AI factories are the new railroads: IBM, Marvell and the inference shift

Key points

Key takeaways from Source theCUBE Pod episode 139 on AI infrastructure, fresh from IBM's analyst council and Marvell Technology's investor day (October 2026):

Data centres are intelligence factories. Furrier argues the industry is misreading the build-out: like railroads and early aviation, AI data centres are infrastructure whose economic value shows up in what runs on them — the intelligence being produced — not in sentiment about valuations.

Training to inference flip. theCUBE cites CoreWeave workload data: a year ago roughly 90% of workloads were training; today it is about 60/40 training versus inference, with expectations that the mix flips to around 10/90 as inference becomes the monetisation layer.

Bubble fears questioned. Furrier and Vellante push back on Ray Dalio's bubble warning, pointing to monetisation at Nvidia and GPU suppliers and productive utilisation at customers such as MasterClass building new paid services on CoreWeave capacity.

IBM's window. IBM's opportunity is the AI operating model — the AI factory concept — plus AI trust, with the mainframe repositioned as a supercomputer-class transaction processor and Confluent's streaming made an always-on system, which Vellante calls a home run.

A Red Hat moment in AI? Vellante argues the Linux/Red Hat playbook — commoditising the underlying layer and getting paid to manage the open abstraction — could repeat in AI, with IBM well placed to run agents trustfully across Nvidia, AMD and other silicon.

Marvell's aggressive push. Marvell's shares jumped about 15 points on its revised numbers at investor day, with custom silicon for hyperscalers and connectivity becoming critical as AI infrastructure scales beyond racks to entire data centres.

Where the alpha accrues. The durable edge will sit in proprietary data, streamlined processes and the operating model — not in models or horizontal software — which is why IBM's pitch resonated with the hosts.

The AI freakout moment. Arvind Krishna's line that the car's speed limit was 4 mph when it launched frames the current panic: every transformative technology has a freakout moment, and the reality inside the industry is more mundane and more constructive.

Read more: theCUBE on YouTube