Former Intel chief executive Pat Gelsinger says the AI investment cycle could last more than two decades, and argues that the industry’s next bottleneck may be memory rather than raw GPU performance. A newly emerged startup, Nagare Labs, is developing an ambitious memory design to take on that challenge, though mass production remains years away.
Watch Desk analysis
What happened
Speaking at a technology summit in Taipei, Gelsinger described AI as a “two-decade-plus” cycle of investment and expansion, while allowing that it will include corrections and disruption. He is now a general partner at Playground Global, which is bringing six portfolio startups to Taiwan to seek partners in areas including memory, power management, optical technology and advanced packaging.
One of those companies is Nagare Labs, which emerged from stealth on Monday. It is developing Z2RAM, an oxide-semiconductor memory architecture intended to combine SRAM-like speed, DRAM-like density and flash-like scalability. Cofounder and interim CEO Suman Datta said a new memory technology typically takes six to seven years to reach high-volume production, and gave no specific commercialisation date. The details and claims were reported by the Taipei Times in its account of the summit.
Why it matters
AI systems need memory not just to store information but to feed processors quickly enough to keep them busy. If memory capacity and bandwidth become a bigger constraint as inference expands, advances beyond the familiar GPU race could matter to the cost and performance of AI infrastructure.
Taiwan is a natural place to watch that effort: Gelsinger said it accounts for about 90% of global AI server production. But a promising design is not a shipping product, and a six-to-seven-year path to high-volume production is a long way from solving today’s bottlenecks.
Our read
The most interesting part is the shift in emphasis from faster processors to the memory feeding them. Nagare’s proposed combination of speed, density and scalability would be a substantial achievement if it works as intended. For now, that is the company’s ambition, not a demonstrated result. Put the idea on the watchlist, but keep the champagne in the cupboard until there is independent technical evidence and a credible route to production.
What to watch
- Whether Nagare publishes technical results or more detail about Z2RAM.
- Whether it secures Taiwanese foundry or ecosystem partners.
- How its commercialisation timeline develops beyond the typical six-to-seven-year estimate.
- Whether AI inference makes memory capacity and bandwidth a more visible industry constraint.
Discussion spark: If memory becomes the main constraint on AI, should the industry put more effort into new memory architectures, or keep prioritising faster processors?
Sources and evidence
- AI cycle to last over 20 years: ex-Intel CEO – Taipei Times (27 October 2026, 16:00 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.