Kawasaki Wires EdgeCortix Silicon Into Aerial Defense

Kawasaki Heavy Industries and EdgeCortix signed a teaming agreement on September 8, 2026, for AI-enabled aerial defense systems. The initial scope is several million U.S. dollars across 2026 to 2028, covering technology development, feasibility studies, system integration and prototype platforms.

The pairing puts a Kanagawa chip company inside Kawasaki’s aerospace group as Japan tells the Self-Defense Forces to field AI and unmanned aircraft faster.

Kawasaki Locks a Japanese Chip Into Aerial Defense

The companies dated the announcement Kanagawa, Japan. EdgeCortix is contributing a next-generation chiplet-based AI computing platform, its MERA compiler and software framework, and related acceleration hardware. Kawasaki is contributing aerospace and defense systems experience that already includes aircraft, engines, unmanned systems and space hardware for Japan’s national defense.

Hideaki Nishimoto, general manager of Kawasaki’s Defense and Aerospace Business Division, appears with EdgeCortix founder and CEO Dr. Sakyasingha Dasgupta on the announcement photo. Kawasaki did not put a quote from Nishimoto in the joint English release. Kawasaki’s edge AI collaboration notice frames the work as a long push toward next-generation defense systems, not a single aircraft buy.

Dasgupta called the pact an expansion of an existing relationship and a commercial and strategic milestone for his firm. He tied it to a problem every airborne program eventually hits: how much thinking the vehicle can do on its own batteries and cooling.

Next-generation aerial defense platforms will require substantial AI computing capability within tightly constrained power, thermal and operational envelopes.

Dr. Sakyasingha Dasgupta, Founder and CEO, EdgeCortix

He said the companies intend to speed intelligent, adaptive and energy-efficient mission systems by combining Kawasaki’s airframe work with EdgeCortix’s chiplet architecture and MERA software. Both sides said they will widen the program as technical milestones land and new deployment chances appear.

THE TEAMING AGREEMENT AT A GLANCE

That table is the whole public contract. There is no production quantity, no radar type, no helicopter or UAV model, and no Ministry of Defense order number. The multimillion-dollar aerial defense program is a foundation, the companies said, for later work across aerospace, defense and other hard environments.

EdgeCortix posted the same news from its own account the morning of the signing.

https://x.com/edgecortix/status/2097132239054328075

What EdgeCortix Is Putting on the Aircraft

EdgeCortix is a fabless chip company founded in July 2019 by Dasgupta, who previously worked at the Max Planck Society, RIKEN, Microsoft and IBM Research. Headquarters are now in Kawasaki City, Kanagawa, with operations in the United States and India. The current production accelerator is SAKURA-II. The Kawasaki program points at the next chiplet-based platform plus the MERA software that already compiles models onto that silicon.

At the May 21, 2024 launch, Dasgupta said SAKURA-II delivers 60 TOPS of effective 8-bit integer performance inside 8W of typical power, with mixed precision and on-chip memory compression. The same device is rated at 30 trillion 16-bit brain-floating-point operations per second. The company ships it as a bare chip, as M.2 modules with different DRAM sizes, and as single- and dual-device low-profile PCIe cards.

SAKURA-II IN NUMBERS

• 8-bit rate: Up to 60 TOPS of integer performance on the DNA-II engine.
• 16-bit rate: Up to 30 TFLOPS of brain-floating-point math.
• Power: Typical envelope of 8W, with on-chip power gating for lower modes.
• Software: MERA ingests PyTorch, TensorFlow Lite and ONNX and scales across one chip, several chips or several cards.

Those figures matter on an aircraft because a cloud GPU never makes the sortie. Radar, cameras and electronic-support sensors produce data on the airframe, and a jammed or silent radio link leaves the crew or the autopilot with whatever computer is bolted in the bay. MERA is the company’s way of taking a trained network and mapping it onto that bay computer without rewriting the model for each board.

Japan has already paid to push the chiplet generation. EdgeCortix said NEDO selected a 40 billion yen project in November 2024 for energy-efficient AI chiplets aimed at post-5G communications, then a further 30 billion yen project in May 2025 for chiplets that do edge inference and on-device learning. By November 2025 the company said total capital raised had passed $110 million, after an oversubscribed Series B. Renesas Electronics is both an investor and a long-term customer. TDK Ventures later joined, calling EdgeCortix its first investment in a Japanese firm.

None of that money is the Kawasaki program. It is why a still-unlisted chip house can walk into a 125-year-old heavy-industry prime and talk about prototypes rather than slideware.

Kawasaki’s Unmanned Support Aircraft Already Assume AI

The teaming text never names an airframe. Kawasaki has, in public, already drawn the class of vehicle that would need this kind of computer.

At DSEI Japan in May 2025 the company showed two Collaborative Support Aircraft concepts, unmanned types meant to fly with Japan Air Self-Defense Force manned jets. A Kawasaki spokesperson said the platforms would work in conjunction with those jets as support aircraft, with planned roles in electronic warfare and intelligence, surveillance and reconnaissance. The company is using the word support on purpose. It has spent decades on patrol, transport and electronic work rather than a pure fighter, and it is not calling these Collaborative Combat Aircraft.

Hiroyoshi Shimokawa, president of Kawasaki’s Aerospace Systems Company, said two types are under study: a missile-like flying body and an aircraft type. Company staff have described the missile-shaped concept as 4 to 5 meters long, based on an existing missile so it can be delivered faster, and able to leave a ground or ship launcher without a runway. The aircraft-type concept is a low-observable, blended-wing, single-engine design. A spokesperson said both would use Kawasaki’s own KJ-series engines so upgrades and buys stay inside the house, and that AI would be integrated so the vehicles can fly with less operator load and take on dangerous missions that would otherwise go to a crewed aircraft.

TWO UNMANNED CONCEPTS KAWASAKI HAS SHOWN

• Missile-like CSA: About 4 to 5 meters, derived from an existing missile, launcher-friendly, meant for a short path to the customer.
• Aircraft-type CSA: Blended-wing, low-observable support UAV, single engine, KJ-series power, open architecture and a modular airframe.
• Stated jobs: Electronic warfare and ISR beside JASDF fighters, with a later look at the P-1 maritime patrol aircraft and even a drop from a C-2 transport.

Those concepts are not the EdgeCortix contract. They are the reason a power-and-thermal speech from Dasgupta is not abstract. A 4-meter missile body and a stealthy support UAV both run out of watts long before they run out of questions for a neural net: classify a contact, pick a jamming waveform, hold formation when the radio dies.

Kawasaki is also talking to Airbus about a Japanese maritime variant of Eurodrone, under a preliminary agreement signed on June 26, 2026. That airframe is a large unmanned patrol type, not a dogfighter, and it would still need onboard inference if Japan wants Japanese sensors and weapons on a European airframe. Separately, Kawasaki already joined a physical AI grouping with Nvidia that also includes Fujitsu, Fanuc and Yaskawa. That is a factory-floor stack. The EdgeCortix work is the airborne one. A prime can hold both.

Tokyo Told the SDF to Field Unmanned AI Faster

Six days before the signing, Prime Minister Sanae Takaichi stood at the Defense Ministry with Defense Minister Shinjiro Koizumi and spoke as commander in chief of the Self-Defense Forces. She said Japan wants to introduce AI and unmanned defense assets even more speedily, because those technologies move on a different clock from ships and fighters. She is pulling forward a rewrite of the three national security documents last set in 2022, a pack that was supposed to cover ten years.

The 2022 set already broke with exclusive defense by adding counterstrike weapons and a spending goal equal to 2 percent of GDP. The new pass, Takaichi said, has to deal with countries that are already fusing AI into command systems and flying unmanned aircraft in bulk. China and North Korea were the named pressures. The speech did not mention EdgeCortix or Kawasaki. It described the market those two companies just walked into.

THE YEAR AROUND THE SIGNING

1. May 12, 2025: EdgeCortix announces a U.S. Defense Innovation Unit Other Transaction contract awarded in February, saying it is the first Japanese company and the first semiconductor company to receive that form of DIU deal.
2. May 21-23, 2025: Kawasaki unveils the two CSA concepts at DSEI Japan and talks about AI on those airframes.
3. January 2026: EdgeCortix shows edge-AI hardware in Japan’s Defense Acquisition Agency booth at the Singapore Air Show.
4. June 26, 2026: Airbus and Kawasaki sign a preliminary agreement on a possible Japanese Eurodrone variant.
5. June 30, 2026: EdgeCortix says SAKURA-II flew in a U.S. Air Force large-force exercise and that DIU issued a Success Memorandum.
6. September 2, 2026: Takaichi tells SDF seniors to field AI and unmanned kit faster, ahead of a year-end document rewrite.
7. September 8, 2026: Kawasaki and EdgeCortix sign the teaming agreement.

Read in that order, the Kawasaki deal is late in a sequence, not a first handshake. EdgeCortix had already been standing in ATLA’s booth. Kawasaki had already published unmanned airframes that assume autonomy. The government had already said the unmanned clock is too slow. The missing piece was a named Japanese inference chip inside a named Japanese airframer’s prototype plan.

SAKURA-II Already Flew in a U.S. Exercise

On June 30, 2026, EdgeCortix said it had finished a U.S. Air Force flight demonstration of SAKURA-II and had a Success Memorandum from DIU covering every prototype objective completed to date. The company said the chip ran high-performance, low-power inference in operationally relevant domains, including aerial operations, on both commercial models and custom defense applications.

Lt. Col. Spencer Liedl, KC-135 operational test director at the Air National Guard Air Force Reserve Command Test Center at Roland R. Wright Air National Guard Base in Utah, described the flight in the company’s release.

The U.S. Air Force and EdgeCortix worked together to integrate SAKURA-II into a relevant mission system and fly it in a large force exercise, validating AI inference in flight with a tactically relevant application in operationally relevant scenarios.

Lt. Col. Spencer Liedl, KC-135 Operational Test Director, AATC

The same DIU project included independent benchmarking support from the Carnegie Mellon University Software Engineering Institute and radiation work that built on NASA heavy-ion testing of SAKURA-II. EdgeCortix said those ion shots showed high radiation resiliency, with no destructive events and only minimal transient effects, which is the language you want if the same family of parts might later ride a missile, a high-altitude UAV or a spacecraft. The company said the airborne tests raised the platform’s technology readiness level for flight.

A Success Memorandum is not a production contract. DIU’s note, in EdgeCortix’s telling, confirms that commercially built edge AI held up in defense-like conditions and opens a path to later transition, deployment and follow-on builds. That is the same shape as the Kawasaki paper: prove the computer in the real envelope, then argue for more money.

Putting a Japanese chip through a U.S. Air Force large-force exercise also answers a quiet question Japanese primes have to ask before they freeze a mission computer. Will this part still be usable if the program later talks to U.S. platforms, U.S. test ranges or U.S. export paperwork? SAKURA-II has at least one data point on that list. The Kawasaki chiplet platform still has to earn its own.

Several Million Dollars Buys Studies, Not a Fleet

Several million dollars, spread across 2026 to 2028, is prototype money at Kawasaki’s scale. It pays for the four workstreams on the table: build the tech, test whether it fits, bolt it to a system, and fly or bench a prototype. It does not buy a squadron of CSA airframes, a Eurodrone line or a helicopter retrofit kit. Anyone reading the announcement as a weapons order is reading a press title, not the scope paragraph.

That small check still changes who writes the software on whatever airframe comes next. If Kawasaki’s mission computer speaks MERA and a Japanese chiplet, model vendors, sensor makers and later export partners have to compile to that stack. The cloud had one dominant inference vendor. An airframe that carries its own accelerator carries that vendor’s compiler for the life of the airframe. That is the part of the deal that outlasts the first invoice.

WHAT WE KNOW

• The paper: A teaming agreement, not a disclosed production contract, with an initial aggregate value of several million U.S. dollars across 2026 to 2028.
• The kit named: A next-generation chiplet AI platform, the MERA compiler and related accelerators, aimed at power, latency and reliability limits on aerial platforms.
• The flight history: SAKURA-II, the current EdgeCortix production part, has already run inference in a U.S. Air Force large-force exercise under a DIU prototype.

WHAT IS UNCONFIRMED

• The airframe: No helicopter, UAV, patrol aircraft or missile is named. CSA, Eurodrone and the P-1 are public Kawasaki programs, not stated targets of this contract.
• The customer: No Ministry of Defense or ATLA award is attached to the teaming agreement in the public text.
• The follow-on: Both companies say the scope can grow after milestones. They have not said what those milestones are, or what a later invoice would buy.

Both companies said they will expand the work if the prototypes hold and if more development and fielding chances appear. Kawasaki has not said which airframe will carry the chips. EdgeCortix has not said which chiplet stepping will sit in the first prototype bay. The computer is being chosen while those questions are still open.