← Back to feed

IBM announced sub-1nm chip technology on June 25, 2026, from its research campus in Yorktown Heights, New York. The architecture, called nanostack, stacks transistors vertically in three dimensions rather than continuing to shrink conventional 2D planar geometry. A chip built on nanostack packs nearly 100 billion transistors into an area the size of a human fingernail — roughly twice the density of IBM’s previous-generation process.

IBM is positioning the technology as the semiconductor industry’s path through what researchers have described as the practical floor of 2D silicon scaling. The stated target market is AI compute.

What Nanostack Does Differently

Conventional transistor scaling works by shrinking the gate — the switch that controls current flow — and placing more of them across the flat surface of a silicon wafer. That approach has been the driver of Moore’s Law since the 1960s, but lithography physics place an increasingly hard limit on how small a planar gate can be made.

Nanostack attacks the problem vertically. Rather than laying transistors flat, the architecture stacks them in a 3D lattice, allowing density to increase without requiring a narrower gate length. The result is a sub-1nm effective node — meaning the transistor pitch, the center-to-center spacing between gates, has broken below 1 nanometer — while maintaining leakage characteristics comparable to current-generation 2nm silicon.

The comparison point matters. TSMC’s N2 node, which entered production in 2025, achieves approximately 20 billion transistors per square millimeter. IBM has not released an equivalent mm² density figure for nanostack, but the 2x claim over its own previous generation suggests a node competitive with or ahead of N2.

Why AI Is the Driver

Frontier model training scales have been doubling roughly every 6-12 months. At the current trajectory, training runs for next-generation models — 5 trillion to 10 trillion parameters — require chips that deliver substantially more on-chip memory bandwidth within the same thermal envelope.

Nanostack’s 3D stacking approach directly benefits transformer inference. Long-context attention operations, which dominate modern workloads at context windows above 100K tokens, are memory-bandwidth-bound. Packing 100 billion transistors per chip means more SRAM, more cache, and faster memory access without enlarging the die — the combination that matters most for inference accelerators.

IBM is treating nanostack as a foundry architecture, not an IBM-manufactured product. The technology is designed to be licensed to chipmakers producing IBM-designed silicon for government and enterprise AI programs. That narrows near-term volume, but it positions IBM as a royalty-bearing IP holder in whatever custom AI silicon generation emerges after the current Blackwell/TPU cycle.

The Roadmap Context

IBM’s 2nm node was demonstrated in 2021 and reached early commercial use by 2025 — a roughly four-year runway from lab to production. If nanostack follows a similar cadence, sub-1nm AI chips would reach commercial deployment around 2030.

That timeline aligns with two pressures already visible in the market. TSMC has published its own roadmap toward 1.4nm (A14) by 2027-2028 using a different approach called complementary FET (CFET). Huawei has claimed 1.4nm chip density by 2031. IBM reaching sub-1nm today with nanostack represents a theoretical benchmark win, though production capacity at that node remains a different problem.

The race matters because whichever fabrication architecture reaches volume production first at sub-1nm density will anchor the AI chip supply chain for the late 2020s and into the 2030s. IBM, TSMC, and implicitly Intel — which is running 18A trials for its Feynman GPU — are all building toward the same inflection point.

Key Numbers

Metric	Value
Process node	Sub-1nm
Architecture	Nanostack (3D stacked transistors)
Transistors per chip	~100 billion
Density vs prior gen	~2x
Announced	June 25, 2026
Commercial target	~2030 (estimated)
Primary market	AI compute accelerators