IBM stacked nearly 100 billion transistors in a chip the size of a fingernail — and it could reshape the next decade of AI computing. On June 25, 2026, IBM unveiled the world's first sub-1 nanometer chip technology, a 0.7 nm (7 angstrom) node built with a revolutionary architecture called "nanostack" that vertically stacks and staggers transistors in three dimensions. The result: up to 50% more performance or 70% greater energy efficiency than IBM's own 2 nm chips from 2021, with nearly double the transistor density.
Key Facts Most People Don't Know
- IBM's 2nm nanosheet transistor announced May 2021 contains gate-all-around channels only 5 nanometers wide, wrapped by metal gates on all four sides unlike FinFET's three-sided design
- The nanosheet stack uses 3 vertically stacked silicon channels per transistor, each 5nm thick with 12nm spacing, achieving 45% better performance than 7nm FinFETs at identical power
- IBM's Albany NanoTech facility in New York spent $3 billion developing extreme ultraviolet lithography tools operating at 13.5nm wavelength to pattern sub-2nm features
This isn't just another incremental shrink. The sub-1 nm node pushes semiconductor manufacturing into what IBM calls the "angstrom era" — where transistor dimensions approach the size of individual atoms. And the technology that makes it possible, nanostack, is the industry's first known three-dimensional nanosheet-based design, a structural leap that goes beyond simply making features smaller.
Why Sub-1 nm Matters More Than You Think
For years, the semiconductor industry has been racing against Moore's Law's perceived end. As transistors shrank below 7 nm, 5 nm, and then 3 nm, the physics got harder. Gate lengths reached just 12 nanometers — only 40 to 50 silicon atoms across their width. At that scale, electrons start tunneling through barriers they shouldn't be able to cross, and heat becomes an existential problem.
IBM's answer isn't to keep squeezing planar transistors. Instead, nanostack vertically stacks transistors and staggers them, taking advantage of 3D sequential integration. This is fundamentally different from existing 3D packaging approaches like chiplets or stacked memory — nanostack builds multiple transistor layers within the same die, with each layer able to use different material combinations optimized independently.
The payoff for AI workloads is enormous. The new architecture provides 40% scaling in SRAM, according to research presented at VLSI 2026 by Chen Zhang et al. That's critical because SRAM bandwidth has become the bottleneck in AI accelerator design. More efficient SRAM means more cache per core, fewer trips to off-chip memory, and faster inference for large language models.
How IBM's Nanostack Architecture Actually Works
Building a sub-1 nm chip isn't a single breakthrough — it's a chain of eight process steps, each of which must work at atomic-scale precision. Here's how IBM does it:
Step 1: Layer the Foundation — 5nm Silicon Sandwiches
The process begins by depositing alternating layers of silicon and silicon-germanium (SiGe) on a substrate. Six layers total, with each silicon layer exactly 5 nanometers thick. The SiGe layers act as sacrificial material — they'll be removed later to create the suspended nanosheet channels. Getting these layers uniform at 5 nm thickness requires molecular beam epitaxy with sub-angstrom control.
Step 2: Pattern with Light 100x Shorter Than the Features
Next, extreme ultraviolet (EUV) photolithography at 13.5 nm wavelength patterns the vertical fin structures, etching through all six alternating layers simultaneously. IBM's Albany NanoTech facility spent $3 billion developing these EUV tools. The upcoming High NA EUV lithography system from ASML — soon to be installed at the Albany site — will push resolution even further for future nodes.
Step 3: Suspend the Nanosheets with Chemistry
The silicon-germanium sacrificial layers are selectively etched away using hot ammonia gas at 600°C. This leaves the silicon nanosheets suspended in air with 12 nm vertical gaps between them. The selectivity of this etch is remarkable — it removes SiGe completely while leaving silicon untouched, a process that took years to perfect.
Step 4: Wrap the Gate Around All Four Sides
A 2 nm thick layer of high-k dielectric hafnium oxide is deposited around all four sides of each suspended nanosheet using atomic layer deposition (ALD), growing at just 1 angstrom per cycle. This gate-all-around (GAA) structure is what distinguishes nanosheet transistors from the older FinFET design — FinFETs wrap the gate around only three sides of the channel.
"IBM's 2nm nanosheet transistor announced May 2021 contains gate-all-around channels only 5 nanometers wide, wrapped by metal gates on all four sides unlike FinFET's three-sided design"
Step 5: Fill with Metal to Complete the Gate
Titanium-nitride metal gate material is deposited via chemical vapor deposition, filling the gaps between nanosheets and wrapping completely around each one. The result is a true gate-all-around structure where the gate electrode completely encircles each silicon channel, providing maximum electrostatic control over current flow.
Step 6: Dope the Source and Drain
Phosphorus ions (for n-type) or boron ions (for p-type) are implanted into the source and drain regions at 5 keV energy, creating junctions with 3 nm abruptness. At these energies, the ions penetrate only a few nanometers into the silicon, and the junction sharpness determines how well the transistor switches off — critical for power efficiency at sub-1 nm.
Step 7: Grow Crystalline Contacts
Epitaxial silicon-phosphorus or silicon-germanium-boron is grown on the source/drain regions to reduce contact resistance below 50 ohm-nanometers. Contact resistance is a major limiter at these scales — as transistors shrink, the contact area shrinks too, and resistance goes up unless you engineer the crystal structure to provide a clean interface.
Step 8: Wire It All Together with 15 Layers of Copper
Fifteen layers of copper interconnects are deposited using a dual-damascene process with ruthenium barrier layers only 0.5 nm thick. These barrier layers prevent copper from diffusing into the silicon while minimizing the resistance penalty in sub-20 nm wire widths. The interconnect stack is what connects billions of transistors into functional circuits — and at these dimensions, even the wires are engineered at near-atomic scale.
The 3D Stacking Breakthrough: Not Just Smaller, Taller
What makes nanostack different from previous nanosheet technology — which IBM itself invented — is the vertical stacking and staggering of transistors. Rather than placing all transistors in a single plane, nanostack builds transistor layers on top of each other using 3D sequential integration. This is validated experimentally through ultra-thin dielectric bonding in CMOS integration, dual-channel engineering capability, and functional CMOS inverter operation with expected switching performance.
The key advantage: each stacked layer can use different materials. One layer might be optimized for n-type transistors with high electron mobility, while the layer above it uses different materials for p-type transistors with high hole mobility. In traditional planar designs, both transistor types share the same substrate and material constraints. Nanostack breaks that limitation entirely.
The IBM-Albany Ecosystem and What Comes Next
IBM doesn't build these chips alone. The work happens at the Albany NanoTech Complex in New York, with partners including Lam Research, Tokyo Electron (TEL), and SCREEN Semiconductor Solutions. Together, they've already developed High NA EUV processes yielding working devices — a critical milestone for a technology that ASML only recently began shipping.
IBM's semiconductor roadmap projects at least a decade of future scaling from nanostack, with the earliest adoption at the sub-1 nm node expected in as few as five years. That timeline aligns with the industry's broader angstrom-era roadmap, where TSMC and Samsung are also pushing toward A10 and A7 nodes.
And IBM recently announced Anderon, the world's first pure-play quantum foundry, which will draw on IBM's semiconductor expertise to position the United States to manufacture most of the world's quantum wafers — a signal that the Albany facility's ambitions extend far beyond classical computing.
What This Means for AI and Cloud Infrastructure
The AI industry's biggest constraint isn't algorithmic — it's physical. Training frontier models requires enormous compute, and the energy cost of running inference at scale is becoming a limiting factor. IBM's sub-1 nm technology promises up to 70% greater energy efficiency at the chip level, which translates directly to lower power consumption per inference and higher throughput per watt.
The 40% SRAM scaling improvement is equally significant for AI. Larger on-chip caches mean fewer memory accesses, and memory bandwidth — not raw compute — is the primary bottleneck in LLM inference. Chips built on nanostack technology could deliver meaningfully faster inference for the same power budget, or the same performance at dramatically lower cost.
Samsung already licensed IBM's earlier nanosheet technology for $500 million back in 2018, implementing it in their 3 nm GAA-FET process that shipped in the Snapdragon 8 Gen 2. If nanostack follows a similar licensing path, the technology could reach consumer devices within the next five to seven years.
But quantum tunneling at 1 nm makes electrons behave like ghosts — how do engineers trap them next?
Top comments (0)