A Peking University team has built a transistor with no silicon in it, and the university’s own summary is not shy: “It is the fastest, most efficient transistor ever.” The claim, published in Nature Materials, is that this two-dimensional device runs 40% faster than Intel’s latest 3nm chips while using 10% less power — numbers that would put it ahead of what TSMC and Samsung currently ship. The tests were run under the same conditions used for leading commercial chips. So the figure is not a demo-day fantasy.
And it still does not kill silicon. Here is my thesis, and I will defend it to the last paragraph: a faster transistor is not a faster chip. Until someone prints a billion of these things with a yield curve attached, Intel’s roadmap is not dead — it has simply been handed a deadline.
What they actually built
The architecture is gate-all-around, GAAFET, where the gate wraps completely around the source instead of covering only part of it as in the FinFET designs that dominate current silicon processors. Full wrap means more contact area between gate and channel, less energy leakage, better current control. Instead of the vertical structures of FinFETs, the new design resembles interwoven bridges. Two new bismuth-based materials do the work: Bi₂O₂Se as the semiconductor and Bi₂SeO₅ as the gate dielectric. Low interface energy, fewer defects, less electron scattering — lead scientist Professor Peng Hailin’s description is water through a smooth pipe.
What exists so far are small logic units. Logic units, not processors. That distinction is the entire gap between a paper and a product.
The sentence that actually matters
The interesting line in this work is not the 40%. It is Peng’s framing: if chip innovations based on existing materials are a “shortcut,” then 2D material-based transistors are “changing lanes.” That is a confession wrapped in a metaphor. Silicon is running into miniaturization limits as the industry pushes below the 3nm threshold, and this team is betting a different lane exists.
The second sentence that matters is the claim that these transistors can be manufactured using current semiconductor infrastructure. That is the only commercially load-bearing statement in the whole announcement, because fabrication compatibility is where promising materials usually go to die. If it holds, the gap between a PKU lab and a production line shrinks dramatically. If it is aspiration, we are reading a very good press release.
The scoreboard nobody is comparing
The comparison on offer is transistor against transistor: Peking University’s device versus Intel’s 3nm parts, with TSMC and Samsung named as the ones being leapfrogged. Backed by density functional theory calculations and validated through physical tests on a high-precision fabrication platform at PKU, the result is serious work.
But the comparison that decides an industry is wafer against wafer, yield against yield, customer against customer. Intel, TSMC and Samsung ship at volume. One lab in Beijing has built small logic units and published in Nature Materials. Both facts are true at once, and only one of them gets a headline with the word “finished” in it.
So no — silicon processors are not officially finished. Officially, silicon remains the incumbent with the fabs, the tooling and the order book, and this paper is a lane change signalled by someone who has not merged yet.
What I would watch next
Two things, and neither is the 40%. First, whether anyone outside PKU reproduces the speed and power figures under their own conditions. Second, whether “current semiconductor infrastructure” survives contact with an actual production line, because that claim carries more weight than any benchmark in the paper.
Silicon has been declared dead more often than it has been replaced. A bismuth compound, a full-wrap gate and a 40% number are a warning shot across the bow, not a coffin lid. Believe the lane change — the moment the yield data lands. Not one day earlier.