Diamond beyond silicon: what Ōkuma and boron compounds mean for AI chips

On 29 May 2026, the town of Ōkuma in Fukushima marked the completion of a factory dedicated to diamond semiconductor devices. Its starting point was a practical problem: how do you keep instruments working where radiation makes the electronics difficult to replace? The need arose during the decommissioning of Fukushima Daiichi. It did not begin with an AI processor. [1] [2] [3]

Diamond reaches AI hardware by a different route. Put a thin diamond plate against a silicon processor and the silicon still does the arithmetic; the diamond gives heat another path out. Make a transistor in diamond and the crystal now carries the signal. A third possibility, studied in boron compounds, is to choose a material for both its heat transport and its electrical behaviour. Experiments in Japan and elsewhere let us separate these three claims and ask which ones have reached a working device.

Illustration of a transparent diamond heat spreader on a silicon chip. A warm-coloured spot marks a local heat source; the colours do not represent measured temperatures.
The transparent upper layer spreads heat; the silicon beneath it does the computing. Colour marks a local heat source for illustration. This is neither a thermogram nor an image of a device made in Ōkuma. © 2026 Juan Antonio Lloret Egea · ITVIA · The Bible of AI (BAI) · CC BY-NC-ND 4.0.

What was completed in Ōkuma?

The town records a ceremony for the completed building. The operator, Ookuma Diamond Device, calls it the world’s first factory with a line dedicated specifically to diamond semiconductor research, development and manufacturing, based on its own survey. Its projects include radiation detectors for work inside damaged reactors. Microprocessors and memories able to withstand radiation appear as development goals. The public records reviewed here establish neither volume shipments from this line nor the manufacture of AI accelerators there. [1] [2] [3]

Size claims need the same care. On 5 November 2026 a different company, Diamond Foundry, announced a single diamond crystal 200 mm across and showed it as grown, before polishing. Growing a large crystal addresses one obstacle. Making useful circuits across it also requires uniform material, reliable contacts and enough working devices per wafer. Diameter alone answers none of those questions. [20]

One row apart: why do carbon and silicon behave differently?

Silicon sits directly below carbon in group 14 of the periodic table. Each has four valence electrons; in both diamond and crystalline silicon, each atom bonds to four neighbours in a tetrahedral arrangement. But a silicon atom is larger. A typical molecular Si–Si single bond is longer and weaker than a C–C single bond, whereas silicon forms particularly strong bonds with oxygen. A bond energy for a molecule is not a measure of how well a whole crystal carries heat. [21] [22]

Life on Earth illustrates the difference. Carbon chains and rings form the molecular framework of sugars, fats, proteins and nucleic acids. Silicon has a more specialised biological role: diatoms build patterned cell walls from silica, a silicon–oxygen material, together with organic molecules. In a chip the relevant question is more specific still: which material should carry the current, and which should carry the heat? [22] [23]

How can diamond carry heat so well and so little current?

Heat and charge need not travel with the same carriers. In diamond, much of the heat moves as vibrations of the tightly bonded carbon lattice, described as phonons. Electrical current needs mobile electrons or holes. At room temperature a very pure diamond has few of either: its valence and conduction bands are separated by about 5.5 eV. It can therefore spread heat efficiently while remaining electrically insulating in practice. In measurements on three high-purity synthetic single crystals, thermal conductivity reached 2,400 W m⁻¹ K⁻¹ at room temperature. That is a property of particular specimens; defects, grain boundaries and the contact with the chip all matter in a real assembly. [4] [5] [15]

Introducing boron changes the electrical question. When a boron atom replaces carbon in the lattice it can accept an electron, leaving a mobile hole. An isolated boron acceptor lies about 0.37 eV above the valence-band edge; the number of mobile holes need not equal the number of boron atoms. An Ōkuma company patent gives a concrete use: two heavily boron-doped diamond regions provide electrical contacts for a transistor. The host material is still diamond. The patent describes a design, not a production run. [5] [6]

Current also generates heat. In a resistive path the power dissipated is I²R. A spreader can lower the temperature reached at a given power without changing how many watts that path dissipates. The heat must still cross a bond and ultimately pass to air or coolant. A processor held back by its temperature limit may then sustain a higher clock; a processor waiting for memory may finish no sooner.

What have Japanese laboratories switched in diamond?

Japan’s National Institute for Materials Science fabricated a diamond transistor whose channel carries electrons, using phosphorus as the donor. At 573 K (about 300 °C), the researchers measured a field-effect mobility of about 150 cm² V⁻¹ s⁻¹. A different collaboration, at Kanazawa University and AIST, demonstrated a diamond transistor with a channel along a trench wall. It was normally off and showed gate control at room temperature and at 300 °C. The first experiment concerns electron transport; the second concerns the geometry and control of a power switch. Neither tested a neural network. [7] [8]

Which boron material does what?

Boron sits in group 13, one column to the left of carbon, and has three valence electrons. Introducing a little boron into diamond is different from making a crystal in which boron is a structural constituent. The three examples below are an arsenide, a carbide and a carbon nitride. None is a carbonate. [6] [21]

Cubic boron arsenide (BAs): a crystal that carries heat and charge

Three teams reported high room-temperature thermal conductivity in 2018. UCLA measured 1,300 W m⁻¹ K⁻¹. A collaboration including UT Dallas and Illinois reported 1,000 ± 90. A third team found local values above 1,000 but a bulk average reaching about 900, in the same units. In 2025 a collaboration involving Houston and UC Santa Barbara reported values above 2,100 in selected regions of purer crystals. The convergence is on unusually good heat transport; the highest local value is not a specification for an entire wafer. [9] [10] [11] [12]

Can the same material carry charge? With an optical transient grating, one team found a thermal conductivity of 1,200 W m⁻¹ K⁻¹ and an ambipolar mobility of 1,600 cm² V⁻¹ s⁻¹ at the same sample locations at room temperature. Another study used transient reflectivity microscopy and obtained 1,550 ± 120 cm² V⁻¹ s⁻¹ under excitation chosen to avoid hot carriers. These optical measurements track electrons and holes diffusing together. They do not measure either the switching speed or the contact resistance of a finished transistor. [13] [14]

Boron carbide: a working junction with a different chemistry

In a Nebraska–Lincoln collaboration, researchers made a diode and a junction field-effect transistor using boron carbide and elemental boron. The film in that experiment was reported as B₅C. It is an example of a semiconducting boron–carbon material with working junctions. The heat-conduction measurements for BAs cannot be assigned to that carbide: the latter contains no arsenic, and the study tested a different property. [16]

Boron carbon nitride (BCN): a one-layer transistor channel

In 2026 a Japanese–Taiwanese collaboration grew a continuous BCN monolayer on a two-inch copper wafer. Carbon incorporated into the boron nitride lattice allowed holes to carry current. The team characterised 224 field-effect transistors from three growth batches. It reported a field-effect hole mobility of about 100 cm² V⁻¹ s⁻¹ and an on-current above 0.9 mA per micrometre of channel width. The journal corrected the unit on the current axis of figure 5d to milliamperes per micrometre. These are transistor results; the wafer was not a two-inch AI processor. [17] [18]

Why can a better crystal make a worse heat spreader?

A comparison on gallium nitride makes the distinction visible. A UCLA-led team fabricated high-electron-mobility transistors from matching semiconductor layers with the same layout. Near 15 W per millimetre of gate width, the measured hot-spot temperature rise was about 60 K with a BAs substrate, 110 K with diamond and 140 K with silicon carbide. These are rises above a reference temperature, not absolute temperatures. In these devices the GaN–BAs interface transferred heat effectively. The result ranks these particular assemblies; it does not say that every BAs crystal conducts heat better than diamond. [15]

What can an AI system claim today?

There is a concrete industrial example outside Ōkuma. In March 2026 MiTAC and Akash announced diamond-cooled AI servers with AMD Instinct MI350X GPUs. The supplier reports up to 10 °C lower temperatures on the GPU and its high-bandwidth memory; it also reports up to 15% more tokens per second at an unusually hot ambient temperature of about 49 °C (120 °F). These are supplier-reported maxima under stated conditions. They establish that diamond cooling has reached an AI server offering, not that a silicon GPU has been replaced, or that a normal data centre will see the same improvement. [19]

A useful independent test would run the same model, precision, batch size, input and output lengths on otherwise matched servers. After warm-up, record GPU and memory temperatures, clocks, throughput and response latency at the same output quality. Meter the whole server, report cooling power separately, and divide energy by completed output tokens. Repeat at comparable coolant inlet or ambient temperatures; otherwise an apparent material advantage can simply reflect a colder inlet. If the temperatures fall but joules per token and sustained throughput do not change, the experiment has demonstrated thermal margin, not faster or more efficient inference.

Gallium nitride already offers a different active material for power switching. Silicon carbide is another option for that job, although it still contains silicon. Experiments on commercial transistors compare switching and conduction losses; they do not tell us which material should perform a neural network’s arithmetic. [24]

Silicon need not perform every job in a computer. We already have evidence for heat spreading in one material and logic in another, for transistors with a diamond channel and for hole conduction in a layer of BCN. Where silicon can give way depends on the job: carrying heat, switching current or completing an inference. Each has a different test.

Sources and scope

This selective review uses material available through 14 November 2026. Journal articles document material and device measurements. The town documents the factory ceremony; the patent describes a possible device design. Statements about facilities, product availability and server performance are attributed to their issuers.

  1. Ōkuma Town. Completion ceremony for the Ookuma Diamond Device Fukushima plant [in Japanese]. 29 May 2026.
  2. Ookuma Diamond Device. Completion of the dedicated diamond semiconductor factory [in Japanese]. Company release, 4 June 2026.
  3. Ookuma Diamond Device. Projects: radiation monitoring, integrated circuits and the first plant. Company project register; consulted 14 November 2026.
  4. Inyushkin, A. V., et al. Thermal conductivity of high purity synthetic single crystal diamonds. Physical Review B 97, 144305 (2018).
  5. Bhattacharya, S., et al. Intervalence plasmons in boron-doped diamond. Nature Communications 16, 444 (2025).
  6. Ookuma Diamond Device. Diamond and semiconductor devices. Japanese patent JP7777876B2; published 1 December 2025.
  7. Liao, M., Sun, H., & Koizumi, S. High-Temperature and High-Electron Mobility Metal-Oxide-Semiconductor Field-Effect Transistors Based on N-Type Diamond. Advanced Science 11, 2306013 (2024).
  8. Nakamura, Y., Nagai, M., et al. Trench-type inversion-channel diamond MOSFET enabled by thermochemical etching based on C solid solution into Ni. Diamond and Related Materials, 114141 (2026).
  9. Kang, J. S., Li, M., Wu, H., Nguyen, H., & Hu, Y. Experimental observation of high thermal conductivity in boron arsenide. Science 361, 575–578 (2018).
  10. Li, S., et al. High thermal conductivity in cubic boron arsenide crystals. Science 361, 579–581 (2018).
  11. Tian, F., et al. Unusual high thermal conductivity in boron arsenide bulk crystals. Science 361, 582–585 (2018).
  12. Niyikiza, A. B., et al. Thermal conductivity of boron arsenide above 2100 W per meter per Kelvin at room temperature. Materials Today 90, 11–14 (2025).
  13. Shin, J., et al. High ambipolar mobility in cubic boron arsenide. Science 377, 437–440 (2022).
  14. Yue, S., et al. High ambipolar mobility in cubic boron arsenide revealed by transient reflectivity microscopy. Science 377, 433–436 (2022).
  15. Kang, J. S., Li, M., Wu, H., Nguyen, H., Aoki, T., & Hu, Y. Integration of boron arsenide cooling substrates into gallium nitride devices. Nature Electronics 4, 416–423 (2021).
  16. Hwang, S.-D., Byun, D., Ianno, N. J., Dowben, P. A., & Kim, H. R. Fabrication of boron-carbide/boron heterojunction devices. Applied Physics Letters 68 (1996).
  17. Tseng, C.-C., Huang, C.-H., et al. Wafer-scale epitaxy growth of high-mobility p-type boron carbon nitride. Nature 658, 380–387 (2026).
  18. Tseng, C.-C., Huang, C.-H., et al. Author Correction: Wafer-scale epitaxy growth of high-mobility p-type boron carbon nitride. Nature, 14 October 2026.
  19. MiTAC Computing. MiTAC Computing Announces World’s First Diamond Cooled AI Servers with AMD and Akash Systems. Company announcement, 3 March 2026.
  20. Diamond Foundry. D200mm: A New Scale for Single-Crystal Diamond. Company announcement, 5 November 2026.
  21. IUPAC. Periodic Table of Elements. Reference table; groups 13 and 14.
  22. Petkowski, J. J., Bains, W., & Seager, S. On the Potential of Silicon as a Building Block for Life. Life 10, 84 (2020), section 3.1 and table 1.
  23. Kröger, N., Deutzmann, R., Bergsdorf, C., & Sumper, M. Species-specific polyamines from diatoms control silica morphology. PNAS 97, 14133–14138 (2000).
  24. Li, K., Evans, P., & Johnson, M. SiC/GaN power semiconductor devices: a theoretical comparison and experimental evaluation under different switching conditions. IET Electrical Systems in Transportation (2018).

More in Horizons of Electronics →

Liked this post? Follow this blog to get more.