Blog

Patents for AI Hardware: What Should You Protect?

Patents for AI hardware can protect a new accelerator architecture, memory controller, chiplet interconnect, or a specific way firmware controls a device. Start with the engineering change. What did your design make possible, and which components had to work differently to get there? The answers help you decide what to file, what to document, and what may be better kept confidential.

If you’re building an inference chip, your best invention may be the way you move data. If you’re building a training system, it may be the way multiple dies stay synchronized. Product names and benchmarks will change. The architecture behind those results is where the patent discussion should start.

What parts of AI hardware can be patented?

Think about the system in layers. Start at the compute unit and follow the data out to memory, across the package, and into the host. At each step, ask what your team changed and why the conventional approach fell short.

Accelerators and memory systems

An accelerator might use a different arrangement of processing elements, a different accumulation path, or a controller that sends only useful data to those elements. A memory invention might involve how weights are staged, how banks are assigned, or how transfers are scheduled to avoid stalling the chip. Describe what happens on a representative workload. Include the state or signal that triggers a decision, the component making that decision, and the resulting change in data flow.

Chiplets, interconnects, and packaging

The inventive work may sit between dies. A chiplet design could use a different coherency mechanism or a path that keeps certain traffic local while sending other requests across the package. Packaging and thermal control can matter too. If the layout changes where heat builds up or how power is delivered, show how the physical arrangement works and what operating conditions make the difference. A good package drawing and data-flow diagram can tell different parts of the same story.

Firmware and workload scheduling

You may find the improvement at the boundary between firmware and hardware. For example, firmware could read memory pressure and thermal signals, then change which work is dispatched to each accelerator cluster. Document the inputs, scheduling rules, and hardware response. Also capture what happens if readings are stale, a cluster becomes unavailable, or two workloads compete for the same resource. Those details show what the control loop actually does.

What does an AI accelerator invention look like on paper?

Say your team built an inference accelerator that stops waiting on external memory as often. “Our chip runs inference faster” tells you why the project mattered. Now trace one batch through the system. An on-chip controller checks which weight tiles are already in local memory, queues the missing tiles, and overlaps the transfers with processing on another bank. Describe what the controller tracks, when it changes priorities, and what happens when a bank fills up.

Capture the same run under the prior design and record the testing conditions. Batch size, model, precision, available bandwidth, and power limits can change what a benchmark shows. If the improvement appears only in bandwidth-constrained configurations, say so. That gives the drafting team a clearer picture of where the invention works and helps distinguish the design from other ways of prefetching data. Use measurements you have, and label simulated results as simulations.

How does the USPTO evaluate AI hardware claims?

The U.S. Patent and Trademark Office evaluates eligibility under 35 U.S.C. § 101, novelty under § 102, nonobviousness under § 103, and disclosure requirements under § 112. Those are separate questions. A claim that describes a particular machine or a technological improvement can still face a rejection based on earlier work or a disclosure that fails to support the requested scope. For how examiners evaluate improvements to computer functionality, see the USPTO’s patent eligibility guidance.

What matters is what the claims cover and what the application teaches. Naming an accelerator in a claim does little if the claim relies on the accelerator to perform a generic calculation. If the improvement comes from a particular memory hierarchy or scheduling relationship, explain it in the application and account for it in the claims. The USPTO’s enablement guidance explains the relationship between what the application teaches and the scope of protection sought.

What should engineering teams collect before filing?

Start with annotated block diagrams and a walkthrough engineers can follow without opening the entire codebase or RTL repository. Bring the baseline, a sequence diagram for the new behavior, and the alternatives the team considered. Keep the details about the alternatives that actually work. One chip revision might make a decision in firmware while another uses a hardware state machine. If both are viable, explain both while the design is fresh.

Talk through what a competitor might ship and where you could see the claimed behavior. Some features are apparent in buyer documentation or can be tested through public interfaces. Others are buried in a chip or confined to a manufacturing process. That affects whether device claims, method claims, or trade secret protection fit the business. BLTG’s article on patent marking for semiconductor products covers another practical issue to plan for once you have a patented product.

Finally, put demos, conference abstracts, customer pilots, product announcements, and standards presentations on the same timeline as filing. Disclosure and commercial activity can affect patent rights. Timing also matters when international protection is part of the plan. Review what will be shared, with whom, and under what terms before your team makes a commitment.

Frequently asked questions about AI hardware patents

Can you patent an AI chip architecture in the U.S.?

Yes. A qualifying AI chip architecture can be claimed in a U.S. utility patent. The claims need to capture a specific invention and satisfy the requirements for eligibility, novelty, nonobviousness, and disclosure. Describe the relevant architecture and the relationships among its components.

Can you patent the firmware that controls an accelerator?

Potentially. Firmware can be part of a patentable invention when it changes how the device operates in a concrete way. Show which hardware states it reads, which commands it issues, and how the device responds to those commands.

Do you need working silicon before filing?

A working chip is generally unnecessary. The application needs to describe the invention and teach a skilled person how to make and use what is claimed. Detailed schematics, design documents, and clearly identified simulations can help explain the design.

Is a provisional application enough to protect a design?

A U.S. provisional application generally gives you up to 12 months to file a corresponding nonprovisional application claiming its benefit. The earlier date only helps for subject matter adequately supported by the provisional. Include the important implementation details up front.

Should we patent a design or keep it as a trade secret?

Look at how hard it would be to discover the design, how long it will matter, and what it takes to keep it confidential. A feature competitors can readily reverse engineer and a closely guarded factory process raise different considerations.

When should a team talk to patent counsel?

Raise the invention before public announcements, demonstrations, or commercial activity, ideally while engineers can still explain why they made the key design decisions. BLTG’s patent services cover filing and prosecution as well as portfolio strategy.