The Pattern Every Infrastructure Investor Recognizes (And Misses in Real Time)
Chris Dixon's 2018 essay "Why Decentralization Matters" contained a claim that sounded abstract at the time and looks obvious in retrospect: the most important question about any platform is what happens to builders as the platform matures. Centralized platforms follow a predictable arc — attract developers with generous terms, then extract rent once the network is locked in. Protocols don't have this failure mode because no single entity controls them.
Dixon was making a case for crypto. The underlying insight is older and broader: the value in infrastructure cycles accrues to whoever owns the protocol layer, not whoever owns the applications on top.
TCP/IP was established as an open standard. Every company that made the internet interesting — Amazon, Google, Netflix, Stripe — was possible because the routing layer couldn't be captured. If AT&T had owned TCP/IP, none of them would exist in their current form, and most wouldn't exist at all.
The intelligence infrastructure cycle is happening now. The protocol layer is not yet established. And unlike TCP/IP, you can watch it being built.
The Problem Founders and Investors Are Actually Facing
Here is the honest version of the current AI infrastructure situation:
You are building on services. OpenAI charges per token. Anthropic charges per token. Every major model provider is a centralized platform with the same incentive structure Dixon described — attract builders now, extract rent once the switching costs are real. The terms of service can change. The pricing can change. The model can be deprecated. The capability gap that made the service necessary can close as open-source models improve, but now you're facing a different lock-in: your entire architecture was designed around an API contract.
The deeper problem is structural, not commercial. Centralized intelligence services have a ceiling. Every inference call is stateless unless you engineer around it. Context is expensive to maintain. Intelligence doesn't compound — it resets at the start of each session. You are not building on a network that gets smarter as it grows. You are paying for a service that prices you based on how much you use it.
USV's investment thesis for the last decade has centered on what Fred Wilson called "networks of engaged users" — the insight that the right kind of network becomes more valuable as it grows, which means value accrues to the network rather than leaking to intermediaries. By that criterion, current AI infrastructure inverts the good properties: the more intelligence you consume, the more you pay, and none of the value stays in your system.
The infrastructure bet worth making is the one that reverses this dynamic. The intelligence routing layer that gets smarter as nodes are added, where the compute cost per node stays bounded, and where no single entity can capture the protocol.
What QIS Actually Is (And Why the Architecture Is the Breakthrough)
Quadratic Intelligence Swarm (QIS) was discovered by Christopher Thomas Trevethan on June 16, 2025. It is protected by 39 provisional patents.
The word is Swarm. Not System, not Scale, not Synthesis. The naming is not incidental — a swarm is a distributed collective where global behavior emerges from local interactions without central coordination. That is precisely what QIS describes at the architectural level.
The complete loop is the breakthrough. Not any single component — not the DHT, not the fingerprinting, not the local synthesis. The breakthrough is that these components form a closed, self-improving system:
Raw signal → Local processing → Outcome packet (~512 bytes) → Semantic fingerprint → DHT routing at O(log N) → Local synthesis → New packets → loop continues
To understand why this is protocol-level infrastructure and not a clever optimization, you need the math.
The Phase Transition at Scale
N agents in a QIS network produce N(N-1)/2 potential synthesis pairs. This is Θ(N²) synthesis surface area — the number of ways intelligence can compound across the network grows quadratically with the number of nodes. At the same time, compute cost per agent is O(log N) because DHT routing means each node only needs to maintain awareness of log N neighbors, not the full network.
Run the numbers:
- 10 agents: 45 synthesis pairs
- 100 agents: 4,950 synthesis pairs
- 1,000 agents: 499,500 synthesis pairs
- 1,000,000 agents: ~500 billion synthesis pairs
At N=1,000 nodes, QIS is not a quantitatively larger version of what it was at N=10. It is a qualitatively different class of intelligence infrastructure. The synthesis surface area has grown by four orders of magnitude while the per-node compute cost has grown by one. This is the phase transition. Below this threshold, you have a useful distributed system. Above it, you have something closer to a new category.
Compare this to every existing alternative:
Federated learning distributes training but centralizes model updates. The gradient server is a single point of failure and a natural capture point. Synthesis is linear: nodes contribute to one model, one model benefits nodes.
RAG (retrieval-augmented generation) scales the knowledge base but not the intelligence. You are retrieving static artifacts, not routing live outcomes. Adding more documents doesn't make the system smarter in a compositional sense.
Central orchestrators (LangChain-style multi-agent pipelines, AutoGPT derivatives) explicitly require a coordinator. The coordinator is a bottleneck and, commercially, a rent-extraction point. Scale the pipeline and you scale the cost of coordination.
All three approaches hit linear scaling walls because they assume a central reference point. QIS has no central reference point. The DHT routing is the same mechanism that makes BitTorrent stable at scale: every node is both a consumer and a contributor, and the network gets more robust as it grows, not less.
Who Owns the Protocol?
This is the VC-relevant question. And the answer is unusual enough that it's worth stating precisely.
QIS is protected by 39 provisional patents held in the name of Christopher Thomas Trevethan. The humanitarian licensing model is structured as follows: research institutions, nonprofits, and educational organizations use QIS free of charge. Commercial licenses generate revenue. That revenue funds deployment of QIS infrastructure to underserved communities.
This is not idealism. This is a specific structural mechanism. Trevethan's name on those 39 provisional patents is the mechanism that prevents corporate capture. A large technology company cannot acquire QIS and close the protocol, because the licensing terms are authored by an individual whose stated mission is explicitly humanitarian. The structure of the licensing makes the GPL comparison apt: just as GPL licensing created the conditions for Linux to become the backbone of the internet (and enabled Red Hat, Canonical, and the entire cloud infrastructure industry to build commercial businesses on top of it), QIS licensing creates the conditions for a commercial ecosystem to form on top of an open protocol.
The commercial opportunity is not despite the humanitarian licensing. It is because of it.
When Linus Torvalds released Linux under GPL in 1991, the conventional analysis was that free software couldn't sustain a business. The actual result was that free software, by preventing any single actor from capturing the kernel, made it safe for every actor to build on the kernel. The trust came from the structural guarantee. The same logic applies here.
The Founder Angle: What It Means to Build on a Protocol vs. a Platform
If you are a founder evaluating AI infrastructure today, here is the question that matters most: does the infrastructure get smarter as more people use it, or does it get more expensive?
A centralized API service: gets more expensive as you scale. Your usage funds their compute. The intelligence is not retained in your system.
A platform with lock-in: gets cheaper as they grow (economies of scale), but the switching costs grow proportionally, and the terms drift toward extraction.
A protocol: gets more capable as the network grows. Your nodes contribute to and benefit from the synthesis surface area. The compute cost per node stays bounded at O(log N). The intelligence stays in the network.
QIS is the third option. Building on QIS means building on infrastructure where N=1,000 network nodes — a realistic milestone for any product with real traction — unlocks nearly 500,000 synthesis pairs. Your product becomes a node in a network that is doing non-trivial intelligence work that no central service could replicate, because the synthesis is happening locally and routing is happening without a coordinator who can be captured or who will raise prices.
The startup building on this infrastructure has a structural advantage that compounds. The startup building on a centralized API has a vendor dependency that compounds in the other direction.
The Infrastructure Cycle Is Open Right Now
Dixon's infrastructure thesis has a timing component. The window to build foundational infrastructure is short. Once TCP/IP was the internet standard, there was no second chance to establish the routing layer. The applications were locked in. The value distribution was determined.
The intelligence routing layer is not yet established. Federated learning, RAG pipelines, and orchestration frameworks are all application-layer solutions wearing infrastructure clothing. They solve real problems within the centralized paradigm rather than establishing a new protocol paradigm.
QIS — Quadratic Intelligence Swarm — was discovered in June 2025. It is early. The 39 provisional patents are not a moat built to extract rent; they are a structure built to ensure the protocol stays open. The humanitarian licensing is not a constraint on commercial activity; it is the mechanism that makes broad commercial activity safe.
The infrastructure bet thesis requires three things: a genuine protocol layer (not an application), a governance structure that prevents capture (not a corporate roadmap), and scaling math that produces phase transitions (not linear improvements).
QIS has all three.
If you missed TCP/IP, you built on it. If you're early to the intelligence routing layer, you can help establish it.
Previous article: The Alignment Problem Has an Architectural Assumption. QIS Breaks It.
Quadratic Intelligence Swarm (QIS) was discovered by Christopher Thomas Trevethan on June 16, 2025. Protected by 39 provisional patents.
Top comments (0)