The controversies sparked by Upstage's Solar-OPEN-100B model at the end of 2025 appear to have been successfully resolved through Upstage's transparent disclosure of information, rigorous verification within Korea's AI industry, and constructive rebuttals. Most encouraging was the fact that, after a long time, engineers and researchers in the domestic AI ecosystem came together to actively communicate.
I previously corrected a statement made during Upstage's YouTube livestream held for clarifying the allegations, that “no patents were filed regarding Transformer." In fact, Google does own patent US10/452,978 (ATTENTION-BASED SEQUENCE TRANSDUCTION NEURAL NETWORKS, Shazeer et al.). The overwhelming response to my LinkedIn post on this matter underscored the industry's intense interest in this issue.
For me, this phenomenon—people being surprised to discover that a core technology they thought was a "public good" actually has an owner—gives me a strong sense of déjà vu.
Eight years ago, when I first brought attention to Google's AI patent (US10/417,562, Ioffe et al.) on Batch Normalization Layers which is an essential component of deep learning model training, the industry's reaction was exactly the same as now.
(Related past columns: In-depth Analysis of Google's AI Patents and Open Source Strategy)
-
Introducing Google's Batch Normalization registered patent: Google Patents & Open Source Software Part 1
-
Google's patent policy and open source license: Google Patents & Open Source Software Part 2
-
The rights relationship of Libraries provided through TensorFlow and AI patents: TensorFlow Patent Analysis
Like the saying, “history repeats itself”, Google's method of securing rights to Transformer technology which sits at the core of today's AI, is identical to its approach to Batch Normalization from 8 years ago. In this column, I would like to outline Google's patent strategy, the effectiveness of open source licenses, and realistic AI patent dispute scenarios we need to prepare for.
Patent Existence: Transformer technology is not a public good. Google holds core patent (US10/452,978) and is strengthening its rights through continuation applications.
Limitations of Open Source Licenses: The Apache 2.0 license is only valid when directly using Google's code. It does not grant unconditional permission to everyone.
Possibility of AI Patent Litigations: While Google is unlikely to immediately start suing for patent infringement, patent disputes between vertical AI companies are already becoming a reality as the AI market matures.
1. Is Google's Patent on the Transformer Architecture Available to Everyone Under the Open Source License (Apache 2.0)?
Upstage CEO Kim Sung-hoon left a comment on my post, acknowledging the patent's existence by saying, "upon verification, there is a patent but it's licensed under Apache 2.0 for use." His statement is half correct and half incorrect.
The Apache 2.0 License is fundamentally a "contract." The effect of a legal contract occurs only between the contracting parties. In other words, those who download and use the source code distributed by Google (such as TensorFlow, JAX) are granted the right to use the patents included in that code.
However, this license does not apply to third parties who independently implemented Transformer architecture by only referring to the paper ("Attention Is All You Need") without using Google's code, or to third parties who developed LLMs based on open source models which Google did not involve. This means Google can still exercise its patent rights against them if it chooses to do so.
Additionally, Google has made an Open Source Non-Assertion Pledge (OPN), promising not to file patent lawsuits against certain open source technologies. However, just as the Batch Normalization patent I analyzed 8 years ago which did not appear on this list, the Transformer patent is also absent from Google's non-assertion pledge.
In conclusion, Google has not waived its rights to this patent "to everyone." It's more like conditional immunity: 'I won't enforce my patent rights if you use my code (but you must enter my ecosystem).'
2. Then Why Isn't Google Exercising Its Rights?
Then why isn't Google enforcing their patent rights by initiating lawsuits despite clearly possessing them? This is not a matter of being "nice" but a matter of "practical benefit." The reasons vary depending on the "weight classes" of Google's targets.
First, exercising rights against the open source community or startups: "It's still time to grow the pie". At this early stage of technology adoption, aggressively exercising patent rights against startups or developers not only violates the spirit of open innovation but also harms Google itself.
To establish their technology (Transformer architecture) as the "standard" in the early ecosystem, Google needs to get as many people as possible to use it. Rather than hastily resorting to lawsuits and stifling the ecosystem, the priority should be helping build a mature market. Patent litigation begins only when the market has sufficiently ripened, damages reach hundreds of billions of won, and the benefits of monopoly exceed the cost of forgoing royalties. At the current stage, there's absolutely no practical benefit in Google suing parties not in direct competition with them.
In other words, attacking opponents of different weight classes in an early stage doesn't help Google's grand strategy of expanding the ecosystem and dominating the market.
Second, exercising rights against OpenAI, Anthropic, and big tech competitors: "Strategic patience and mutually assured destruction". Then what about Google's direct competitors?
First, there are emerging competitors like OpenAI and Anthropic. While they're innovative companies, they lack the patent portfolio to counter Google due to their short history. They might seem like easy prey for Google at first glance. However, Google is likely withholding patent enforcement through a kind of "strategic patience," under the thought that the market hasn't yet ripened enough to make litigation worthwhile.
On the other hand, what about companies of equal weight to Google, like Microsoft or Amazon? There's a very high probability that they're all substantially infringing on each other's extensive technology patents. Therefore, if Google attacks, immediate large-scale retaliation is expected.
This is a typical Mutually Assured Destruction (MAD) scenario. Because the balance of terror—"if you shoot, I shoot too"—is maintained, all-out war between giant corporations doesn't happen easily. It's worth recalling that even the epic patent litigation between Apple and Samsung began only when the smartphone market had fully matured and mutual profit infringement reached a level that could no longer be overlooked.
3. Why Does Google Persistently Secure Patents It Won't Even Use?
Google is currently refining the scope of Transformer patent rights through at least 7 confirmed continuation applications. Why invest so heavily in this patent when it's openly accessible to the open source community, yet they're unlikely to actively assert their rights against competitors?
This is because it's an "eat or be eaten" situation—if I don't own it, I become the prey. A patent portfolio is like nuclear weapons during the Cold War. It exists not to actually launch but to have deterrence—"if you attack me, I strike back." Peace between large corporations is maintained through a balance of power.
4. Can AI Companies Afford to Overlook Patents?: Patent Disputes Between Vertical AI Companies Become Reality
Eight years ago, when I revealed that a patent on Batch Normalization Layers had been granted to Google, many startup founders worried, asking "Will Google sue us?". My answer then was "the possibility is extremely low." For the same reasons mentioned earlier, most startups don't need to worry about Google-initiated lawsuits regarding the current Transformer patent.
However, this doesn't mean startups are safe from patent disputes. The real threat is not Google but your competitors next door. Patent wars between vertical AI companies competing for mature markets have already begun.
The AI-markets of specific industry sectors like healthcare, finance, and law are likely to mature much faster than the general AI market. When technology is standardized and a "zero-sum game" begins where you must push out competitors to survive, patents become the most powerful and lethal weapon.
We're already witnessing the evidence of this through a recent patent dispute in Korea's medical AI sector that stirred significant controversy. What was merely a "theoretical scenario" 8 years ago has now become reality, and PI IP LAW is currently representing a company in the midst of that fierce dispute.
For vertical AI companies, battles with giant big tech may sound like what may happen only in the distant future. But as market competition intensifies, lawsuits between competitors of similar size could soon become a reality.
I hope the discussion triggered by Upstage’s AI plagiarism controversy becomes an opportunity for AI companies to seriously consider not just the “performance” of technology but also the importance of the "rights" to protect and wield that technology.
[FAQs]
Does Apache 2.0 license give me free access to Google's Transformer patent?
No—only partially. The Apache 2.0 license grants patent rights exclusively to those who download and use Google's actual source code (TensorFlow, JAX). If you independently implemented Transformer by reading the "Attention Is All You Need" paper without using Google's code, or if you built your LLM on third-party open source models, Apache 2.0 does not protect you. Google retains full patent enforcement rights against such implementations.
Why hasn't Google sued anyone over the Transformer patent yet?
It's strategic timing, not generosity. For startups and small players, Google benefits more from ecosystem expansion than early litigation—patent lawsuits only make economic sense when the market matures and damages reach significant scale. For big tech competitors like Microsoft and Amazon, it's mutually assured destruction: all parties likely infringe each other's patents, so attacking triggers immediate retaliation. Google is playing the long game.
Should AI startups worry about Transformer patent disputes?
Yes—but not from Google. The real threat comes from vertical AI competitors in your own industry sector. As healthcare AI, fintech AI, and legal AI markets mature faster than general AI, patents become the primary weapon in zero-sum competition. PI IP LAW is currently handling exactly such disputes in Korea's medical AI sector. Your direct competitor, not Google, will likely be the one who sues you.
