Beyond the GPU: AMD's Jeremy Gerbin on right-sizing AI infrastructure at Cloudera Evolve
Key points
Key takeaways from the theCUBE (SiliconANGLE) interview at Cloudera Evolve 2026 in New York City with Jeremy Gerbin, AMD's AWS technical enablement manager (October 2026):
The CPU flip. When AI took off, the story was GPU-centric — roughly one CPU managing eight GPUs. Gerbin says the industry is now inverting: one GPU is being serviced by multiple CPUs, and CPU has become the more heavily focused product. He cautions the picture is not fully settled.
Cost is the #1 pain point. Customers are struggling to move AI from pilot to production and the dominant concern is cost. Much of Gerbin's one-of-one role is translating AMD's performance numbers into business value, since companies talk about performance but pay for outcomes.
Requirements first, hardware second. Not every AI workload needs the best model or most expensive hardware. Right-sizing starts by pinning down requirements — latency, acceptable time-to-first-token — and discovering that what looks like an AI problem sometimes is not. If a user can wait 30 seconds, cheaper hardware suffices.
AMD + Cloudera. The two companies have been partners for years, validating AMD products on Cloudera so customers can run AI where their data lives. AMD brings the infrastructure story, Cloudera the application side.
Governance as the foundation. Agents can technically touch anything on the network; without security and governance policies as the framework, exposure like HR documents to the wrong employee becomes possible. Governance comes before laying AI on top.
Hidden costs bite. Cloud egress fees frequently surprise customers who never budgeted for them — a reason to keep data close to processing or on-prem. With average CPU/RAM utilization around 30–40%, most systems are oversized; in pay-for-consumption clouds, oversubscription is money thrown away. Gerbin recommends a monitor-resize flywheel, and notes AI solutions remain bespoke — cookie-cutter approaches are still about five years away.
Advice for CIOs. Talk to a peer in your industry who has already gone down the AI path — they know the rough edges. And expect AI to eventually recede into the background: like IoT in 2015, it will become magic that just works.
Read more: theCUBE
