Can gallium arsenide replace silicon? From a working processor to the optical links that carry AI data

In 1993, engineers had a gallium arsenide processor on the test bench and a surprisingly hard question to answer. Its arithmetic unit and register file could run together at 137 MHz on the best specimen. A problem with the package connections kept the team from testing branch instructions at speed. What clock rate, then, could they promise someone who wanted to run a program? [1]

That question still matters to AI. An accelerator must fetch numbers, combine them and sometimes send a result to another machine before it can continue. Gallium arsenide (GaAs) has been used to build digital logic and is a material platform for the lasers that carry those results over fibre. Each application draws on a different property of the crystal. Which one makes a difference to the time or energy needed for a prediction?

A ceramic processor package beside a separate fibre-optic module plugged into a server card.
Two uses of gallium arsenide: digital logic in a processor and light emission inside an optical transmitter. Original conceptual illustration; the package and module represent classes of devices. © 2026 Juan Antonio Lloret Egea · ITVIA · The Bible of AI (BAI) · CC BY-NC-ND 4.0.

How far does electron mobility carry a processor?

Apply a weak electric field to a semiconductor and its mobile electrons acquire a mean drift velocity. Mobility tells us how much velocity the field produces. A Jet Propulsion Laboratory handbook tabulates electron drift mobilities at 300 K of about 8,500 cm²/(V·s) for GaAs and 1,500 for silicon. For holes, the positive carriers involved in the other half of a complementary logic circuit, the corresponding values are 400 and 450. A design that alternately charges and discharges a wire therefore inherits quite different starting conditions on its two transitions. [2]

At the same weak field the tabulated electron values imply a ratio of about 5.7. To use that as a ratio of processor speeds would require several more measurements. A gate must deliver enough current to charge the next gate and its wiring. Contacts add resistance, wires add capacitance and a short channel can operate well beyond the weak-field regime. In the Michigan experiment, a 32-bit processor containing 60,500 transistors put these circuit questions within reach. Even on a fabricated chip, the answer depended on which paths the team could exercise. [1] [2]

Where do the electrons actually travel?

At Fujitsu Laboratories in Japan, researchers demonstrated a high-electron-mobility transistor in 1980 by growing two different semiconductors together. Donor atoms in aluminium gallium arsenide supplied electrons that could move along the adjoining GaAs side of the interface. Separating the mobile charge from many of the charged donors reduced scattering. A gate controlled this thin conducting region. The same materials system later yielded a mobility of about 57 million cm²/(V·s) in a Princeton two-dimensional electron sample at 0.3 K. The cold sample answers a question about scattering under carefully controlled conditions. A room-temperature processor has a different thermal environment, geometry and electrical load. [3] [12]

A Purdue–AmberWave experiment made the material substitution quite literal. The researchers grew GaAs in patterned regions on a silicon substrate and formed a transistor with a GaAs channel and an aluminium oxide gate dielectric. For a 10 µm gate they estimated a peak surface mobility of about 500 cm²/(V·s) from the drain current. This figure belongs to carriers moving near the interface of that particular device. The experiment also shows how much silicon can remain when the channel material changes. [4]

The layer names matter. In Beijing, researchers measured a Hall mobility of 4,732 cm²/(V·s) and an effective mobility of 3,305 at room temperature for a technology using indium gallium arsenide (InGaAs) on silicon. In Korea, KAIST and the Korea Advanced Nano Fab Center investigated another arrangement: a GaAs substrate supported an InGaAs-channel transistor. Assigning both results to electrons travelling through pure GaAs would erase the feature that the teams designed. [6] [7]

Which part of the circuit can GaAs take over?

The Nanjing Electronic Devices Institute built a broadband switch with two jobs. A silicon CMOS circuit decoded the digital control signal. Above it, a bonded layer of GaAs-platform transistors switched the radio-frequency path. The authors reported a 15% smaller circuit area than their reference arrangement. That is an area result for the integrated switch. A designer assessing a radio front end would also measure how much signal is lost when the switch is on and how well the paths are isolated when it is off. [5]

Replacing the logic fabric of an AI accelerator would entail a wider comparison. Arithmetic units need dense local storage, reliable logic levels across many gates and enough interconnect to supply their operands each cycle. A silicon design can itself spend much of its time waiting for memory. The GaAs processor establishes that digital computation in the compound is possible. The reviewed experiments leave open how a contemporary GaAs accelerator would perform on the same trained model, at the same numerical accuracy, with the memory system included. [1] [4] [5]

The short optical link between two machines

On a short fibre link the useful output is a light pulse. A vertical-cavity surface-emitting laser, or VCSEL, sends light out through the surface of a semiconductor structure. Its active layers convert injected electrical carriers into light, and mirrors above and below them form a small optical cavity. This application takes advantage of light emission in direct-bandgap semiconductor layers. The electronic drift-mobility ratio alone cannot predict the transmitter’s data rate. Broadcom specifies a four-emitter, GaAs-based 850 nm VCSEL array for four-level signalling at 53 billion symbols per second. Four distinguishable levels encode two bits per symbol, giving 106 gigabits per second on a lane before allowing for coding overhead. [8]

Distance changes the result. At Chalmers in Sweden, an 850 nm VCSEL experiment reached 57 Gb/s with transmitter and receiver connected back to back, 55 Gb/s over 50 metres of fibre and 43 Gb/s over 100 metres, under the paper’s error criterion. A Changchun study published in the 2026 volume of the Journal of Semiconductors reported 104 Gb/s with four-level modulation over 100 metres of OM5 fibre, and 60 Gb/s with two-level NRZ signalling. Its eye diagrams show how clearly the receiver could distinguish the levels at 25 °C. The fibres, signalling and receivers belong to their respective experiments; a comparison of the lasers alone would require a matched test. [9] [10]

Consider two accelerators exchanging partial results. If the optical link is already idle while the machines compute, increasing its line rate buys little time. If they regularly wait for a transfer, a faster usable link can shorten that wait. An informative test would clock the same exchange at the same distance and error target, meter the transmitter, receiver and signal processing, then repeat the model run. Useful bits per joule and time per completed inference would connect the component to the performance of distributed AI.

Can the laser itself do some of the computing?

A French–German experiment gives a third answer. Researchers at FEMTO-ST and TU Berlin used different optical modes of a single semiconductor laser to transform an input pattern. Programmable micromirrors encoded inputs and selected how much of each part of the output reached a detector. The team trained the readout to perform tasks including two-bit exclusive OR: its output is one when precisely one input is one. The laser was grown on GaAs, while its light-generating quantum wells were InGaAs. Reproducing this operation used the optical setup around the chip, including an injection laser. Any energy comparison with a digital implementation must count those components. [11]

Of the three routes, emitters for short optical links have the clearest industrial footing in the evidence reviewed here. Digital GaAs has a working processor behind it, and optical neural processing has a measured learning experiment. The most informative next result for AI would put one of these devices into a complete system and show which wait, transfer or calculation it improves, at a specified accuracy and total energy. [1] [8] [11]

Sources and scope

Evidence reviewed on 17 November 2026. Material reference values, device experiments and a manufacturer’s component specification are identified separately. Data rates refer to the signalling conditions of each link. The processor result refers to the paths exercised on the test bench.

  1. Brown, R. B., et al. (1993). Gallium-arsenide process evaluation based on a RISC microprocessor example. IEEE Journal of Solid-State Circuits, §III and Table I.
  2. Anspaugh, B. E. (1996). GaAs Solar Cell Radiation Handbook. JPL Publication 96-9, Appendix A, Table A-2. Reference properties at 300 K.
  3. Mimura, T., et al. (1980). A new field-effect transistor with selectively doped GaAs/n-AlGaAs heterojunctions. Japanese Journal of Applied Physics, 19, L225–L227.
  4. Wu, Y. Q., et al. (2008). Atomic-layer-deposited Al₂O₃/GaAs metal-oxide-semiconductor field-effect transistor on Si substrate using aspect ratio trapping technique. Applied Physics Letters, 93, 242106.
  5. Wu, L.-S., et al. (2016). Heterogeneous integration of GaAs pHEMT and Si CMOS on the same chip. Chinese Physics B, 25, 067306.
  6. Li, Q.-C., et al. (2025). High-Mobility InGaAs HEMT Epitaxially Grown on Silicon. Chinese Physics Letters, 42, 120802.
  7. Jeong, J., Kim, J., and Kim, S. (2025). High-performance InGaAs HEMTs on GaAs substrate using AlAs/GaAs superlattice buffer-assisted heterostructure growth. Journal of Alloys and Compounds, 178699.
  8. Broadcom (28 February 2025). AFCD-V84LP, 106-Gb/s (53-Gbaud) 850-nm 1 × 4 Array Oxide VCSEL. Manufacturer data sheet.
  9. Westbergh, P., et al. (2013). High-speed 850 nm VCSELs operating error free up to 57 Gbit/s. Electronics Letters.
  10. Tian, S.-C. (2026). High-speed single-mode 850 nm vertical-cavity surface-emitting laser. Journal of Semiconductors, 47, 032401, Figs. 6–7.
  11. Porte, X., et al. (2021). A complete, parallel and autonomous photonic neural network in a semiconductor multimode laser. Journal of Physics: Photonics, 3, 024017, Fig. 1 and device description.
  12. Chung, Y. J., et al. (2022). Understanding limits to mobility in ultrahigh-mobility GaAs two-dimensional electron systems. Physical Review B, 106, 075134.

Explore the Horizons of Electronics collection.

Liked this post? Follow this blog to get more.