The CTO Advisor

Calvin Hendryx-Parker, co-founder and CTO of Six Feet Up, is back with Keith Townsend to ask whether the AI bubble is bursting and who is left holding the bag. They start with Oracle's Stargate commitments and Keith's objection to funding that build-out with resources from its database and ERP customers, then get to the constraint underneath it all: power. Keith explains how colocation contracts went from floor space to power, why a rack now holds only a few AI servers, and what that means for ordinary enterprise clusters. From there, supplier continuity when your GPU provider is carrying heavy debt or gets acquired, why software teams abstracted from the hardware should still care, and what happened when Keith asked three clouds for GPU access for his labs. Keith closes with a warning about sticky services like CoreWeave's Weights & Biases tooling: plan for hybrid workflows, not just hybrid cloud. Calvin wraps with Six Feet Up's news and the XPRIZE Visioneering event.

Calvin Hendryx-Parker, co-founder and CTO of Six Feet Up, joins Keith Townsend to ask a question CTOs should be asking now: who gave your AI agent that authority? They start with two recent stories, the reported AWS AgentCore case where an agent fetched IAM session credentials from the metadata server on a user's behalf, and a Codex CLI update that leaves a self-updating daemon running after you quit, restarting itself mid-turn when a new version lands. From there, enterprise persistent memory and which principal an agent acts as when it changes how your business works, Slack bots invited into federated channels, and who is accountable when an agent chases a goal in a way nobody instructed. Calvin's practical advice is granular, explicit roles on the containers agents run in, using the Paperclip meta harness with a Kubernetes agent operator. Keith closes with four governance tests: authorization boundaries, state exportability, behavior traceability, and memory independence.

Rob Lee, Chief Technology and Growth Officer at Everpure (formerly Pure Storage), is back on the show. Keith Townsend asks how a CTO tells a real compute shortage from a data problem dressed up as one. Rob's answer: stop chasing GPU utilization. Measure how long a task takes and how accurate the result is, and utilization follows. From there, data gravity and why agents working inside transactional flows pull compute toward the data, why AI pushes enterprises to reason across every silo at once instead of one application stack like SAP, and KV cache offload as the first layer of a new memory hierarchy for inference. Rob closes with what Everpure has shipped and what's still on the roadmap.

Episode 700, and a reboot. Keith Townsend welcomes new co-host Calvin Hendryx-Parker, CTO and co-founder of Six Feet Up, pairing an infrastructure view and a development view on the same CTO problems. First up: Jev, TypeSafe AI's System One classifier that returns typed answers in milliseconds for a fraction of a language model's cost, and what it means to hand a closed-weight third party 200,000 rows of your data. Then Qwen3.8 running locally, and the open-weight classifiers chasing Jev. Next week: local versus cloud inference.

A replay of Layer2C Labs episode 001 on The CTO Advisor feed. The managed data plane is cheap and fast; the cost it hides is the chunking judgment it takes from you. Keith Townsend built a retrieval pipeline over thirteen years of Tech Field Day transcripts, about 237,000 segments, split between Amazon S3 Vectors and an NVIDIA DGX Spark on his desk, and traced where the decisions actually landed. Written by Keith Townsend and narrated by AI — the voice isn’t his and isn’t pretending to be. The full lab, with every number: https://labs.layer2c.com/labs/spark-s3vectors