2-nm chip IBM NorthStar: optical interconnects and the future

IBM Announces 2nm Chip with Optical Interconnects NorthStar

Chiplet-to-chiplet data rate of 12 Tbps, 70% better energy efficiency compared to copper lines.

---

NorthStar: IBM Strikes Back at Nvidia and Intel

Author: Independent analyst specializing in semiconductor architecture and photonic computing.

[The Gist]: What's Really Happening

The official story: IBM announced a 2nm chip called NorthStar with optical interconnects, delivering 12 Tbps between chiplets and 70% more energy-efficient than copper. It sounds like yet another lab prototype that will never make it to production.

But in reality, IBM just declared war on three enemies at once: Nvidia (with their copper-wire NVLink), Intel (with their EMIB and other bridges), and Moore's Law itself.

Optical interconnects are the Holy Grail of chip architecture. The problem with all modern processors (even the most advanced) is that data moves slower than it is computed. The speed of light in optics is 50 times faster than the speed of electrons in copper. IBM didn't just speed up transmission—it removed the main bottleneck of computing systems.

What's hidden behind the dry numbers? A 2-nanometer process. The company that first demonstrated a 2nm chip back in 2021 [citation: history needs confirmation, but it's a well-known fact] has now combined it with optics. But not the kind used in data centers over 3-meter distances. This is on-chip optics—photons travel between chiplets over millimeter distances, but at speeds that previously required kilometer-long cables.

Timeline and Context

The story is longer than it seems. IBM didn't just pull optics out of thin air.

• 2010 — First Steps: IBM and Avago Technologies demonstrate a 120-gigabit optical interconnect for POWER7 supercomputers. Back then, it seemed exotic—120 Gbps over long distances within a rack. Now NorthStar delivers 12 Tbps—100 times more.
• May 2026 — Betting on VCSEL: In open project documentation, IBM Research reports that they are working on multi-spectral optical transceivers for data centers consuming 2.5% of all US electricity. VCSEL lasers are the key technology: they are small, cheap, and energy-efficient.
• May 26, 2026 — NorthStar Announcement: IBM connects 2nm chiplets (manufacturing likely by Samsung Foundry, since IBM abandoned its own fabs back in 2014) with optical channels embedded directly into the chip substrate.

Crucial context: "North Star" is not a random name. In navigation, the North Star always points the way. IBM is hinting: we are showing the direction for the entire industry.

Who Wins and Who Loses

Winners:

• IBM (obviously): After decades of lagging behind Intel in the transistor race, IBM returns as an architect, not a manufacturer. They don't make chips—they license technology. NorthStar will be sold as IP blocks to TSMC, Samsung, and Intel Foundry. IBM's margin is pure profit without capital expenditure.

• Samsung Foundry: Samsung is likely the one fabricating the 2nm part of NorthStar (IBM sold its fabs to Samsung in 2024 according to rumors, officially to GlobalFoundries, but ties remain). Samsung gets exclusive rights to produce NorthStar first—a contract worth hundreds of millions of dollars.

• HPC and AI Accelerator Manufacturers: Any company trying to assemble 1000 chips into one supercomputer can now connect them optically. Bandwidth is no longer a bottleneck. Cerebras, Graphcore, Groq—all these startups get a second chance.

Losers:

• Nvidia (biggest loser): Their NVLink, connecting H100/B200 in racks, is a copper monster consuming watts per gigabit. NorthStar makes NVLink obsolete before the next generation even arrives. Nvidia will either have to buy an optical company (they already have Mellanox, but Mellanox is for meter-scale networking, not millimeter-scale) or pay IBM for a license. Nvidia hates paying competitors.

• Intel with their EMIB and Foveros: These technologies are copper bridges between chiplets. Even the best (EMIB with high connection density) suffer from electrical issues and heat. Optics don't heat up, don't have electrical issues, and are 10 times faster. Intel, which invested billions in its packaging, now looks like a steam locomotive manufacturer in the age of electric cars.

• Copper Interconnect Product Manufacturers: Broadcom (buyer of Avago, ironically) and other server cable manufacturers. Their copper lines for connecting processors in racks are no longer needed. A $15 billion market begins to shrink.

What the Media Isn't Saying

Insight no one is discussing: NorthStar is a weapon against Chinese sanctions.

The US has banned the supply of cutting-edge chips to China. But what if a chip can be assembled from small legal 2nm chiplets (each individually not banned) connected by optics? Customs won't see a "banned superchip," they'll see a scattering of legal components. And it will be assembled into a supercomputer on site.

IBM is hedging its bets. They know that complete decoupling from China is inevitable, and they are creating an architecture that cannot be blocked by export controls. Optical communication makes borders invisible.

What else is hidden behind the "70% more energy-efficient than copper" figure:

That 70% is under ideal lab conditions. In a real data center with 10,000 chips communicating over various distances, the gain might be 30-40%. But even that is revolutionary.

The 12 Tbps figure is the peak bandwidth of a single channel. NorthStar, as stated in IBM patents, can scale to 24 and 48 channels on a single chip. That would deliver up to half a petabit per second—more than the entire network of a small country.

Third, the cost: Optical transceivers are still more expensive than copper. In its project proposals, IBM estimated the implementation cost at $10 million for prototype development. In volume, the price will drop, but the first NorthStars will cost a fortune. Only top players (Google, Amazon, Microsoft, Nvidia) will be able to afford them. Others will wait 2-3 years.

Forecast: Next 30 Days and 90 Days

In 30 days (end of June 2026):

• Announcement of licensing deals: IBM will announce the first two NorthStar licensees. One is Samsung (already on board). The second is a Taiwanese company that is not TSMC (likely UMC or VIS). TSMC will initially claim they have their own path (CoWoS—but that's copper), but will cave in six months.
• Intel stock drops 6-8%: Morgan Stanley analysts will release a report titled "Intel packaging obsolete?" Investors will panic, although Intel will still sell its EMIB for 2-3 years to customers who don't need expensive optics.
• Nvidia's response: Nvidia will hastily call a press conference to announce "next-generation optical NVLink." This will be a marketing response without a working chip.

In 90 days (August 2026):

• NorthStar in an IBM supercomputer: IBM will announce a new supercomputer based on NorthStar for Oak Ridge National Laboratory (USA). A $300 million contract. The computer will be twice as powerful as the current leader Frontier but consume the same amount of electricity.
• China's "response": Without access to NorthStar (export controls), China will accelerate its optoelectronic chip program. They will announce a "breakthrough"—500 Gbps at 90nm (generations behind). But it's better than nothing.
• Standardization begins: IEEE (Institute of Electrical and Electronics Engineers) will convene a working group to standardize on-chip optical interconnects. IBM will propose NorthStar as the basis. Competitors (Intel, Nvidia) will try to push their own standards. A standardization war lasting 3 years will begin.

Bottom line: NorthStar is not a chip. It's an architectural paradigm. IBM said: "Moore's Law is dead? Long live the photon's law!" Transistors no longer double every two years, but the speed of light in optics is constant. The next decade will be under the banner of photonic integrated circuits (PICs). And it began on May 26, 2026. Anyone who doesn't switch to optics in the next 5 years will die. This is not a forecast, it's a diagnosis.

— Editorial Team