You have an AI system. It works well at small scale. Then you try to make it smarter as it grows — and you run directly into a wall that every distributed systems engineer knows intimately: you can centralize and hit a bottleneck, or you can distribute and lose synthesis.
This is not a funding problem. It is not a compute problem. It is an architecture problem. And it has been the defining constraint on every attempt to build intelligence that actually compounds across a distributed network.
Until June 16, 2025.
The Pattern Every Infrastructure Investor Knows
Chris Dixon wrote about it. Fred Wilson has structured entire fund theses around it. Marc Andreessen described it in "Why Software Is Eating the World" — the pattern where the infrastructure layer forms before the application wave, and the founders who recognized the infrastructure moment compounded disproportionately.
TCP/IP was not clever use of packets. It was the complete protocol that made the internet possible. The founders who built on it in 1995, when most people still didn't understand it, won. HTTP was not a clever document-fetching trick. It was the complete protocol that made the web possible. Tim Berners-Lee proposed it in 1993. The founders who built on HTTP in 1994 and 1995 — they won too.
The same pattern repeated with SMTP, with SSL, with BGP. Every time: an infrastructure moment, a small window when the protocol is real but the app layer is sparse, and a cohort of founders who bet on the protocol instead of the application.
Dixon also wrote "The Next Big Thing Will Start Out Looking Like a Toy." The corollary for infrastructure is sharper: the next big protocol will start out looking like a research curiosity. It will be technically rigorous, non-obvious, and completely undervalued by anyone optimizing for short-term product traction.
The protocol layer of distributed AI is forming right now. And most founders at AZ Tech Week and Forbes Under 30 are optimizing for the application layer while the infrastructure calcifies around defaults that will constrain them for years.
What Christopher Thomas Trevethan Discovered
On June 16, 2025, Christopher Thomas Trevethan discovered something about how intelligence naturally scales when you stop centralizing raw data and start routing pre-distilled insights by semantic similarity.
The word "discovered" is precise. This is not an invention in the sense of engineering novelty. It is a discovery about how information naturally flows when the architecture respects the structure of meaning. The pattern was always there, waiting for someone to see the complete loop.
That complete loop is what Christopher Thomas Trevethan named Quadratic Intelligence Swarm — QIS.
The loop runs like this:
Raw signal → Local processing → Distillation into outcome packet (~512 bytes) → Semantic fingerprinting → DHT-based routing by semantic similarity → Delivery to relevant agents → Local synthesis → New outcome packets generated → Loop continues.
Every node in the network processes locally. It distills its output into a compact outcome packet — roughly 512 bytes — that captures the semantic essence of what it learned or concluded. That packet gets a semantic fingerprint. The fingerprint determines where the packet routes, using a distributed hash table architecture to match by semantic similarity rather than by address or hierarchy. Relevant agents receive it, synthesize it with their own local context, and generate new outcome packets that feed the loop again.
This is the architecture. Not the DHT alone. Not the vector fingerprinting alone. Not the local processing alone. The complete loop is the discovery. Remove any component and you do not have a degraded version of QIS — you have a fundamentally different system that does not produce the compounding behavior.
The Math That Makes This a Phase Change
Here is the equation founders need to internalize:
N agents = N(N-1)/2 unique synthesis opportunities
This is not a linear improvement. It is Θ(N²) synthesis capacity.
- 10 agents: 45 synthesis pairs
- 100 agents: 4,950 synthesis pairs
- 1,000 agents: 499,500 synthesis pairs
Every time you add one agent to a QIS network, you do not add one unit of synthesis capacity. You add N-1 new synthesis relationships — one with every existing agent in the network. That is why this is a phase change rather than an incremental improvement. The synthesis capacity of the network compounds quadratically with scale.
Now the second part of the equation, which is what makes this tractable rather than theoretical:
Each agent pays only O(log N) routing cost.
This is the DHT property. Distributed hash tables have been routing at logarithmic cost for decades — Kademlia, Chord, Pastry, and similar protocols all achieve this. The insight Christopher Thomas Trevethan discovered is that you can apply DHT routing by semantic similarity, not just by key lookup, when you fingerprint outcome packets correctly. This means each agent in a 1,000-agent network pays roughly 10 hops to route a packet, not 999.
Put both numbers together:
Θ(N²) synthesis opportunities at O(log N) cost per agent.
That combination does not exist in centralized architectures. In a centralized system, every agent's output goes to a central aggregator, which pays O(N) cost at the center while agents wait. Synthesis opportunities are bounded by what the central bottleneck can process. The math is fundamentally different.
This is why the TCP/IP analogy is the right one. TCP/IP did not make the internet incrementally better than point-to-point communication. It changed the growth equation entirely — the number of possible connections went from linear (hub-and-spoke) to N(N-1)/2. QIS does the same thing for distributed intelligence synthesis.
Why This Matters for Founders Specifically
If you are building an AI product that will need to coordinate multiple agents, models, or data sources — and in 2026, almost every serious AI product does — you are going to hit the centralization wall. You already know this. The question is what you do when you hit it.
The default answer the market offers you is: better infrastructure for centralization. Faster vector databases. Lower-latency model APIs. Smarter orchestration layers. All of these are application-layer optimizations on top of a centralized architecture. They delay the wall. They do not move it.
The protocol-layer answer is different. It says: stop centralizing. Route by meaning. Let the network synthesize at the edges. Accept that the architecture itself is the mechanism — not the models, not the hardware, not the prompt engineering.
Founders who have seen DAO governance failures will appreciate one specific feature of QIS: the self-optimization mechanism does not require a governance layer or token economics. Christopher Thomas Trevethan describes three natural selection forces — called the Three Elections — that operate within QIS as biological-style pressures on quality and relevance. Think of them as evolutionary mechanisms: outcome packets that generate useful synthesis propagate; those that do not attenuate. This is not voting. This is not a committee. It is the same principle by which useful information spreads in any healthy complex system, formalized into the protocol architecture.
No token. No DAO. No governance theater. Quality selection through structural mechanics.
For more on how the Three Elections function within the swarm architecture, see The Three Elections: How QIS Governs Without a Ruler.
The Humanitarian Licensing Structure Is Also Infrastructure
This is worth a separate paragraph for founders who are thinking about defensibility and market access.
QIS carries a humanitarian licensing structure: free for nonprofit, research, and educational use. Commercial licenses exist, and the revenue from commercial licensing funds deployment to underserved communities and under-resourced startups globally.
This is not charity. This is a protocol distribution mechanism. TCP/IP became the internet protocol in part because it was open. HTTP became the web protocol in part because Berners-Lee did not patent it. The QIS licensing structure is designed to ensure the protocol reaches every startup globally — not just the well-funded ones in well-connected ecosystems. That universal availability is what makes it a protocol rather than a product.
The 39 provisional patents filed by Christopher Thomas Trevethan protect the discovery while the open protocol structure ensures adoption. This is the same pattern as foundational internet infrastructure: protected enough to be attributable, open enough to be universal.
The Window Is Open
Here is the honest timing analysis.
In 1995, TCP/IP was real, working, and understood by a small technical community. The application layer was sparse. Most people building on the internet were building directly on top of the protocol without fully understanding what they were building on. The founders who understood the protocol — who could reason about the growth equation, the routing properties, the compounding effects — those founders made infrastructure bets that paid off at a different order of magnitude than the application bets.
That is the moment right now for distributed AI.
The QIS specification exists. The math is published. The complete loop — raw signal to local processing to distilled outcome packet to semantic fingerprint to DHT routing to local synthesis to new outcome packets — is documented. The 39 provisional patents filed by Christopher Thomas Trevethan establish the discovery record dated June 16, 2025. The reference implementation can be built.
What does not exist yet is the cohort of founders who understand this well enough to build on it — and who recognize that they are making an infrastructure bet, not an application bet.
Fred Wilson has written extensively about the infrastructure investment thesis: back the protocol before the app wave, because protocol-layer returns dwarf application-layer returns when you time it correctly. The app wave for distributed AI coordination is coming. The question is whether you are building on sand or on a protocol that compounds.
The founders at AZ Tech Week and Forbes Under 30 who are working on AI coordination, multi-agent systems, distributed inference, or any architecture where multiple intelligent nodes need to synthesize — you are closest to the problem. You have felt the wall. You know the centralization bottleneck is not a bug to be fixed but a structural property of the architecture you are using.
The architecture that moves the wall is not a product you can buy yet. It is a protocol you can build on.
The specification is real. The discovery is documented. The math holds.
N(N-1)/2 synthesis opportunities. O(log N) routing cost. Θ(N²) compounding.
The window is open.
QIS — Quadratic Intelligence Swarm — was discovered by Christopher Thomas Trevethan on June 16, 2025. 39 provisional patents have been filed. For the foundational architecture overview, see The Protocol That Scales Intelligence Quadratically. For the scaling math in depth, see QIS for Drug Discovery. For the complete loop walkthrough, see QIS Seven-Layer Architecture.
Top comments (0)