get eaten by engineer salaries.
Impact on regular people
For enterprise IT: An "AI project budget" should not only count GPU costs, but also the hidden costs of compliance and disaster recovery. Before presenting the 8x cost reduction to the board, first confirm how many 9s of SLA (Service Level Agreement — the committed availability number, e.g., 99.99%) you can actually sign for.
For individual careers: Once AI Agents go live, "the system is down again" will become a more common line item in your workday. Reliability problems are spreading beyond internet-company IT departments and into every business unit that uses AI.
For the consumer market: Once AI is embedded in customer service, banking, and government services, the impact of a single data center outage gets amplified. Our tolerance for "AI downtime" will end up far lower than for traditional software.