For years, AI infrastructure companies heard the same frustrating refrain: "Your machine learning innovation is just an abstract algorithm—not patentable." The Alice decision cast a long shadow over software and AI patents, leading to widespread pessimism. But Director John Squire's USPTO is signaling a dramatic shift. AI infrastructure companies now have a window to patent innovations that would have been rejected just months ago.

The In re Desjardins Decision Changes Everything

The precedential In re Desjardins decision directly addressed the treatment of AI innovations under Section 101. The USPTO Appeals Review Panel explicitly warned against "categorically excluding AI innovations from patent protection," noting this "jeopardizes America's leadership in this critical emerging technology."

Critically, the panel rejected the practice of equating "any machine learning with an unpatentable 'algorithm'" and dismissing additional elements as "generic computer components without adequate explanation." This reverses years of examiner practice that was hostile to AI infrastructure patents.

What's Now Patentable for AI Infrastructure

Under the new guidance, AI infrastructure companies should reconsider patenting:

• Novel training architectures: Improvements to how models are trained, including data pipelines, optimization techniques, and distributed training methods
• Inference optimization: Technical improvements to model deployment, quantization, pruning, and serving infrastructure
• Model distillation innovations: Novel approaches to knowledge transfer between models that improve efficiency or capability
• Hardware-software co-design: Optimizations at the intersection of AI software and underlying compute infrastructure
• Novel attention mechanisms: Architectural improvements to transformer models and their successors

The Mental Process Limitation Works in Your Favor

The August 2025 memo reminds examiners not to "expand the 'mental process' grouping of abstract ideas to encompass claim limitations that cannot practically be performed in the human mind." This is crucial for AI infrastructure: no human can mentally execute a trillion-parameter model inference, implement backpropagation across millions of weights, or perform the matrix operations underlying modern AI.

When drafting claims, emphasize the computational scale and technical specificity that places your innovation firmly outside what any human mind could perform.

Acting Before the Window Closes

USPTO policy can shift with administrations. The current favorable environment represents an opportunity—not a permanent state. AI infrastructure companies should:

• Audit your innovation backlog: Review technical innovations from the past 2-3 years that were deemed "too risky" to patent
• Prioritize continuation applications: If you have pending applications with broader parent claims, consider continuations that capture newly-patentable subject matter
• Document technical improvements: Frame innovations around specific technical problems solved—latency, accuracy, efficiency, scalability
• File provisionally now: Secure priority dates while the favorable guidance is in effect, then refine claims as prosecution practice develops

Conclusion: The Competitive Imperative

In an industry where model weights can be distilled in days and trade secrets offer limited protection, patents remain the most durable form of IP protection. The USPTO's expanded eligibility guidance removes barriers that have held AI infrastructure companies back. Those who act now will build portfolios that create real competitive moats; those who wait may find the window closed.