In November 2022, getting GPT-3.5-level AI performance cost about:

$20 per million tokens.

By October 2024:

$0.07.

Same benchmark threshold.

Roughly:

286× cheaper.

Now consider software development.

On SWE-bench, AI systems solved:

4.4% of coding problems in 2023.

One year later:

71.7%.

That is a:

16.3× jump.

Meanwhile, the smallest model able to cross a 60% MMLU threshold shrank from:

540 billion parameters

to:

3.8 billion.

142× reduction.

And by March 2026, the top four AI companies on the Arena leaderboard—Anthropic, xAI, Google, and OpenAI—sat within only:

22 Elo points

of one another.

AI is becoming:

cheaper.

smaller.

easier to access.

better at writing software.

more interchangeable.

That creates a problem for technology moats.

If your competitive advantage can be recreated simply by giving a good engineer a frontier model, enough context, and a few weeks, how much of a moat is it?

This is why patents are becoming more—not less—strategically important in AI.

Not because every AI feature should be patented.

Not because patents magically create a good company.

And certainly not because saying “AI” inside a patent application makes it valuable.

The reason is much more fundamental:

AI is rapidly lowering the cost of independently recreating technology. Patents are one of the few forms of intellectual property that can still restrict an independently built implementation of the claimed invention.

That distinction could become enormously important over the next decade.

PatentPC analyzed AI performance, development costs, open-source activity, investment, patent filings, litigation, trade-secret law, copyright law, and recent U.S. AI patent decisions to understand how the economics of the technology moat are changing.

The result is what we call the AI Moat Squeeze.

PatentPC’s AI Moat Snapshot

Here is the argument in numbers.

What is becoming easier to reproduce?	What the public data says
GPT-3.5-level inference	~286× cheaper in ~18 months
Model size for >60% MMLU	142× smaller from 2022 to 2024
AI software-engineering performance	16.3× SWE-bench improvement
Open-vs-closed model performance gap	Fell 78.9% in ~13 months
Top-four frontier-provider spread	Just 22 Arena Elo points by March 2026
Open-source AI projects	~5.6 million on GitHub in 2025
Global corporate AI investment	~40× larger in 2025 than 2013
Newly funded AI companies	+71% in 2025
Granted AI patents	33.9× 2010 level by 2024
GenAI patent families	+168.5% from 2023 to 2025
GenAI software/code families	+376.7% from 2023 to 2025
2024–25 GenAI families vs. entire 2014–23 decade	4.3% more in just two years

These figures come from different datasets and should not be combined into one synthetic “moat score.”

But the direction is difficult to miss.

Technological reproduction is getting cheaper.

At the same time:

legal claims over AI technology are accumulating faster.

That combination changes strategy.

Original Research #1: AI Has Cut the Cost of Replicating Capability by About 99.65%

Stanford’s AI Index found that the price of querying a model that achieved GPT-3.5-level performance on MMLU fell from:

$20 per million tokens in November 2022

to:

$0.07 by October 2024.

Stanford AI Index — Research and Development

PatentPC calculation:

$20 ÷ $0.07 =

285.7× cheaper

Price decline:

($20 − $0.07) ÷ $20 =

99.65%

Think about what that means commercially.

A fixed volume of benchmark-equivalent inference that would have cost:

$100,000

at $20 per million tokens would cost:

$350

at $0.07.

That is not a prediction of your actual API bill.

Models differ.

Applications differ.

Token use changes.

But it shows the magnitude of the economic collapse.

And falling inference cost does something important to moats:

It allows many more competitors to buy roughly comparable intelligence.

Five years ago, access to strong machine intelligence could itself be scarce.

Increasingly, it is an input.

A Model Is Starting to Look More Like Cloud Infrastructure

Nobody builds a durable SaaS moat by saying:

“We use AWS.”

Why?

Because competitors can use AWS too.

AI models are moving in the same direction.

A company can still have a frontier-model advantage.

Building state-of-the-art models remains extraordinarily expensive.

But for most AI businesses, the model is not proprietary.

It comes through:

an API;
an open-weight model;
a cloud provider;
a licensed model;
several interchangeable providers.

That means the moat moves elsewhere.

Into:

proprietary architecture

specialized data

distribution

workflow integration

trust

regulatory position

network effects

switching costs

brand

and increasingly:

intellectual property.

PatentPC explored the broader issue in its separate study, Your AI Is Not Your Moat: What Actually Makes an AI Company Defensible.

Original Research #2: Powerful Models Became 142× Smaller

Cost is only one side of replication.

The models themselves are getting much smaller.

Stanford reports that in 2022, the smallest model scoring above 60% on MMLU was PaLM:

540 billion parameters.

By 2024, Phi-3-mini achieved the threshold with:

3.8 billion.

PatentPC calculation:

540 ÷ 3.8 =

142.1× smaller

A capability that once required giant infrastructure increasingly fits into much smaller systems.

That means:

cheaper hosting;
cheaper inference;
easier fine-tuning;
more edge deployment;
more private deployment;
more potential competitors.

Again:

the capability becomes less scarce.

Original Research #3: AI’s Ability to Build Software Improved About 16× in One Year

Now we reach the part that should worry software founders most.

AI is not merely getting cheaper to use.

It is getting better at building the software around itself.

Stanford reports SWE-bench performance of:

4.4% in 2023

and:

71.7% in 2024.

PatentPC calculation:

71.7 ÷ 4.4 =

16.3×

in one year.

That does not mean AI can independently recreate every sophisticated product.

It cannot.

Real systems contain:

undocumented assumptions;
messy integrations;
production constraints;
security problems;
domain knowledge;
customer-specific workflows;
organizational knowledge.

But the direction matters.

The cost of writing software is falling.

The cost of understanding software is falling.

The cost of prototyping competing features is falling.

The cost of experimenting with architectures is falling.

The cost of translating an idea into functioning code is falling.

That weakens a moat based mainly on:

“Our competitors would have to build this.”

Increasingly, they can.

And increasingly, AI can help them.

Original Research #4: Open-Source AI Development Has Reached 5.6 Million GitHub Projects

Stanford’s 2026 AI Index counted roughly:

5.6 million AI-related GitHub projects

in 2025.

The number grew another 23.7% in that year alone.

There were still more than:

206,000 projects with at least 10 GitHub stars

after filtering out much of the experimental noise.

That is not merely “more open source.”

It is an enormous global experiment in:

agents;
RAG;
inference;
model routing;
multimodal systems;
vector databases;
evaluation;
speech;
computer vision;
tool use;
orchestration;
synthetic data;
fine-tuning.

Every public implementation teaches the next competitor something.

That is good for innovation.

It is harder for weak moats.

Open Models Are Also Closing the Gap

At the start of January 2024, Stanford measured an 8.04% Chatbot Arena performance advantage for the leading closed-weight model over the leading open-weight model.

By February 2025:

1.70%.

PatentPC calculation:

1 − (1.70 ÷ 8.04) =

78.9% reduction in the gap

in about thirteen months.

The exact gap has moved since then, and leading closed models still hold important advantages in some areas.

The broader point survives:

Access to capable models is widening.

Original Research #5: Even the Frontier Labs Are Clustering

By March 2026, Stanford reported these Arena Elo ratings:

Company	Elo
Anthropic	1,503
xAI	1,495
Google	1,494
OpenAI	1,481

Spread:

1,503 − 1,481 =

22 points

As a percentage of the leader’s score:

22 ÷ 1,503 =

1.46%

Elo scores are not percentage performance scores, so do not interpret that 1.46% as a literal capability gap.

The useful fact is the clustering itself.

Stanford describes four companies as sitting within 25 Elo points.

This matters strategically.

Suppose your pitch is:

“Our product is better because we use the best model.”

What happens when four vendors become very good?

Your customers can switch.

Your competitors can switch.

You can switch.

Model access becomes less of a durable differentiator.

This Is the AI Moat Squeeze

Put the pieces together.

Intelligence becomes cheaper.

~286× cheaper for one benchmark-equivalent inference threshold.

Models become smaller.

142× fewer parameters for one MMLU threshold.

Models converge.

Four frontier providers within 22 Arena Elo points.

Open alternatives improve.

The measured open/closed gap collapsed sharply.

Software gets easier to build.

16.3× SWE-bench improvement.

Public implementation knowledge explodes.

5.6 million AI GitHub projects.

This squeezes traditional technology moats from several directions at once.

We can express the basic economic problem simply:

If competitors can buy the intelligence and AI can help them rebuild the implementation, the proprietary technical mechanism becomes more important than the visible feature.

And this is where patents become unusually relevant.

Why Patents Are Different From Copyright and Trade Secrets

This is the most important legal point in the article.

Imagine your company invents a new AI inference-routing architecture.

A competitor never sees your source code.

Nobody steals anything.

Their engineers independently solve the same problem.

AI tools help them.

They write completely different code.

What happens?

It depends on the IP right.

Protection	Exact code copied	Independently recreated functionality	Lawful reverse engineering
Copyright	Stronger protection	Usually weak/no protection over function itself	Limited
Trade secret	Can help if information was misappropriated	No protection against independent development	Generally cannot stop lawful reverse engineering
Patent	Can reach claimed invention	Potentially yes	Potentially yes if claims are practiced

The U.S. Copyright Office expressly says copyright in computer programs does not extend to program logic, algorithms, systems, methods, or concepts.

WIPO says a trade-secret owner generally cannot prevent somebody from using the same information if they developed it independently or obtained it through lawful reverse engineering.

U.S. patent law works differently. A patent provides a right to exclude others from making, using, selling, offering for sale, or importing the claimed invention. Direct patent infringement generally does not require proof that the defendant copied the patent or intended to infringe.

That difference becomes more valuable when independent recreation gets easier.

The Legal Asymmetry AI Makes More Important

Before AI, independently recreating a complicated technical system could take:

years;
a large engineering team;
difficult reverse engineering;
scarce specialized knowledge.

That practical difficulty itself acted like a partial moat.

AI reduces some of that friction.

Now imagine the same outcome after the code can be generated, analyzed, rewritten, benchmarked, and debugged much faster.

Trade secrets still protect secrets.

But if nobody copied your code and nobody stole your secret:

those rights may not stop an independent implementation.

A valid patent with the right claim may.

That is the strategic asymmetry.

PatentPC has deeper guides on patents versus trade secrets, trade-secret protection for AI and deep-learning code, and protecting intellectual property for AI inventions.

That Does Not Mean “Patent Everything”

Patents have serious costs.

You disclose the invention.

Protection is territorial.

Claims can be invalidated.

Competitors may design around them.

Enforcement can be expensive.

The patent eventually expires.

Some technical advantages are almost impossible to detect in competitors.

For those, secrecy may be stronger.

The point is not:

Patents beat every other moat.

The point is:

AI increases the strategic value of the one major IP right designed to survive independent invention.

That makes patents more important relative to some of the moats AI is weakening.

Original Research #6: Businesses Are Already Responding—AI Patenting Grew About 34×

The acceleration is visible in the patent data.

Stanford’s 2026 AI Index reports:

Granted AI patents in 2010

3,866

Granted AI patents in 2024

131,121

PatentPC calculation:

131,121 ÷ 3,866 =

33.9×

Approximate compound annual growth over fourteen years:

28.6% per year

Global AI patent grants still grew another 8.2% from 2023 to 2024.

That does not prove companies are filing because moats are weakening.

Many causes drive patenting.

But it shows something important:

Companies are not reacting to AI commoditization by abandoning patents. They are accumulating more of them.

Original Research #7: GenAI Patenting Increased 168.5% in Two Years

WIPO’s data makes the post-ChatGPT shift even clearer.

Published GenAI patent families:

2023

approximately 14,080

2024

18,862

2025

37,808

PatentPC calculation:

37,808 ÷ 14,080 =

2.685×

Increase:

168.5%

Two-year CAGR:

63.9%

And 2024 plus 2025 produced:

18,862 + 37,808 =

56,670 patent families

WIPO’s prior-decade total was roughly 54,000.

In other words:

More GenAI patent families appeared in two years than during the entire previous ten-year period.

WIPO independently highlights the same milestone.

PatentPC previously analyzed the wider surge in its AI patent boom research.

Original Research #8: GenAI Software Patenting Grew Even Faster—4.77×

This may be the most relevant number for software founders.

WIPO reports GenAI patent families in the software/code category growing from:

339 in 2023

to:

1,616 in 2025.

PatentPC calculation:

1,616 ÷ 339 =

4.77×

Increase:

376.7%

in two years.

Now place that beside our earlier number:

AI coding benchmark performance: 16.3× improvement in one year.

These datasets do not establish causation.

But together they reveal an important market transition.

AI is simultaneously getting better at creating software while companies are rapidly filing patents around AI-generated and AI-related software technology.

The race is no longer merely:

Who can build it first?

It increasingly includes:

Who can own the technical architecture once everyone can build faster?

Original Research #9: At the 2025 Pace, About 104 GenAI Patent Families Became Public Every Day

Take WIPO’s 2025 figure:

37,808 patent families.

Divide by 365:

103.6 per day

Or:

about 104 GenAI patent families every day.

Six months at that pace:

37,808 ÷ 2 =

18,904 families

That is remarkable for another reason.

All of 2024 produced:

18,862.

So at the 2025 publication pace:

Six months creates slightly more published GenAI patent families than the entire year of 2024 did.

Again, 18,904 global patent families are not 18,904 patents that threaten your company.

Most will be irrelevant.

But the number measures how quickly the searchable landscape is moving.

And patent publications normally lag filing by roughly 18 months, meaning today’s public database does not even show the entire current filing landscape.

That gives founders a practical rule:

Do not do an AI patent search once and assume the landscape freezes.

PatentPC explains the search side in Patent Search Strategies for Startups and its guide to AI-powered patent searching in deep tech.

The Patent Race Is Not Just Defensive

Patent moats have several possible economic functions.

A patent can help a company:

Exclude

Stop certain competing implementations.

License

Turn technical exclusivity into revenue.

Cross-license

Create bargaining power against another patent owner.

Negotiate

Change leverage in partnerships or disputes.

Raise money

Give investors something legally ownable around core technology.

Sell

Create an asset that may survive product versions, employees, and code rewrites.

Deter

Make copying or close imitation legally riskier.

None of these requires the patent to be your entire moat.

The best companies usually combine several moats.

There Is Already Real Money Behind AI-Related Patent Rights

Cerence provides a useful public example.

Cerence brought two patent cases against Samsung asserting nine Cerence patents.

Samsung responded with patent counterclaims of its own.

The parties ultimately entered a cross-license agreement in October 2025.

Samsung agreed to pay Cerence:

$49.5 million

as a lump-sum IP license payment.

Cerence later disclosed approximately:

$20.7 million in related legal fees.

That gives a useful reminder in both directions.

Patent rights can produce material commercial leverage.

Patent enforcement can also be expensive.

A patent becomes a better moat when it creates negotiating leverage before a decade-long courtroom fight is required.

PatentPC examines these dynamics further in The Biggest Latest AI Patent Lawsuits and Handling AI Patent Infringement Cases.

Direct GenAI Patent Litigation Has Arrived

This is no longer theoretical.

On August 31, 2026, Many Worlds 2T Innovations filed a patent infringement case against OpenAI in the Eastern District of Texas. The complaint remains an allegation; no finding of infringement follows merely from filing a case.

One day later, another patent case was filed against ElevenLabs.

The important point is not whether those plaintiffs ultimately win.

It is this:

GenAI products have become valuable enough—and mature enough—to sit directly inside patent litigation.

AI patent strategy is moving from hypothetical portfolio planning into real commercial conflict.

But Bad AI Patents Can Be Worth Very Little

This is where founders often misunderstand the argument.

If patents are becoming more useful as moats, you might conclude:

File as many AI patents as possible.

That is wrong.

WIPO gives us a useful warning.

It reports:

37,808 GenAI patent families in 2025.

But only:

3,297 international patent families

were published across two or more jurisdictions.

That is about:

8.7%

or roughly:

1 in 11.5

GenAI families.

WIPO treats international patent families as a useful indicator of inventions applicants considered important enough to justify multi-country filing costs, although recent families may later gain foreign members.

This is not a patent-quality score.

But it exposes the main lesson:

Patent volume is not the same thing as patent moat.

A Hundred Weak Patents Can Be Less Valuable Than Five Choke-Point Patents

Imagine Company A has:

100 AI patents.

They cover:

minor UI features;
specific prompts;
narrow workflows;
outdated implementations.

Company B has:

5 patents.

They cover:

model routing;
memory architecture;
inference efficiency;
a key verification loop;
a technical mechanism every competitor needs.

Which company has the stronger moat?

Patent count cannot answer.

The real questions are:

What do the claims cover?

Will competitors need the technology?

Can they design around it?

Can infringement be detected?

Will the technical concept survive model changes?

Is the patent likely to survive validity attacks?

PatentPC’s guide to How to Get Strong Software Patents goes much deeper into this distinction.

The Most Valuable AI Patent Usually Does Not Protect the Visible Feature

Suppose your product says:

“Upload a contract and our AI identifies risk.”

That feature is easy to understand.

It may also be easy to imitate.

The valuable invention may instead be:

document segmentation

→ risk-dependent retrieval

→ multi-model extraction

→ confidence calibration

→ claim verification

→ exception routing

→ structured output

→ human escalation

The product feature is:

Analyze contracts.

The moat may sit inside:

how the machine performs the analysis.

That is where sophisticated software patent drafting begins.

PatentPC’s guides to technical improvements in software patentability and technical effects in software patents explain the distinction.

Recentive Shows What Not to Patent

The Federal Circuit’s 2025 decision in Recentive Analytics v. Fox should be required reading for AI founders.

Recentive asserted patents using machine learning to generate television network maps and schedules.

The Federal Circuit held the claims patent ineligible.

Its central problem:

generic machine learning applied to a particular problem environment, without an inventive technological improvement.

This is weak:

Use AI to price houses.

Use AI to hire employees.

Use AI to draft documents.

Use AI to schedule trucks.

Those statements describe applications.

Not necessarily inventions.

Aon Shows the Other Side

In Aon Re v. Zesty.ai, the defendant also attacked machine-learning patents under §101.

But the Delaware district court refused to dismiss the patents at that stage.

The court focused on a specific architecture involving:

two independently trained classifiers

used to analyze property characteristics and conditions from aerial images.

Compare the pattern:

Weaker story	Stronger story
AI performs task X	Specific machine architecture
Desired outcome	Technical mechanism
Generic ML	Defined arrangement of components
“Use a model”	Explain how the model/system works
Business benefit	Technical improvement

That does not guarantee validity.

But it is a much stronger patent starting point.

Desjardins Makes the Lesson Even Clearer

The USPTO’s precedential Ex parte Desjardins decision involved machine-learning technology.

The Appeals Review Panel found patent-eligible technological improvements involving AI.

The USPTO highlighted benefits including:

less storage use;
lower system complexity;
continual learning while protecting prior knowledge.

The USPTO later updated its examination guidance to emphasize evaluating technological improvements to computer functionality, data structures, learning models, and related applied fields.

The lesson is not:

“AI patents are easy now.”

It is:

A defensible AI patent should explain the technical machine, not merely the AI-powered business result.

PatentPC covers this further in AI Patent Eligibility for Machine-Learning Innovations and Machine-Learning Algorithms: Overcoming Patentability Hurdles.

Where the Real AI Patent Moats Are Forming

Think in layers.

1. Training Architecture

Potential patent targets:

continual learning;
model adaptation;
knowledge retention;
efficient fine-tuning;
distributed training;
training-memory reduction;
new data-selection mechanisms.

2. Inference

Often more commercially visible:

speculative decoding;
model selection;
batching;
caching;
inference scheduling;
compute allocation;
latency reduction;
edge inference.

3. Retrieval and Memory

A major GenAI battlefield:

retrieval architecture;
context selection;
memory creation;
memory deletion;
ranking;
retrieval verification;
multimodal retrieval;
adaptive context assembly.

4. Agent Orchestration

Agents may become highly commoditized at the surface.

The defensible layer could be:

tool selection;
permission control;
action verification;
multi-agent coordination;
recovery;
state management;
task decomposition.

5. Reliability

A flashy AI demo is not necessarily valuable technology.

Commercial systems need:

hallucination controls;
confidence estimation;
grounding;
audit trails;
deterministic recovery;
output validation;
failure escalation.

These systems may prove more durable than the model underneath them.

6. Security and Privacy

Potentially critical for enterprise AI:

private inference;
secure model execution;
prompt-injection defenses;
data-isolation architecture;
federated systems;
access-control mechanisms.

7. AI Infrastructure

Not every valuable AI patent is a model patent.

Moats may sit in:

GPUs;
DPUs;
memory;
networking;
acceleration;
distributed inference;
data centers;
cooling;
hardware/software coordination.

8. Multimodal Interaction

Speech.

Vision.

Gesture.

Documents.

Video.

Sensor data.

The valuable invention may be the architecture connecting those modalities rather than the foundation model.

9. Vertical Technical Systems

AI for:

medicine;
robotics;
manufacturing;
cybersecurity;
autonomous systems;
semiconductors;
telecom;
scientific computing.

The strongest patents usually protect a technical improvement inside the vertical, not simply “AI applied to industry X.”

The PatentPC AI Moat Test

Here is a practical framework.

Score a technical invention from 0 to 2 on seven factors.

Factor	0	1	2
Competitors need it	No	Maybe	Likely
AI makes it easy to recreate	Hard	Moderate	Easy
Technical effect	Weak	Useful	Strong
Product importance	Peripheral	Important	Core
Design-around difficulty	Easy	Moderate	Difficult
Infringement detectability	Hidden	Partial	Visible
Technical durability	Short-lived	Medium	Survives product generations

Maximum:

14 points

This is not a legal patentability test.

It is a moat-priority test.

How to Read the Score

0–4

Probably do not spend much money patenting it.

Maybe it is ordinary engineering.

Maybe secrecy is enough.

Maybe it will disappear next quarter.

5–9

Investigate.

Run prior art.

Understand business value.

Consider whether a provisional application makes sense.

10–14

This is where patent counsel should pay attention quickly.

Especially when:

the invention is about to become public;
competitors will need it;
it can be detected;
it survives model changes.

PatentPC explains early filing strategy in How to Write a Provisional Patent Application and Provisional Patents for Growing Companies.

Patent the Choke Point, Not Every Feature

Suppose you build an AI customer-support platform.

It has:

100 features.

Maybe only three create the moat.

Feature A

Beautiful chat interface.

Easy to copy.

Feature B

Special prompt.

Easy to modify.

Possibly better as secret know-how.

Feature C

A new state-reconciliation architecture that lets hundreds of autonomous tools operate on customer records without conflicting updates.

Now we have something interesting.

Competitors may need it.

It solves a technical problem.

It may survive changes in LLM vendor.

It may be detectable.

It may be difficult to design around.

That is closer to a patent moat.

The Best Test: Remove the Model Name

Take your invention disclosure.

Delete:

GPT-5

Claude

Gemini

Llama

OpenAI

Anthropic

Now read it again.

Does the invention still exist?

If no, your “moat” may simply be model access.

If yes—because the architecture, mechanism, control process, data flow, security method, or technical system remains valuable—you may have found the durable invention.

Patents Become More Valuable When the Model Is Replaceable

This sounds backwards.

But consider:

Model A today.

Model B next year.

Open model C in two years.

A patent covering only today’s model implementation may become stale.

A patent covering the durable architecture around interchangeable models may become more valuable as model competition increases.

Examples:

dynamic model routing

provider-independent memory

cross-model confidence verification

model fallback

cost-aware inference

multi-model orchestration

These systems become more useful when model suppliers proliferate.

Original Research #10: AI Investment Has Grown About 40× Since 2013

Competition is not merely technical.

Money is flooding into the field.

Stanford reports total global corporate AI investment reached:

$581.69 billion in 2025

and had grown approximately fortyfold since 2013.

Its underlying series gives about:

$14.57 billion in 2013.

PatentPC calculation:

581.69 ÷ 14.57 =

39.9×

Approximate CAGR:

36.0%

Stanford also reports:

newly funded AI companies rose 71% in 2025.

That means even if the technology itself becomes easier to reproduce, there is more capital available to the companies trying to reproduce it.

A product moat now has to survive:

better AI

more engineers

more capital

more open source

cheaper compute.

That is a much harsher competitive environment.

The Economy Itself Is Becoming More Intangible

This shift extends well beyond AI.

WIPO reports that intangible investment exceeded:

$10 trillion in 2025

across the economies in its study.

Since 2008, intangible investment grew at:

3.5% per year

versus only:

0.98%

for tangible investment.

That is:

3.6× faster

growth.

Software and databases were the fastest-growing major intangible category in WIPO’s 2013–2023 data at:

7.3% per year.

Modern company value increasingly sits in things that cannot be touched:

software;
data;
algorithms;
patents;
brands;
know-how;
organizational systems.

AI accelerates that transition.

And the more enterprise value shifts into intangibles, the more important the legal architecture around those intangibles becomes.

PatentPC discusses this investor perspective in Why Investors Care About Your IP Portfolio and How Patents Help in Valuation Talks With Investors.

But Some AI Moats Should Absolutely Remain Trade Secrets

Suppose your advantage is:

a private evaluation dataset.

A competitor cannot inspect it.

Keep it secret.

Or:

model-routing weights derived from millions of production events.

A competitor cannot easily determine them.

Maybe secret.

Or:

a proprietary customer-specific fine-tuning dataset.

Potentially secret.

Or:

a manufacturing parameter invisible from the product.

Secret protection can be powerful.

The useful test is:

Will a competitor independently discover this anyway?

If no:

trade secrecy becomes more attractive.

If yes:

patenting deserves much more serious consideration.

PatentPC has a detailed guide to Everything You Need to Know About Trade Secrets.

The Patent-or-Secret Matrix for AI

Technology	Usually start by considering
Public-facing technical architecture	Patent
Invisible tuning parameters	Trade secret
New inference mechanism detectable by benchmarking/reverse engineering	Patent
Proprietary dataset	Trade secret / contract / database rights depending on facts
Source-code expression	Copyright + secrecy
New technical model architecture	Patent
Internal prompt library	Often secrecy
Security architecture visible through operation	Patent + secrecy around implementation
Production heuristics	Trade secret
Hardware architecture	Patent
Hidden fabrication process	Often patent + trade-secret split
Brand	Trademark

The answer can be both.

Good IP portfolios overlap rights.

They do not force every innovation into one bucket.

AI Also Changes the Prior-Art Problem

Generative AI accelerates invention.

That means it also accelerates other people’s invention.

Remember:

about 104 GenAI patent families were published per day at the 2025 pace.

You cannot freeze the competitive landscape while you think.

That creates another reason to move quickly when an invention is important.

Not:

File junk immediately.

Instead:

Identify strategic inventions quickly enough to search, document, and file properly before the world moves past you.

PatentPC’s guide to building an IP portfolio for emerging-tech companies explains how to turn continuous technical development into an IP pipeline.

AI Can Help Invent—But AI Cannot Be Named the Inventor

Another misconception:

“If we used AI while inventing it, we cannot patent it.”

That is not the USPTO’s rule.

The USPTO’s revised 2025 guidance states that the normal inventorship standard applies regardless of whether AI tools were used.

Only natural persons may be named as inventors.

AI is treated as a tool in the inventive process.

The practical problem is therefore documentation.

When engineers use AI heavily, record:

what problem the human identified;
what architecture the human conceived;
which alternatives were generated;
what decisions humans made;
who conceived each claimed feature.

The more AI becomes part of engineering, the more important invention records become.

Your Company’s AI Usage May Be Creating Patentable Technology Without Anyone Noticing

Where should you look?

Not only inside “R&D.”

Look inside:

latency fixes

security fixes

inference-cost work

model migration

agent reliability

data pipelines

database synchronization

evaluation systems

deployment infrastructure

memory

retrieval

fallback logic

multimodal processing

privacy architecture

An invention often arrives disguised as:

“We had to solve this annoying production problem.”

PatentPC’s earlier article on PAT testing explains why production testing can become an invention-capture checkpoint.

The PatentPC AI Moat Audit

Once per quarter, ask engineering seven questions.

1. What did we build that was technically difficult?

Not:

“What did we launch?”

Ask:

“What was difficult?”

2. What did we build because normal approaches failed?

This often identifies the inventive step.

3. What did we make faster, cheaper, safer, smaller, or more reliable?

Technical effects matter.

4. Which technology would hurt most if a competitor copied it?

That is the commercial filter.

5. Which technology will competitors inevitably need?

That is the choke-point filter.

6. Which advantage will remain useful when today’s model is obsolete?

That is the durability filter.

7. Which important invention is about to become public?

That is the filing-clock filter.

You can do this in:

30 minutes per quarter.

That is far cheaper than discovering the invention during investor diligence two years later.

An AI Company Should Have an Invention Pipeline, Not an Annual Patent Meeting

Old process:

Once a year → ask engineers for invention ideas → file several patents.

That is too slow for AI.

Better:

Engineering

↓

technical event

↓

short invention disclosure

↓

business/IP triage

↓

prior-art scan

↓

patent / secret / ignore decision

↓

draft

↓

portfolio review

That turns patents into part of product strategy.

Not legal archaeology.

Investors Should Stop Asking “How Many Patents?”

Ask these instead.

1. Which product revenue does each family protect?

A patent disconnected from the product may add little.

2. Which claims would a serious competitor have to avoid?

That identifies leverage.

3. Can you detect infringement?

If not, enforcement becomes difficult.

4. What remains pending?

Pending continuation applications can create strategic flexibility.

5. What is kept secret instead?

A mature IP strategy has both.

6. Is ownership clean?

Employees.

Contractors.

Founders.

Universities.

Partners.

All matter.

7. Has somebody else already patented the choke point?

That is freedom to operate.

PatentPC explains the valuation side in How to Conduct an IP Valuation Investors Actually Trust and its guide to patent portfolio management in financing.

Your Own Patent Does Not Give You Freedom to Operate

This is one of the most common founder mistakes.

Suppose you patent:

B + C

But another company owns a valid earlier patent covering:

A + B

Your patent does not automatically let you use A+B+C.

A patent is principally a right to exclude.

Not a government license to operate.

That is why an AI company’s two patent questions are different:

Offensive

What can we own?

Defensive

What can stop us?

PatentPC’s FTO guide for technology CEOs and guide to conducting a freedom-to-operate analysis explain the distinction.

The Strongest AI Companies Will Probably Use Layered Moats

Patents alone are rarely enough.

Imagine two businesses.

Company A

Great patents.

No distribution.

No customers.

No data advantage.

No execution.

That is not a great business.

Company B

Great product.

Strong brand.

Customer workflow lock-in.

Unique data.

Trade secrets.

Powerful patent families around the technical choke points.

That is much harder to attack.

The future AI moat probably looks like:

distribution + data + workflow + brand + trade secrets + patents

Not:

patents instead of business execution.

That distinction matters.

PatentPC’s AI Moat Stack

A strong AI company should ask what protects each layer.

Layer	Potential moat
Customer acquisition	Brand / distribution
Workflow	Switching costs / integration
Data	Contracts / secrecy / exclusive access
Code expression	Copyright
Hidden know-how	Trade secret
Technical architecture	Patents
Community	Network effects
Regulatory position	Licenses / compliance
Hardware	Patents + know-how
Technical standards	Patents / standards position

Do not try to make one moat do every job.

Original Research #11: Only ~9% of 2025 GenAI Patent Families Were International Families

This deserves another look because it changes the filing strategy.

WIPO reports:

3,297 GenAI international patent families

in 2025.

Against:

37,808 total GenAI published families.

PatentPC calculation:

3,297 ÷ 37,808 =

8.72%

Or:

roughly one international family for every 11.5 published families.

International filing is expensive.

Companies choose markets.

That is another reason the right strategy is not:

Patent everything everywhere.

It is:

Patent the inventions that actually create exclusion value, in the markets where that exclusion matters.

Patent Portfolio Quality Is About Economic Geometry

Think about a competitor trying to reach your customers.

A useful patent portfolio changes the routes available to them.

Without patents:

Route A

Route B

Route C

Route D

With strategically placed patents:

~~Route A~~

~~Route B~~

Route C — expensive

Route D — poor performance

That is a moat.

Not because competitors cannot compete.

Because the legally available design space becomes worse.

The stronger the commercial choke point, the stronger this effect can become.

The Perfect AI Patent Moat Has Five Properties

1. Necessary

Competitors actually want the technology.

2. Durable

It survives several product generations.

3. Detectable

You can tell when somebody uses it.

4. Difficult to design around

Alternatives impose real cost or lost performance.

5. Legally defensible

The claims have meaningful support and are not merely an abstract result.

Finding technologies with all five is hard.

That is why they are valuable.

The Worst AI Patent Has the Opposite Five

It protects:

something nobody needs

in one obsolete implementation

that cannot be detected

that can be bypassed in a day

with claims vulnerable under §101 or prior art.

That patent may technically exist.

It is not much of a moat.

This Is Why Patent Drafting Quality Matters More in AI

AI evolves too quickly for sloppy applications.

A patent application drafted around:

“The system sends prompt X to Model Y”

may age badly.

The durable invention may instead be:

a model-independent mechanism that performs specific transformations, decisions, validation steps, and control operations.

The specification should explore:

alternate models;
alternate architectures;
alternate data structures;
alternate execution orders;
alternate deployment modes;
edge and cloud variants;
failure states;
fallback mechanisms.

Not to create meaningless bulk.

To capture the actual inventive concept.

PatentPC explains this approach in Software Patent Attorney: Protect Your Key Features.

A 90-Day AI Moat Plan

You do not need a two-year IP strategy exercise.

Days 1–15: Map the product

Draw:

model

data

retrieval

memory

agents

tools

security

infrastructure

interfaces

Do not use the pitch deck.

Use the real architecture.

Days 16–30: Mark the moat

For every technical layer ask:

easy to copy?

hidden?

patentable candidate?

trade-secret candidate?

third-party technology?

Days 31–45: Harvest inventions

Interview the engineers.

Ask:

“What was unexpectedly hard?”

That question is better than:

“Did you invent anything?”

Days 46–60: Run the PatentPC AI Moat Test

Score:

competitor necessity;
replication ease;
technical effect;
commercial importance;
design-around cost;
detectability;
durability.

Patent only what deserves the cost.

Days 61–75: Run prior art and FTO

Ask both:

Can we get a patent?

and:

Can we safely build this?

They are different questions.

Days 76–90: Build the layered moat

For each important asset choose:

patent

trade secret

copyright

contract

trademark

or:

do nothing.

Then repeat quarterly.

Why This Matters More for Startups Than Big Tech

A giant company can sometimes survive technology commoditization through:

distribution;
installed base;
brand;
capital;
ecosystem;
bundled products;
compute infrastructure.

A startup may have none of those.

What does it own?

Sometimes the answer is:

the technology.

That makes protecting the right technology disproportionately important.

A startup does not need 300 patents.

It may need:

five exceptional ones.

PatentPC explains the startup portfolio problem in Building an IP Portfolio for Emerging-Tech Companies.

Why PatentPC Is Built Around This Problem

Software and AI patents are difficult because there are two ways to fail.

Failure 1

The lawyer does not understand the technology.

The application describes the product at a high level.

The important architecture disappears.

Failure 2

The lawyer understands the technology but not the business moat.

The patent protects an interesting engineering detail that no competitor needs.

Neither produces strong strategic IP.

For AI and software companies, the patent attorney needs to understand:

the technical architecture

and:

why that architecture matters commercially.

PatentPC focuses heavily on software, AI, and advanced technology patents. The objective is not to turn every feature into a filing.

It is to identify the technical bottlenecks that are actually worth owning.

Explore PatentPC’s guides to AI patent strategy and business value, AI patent portfolio management, and the impact of AI on intellectual-property rights.

The Bottom Line

AI is destroying one of technology’s oldest informal moats:

the difficulty of building things.

Consider what has happened already.

Benchmark-equivalent AI inference became:

286× cheaper.

A similar language benchmark threshold became achievable with a model:

142× smaller.

AI coding performance jumped:

16.3× in one year.

The leading open/closed model gap fell:

78.9%.

Open-source AI development reached:

5.6 million GitHub projects.

The top four frontier AI companies now sit:

22 Arena Elo points apart.

Meanwhile:

granted AI patents have grown roughly:

34× since 2010.

GenAI patent-family publications increased:

168.5% in two years.

And GenAI software/code patent families rose:

376.7%.

Those trends are connected by a simple strategic truth:

When intelligence becomes abundant, ownership becomes scarce.

Code can be rewritten.

Models can be swapped.

Features can be cloned.

Engineers can move.

APIs can commoditize.

Open source can spread.

But a well-drafted patent around a commercially necessary technical mechanism creates something different:

a legally defined area competitors cannot simply reproduce because they figured out how to build it too.

That is why patents are becoming more important as AI moats.

Not because every AI company needs hundreds of them.

Because the best AI companies need to identify the few technical advantages that will still matter after everything else becomes easier to copy.

Those are the inventions worth owning.

Talk to PatentPC about building an AI or software patent moat.

PatentPC Original Research Methodology

The calculations in this article combine several public datasets. They measure different phenomena and should not be merged into a single index.

Inference-cost compression

Stanford:

$20 per million tokens → $0.07

PatentPC:

20 ÷ 0.07 = 285.7×

Price decline:

1 − (0.07 ÷ 20) = 99.65%

Model-size compression

540 billion ÷ 3.8 billion =

142.1×

SWE-bench improvement

71.7 ÷ 4.4 =

16.30×

Open-vs-closed model gap reduction

1 − (1.70 ÷ 8.04) =

78.86%

March 2026 top-four Arena spread

1,503 − 1,481 =

22 Elo points

22 ÷ 1,503 = 1.46% of the leader’s numerical Elo score

This percentage is only a scale comparison. Elo is not a percent-performance metric.

AI patent-grant growth

131,121 ÷ 3,866 =

33.92×

2010–2024 CAGR:

approximately 28.6%

GenAI patent-family growth

37,808 ÷ 14,080 ≈

2.685×

Increase:

168.5%

Approximate two-year CAGR:

63.9%

2025 GenAI publication velocity

37,808 ÷ 365 =

103.6 families/day

37,808 ÷ 2 =

18,904 per six months

GenAI software/code growth

1,616 ÷ 339 =

4.77×

Increase:

376.7%

International GenAI-family share

3,297 ÷ 37,808 =

8.72%

Equivalent:

1 international family per ~11.5 total published families

Corporate AI investment growth

$581.69B ÷ $14.57B ≈

39.9×

Approximate 2013–2025 CAGR:

36.0%

Important limitations

These calculations measure:

benchmark results;
inference prices;
repository activity;
investment;
patent publications;
patent grants.

They do not directly measure economic moat strength.

Patent families are not equal in value.

Benchmarks are not identical to real product performance.

A patent publication does not mean the patent is valid, infringed, commercially useful, or even granted.

GitHub project counts include experimental projects.

Therefore, the article uses these datasets to establish structural trends, not to claim that patents mathematically cause defensibility.

Primary External Research Sources

Recommended PatentPC Deep Dives

General information only

This article is provided for general informational purposes and does not constitute legal advice. Reading it or using this website does not create an attorney-client relationship. Consult qualified counsel about the facts and law applicable to your situation.