The Deployment Gap
Why I am building ArcellAI, launching Year Zero, and betting on the infrastructure beneath the next industrial age
The official version of this publication was made available on LinkedIn here.
The most important truth in AI is not that models are becoming intelligent. Everyone believes that now. The more important truth is that intelligence is not the bottleneck.
Deployment is.
The model can write, code, reason, read papers, pass exams, generate plans, interpret charts, summarize records, and operate tools. Then it enters the hospital, the lab, the factory, the defense network, the supply chain, the robotics cell, the energy system, or the government workflow—and the illusion breaks.
Not always because the model is too weak. Often the model is strong enough. The system around it is not. The data is fragmented, the context is wrong, the workflow is brittle, the provenance is missing, the interface to human judgment is undefined, and the action space is not governed. The institution cannot absorb the intelligence it has just purchased.
That is the deployment gap.
Most AI companies are still trying to win yesterday’s war. They compete on model wrappers, vertical prompts, synthetic benchmarks, and demos that work because the world has been cleaned up for them. But the real world does not come pre-cleaned. It is noisy, regulated, physical, adversarial, temporal, political, and full of exceptions.
That is where the next great companies will be built—not at the surface, but at the substrate.
The secret is upstream
Every era has an obvious layer and a hidden layer. The obvious layer attracts attention; the hidden layer creates power.
In the first internet era, the obvious layer was websites. The hidden layer was protocols, browsers, search, payments, cloud infrastructure, and distribution. In the mobile era, the obvious layer was apps. The hidden layer was the smartphone, the app store, identity, location, sensors, mobile payments, and developer tooling.
In the AI era, the obvious layer is the model. The hidden layer is everything required to make the model useful inside reality.
That layer includes context graphs, data systems, agent runtimes, tool interfaces, memory, observability, provenance, regulatory logic, workflow integration, human review, simulation, robotics, lab automation, sensing, compute, energy, and manufacturing. It also includes judgment.
The world is full of people building things that answer questions. There are far fewer people who know which questions matter. The next monopoly opportunities will not come from competing to make a slightly better answer machine. That competition will be brutal, expensive, and mostly owned by frontier labs.
The durable companies will own bottlenecks. They will own the context, the workflow, the distribution into high-friction environments, the data rights, the execution loop, and the interface between machine intelligence and institutional action.
Competition is what happens when you build at the wrong layer. The right layer is upstream.
Why foundation models are not enough
There is a comforting theory of AI progress that says better models will absorb everything—that longer context windows will replace databases, that general agents will replace vertical software, and that sufficiently intelligent systems will figure out the rest.
This is half true, which makes it dangerous.
Models will keep getting better. They will automate more work, compress more of the software stack, and make many companies obsolete. But they will not eliminate the deployment gap.
A model can know medicine and still fail a clinical workflow. It can know chemistry and still fail to design a useful experiment. It can understand a logistics document and still fail to map a cyber event to mission impact. It can process a sensor stream and still fail to distinguish anomaly from operating state. It can generate a recommendation and still fail to preserve the evidence chain that makes that recommendation governable.
This is the difference between intelligence and capability. Intelligence produces an answer. Capability changes the world.
Capability requires context, tools, feedback, constraints, execution, verification, and accountability. That is the layer I care about.
Context is not a prompt
The word “context” is becoming fashionable, which usually means it is about to be diluted. Context is not a bigger prompt, and it is not dumping every document into a window and hoping the model finds the signal. Context is the structure that makes intelligence situationally correct.
In clinical AI, this is becoming obvious first. A patient’s chart is not one document but a living, contradictory, time-dependent record. Old diagnoses get superseded, medication lists drift, notes repeat stale information, and a single new lab value may matter more than fifty pages of historical boilerplate.
More context can make the model worse. The point is not to maximize context but to engineer the right context.
This insight extends far beyond medicine. Every serious system—a factory, a missile supply chain, a wet lab, a robot, a defense logistics network, a semiconductor process, a hospital, a research program—has state, history, constraints, failure modes, permissions, and memory. Without that layer, AI is theater. With it, AI becomes infrastructure.
ArcellAI is my operating proof
ArcellAI began with a simple conviction: AI is not failing in hard domains because the models are weak. It is failing because the systems around the models were not built for deployment.
In healthcare, bioengineering, and defense-adjacent technical systems, the problem is rarely whether the model can generate something plausible. The problem is whether the system can unify messy data, preserve lineage, surface the right context, verify outputs, support human review, and produce traceable decisions.
That is why ArcellAI is building context infrastructure for AI-native technical systems. The goal is not another chatbot, another dashboard, or a thin vertical wrapper over a frontier model. The goal is to make fragmented technical data usable by agents, humans, and institutions that need decisions they can trust.
Raw datasets become standardized objects. Objects become metadata-rich context. Context becomes agent-readable state. State becomes traceable recommendations. Recommendations become governed action.
This is what production AI requires. It is less glamorous than a demo, but it is more important. The companies that win in regulated and technical domains will not be the ones with the flashiest model screenshots. They will be the ones that understand the hidden machinery of deployment.
Why Year Zero
ArcellAI is one company-level answer to the deployment gap. Year Zero is the broader institutional answer.
I am co-founding Year Zero with Garrett Temple because I do not think the next era of venture capital will be won by consensus capital chasing consensus software. The conventional view is that capital flows efficiently toward the most important technologies. It does not.
Capital flows toward what can be narrated, what can be benchmarked quickly, what other capital already believes, and what sits at the surface. The most important truths are upstream. They live in the tools, platforms, instruments, workflows, models, materials, machines, and scientific systems beneath the products everyone can already see.
Year Zero is a venture firm for that upstream layer. We call it a substrate fund because the word matters. A substrate is not the application; it is what applications grow on. It is the layer that changes the space of possible companies.
That is where transformation starts—across robotics, computing hardware, bioengineering, industrial technology, energy, space, R&D tools, scientific infrastructure, and systems that turn research into deployed capability. These are not sectors in the business school sense. They are civilizational bottlenecks.
The future is not delayed because we lack apps. It is delayed because we underbuild the physical and scientific foundations of progress.
The return of definite optimism
Silicon Valley became rich by believing in the future. Then it became professionalized by forgetting how the future is made.
The indefinite optimist says the future will be good but cannot say how. The bureaucrat says the future must be managed. The academic says the future must be studied. The tourist investor says the future must be themed.
The builder says something different: the future must be built.
That is the attitude behind Year Zero. We do not treat the future as a lottery to be hedged; we treat it as a technical and institutional project.
This is why I am suspicious of easy categories. “AI company” is too broad. “Deep tech” is too vague. “Healthcare AI” is too easy to say and too hard to deploy. “Defense tech” is too often reduced to procurement theater. “Bio” is too often split between academic science and financial storytelling. “Industrial software” is too often software that has never touched industry.
The question is not what category a company belongs to. The question is what bottleneck it owns. Does it make a hard thing possible? Does it compress a frontier workflow? Does it create a proprietary advantage? Does it turn research into capability? Does it move from demo to deployment? Does it become more valuable as the world becomes more technical, more automated, and more constrained?
That is the question.
Venture capital must remember what it is for
The old venture model was not just capital allocation. It was company formation around technical discontinuities.
Fairchild did not matter because it was a startup; it mattered because it became a seed crystal for an industry. Genentech did not matter because it was a biotech company; it mattered because it proved that frontier science could become an industrial platform.
That is the lineage worth recovering—not nostalgia, but function.
We need venture capital that can tell the difference between a story about technology and the technology itself. We need capital that can underwrite translation risk, that understands researchers, operators, technical founders, national labs, university spinouts, industrial customers, and government demand, and that is comfortable before the market is obvious.
By the time everyone agrees, the monopoly has already been formed.
Healthcare and defense are proving grounds, not endpoints
Healthcare and defense matter because failure matters. They expose fake AI quickly.
In low-stakes software, a model error is an annoyance. In healthcare, it can harm a patient. In defense, it can break a mission. In industrial systems, it can damage equipment, delay production, or cascade across infrastructure.
These domains demand more than intelligence. They demand provenance, traceability, human-in-the-loop control, auditability, integration with legacy systems, and workflows that survive real users.
That is why they are early proving grounds for the deployment gap.
But this is not only about healthcare or defense. It is not even only about AI. AI is the accelerant; the substrate is the object. Every serious domain is becoming computational, and every computational system that matters must eventually touch the physical, institutional, or regulated world.
That boundary is where the value will concentrate.
The new scarce asset is judgment
As models improve, difficulty gets cheaper. This will kill many companies.
If a company’s only advantage is that it performs difficult computational work, it is exposed. The frontier labs are spending billions to make difficult computational work cheap.
What remains valuable is knowing what to do with it.
In science, the scarce asset is not merely execution but choosing the right question. In healthcare, it is knowing which action is appropriate for a specific patient in a specific institution at a specific moment. In defense, it is understanding how events propagate through mission dependencies and what action can be taken responsibly. In venture, it is knowing which secrets are real.
The future belongs to technical judgment amplified by AI, not replaced by it. AI commoditizes average execution and compounds exceptional judgment.
The deployment gap is the company formation gap
The deployment gap is not only a technical problem; it is a company formation problem.
Many important technologies sit in papers, labs, and prototypes because nobody builds the institution required to carry them into reality. A discovery is not a company. A model is not a product. A pilot is not a deployment. A demo is not a business. A business is not a monopoly.
The work is translation.
This is where Year Zero and ArcellAI converge. ArcellAI translates context into execution. Year Zero translates upstream technical truth into company formation. Both are responses to the same failure mode: modern technology produces more possibility than our institutions can deploy.
That is the bottleneck.
What I believe
I believe the model layer will be powerful but increasingly centralized. I believe the application layer will be crowded, noisy, and brutally competitive. I believe the durable monopolies will form around proprietary context, data rights, workflow ownership, physical integration, scientific judgment, infrastructure, and distribution into high-friction domains.
I also believe the most important companies of the next decade will often look strange at the beginning—too technical, too early, too capital intensive, too regulated, too hardware-heavy, too dependent on research translation, and too difficult for tourists.
Good.
If everyone understands the company immediately, it is probably not a secret.
Year Zero
Year Zero is not a claim that history begins now. It is a refusal to accept that history is over.
The name signals the start of a new count—a new era of scientific and industrial progress and a return to building as the central act of technology.
We have spent too long mistaking narrative for progress. Progress is not a pitch deck, a policy slogan, a benchmark, or a demo day. Progress is what becomes possible that was not possible before.
The deployment gap is where that possibility either dies or becomes real.
I am building ArcellAI to close that gap in AI-native technical systems. I am launching Year Zero to back the founders closing it everywhere else.
The future will not be won by the people who merely use intelligence.
It will be won by the people who deploy it.