BrainChip Startup Launches Project to Create AKD2500 Neuromorphic Chip
BrainChip invests $2.5 million in developing a silicon version of its Akida 2.0 neuromorphic architecture. The chip will be manufactured by TSMC using a 12nm process for ultra-efficient edge AI.
The Gist: What's Really Happening
BrainChip is launching the AKD2500 development not for the sake of another chip—the company is creating a physical vehicle for its IP licensing strategy. The neuromorphic computing market is stuck in a trap: everyone talks about the potential, but no one commits to mass adoption without being able to "touch" the technology. AKD2500 solves precisely this problem—it will serve as a demonstration platform where potential licensees can evaluate the Akida 2.0 architecture in real-world conditions before embedding the IP into their own SoCs.
At first glance, $2.5 million for development is a modest sum by semiconductor industry standards. However, this figure hides a fundamental shift: BrainChip deliberately chose TSMC's mature and inexpensive 12nm process instead of chasing nanometers. This shows that the company is not focused on demonstrating technological superiority but on a practical goal—showing clients a working solution with acceptable cost. With BrainChip's current market cap around A$323-329 million and an annual loss of A$20.47 million, every dollar spent must deliver measurable commercial results.
Timeline and Context
BrainChip has come a long way to reach this point. The company became the world's first commercial producer of neuromorphic AI technologies, releasing Akida 1 and the AKD1000 chip—the first fully digital event-based neuroprocessor. The model was simple: create a reference chip, prove its viability, then sell IP for integration into other systems. Akida 1 found its first customers—licenses were acquired by Megachips and Renesas.
The second generation of the architecture, Akida 2.0, brought four key improvements: Temporal Event-based Neural Networks (TENN), hardware acceleration for Transformer encoders, INT8 support, and long skip connections. These updates take the technology beyond niche sensor applications into the realm of language models and streaming video processing.
In February 2026, BrainChip officially launched the AKD2500 project. The company engaged ASICLAND for design, coordination with TSMC, packaging, and testing. A prototype is expected in Q3 2026, and a Technology Roadmap detailing the strategy was scheduled for May 14.
The financial situation is gradually improving. Annual revenue for 2025 grew 374% to $1.89 million, and in Q1 2026, customer inflows reached $700,000—up from $400,000 the previous quarter. Cash reserves at the end of 2025 stood at $31.71 million. Concurrently, the company launched cloud-based FPGA access to Akida Pico IP, allowing developers to test the technology remotely without their own hardware.
Who Wins and Who Loses
Winners:
BrainChip gains a key tool for converting technical interest into licensing deals. Two strategic agreements have already been signed: a global Akida 2.0 license with Korea's EDGEAI and a collaboration with ForwardEdge ASIC—a Lockheed Martin subsidiary focused on defense and aerospace applications. The defense sector is particularly important—its requirements for energy efficiency and autonomous operation without cloud connectivity align perfectly with neuromorphic chip capabilities.
TSMC wins even with the project's modest budget. Loading the 12nm line with an additional client, even at the multi-project wafer stage, means capacity utilization in a mature process where margins remain high. And if BrainChip's licensing model works, production volumes will multiply through orders from licensees.
Potential customers in industrial automation, IoT, and wearable electronics gain an alternative to traditional AI accelerators. Akida's event-based architecture processes only non-zero activations in the network's internal layers, reducing computation by an order of magnitude compared to standard approaches. For battery-powered devices, this is critical.
Losers:
Traditional edge AI chipmakers face a new competitor. BrainChip claims it has no direct competitors in event-based IP. If the commercial model succeeds, companies like Hailo or Kneron, which offer traditional inference accelerators, will face competition in the most attractive segment—ultra-low power consumption.
Chinese IoT chip manufacturers lose a potential market. Since BrainChip sells IP, not finished chips, its clients are large integrators with their own production chains. Small manufacturers, especially from countries with limited access to advanced processes, will not be able to compete with solutions built on TSMC's 12nm technology.
Startups without their own development ecosystem lose implicitly. BrainChip invests in tooling and cloud access to its IP, recognizing that ease of adoption is the main barrier for new architectures.
What the Media Isn't Saying
Financial publications focus on the $2.5 million budget but miss the key issue: scalability. BrainChip burns about $3-4 million per quarter while customer inflows are only $700,000. Even with $31.71 million in reserves, the company must double or triple licensing revenue by the end of 2027 to avoid another dilutive offering. And here lies the main risk: IP licensing is a long process, especially in the defense sector, where procurement cycles stretch over years.
The second overlooked point is that architectural uniqueness is a double-edged sword. The lack of competitors in event-based IP means BrainChip must build the market itself: train developers, create tooling, and prove the approach's advantages. TechInsights directly points out: uniqueness makes the ecosystem critical so that clients can deploy networks without diving into technology details.
The third aspect concerns the demonstration at Embedded World 2025. BrainChip showed a billion-parameter LLM running fully autonomously on an FPGA with Akida 2.0, consuming so little power that the chip could run on a watch battery. This is impressive, but the FPGA runs at one-tenth the speed of an ASIC implementation. Moving to silicon promises a tenfold performance leap, but compatibility with real-world use cases remains an open question.
Finally, the alliance with Lockheed Martin via ForwardEdge ASIC is a double-edged sword. On one hand, it validates the technology with a strategic customer. On the other, defense contracts often come with export restrictions and commercial use limitations, which could narrow the market for civilian licensing.
Forecast: Next 30 Days and 90 Days
30 days (by June 10, 2026):
On May 14, BrainChip held a Technology Roadmap presentation, likely revealing details of the Akida strategy, including timelines for AKD2500 and plans for generative AI tools. Market reaction to this presentation may determine the stock's trajectory for the next quarter.
The annual shareholder meeting in late May 2026 will bring detailed financial reports. I expect announcements of progress on new licensing agreements, as the company needs positive signals to maintain shareholder value.
90 days (by August 9, 2026):
The AKD2500 prototype will be released in Q3 2026, and this will be the main event. If initial tests confirm the claimed performance and energy efficiency on real tasks, the number of licensing inquiries could grow exponentially.
EDGEAI should begin integrating Akida 2.0 IP into its own SoCs. Initial results from this collaboration will show how quickly BrainChip's licensing model converts into real revenue.
Strategic takeaway: BrainChip is betting that energy efficiency will become the main criterion for edge AI in 2027-2028. If this forecast holds, the company will be in the right place at the right time with the right architecture and early customers in strategic sectors. There are no direct competitors in this niche today: some companies make conventional edge accelerators and ignore the neuromorphic approach, others develop neuromorphic chips but use analog circuits that are less convenient for mass production. BrainChip is the only company offering fully digital, event-based neuromorphic IP that can be embedded into standard SoC designs without exotic technology requirements. The next 90 days will show whether this uniqueness turns into commercial success or remains a technological curiosity.
— Editorial Team
No comments yet.