Programmable Terahertz Metasurface for Optical Computing Created
The development enables high-precision amplitude modulation and optical logic operations, opening new possibilities for next-generation communication systems and dynamically tunable optical computers.
Programmable Terahertz Metasurface: How One Chip Replaces an Entire Optics Assembly in Your Data Center
The Gist: What's Really Happening
On May 6, 2026, a paper published in the journal Light: Science & Applications described in dry academic language a "sub-matrix programmable metasurface." Behind this term lies a breakthrough that redefines the boundary between optics and computing architecture: researchers have created a chip capable of performing Boolean algebra (AND, OR, XNOR) directly in the terahertz wavefront—without optoelectronic conversion, without a DSP chip in between, without a software stack. Light goes in—a computed answer comes out.
This is not just another "step toward 6G." It is a direct incursion of programmable metasurfaces into territory monopolized by specialized integrated circuits. AlGaN/GaN HEMT transistors are integrated directly into the meta-atoms of the surface, modulating the two-dimensional electron gas density and controlling the amplitude of transmitted radiation in the 170–260 GHz range. The result: the same surface switches in real time between the role of a PAM-4 modulator (four amplitude levels, up to 6 GHz carrier) and an optical logic gate operating at 200 MHz.
The key architectural decision is the rejection of per-pixel addressing in favor of sub-matrix addressing. Instead of driving each of thousands of meta-atoms individually, the researchers grouped them into four independently controlled sub-matrices. This yielded gains in routing complexity and timing constraints while preserving spatial combinatorial code for logic and multi-level modulation.
Timeline and Context
Over the past five years, terahertz metasurfaces have evolved along two dead-end trajectories. The first is per-pixel control on CMOS arrays, which produces beautiful holograms but chokes on routing complexity when scaling up. The second is global control of the entire aperture, which is cheap to implement but supports only binary OOK modulation. Neither approach met the practical requirements of 6G systems, which need simultaneous support for sensing, communication, and computing.
The race for the terahertz band is a race for spectral resources for post-5G networks. The European TERRAMETA project has invested millions of euros in developing reconfigurable intelligent surfaces for terahertz communications. Meanwhile, Chinese groups are pushing alternative approaches: in January 2026, the team of Zhang and Zhang demonstrated a graphene metasurface with fivefold rotational symmetry for independent amplitude and phase control without crosstalk. In early May 2026, Wu et al. proposed a geometric-phase approach with simultaneous amplitude-phase modulation for reconfigurable intelligent surfaces.
But the work of Wang, Gong, and Xia is fundamentally different: it does not optimize modulation or phase but embeds logic operations directly into the physical layer of the metasurface. This is a shift from the "metasurface as antenna" paradigm to the "metasurface as coprocessor" paradigm.
The technical foundation is AlGaN/GaN HEMT transistors on a SiC substrate with carrier mobility above 2200 cm²/V·s and 2DEG concentration over 10¹³ cm⁻². At zero bias, the transistor is on, and the surface is in a low-transmission state. When a negative bias is applied, the depletion region interrupts the channel, suppressing the collective resonance and sharply increasing transparency. The transition between states takes about 10 ns.
Who Wins and Who Loses
There are two classes of winners.
6G Infrastructure Vendors. Manufacturers of reconfigurable intelligent surfaces get a "function killer": the same metasurface simultaneously serves as a modulator, a physical-layer logic element, and a beam-steering interface. Cellular operators investing billions in 6G licensing will see this as a way to shorten the equipment chain: one chip instead of the "antenna + modulator + DSP" bundle.
AlGaN/GaN Heterostructure Manufacturers. Wolfspeed, Infineon, and NXP have invested years in GaN-on-SiC for power electronics and radar. Now their technology stack becomes the foundation for computational metasurfaces. Every new metasurface design creates additional demand for HEMT wafers with high-quality AlGaN/GaN interfaces. For fabs that risked being left with excess GaN capacity after the EV market stabilizes, this is a new lease on life.
The losers are equally obvious.
Silicon Photonics in Short Links. Startups like Ayar Labs and Lightmatter, promising chip-to-chip optical interconnects via silicon photonics, now see an alternative: a metasurface that not only transmits a signal but performs logic operations on it in transit. If computation happens at the wave propagation stage, the argument for a dedicated optical transceiver weakens.
DSP Manufacturers for Terahertz Transceivers. The traditional architecture: antenna → LNA → mixer → ADC → DSP → reverse path. Each stage adds latency, power consumption, and cost. A metasurface performing PAM-4 modulation and logic "in the air" makes part of the DSP processing redundant.
Chinese Groups That Invested in the PB-Phase Approach. The work by Wu et al. on geometric phase for terahertz RIS, published just days before Wang's breakthrough, risks being overshadowed. The sub-matrix HEMT solution offers fundamentally higher functionality at comparable fabrication complexity.
What the Media Isn't Saying
Here is the main insight missing from 90% of publications.
The team of Wang, Gong, and Xia deliberately limited the design to four sub-matrices—2x2. This is not engineering caution but an architectural gambit. Four sub-matrices yield 16 amplitude states, sufficient for PAM-4 and a full set of three Boolean functions. But the authors explicitly state that the architecture scales to N×N. With 4×4 sub-matrices, PAM-8 and more complex logic functions become possible. With 8×8, a full-fledged optical ALU emerges.
Why is this more important than headlines suggest? Because a metasurface with 64 sub-matrices is a 6-bit optical coprocessor operating at frequencies above 200 MHz without clocking, without the heat dissipation of traditional logic, and with latency determined only by wave propagation time. For signal classification tasks, physical-layer security, and edge AI on the radio interface, this redefines the very concept of a "computing element."
Second non-obvious point: No publication discusses how the group solved the problem of nonlinear amplitude saturation. The paper honestly notes that as the number of active sub-matrices increases, amplitude does not sum linearly: growth slows due to collective coupling between sub-matrices. For PAM-4, this is addressed by post-distortion or a coefficient lookup table. But for PAM-8, these nonlinearities will become a critical barrier. The researchers are aware of this, pointing to digital predistortion (DPD) and heterogeneous integration as future directions. This means the next step is not scaling the metasurface but developing a dedicated nonlinear correction controller, which will become the main intellectual property asset.
Third insight: The simultaneous publication of several papers on terahertz metasurfaces in May 2026 is no coincidence. Research groups sense the approaching 6G standardization window and are rushing to establish architectural priority. The design that enters the 3GPP white paper on terahertz radio access will earn royalties from billions of devices. That is why the competition between Wang's HEMT approach, Zhang's graphene approach, and Wu's PB-phase approach is so fierce.
Forecast: Next 30 Days and 90 Days
30 Days (by June 9, 2026)
A patent race will begin. Wang's team has likely already filed provisional patents on the sub-matrix control architecture for HEMT metasurfaces performing logic operations. Within a month, patent publications will follow, triggering a wave of counter-filings from groups working with graphene and phase-change metasurfaces.
Simultaneously, negotiations with HEMT wafer manufacturers will intensify. AlGaN/GaN-on-SiC technology is well-established, but no commercial foundry offers a process optimized for metasurfaces with integrated HEMT meta-atoms. The first to announce such a PDK will gain a temporary monopoly in the emerging market for terahertz computational surfaces.
In academia, replication efforts will begin. Groups from Berkeley, MIT, and IMEC will attempt to reproduce the 200 MHz Boolean logic on their own setups. The success or failure of these attempts will determine the credibility of the approach.
90 Days (by August 9, 2026)
The key event will be an attempt to scale to 4×4 sub-matrices. If Wang's group demonstrates PAM-8 and an 8-bit set of logic operations, it will become a technology trigger for the industry. Expect an announcement of a startup licensing the architecture for commercialization.
Penetration into the 6G standardization agenda will begin. ITU and 3GPP hold working meetings in August on radio interfaces beyond 100 GHz. Wang's work is a ready-made case for including "optical logic surfaces" in the 6G roadmap as a candidate technology.
For the semiconductor foundry market, this is a moment of strategic choice. GaN-on-SiC is a mature technology, but production volumes are driven by demand for power electronics and radar. The emergence of computational metasurfaces as a new application domain could shift investment balance in favor of fabs with available GaN lines (Infineon, Wolfspeed) versus those focused exclusively on CMOS (TSMC, Samsung).
This chip is a watershed. Before it, metasurfaces were antennas: passive or slightly tunable. After it, they become computing elements capable of performing logic operations on a signal without conversion to the electrical domain. The difference is roughly like between a shoulder and a multi-lane highway. The industry will need years to digest the implications—but the direction is set. And those who today consider metasurfaces a niche academic toy will, in five years, be buying production licenses from today's graduate students.
— Editorial Team
No comments yet.