Old-school tech kit is starting to look less knackered as AI agents start poking around in business computing.
Judging by the recent bounce in stocks that were previously out of favour, it may be time to relearn some older tech terminology.
These are the technologies that ruled before generative AI barged in: servers running central processing units for general-purpose computing, storage, and other components in the old tech “stack,” along with the software running on them.
That lot has been shoved into the shadows during the ChatGPT era. Yet as AI digs deeper into business, the pendulum is showing signs of swinging partly back towards Old IT.
That has left the cocaine nose jobs of Wall Street trying to work out how much IT spending these older technologies can still grab.
The answer depends on how central they become to AI-dominated computing workloads, how much value Old IT suppliers can claim and whether their ageing business models still work.
One reason for the shift is demand moving from training large language models to running queries against them, known as inference. CPUs play an important role in managing inference, such as loading an AI model and handling its output.
The latest evidence came from surprisingly strong CPU demand reported this week by ARM and AMD, which gave both companies’ shares another shove upwards.
AMD said it now expects compound growth of 35 per cent from CPU sales during the next few years, roughly twice its forecast from six months ago. Old IT could get a bigger boost from wider use of AI agents.
As agents take on more work previously done by humans, Nvidia chief executive Jensen Huang claims they will use the same digital tools.
That means the online software applications that already do much of the donkey work in business.
Intel’s chief financial officer David A. Zinsner told investors last month that AI training needs only one CPU for every eight GPUs. For agent workloads, the Intel finance boss said they will need as many CPUs as GPUs.
That hope has helped fuel the daft 400 per cent bounce in Chipzilla’s shares since the US government took a stake last summer.
The excitement is spreading to other old-line hardware names. Shares in disk drive maker Seagate have jumped 60 per cent during the past month.
The key question is how value from agent-led computing workloads gets carved up. If companies involved in Old IT are left playing a supporting role in the background, they may struggle.
If they can present their technology as central to the agent-led era, they are more likely to keep pricing power. This is a particular problem for the most battered tech sector of the lot: software.
Last month, Salesforce became the first big company to offer its software without a browser-based interface for human workers, known as “headless” software.
That is an important step towards opening its systems to AI agents. It carries a nasty risk because Salesforce, like most cloud software outfits, has long depended on charging by the number of people using its service.
Salesforce is betting that its data advantage will keep it essential. That comes from years of information stored in its systems by customers.
The company claims this will let it understand the context of business activity and orchestrate battalions of AI agents. That claim is about to get a kicking in the real world.
AI companies lack historical data, but by working alongside humans, they may develop a sharper sense of how businesses run and how decisions are made. Salesforce is betting that probabilistic AI models alone will never deliver the guaranteed performance companies demand.
This points to an important tech battleground: how raw AI horsepower gets turned into something useful. The cocaine nose jobs of Wall Street are not yet ready to give companies like Salesforce the benefit of the doubt. As a new software supply chain around AI takes shape, that scepticism may start to wobble.
