.png){kind=logo}
The Real IP Isn’t the Model
- May 13
- 5 min read
Cyber Thoughts Newsletter
MAY 2026
IP Theft Then and Now.
The value of intellectual property has always mattered and historically companies have gone to extraordinary lengths to protect it, even when the underlying product could eventually be reverse engineered.
Take Coca-Cola and the “secret formula.”
Even though you can analyze Coke with modern chemistry equipment, Coca-Cola still fiercely protects the recipe. The formula itself may not be impossible to replicate, but the idea of the secret formula still has enormous value.
The recipe, often referred to as “Merchandise 7X,” has never been publicly confirmed by the company. It is famously said to be known by only a tiny number of people at any given time, and the written formula is reportedly stored in a vault at the World of Coca-Cola museum in Atlanta. Some of that is probably genuine security practice, and some is part marketing mystique.
Fun fact: patents expire after 20 years. Trade secrets can last forever, as long as they remain secret.
Another company could, and almost certainly has, reverse engineered Coke. Heck, the “original recipe” was discovered decades ago in a leather notebook allegedly owned by a friend of Coke’s inventor. But that barely matters because the “secret formula” itself became part of the brand and efforts to preserve the “secret formula” are correctly understood as efforts to preserve the brand, not the formula itself.
The Modern Version: TSMC
Now let’s talk about a more modern example: TSMC.
While competitors can reverse engineer the chips coming out of TSMC fabs, that is not where the real IP lives. Two things prevent competitors from easily reproducing what TSMC does:
The tools of creation
The process itself
The magic tool.
To manufacture sub-5 nanometer chips requires extraordinarily specialized lithography machines. Think of them as darkroom equipment for silicon.
Except that one machine costs roughly $200 million.
These EUV lithography machines are built by the Dutch company ASML. They are the size of a bus, staggeringly complex, and heavily export restricted. One reason Chinese fabs have struggled to catch up to TSMC is that access to these machines is tightly controlled.
Owning the recipe is one thing. Owning the machine that makes the recipe possible is another.
The Process
But the machines are only part of the moat.
The real IP may be the manufacturing process itself.
Chip manufacturing requires incredibly controlled clean rooms, but that is just the beginning. The harder part is tuning thousands of variables that affect yield rates across wafers. Some chips will always fail. The goal is to maximize how many survive.
By way of example, TSMC’s new fabs in Arizona are improving rapidly, but yields still reportedly lag behind those of its Taiwanese facilities. Dialing in those variables over years of iteration is where the real advantage lives.
Rumor has it that SMIC, China’s leading semiconductor manufacturer, has aggressively recruited former TSMC engineers in an attempt to close the gap. Stories circulate that some engineers even use pseudonyms internally because of the sensitivity around trade secrets and process knowledge.
IP for Frontier Models
Great history lesson. What’s the point?
A few weeks ago Elon Musk admitted in court that xAI used distillation techniques involving OpenAI models while training Grok. Distillation is essentially using one model (say, ChatGPT) to help train another (Grok), and many frontier labs consider it a form of IP theft that violates API terms of service.
Chinese labs are widely suspected of doing the same thing. Some observers believe DeepSeek’s rapid progress and low training costs were heavily aided by distillation.
But this raises an interesting question:
How does a frontier AI lab protect its IP when it gives developers API access to the product itself?
One answer is that, much like TSMC, the real IP may not be the model itself. It may be the process used to create the model.
Other labs can imitate currently released models. See DeepSeek, or even xAI. But the leading labs have already built the next model internally, and are using it to help build the next-next model before anyone else ever sees it.
That creates a compounding advantage.
The best models help create even better models. Those models help create the next generation faster still. Over time that gap compounds.
And if that dynamic holds, the real moat in AI may not be the weights themselves.
It may be the factory.
Lastly, if you appreciate our highlights and heresies, follow us on Twitter and LinkedIn, we post regularly about real things worthy of your attention.
What We're Reading
Here's a curated list of things we found interesting.
Mythos and Early-Stage Cyber Investing: The Sky Is Not Falling
This is our take on Mythos and what it means for cybersecurity investing.
Cybersecurity stocks cratered on April 11. As a headline read, "Anthropic's Mythos Just Broke Cybersecurity's Business Model." The prevailing narrative condensed overnight into something clean and terrifying: Anthropic's Mythos model can find every vulnerability in every codebase, so everything we knew about cybersecurity is dead.
Napster Abruptly Pivots to AI As Its Streaming Service Goes Down
Wait! Napster is still a thing?!? And now it’s an AI agent platform??? 1999 is calling and it wants its relevant company back.
Pioneer streaming service Napster shifts away from formally streaming music to an AI-centered experience as of January 2026.
Six Reasons Claude Mythos Is an Inflection Point for AI—and Global Security
Gordon Goldstein, a Lytical advisor, wrote a great article on Mythos and what it means for global cybersecurity.
Anthropic’s new AI model has taught itself to hack into software infrastructure systems believed to be among the most secure in history. While there is no question the technology is profoundly dangerous, it is unclear if defenders will win a race against time to protect a sea of vulnerable targets.
Transactions
Deals that caught our eye.
Cisco acquires AI security startup Astrix for $400 million
The Israeli startup, backed by Menlo Ventures and Anthropic, targets growing risks from non-human identities and autonomous AI agents.
Podcasts
What we’re listening to.
What the Investment Landscape Tells CISOs About Where to Focus Next | The AI Security Circle Podcast
In this episode of The AI Security Circle, Matthias sits down with Lucas Nelson, Partner at Lytical Ventures in New York.
They talk about:
Where smart cybersecurity capital is flowing in 2026 and what it signals about the threat surface
How a hacker-turned-VC evaluates security startups
Separating signal from noise on "AI-powered" vendor claims
The areas Lucas would prioritize if he were a CISO today
About Lytical
Lytical Ventures is a New York City-based venture firm investing at the intersection of Cybersecurity and AI. We aim to be the most connected, most helpful team for founders, investors, and anyone else who cares about cybersecurity and its adjacencies.