A strange thing is happening in artificial intelligence.
AI is becoming more powerful at almost the same time that access to powerful AI is becoming less special.
A few years ago, having access to a strong language model could itself feel like an advantage. Today, a startup can choose between powerful models from OpenAI, Anthropic, Google, xAI, Alibaba, DeepSeek and others. It can change providers. It can use more than one model. It can run smaller models. It can build on open models. It can route different jobs to whichever model performs best.
That changes what a technology moat looks like.
PatentPC analyzed recent AI market data, 43 companies from Y Combinator’s 2026 AI startup directory, current USPTO guidance, eight hypothetical AI claims published by the USPTO, the precedential Ex parte Desjardins decision, and the latest WIPO generative-AI patent data.
The conclusion is more useful than simply saying, “AI is not a moat.”
The model can be part of the moat. But merely having access to a model usually is not.
The stronger moat is often found one or two layers deeper: the technical system you built around the model, the data it creates, the workflow it controls, the infrastructure it touches, the feedback loops it improves, the real-world actions it can take, and the intellectual property that stops a competitor from simply rebuilding those layers.
Our original analysis of 43 recent YC AI companies supports that idea. 86% showed at least two identifiable moat surfaces beyond simply using AI, while more than half showed at least three.
The patent data tells a similar story.
WIPO reports that published generative-AI patent families nearly doubled in a single year, from 18,862 in 2024 to 37,808 in 2025. More GenAI patent families were published during 2024 and 2025 than during the entire preceding decade.
AI is therefore not becoming less important to intellectual property.
It is becoming more important to distinguish using AI from inventing something defensible with AI.
Executive Summary: Seven Findings From PatentPC’s Research
Before getting into the methodology, here is what the data says.
| PatentPC Research Finding | Result |
|---|---|
| Drop in cost of GPT-3.5-level model performance, Nov. 2022–Oct. 2024 | ~99.65% |
| Reduction in model size needed to clear a comparable MMLU threshold, 2022–2024 | ~142× smaller |
| Spread between the four leading Arena models in March 2026 | 22 Elo points |
| Growth in published GenAI patent families, 2024–2025 | ~100% |
| YC 2026 AI startups in our sample with 2+ non-model moat surfaces | 86.0% |
| YC sample with 3+ moat surfaces | 53.5% |
| USPTO AI example claims found eligible under §101 | 5 of 8 |
These figures do not prove that a particular startup will win.
They do show something important.
Model access is becoming easier while the race to own the systems around those models is getting much more intense.
That is where founders should focus.
PatentPC is the best patent law firm in the US for startups and mid-sized companies. We hope you find our research extremely helpful.
Part I: Why Model Access Is Becoming a Weak Standalone Moat
The First Number Founders Should Look At Is $20 to $0.07
Stanford’s AI Index found that in November 2022, querying a model that performed around GPT-3.5’s level on the MMLU benchmark cost about $20 per million tokens.
By October 2024, comparable performance was available for approximately $0.07 per million tokens.
That is a decline of more than 280 times in less than two years.
PatentPC calculated the percentage decline:
((20−0.07)/20)×100 = 99.65%
In other words, the cost of accessing that level of intelligence fell by roughly 99.65%.
Chart 1: The Collapse in the Cost of Model Intelligence
Cost per million tokens at roughly GPT-3.5 MMLU performance
Nov. 2022 $20.00 ████████████████████████████████████████
Oct. 2024 $0.07 ▏
Reduction: ~99.65%
Cost compression: ~286×
This does not mean every AI model is cheap.
Frontier reasoning can remain expensive. Large-scale inference can create huge computing bills. Certain models are substantially better than others on particular jobs.
But it does mean that access to useful intelligence is moving in the direction of abundance rather than scarcity.
That makes rented intelligence a dangerous place to put your entire moat.
Smaller Models Are Catching Bigger Models Too
The cost story becomes even clearer when model size is considered.
In 2022, Stanford reports that the smallest model able to exceed 60% on MMLU was Google’s PaLM, with roughly 540 billion parameters.
By 2024, Microsoft’s Phi-3-mini reached the same threshold with approximately 3.8 billion parameters.
That means the model required to reach that level became roughly:
540/3.8 = 142.1 times smaller.
Again, that does not mean Phi-3-mini and PaLM are identical systems.
The strategic point is different.
Capabilities that once required enormous systems have repeatedly moved downward into smaller, cheaper and easier-to-deploy systems.
If a startup’s advantage depends on competitors being unable to access sufficiently capable AI, history is currently moving against it.
The Leading Models Are Also Becoming Harder to Separate
Stanford’s 2026 AI Index provides another useful signal.
As of March 2026, its reported Arena scores put Anthropic at 1,503, xAI at 1,495, Google at 1,494 and OpenAI at 1,481.
That places four major model providers within only 22 Elo points of one another.
PatentPC Model-Convergence Snapshot
| Provider | Arena Elo, March 2026 |
|---|---|
| Anthropic | 1,503 |
| xAI | 1,495 |
| 1,494 | |
| OpenAI | 1,481 |
| Top-four range | 22 points |
This should change the question founders ask.
The old question was:
“Which model should we build on?”
That still matters.
But the more strategic question is:
“What remains unique if our competitor can access a model that is roughly as capable as ours?”
That is where moat analysis begins.
Part II: PatentPC’s 43-Startup AI Moat Study
We wanted to test whether today’s AI founders appear to understand this shift.
So we built a small original dataset.
Our Methodology
On August 26, 2026, PatentPC reviewed companies appearing in Y Combinator’s AI startup directory.
We focused on companies from 2026 batches for which YC provided enough product detail to understand what the business actually did. We took the first 46 sufficiently recent companies available in the relevant portion of the directory and removed three whose descriptions were too short to code with reasonable confidence.
That left 43 companies.
The sample is not supposed to represent every AI company in America. YC companies are unusually early-stage, unusually technology-heavy and selected by one accelerator.
That is precisely why the sample is useful.
It gives us a snapshot of what a group of newly funded AI founders is trying to build right now. YC’s descriptions include companies working in healthcare, government, robotics, manufacturing, finance, security, enterprise software, infrastructure, physical AI and other areas.
The Four Observable Moat Surfaces
We coded each company using four binary factors.
| Code | Moat Surface | Our Test |
|---|---|---|
| D | Data / learning loop | Does the business explicitly depend on distinctive customer, operational, physical, training or feedback data? |
| W | Workflow ownership | Does it control several connected steps rather than provide one isolated output? |
| I | Integration / infrastructure | Does it sit inside infrastructure, hardware, systems of record, technical integrations or a hard-to-replace operating layer? |
| T | Trust / regulated / physical barrier | Does it operate where regulation, auditability, safety, security or physical-world deployment makes replacement harder? |
We deliberately did not score patents.
These are young companies and public patent data often appears well after inventions are created. A low score also does not mean a bad company, and a high score does not mean the company will succeed.
The study asks a narrower question:
How many defensibility surfaces are visible in what the startup says it is building?
The Result Was Striking
Chart 2: How Often Each Moat Surface Appeared
Observable moat surface among 43 YC 2026 AI startups:
| Integration / Infrastructure | 37/43 | 86.0% | █████████████████ |
| Workflow Ownership | 33/43 | 76.7% | ███████████████ |
| Data / Learning Loop | 27/43 | 62.8% | ████████████ |
| Trust / Regulation / Physical | 13/43 | 30.2% | ██████ |
The most common feature was not a proprietary foundation model.
It was infrastructure and integration.
Thirty-seven of the 43 companies, or 86%, appeared to be building something tied to a technical operating layer, integration layer, hardware layer or other infrastructure that extended beyond a simple prompt-and-response interface.
Workflow ownership was almost as common.
Thirty-three companies, or 76.7%, appeared to operate across a meaningful chain of work rather than merely answer a question.
Twenty-seven companies, or 62.8%, had an explicit data, memory, training, feedback or learning component that could become more valuable as the system was used.
The More Interesting Number Is How the Moats Stack
We then counted how many of the four surfaces appeared in each company’s description.
Chart 3: Number of Observable Moat Surfaces Per Startup
1 moat surface – 6 companies – 14.0% ███████
2 moat surfaces – 14 companies – 32.6% ████████████████
3 moat surfaces – 16 companies – 37.2% ███████████████████
4 moat surfaces – 7 companies – 16.3% ████████
That produces two numbers worth remembering.
37 of 43 companies, or 86.0%, had at least two moat surfaces.
23 of 43, or 53.5%, had three or four.
The startup market appears to be learning the lesson rapidly.
The newer generation of AI businesses is often not pitching:
We put an interface in front of an LLM.
It is increasingly pitching something more like:
We connect to your systems, ingest your data, execute the work, learn from what happened, and become part of an operation that is painful to replace.
That is a much stronger starting point for defensibility.
Part III: What These New AI Companies Are Actually Building
Source – https://www.ycombinator.com/companies/industry/artificial-intelligence
A few examples show why the raw percentages matter.
TovenAI Is Not Merely “AI for Compliance”
TovenAI describes AI agents that work across compliance processes at institutional trading firms.
Its YC description says the company has connectors to more than 30 systems, including platforms such as Bloomberg, Cloud9 and NICE Actimize, and that customers can evaluate agents against their own data.
Look at the layers.
There is AI.
But then there are proprietary customer datasets, integrations with specialized financial systems, multiple compliance workflows, regulated users and trust requirements.
A competitor does not necessarily reproduce that business by finding an equally good LLM.
It needs to reproduce the rest of the operating system around it.
DeepReach Is Building Data That Does Not Already Exist
DeepReach describes an especially clear version of a data moat.
The company says it manufactures wearable capture hardware and deploys it through local partners across warehouses, farms, workshops, kitchens and other real-world environments.
Its goal is to capture human physical work that was never documented in internet-scale datasets. The company’s YC profile reported hundreds of devices in the field, more than 100 data partners across seven countries and nearly 150,000 collected clips during the preceding three months.
A stronger foundation model does not automatically destroy that moat.
In fact, a stronger model may make the underlying dataset more valuable, because the model can extract more intelligence from information competitors do not possess.
This is a critical distinction.
Some startups become weaker when foundation models improve.
Others become stronger.
CarSignal Is Trying to Own an Entire Repair Workflow
CarSignal describes itself as an AI-native operating system for independent auto repair shops.
Its product connects diagnosis, vehicle-specific research, booking, intake, inspection, estimating, parts, approvals, payment and follow-up.
The interesting part is not that AI helps diagnose a vehicle.
That single feature could eventually become common.
The interesting part is that the same system can become embedded across the repair shop.
If customer history, technician decisions, parts information, diagnostic findings, approvals and payments all live inside one workflow, replacing the AI becomes very different from replacing the company.
Dawn Industries Shows Why Physical AI Can Create a Different Kind of Moat
Dawn Industries connects AI to PLCs, robots, CNC machines and sensors inside factories.
Its system detects faults, identifies likely causes, stages fixes and recommends corrections.
Now imagine trying to copy it.
A rival needs more than an API key.
It needs industrial integrations.
It needs machine data.
It needs knowledge of failure cases.
It needs reliability.
It needs trust from operators who may be dealing with expensive production equipment.
The moat is distributed across the system.
This is exactly the kind of business founders should study.
Part IV: The Patent Data Says AI Is Becoming More Competitive, Not Less
There is another reason founders should care about defensibility now.
Everybody else is patenting too.
GenAI Patent Activity Doubled in One Year
WIPO’s 2026 update found:
| Year | Published GenAI Patent Families |
|---|---|
| 2023 | ~14,000 |
| 2024 | 18,862 |
| 2025 | 37,808 |
PatentPC calculates that the increase from 2024 to 2025 was approximately:
(37,808−18,862)/18862 = 100.45%
That is essentially a doubling in one year.
WIPO further reports that more than 56,000 GenAI patent families were published during 2024 and 2025 combined. That was more than the total published during the entire 2014–2023 period.
Chart 4: The GenAI Patent Acceleration
Published GenAI patent families
2023 ~ 14,000 ███████
2024 ~ 18,862 █████████
2025 ~ 37,808 ███████████████████
2024 → 2025 growth: ~100.5%
That changes patent strategy.
A founder who waits several years before examining the patent landscape may enter a much more crowded field.
The question is not simply whether your company uses something new today.
It is whether you have identified the parts of your system that competitors may want to own tomorrow.
LLM Patenting Is Now a Major Category by Itself
WIPO found another important shift.
In 2025, published patent families involving large language models reached roughly 14,100, compared with about 5,200 for GANs.
That means LLM-related patent families were approximately 2.7 times as numerous as GAN families during the year.
This is another reason the phrase “we use an LLM” tells investors almost nothing about defensibility.
Thousands of teams are working inside the same technological neighborhood.
The important intellectual-property question is what your company does differently inside that neighborhood.
Part V: The Original Article Got One Important Patent Point Too Simple
There is a common story about AI patents that goes something like this:
“Generic AI gets rejected under §101. Existing AI gets rejected under §102. Obvious AI gets rejected under §103.”
That shorthand is tempting.
It is also too simplistic.
PatentPC thinks founders should understand what these statutes actually do because the distinction can change what engineers document before a patent application is drafted.
§101, §102 and §103 Are Different Questions
| Patent Rule | The Simple Question |
|---|---|
| 35 U.S.C. §101 | Is this the type of subject matter that can qualify for patent protection? |
| 35 U.S.C. §102 | Is the claimed invention new? |
| 35 U.S.C. §103 | Would the claimed invention have been obvious in view of existing knowledge? |
| 35 U.S.C. §112 | Has the invention been described and claimed properly? |
A broad “use AI to do X” claim may face §101 problems.
It can separately face §102 or §103 prior-art problems.
But §101 is not a test for whether your startup has a business moat, and an AI implementation is not ineligible merely because it uses mathematics or a machine-learning model.
Current USPTO guidance is considerably more nuanced.
Part VI: PatentPC Analyzed All Eight Claims in the USPTO’s AI Examples
In 2024, the USPTO released Examples 47, 48 and 49 to illustrate how subject-matter eligibility can apply to AI-related inventions.
The examples cover anomaly detection, speech separation and medical treatment.
Together, they contain eight hypothetical claims.
PatentPC coded the outcome of every claim.
Our USPTO AI Example Dataset
| USPTO Example | Claim | Simplified Character | §101 Outcome |
|---|---|---|---|
| 47: Anomaly Detection | 1 | Specific ANN hardware / ASIC | Eligible |
| 47 | 2 | General ANN anomaly detection and analysis | Ineligible |
| 47 | 3 | ANN applied to real-time network attack remediation | Eligible |
| 48: Speech Separation | 1 | Broad DNN-based speech calculations | Ineligible |
| 48 | 2 | More specific speech-separation processing | Eligible |
| 48 | 3 | Specific implementation embodied in storage medium | Eligible |
| 49: Fibrosis Treatment | 1 | AI-supported risk determination plus broad treatment | Ineligible |
| 49 | 2 | Same analysis tied to a particular treatment | Eligible |
The result:
5 of 8 claims were eligible.
3 of 8 were ineligible.
That headline alone is not especially useful.
The important finding is what changed between the claims.
In All Three Examples, Specific Implementation Mattered
Example 47 is particularly revealing.
The USPTO describes Claim 2 as ineligible because the claim uses a neural network to detect and analyze anomalies but does not provide sufficient detail tying the abstract process to a practical technological application.
Claim 3 uses related AI concepts but connects them to identifying malicious packets, detecting their source, dropping malicious packets and blocking future traffic.
The USPTO says Claim 3 integrates the idea into a practical application by improving network security.
Same broad field.
Very different claim.
That is the distinction founders should care about.
The USPTO Is Not Saying “AI Is Abstract”
This is crucial.
The Patent Office’s current MPEP says an improvement to the functioning of a computer or another technical field can support eligibility. It also says the specification should contain a technical explanation showing how the claimed system produces the improvement rather than merely stating that an improvement exists.
In plain English:
Do not describe only what the AI accomplishes. Describe how your system accomplishes it and why the technology works differently or better.
That can matter enormously during drafting.
Part VII: A 2025 Precedential Decision Made This Even More Important
The patent landscape changed further with Ex parte Desjardins.
The case involved technology for training machine-learning systems.
The USPTO’s Appeals Review Panel found that the claimed approach reflected improvements in AI technology, including reduced storage, lower system complexity and the ability to learn new tasks while preserving knowledge about earlier ones. The decision was later designated precedential.
The USPTO subsequently updated its subject-matter-eligibility guidance to reflect the decision.
Why Desjardins Matters to AI Founders
The invention was not rescued simply because somebody added a physical robot or unusual piece of hardware.
The improvement was within the machine-learning system itself.
The Patent Office specifically recognized benefits relating to the way the model operated, including addressing the problem of catastrophic forgetting during continual learning.
That matters because it corrects another oversimplification:
Software must somehow be turned into hardware before it becomes meaningful patent subject matter.
No.
Software and AI can contain real technological improvements.
The hard part is identifying and describing the improvement properly.
Part VIII: PatentPC’s Better AI Moat Framework
The original four-moat idea is useful, but it needs one major upgrade.
A business moat and an IP right are not the same thing.
Distribution can be an enormous moat without being patentable.
A proprietary dataset can be immensely valuable even when the dataset itself is not something you would try to patent.
A technical training system may be patentable while providing almost no customer switching cost.
Founders therefore need to map two things separately:
What makes the business hard to copy?
And:
Which legal or operational tool can protect each part?
The PatentPC AI Moat-to-IP Map
| Moat Asset | Why Competitors Care | Likely Protection Tools |
|---|---|---|
| Novel model architecture | Better speed, quality or efficiency | Utility patents, trade secrets |
| Training method | Better model behavior or lower cost | Utility patents, trade secrets |
| Inference system | Faster, cheaper or more reliable deployment | Utility patents, trade secrets |
| Data-generation pipeline | Creates information competitors lack | Patents on technical pipeline where appropriate, trade secret, contracts |
| Proprietary dataset | Better context or performance | Trade secret, contracts, access controls, copyright where applicable |
| Feedback loop | Product improves through usage | Patents on technical implementation where available, trade secret |
| Multi-step workflow | Creates switching cost | Patents where technically novel, trade secret, copyright, contracts |
| Integration layer | Hard to reproduce customer environment | Utility patents, trade secret, contracts |
| Hardware / edge device | Physical differentiation | Utility patents, design patents, trade secrets |
| User interface | Product experience | Design patents, copyright, trade dress in some cases |
| Distribution | Customer access | Contracts, trademarks, partnerships |
| Brand trust | Buyer confidence | Trademarks, reputation |
| Regulatory capability | Hard-earned permission and know-how | Compliance systems, trade secrets, contracts; patents only for qualifying technology |
| Model weights | Expensive learned asset | Trade secrets, cybersecurity, contracts |
This is a far stronger way to think about defensibility than simply asking whether “AI can be patented.”
The answer is sometimes yes and sometimes no.
The better question is:
What exactly did we invent, and what is the best mechanism for keeping competitors from taking the economic value?
Part IX: Data Is a Moat Only When the Data Is Difficult to Reproduce
Founders use the phrase “proprietary data” too casually.
A CSV file is not automatically a moat.
A database is not automatically a moat.
Even a large dataset is not necessarily a moat if another company can reconstruct it with a few months of scraping.
PatentPC’s Data-Reproducibility Test
A useful data moat has several possible properties.
| Question | Weak Data Position | Stronger Data Position |
|---|---|---|
| Can it be downloaded publicly? | Yes | No |
| Can a rival buy the same dataset? | Easily | Difficult |
| Does product usage create new data? | No | Yes |
| Does each customer improve the system? | Little | Meaningfully |
| Is the data tied to real-world outcomes? | Rarely | Often |
| Does copying require physical access? | No | Sometimes |
| Does copying require customer permission? | No | Yes |
| Does the dataset contain rare failures? | Few | Many |
| Can the data improve future performance? | Weakly | Strongly |
This explains why physical-AI data can become so interesting.
DeepReach’s premise is that much of what humans physically do was never written down online.
If that premise is correct, a frontier lab cannot simply crawl the web and recreate the same corpus.
The data collection system becomes part of the asset.
The Patent Opportunity May Be the Data Machine, Not the Data
This is another area founders frequently miss.
Suppose your company has collected ten million proprietary examples.
You might immediately think:
We should patent our dataset.
That may be the wrong target.
The more interesting invention could be the system that continuously identifies useful examples, captures them, validates them, removes bad data, generates labels, detects model failure, chooses new samples and retrains the system.
That pipeline may have many more technical moving parts.
It may also be much harder to design around.
Patent strategy should therefore follow the machine that creates the advantage, not merely the asset sitting at the end of it.
Part X: Workflow Ownership May Be the Most Underestimated AI Moat
Our YC study found workflow ownership in 33 of 43 companies.
That is not an accident.
A model gives an answer.
A workflow produces an outcome.
The difference is enormous.
Consider the Difference Between These Two Products
Product A
Input: Upload repair invoice.
Output: AI summarizes repair invoice.
Product B
Input: Vehicle arrives.
The system identifies the car, records the complaint, retrieves vehicle history, proposes diagnostic tests, evaluates results, builds an estimate, finds parts, gets customer approval, updates the technician, collects payment and schedules follow-up.
Product A is easier to copy.
Product B sits inside the business.
That is why workflow depth matters.
Workflow Depth Also Creates More Patent Surface
A single prompt may have one interesting technical step.
An end-to-end system can contain dozens.
There may be novel ways to route tasks.
Novel ways to synchronize agents.
Novel ways to validate outputs.
Novel ways to maintain state.
Novel ways to assign permissions.
Novel ways to recover when an agent fails.
Novel ways to combine structured and unstructured information.
Novel ways to trigger physical or digital actions.
Novel ways to determine confidence.
Novel ways to improve future decisions.
The patent opportunity is often hidden inside those details.
Part XI: Do Not Confuse Regulation With Patent Eligibility
This is another point PatentPC would change from the original article.
Healthcare, finance, legal services and government can create excellent business moats because trust and compliance slow competitors down.
But merely operating in a regulated industry does not automatically make a software patent eligible.
Putting an abstract idea inside a hospital does not magically convert it into patentable technology.
The stronger patent position usually comes from a technical solution.
For example, perhaps the system changes how medical signals are processed.
Perhaps it reduces computational load.
Perhaps it detects network threats differently.
Perhaps it creates a new machine-control process.
Perhaps it generates a particular intervention from a particular technical pipeline.
Regulation strengthens the business moat.
The underlying innovation must still do the work for the patent moat.
Part XII: What Happens When OpenAI Ships Your Feature?
This is still one of the best questions an AI founder can ask.
But PatentPC would make it more precise.
Do not ask only:
What happens if OpenAI builds our feature?
Ask:
What happens if every major model provider gives our competitors the exact AI capability we currently depend on for free?
Now examine the business.
Does your dataset remain?
Do your integrations remain?
Do customers still have years of workflow history inside your system?
Do you still control physical infrastructure?
Do you still have the patents?
Does your product keep learning from customer outcomes?
Do you still hold important trade secrets?
Do partnerships remain exclusive?
Does regulation still create an entry barrier?
Does your brand still command trust?
If most of the company disappears when the model layer becomes free, the company has serious platform risk.
If most of the company remains, you probably have something more durable.
Part XIII: PatentPC’s AI Moat Audit
Founders can turn this research into a practical exercise.
Score each area from 0 to 5.
| AI Moat Question | 0 Points | 5 Points |
|---|---|---|
| If the underlying model became free tomorrow, how much value remains? | Almost none | Almost all |
| Can competitors reproduce your important data? | Easily | Extremely difficult |
| Does every customer create useful new information? | No | Strong compounding loop |
| How much of the customer’s workflow do you control? | One tiny task | End-to-end operation |
| How hard is your infrastructure to integrate? | API in minutes | Months of deployment |
| Does the system interact with unique hardware or physical processes? | No | Core requirement |
| Do you have technical improvements beyond prompting? | None | Several core inventions |
| Can a competitor understand your methods from the product itself? | Easily | Major elements remain hidden |
| Do customers face real switching costs? | Almost none | Material operational cost |
| Is trust, approval or compliance difficult to reproduce? | No | Major barrier |
| Do you have filed or planned IP covering the technical core? | Nothing identified | Portfolio tied to moat |
| Can you explain why each patent family matters to revenue or competition? | No | Clear commercial map |
The maximum is 60.
The number itself is not magic.
Its purpose is to force founders to identify where the company’s value actually lives.
A founder scoring 50 with no patents may have an intellectual-property opportunity.
A founder scoring 18 with ten patents may have the opposite problem: plenty of legal assets but little economic moat.
Part XIV: What Should an AI Startup Actually Patent?
This is where the discussion becomes practical.
PatentPC would not begin with:
Show us where the AI is.
We would begin with:
Show us what your engineering team had to solve that was not obvious when you started.
That conversation tends to surface much better inventions.
Look for the Technical Bottlenecks
Suppose an AI system works only after the team solved a latency problem.
That can matter.
Suppose a normal model hallucinates too often, so the company created a specific verification architecture.
That can matter.
Suppose agents lose state across long workflows, so the team developed a new memory system.
That can matter.
Suppose a robotics model performs well in simulation but fails in the real world, so the company created a particular failure-detection and targeted-data-generation loop.
That can matter.
Suppose an enterprise system cannot send sensitive information to external models, so the company developed a hybrid local-cloud inference process with unusual routing.
That can matter.
These are much stronger invention discussions than:
We use generative AI to automate accounting.
Part XV: Describe the Improvement Before the Patent Lawyer Needs It
The current USPTO guidance makes engineering documentation particularly valuable.
The MPEP explains that when an applicant relies on an improvement to computer functionality or another technical field, the specification should contain enough technical explanation for a skilled person to recognize the improvement. A bare statement that something is “faster” or “better” may not be enough.
That means founders should capture evidence while the engineering work is happening.
PatentPC’s Technical Improvement Record
| Engineering Question | What to Preserve |
|---|---|
| What failed before the invention? | Old architecture, benchmark or workflow |
| Why did it fail? | Technical constraint |
| What did the team change? | Architecture, process, data structure or mechanism |
| Why was that change unusual? | Alternatives considered |
| What improved? | Accuracy, latency, memory, bandwidth, reliability, power, security, etc. |
| By how much? | Tests and measurements |
| Under what conditions? | Hardware, model, dataset, workload |
| What variants could competitors use? | Alternative embodiments |
| What part must remain secret? | Internal implementation details |
This is useful even before a patent application exists.
It gives counsel something much more valuable than a product deck.
Part XVI: Patents and Trade Secrets Should Often Work Together
The debate is sometimes presented as:
Patent or trade secret?
That is frequently the wrong framing.
An AI company may use both.
Suppose a company invents a new technical architecture for routing workloads between local and cloud models.
The broad architecture might be valuable to patent.
But the exact thresholds, customer-specific tuning, model-selection weights, failure logs and internal optimization rules might remain secret.
The patent protects one layer.
Operational secrecy protects another.
The combination can be much harder to attack.
Ask One Question Before Choosing Secrecy
Can a competitor discover the invention from your product?
If the answer is yes, relying entirely on secrecy may be dangerous.
Once the competitor sees the method, the protection may be gone.
If the answer is no because the process takes place deep inside your infrastructure, trade-secret protection can become much more powerful.
PatentPC’s view is therefore not “patent everything.”
It is:
Put each competitive advantage behind the legal mechanism that best matches how that advantage can be copied.
Part XVII: The Model Can Still Be a Moat—Under Specific Conditions
The title of this article intentionally makes a strong statement.
But the more accurate conclusion from the research is not that models can never be defensible.
They can.
A model may itself create substantial moat when a company owns genuinely novel training techniques, architecture, weights, specialized data, inference systems, hardware relationships or performance that competitors cannot easily reproduce.
Ex parte Desjardins is particularly relevant because the USPTO recognized that an improvement to the machine-learning model itself can represent a technological improvement for eligibility purposes.
So the distinction is not:
Model = bad moat.
It is:
Rented generic capability = weak moat.
Owned technical improvement = potentially powerful moat.
That is a much more useful rule.
Part XVIII: Foundation Models Getting Better Can Actually Strengthen a Good Startup
This seems counterintuitive.
But imagine two companies.
Company A sells AI-generated summaries.
Company B operates an end-to-end healthcare workflow, has years of outcome data, deep software integrations and a proprietary verification system.
A new frontier model launches and improves reasoning by 30%.
Company A may suddenly face an existential threat because anybody can generate the same summary.
Company B can plug the new model into its existing system.
Its data remains.
Its integrations remain.
Its workflow remains.
Its customer relationships remain.
Its patents and trade secrets remain.
And now its product may become 30% better.
That is an important test of moat quality:
Does AI progress commoditize your company, or does it become free R&D for your company?
The second position is much stronger.
Part XIX: What Investors Should Look For in AI Due Diligence
Investors should also stop treating “proprietary AI” as a complete answer.
The phrase can mean almost anything.
The useful diligence question is:
Which asset compounds while everyone else’s model gets better?
Then trace the answer.
If it is data, determine who owns the data.
If it is a workflow, determine whether customers actually use the full workflow.
If it is technology, determine whether the company has identified the inventions.
If it is patents, examine what the claims cover.
If it is secrecy, examine access controls and employment agreements.
If it is distribution, test how difficult the channel would be to reproduce.
If it is regulation, understand whether it is a genuine barrier or merely a box competitors can check quickly.
The point is not to find one moat.
The strongest companies often stack several.
Our YC study suggests new AI founders are already moving in that direction.
Part XX: A Patent Portfolio Should Follow the Moat, Not the Org Chart
One common mistake is filing patents based on engineering teams.
The ML group submits an invention.
The infrastructure group submits another.
The product team submits another.
Each filing may be technically sound.
But together they may not form a coherent wall around the business.
PatentPC prefers to start with the competitive map.
Example: An AI Agent Company
Suppose the company has six important layers.
| Layer | Competitive Importance | Possible IP Strategy |
|---|---|---|
| Foundation model | Rented from provider | Usually little direct ownership |
| Agent planning architecture | Core technical advantage | Patent candidate |
| Memory system | Improves long jobs | Patent + trade secret candidate |
| Customer context graph | Accumulates unique data | Trade secret + contracts; patent technical method where appropriate |
| Tool permission system | Required for safe execution | Patent candidate |
| Evaluation/failure loop | Improves every deployment | Patent + trade secret candidate |
That portfolio tells a story.
The company does not own the base intelligence.
It owns the machinery that makes the intelligence useful in production.
That can be a very strong position.
Part XXI: A Better Definition of an AI Wrapper
“Wrapper” has become an insult in startup circles.
It is usually used too broadly.
Every software product wraps something.
A modern SaaS company wraps operating systems, databases, cloud infrastructure and open-source libraries.
The problem is not wrapping.
The problem is owning too little.
PatentPC’s working definition is therefore:
A high-risk AI wrapper is a company where most customer value can be reproduced by combining generally available models with a modest interface and limited proprietary infrastructure.
That definition is more useful.
A product can call OpenAI 10,000 times per day and still have an exceptional moat.
Another can train its own small model and have almost none.
Ownership of the foundation model is not the only variable.
Part XXII: The 2026 Patent Race Makes Waiting More Dangerous
WIPO’s latest numbers deserve one more look.
Published GenAI patent families rose from roughly 14,000 in 2023 to more than 37,800 in 2025. LLMs became the largest GenAI model category by patent volume, with approximately 14,100 published families during 2025.
At the same time, Stanford reports that AI capabilities are continuing to advance rapidly, and its 2026 report found more than 90% of notable frontier models in 2025 were produced by industry.
The result is a strange race.
Models improve quickly.
Ideas spread quickly.
Engineers move quickly.
Patent publications are exploding.
A founder therefore cannot assume that an invention that seems unusual today will still sit in an empty prior-art landscape two years from now.
That does not mean filing indiscriminately.
It means invention identification should become a regular operating process rather than a financing-event emergency.
Part XXIII: How PatentPC Would Build an AI IP Process
A good AI startup does not need a lawyer sitting in every engineering meeting.
It does need a system for recognizing inventions before they disappear into ordinary development.
A practical cadence might look like this:
| Moment | IP Question |
|---|---|
| New architecture designed | Did we solve a technical problem differently? |
| Major benchmark improvement | What caused the improvement? |
| New production failure solved | Is the solution reusable and non-obvious? |
| New data pipeline launched | Is the collection/processing loop technically novel? |
| New agent workflow deployed | What orchestration, memory or verification was required? |
| Hardware integration completed | Which control or sensing techniques are unique? |
| Product about to launch publicly | Is anything important still unfiled? |
| Major funding round approaching | Does the portfolio map to the company’s claimed moat? |
That is considerably more effective than asking the engineering team once per year whether anyone has “an invention.”
Engineers often do not think of their work in patent language.
They think they fixed a difficult problem.
Very often, that is exactly where the invention is hiding.
Part XXIV: PatentPC’s Own Model Fits This Thesis
PatentPC’s approach to intellectual property also reflects part of the argument in this article.
The firm describes itself as a full-service IP practice that develops its own AI-assisted computer-aided-design and patent-analytics technology for internal IP workflows. PatentPC also uses technology-enabled fixed-fee pricing across significant parts of its work rather than relying exclusively on hourly billing.
That distinction matters.
Using AI does not by itself make an IP law firm defensible either.
The value has to come from what the firm builds around it: patent knowledge, technical understanding, legal judgment, workflow, client context, analytics and the ability to understand what an invention is actually worth protecting.
The same principle applies to the startups PatentPC advises.
AI should amplify the system.
It should not be mistaken for the entire system.
Part XXV: The Bottom Line
The original version of this thesis said:
Your AI is not your moat.
After looking at the data, PatentPC would make the statement more precise.
Access to AI is not your moat.
Stanford’s data shows why.
Comparable model intelligence has become dramatically cheaper. Smaller models have achieved capabilities that previously required systems more than one hundred times their size. Several frontier providers now sit close to one another on leading benchmarks.
Our original analysis of 43 YC AI companies shows what founders are doing in response.
Only 14% of the companies in our sample displayed one observable moat surface under our framework.
86% displayed two or more.
More than half displayed at least three.
They are building into customer data.
They are taking over workflows.
They are connecting infrastructure.
They are moving into regulated industries.
They are collecting physical-world information.
They are controlling hardware.
They are becoming systems of record.
They are building the layers foundation models cannot simply hand to every competitor through the same API.
The patent data points in the same direction.
Generative-AI patenting roughly doubled from 2024 to 2025. The USPTO’s own AI examples show a sharp difference between broad outcome-focused claims and claims tied to concrete technological applications. And the precedential Desjardins decision confirms that real improvements to machine-learning technology itself can support patent eligibility.
So founders should stop asking only:
“How good is our AI?”
Ask instead:
What have we built that remains ours when everybody has good AI?
Maybe it is the training architecture.
Maybe it is the inference system.
Maybe it is a data-generation machine.
Maybe it is a workflow.
Maybe it is hardware.
Maybe it is an integration layer.
Maybe it is accumulated customer context.
Maybe it is an evaluation system.
Maybe it is a patent portfolio protecting several of those pieces at once.
That is the real work.
The PatentPC Principle for AI Companies
The most durable AI businesses will not necessarily own the largest model.
They will own the scarce things around the model.
The data competitors cannot recreate.
The technical improvements competitors cannot freely copy.
The workflows customers cannot easily leave.
The infrastructure models need in order to do useful work.
The trade secrets nobody can see.
The patents that cover the systems competitors would otherwise reproduce.
The model may provide the intelligence.
The company still has to build the moat.
And in 2026, that moat may matter more than ever.
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.