papers/defense/hanzo-vs-claude.tex

\documentclass[11pt,letterpaper]{article}
\usepackage{hanzo-defense}

\doctitle{hanzo.ai vs.\ claude.ai:\\Owned, Self-Hostable AI versus a Centralized Cloud}
\docid{HAI-2026-011}
\docversion{Version 1.0}
\docdate{2026}
\docauthors{Zach Kelling --- Hanzo Team}
\docorgs{Hanzo Industries\\[email protected]}

\begin{document}
\makecover

\begin{abstract}
We compare two structurally different ways to obtain frontier AI: a closed,
centrally hosted model accessed as a metered cloud service (typified by
Anthropic's Claude), and a sovereign, self-hostable stack with owned model
weights (Hanzo). The distinction is not primarily one of model quality---closed
frontier models are excellent---but of \emph{deployment model and control}.
For regulated industries, critical infrastructure, and government (including
defense), the binding constraints are data residency,
operation in disconnected and critical environments, supply-chain provenance,
post-quantum security, auditability, and cost at scale. A closed cloud service
cannot satisfy several of these by construction: it cannot run air-gapped, its
weights cannot be inspected or owned, and sensitive data must leave the
enclave to be processed. We argue that for these workloads the relevant choice
is between foreign open-weight models and an American open-source alternative,
and we position Hanzo as the latter---owning both the model (the Zen family)
and the cryptography (a production post-quantum stack). We give a dimension-by-dimension
comparison and a decision framework for when each model of consumption fits.
These constraints are most acute in financial services, healthcare, critical
infrastructure, and government systems---defense among them---but the framework
applies to any organization that must control where its data and models live.
We also make the productivity case concrete: on a regulated capital-markets build
where every engineer had the same frontier chat model, the client's own
engineering audit found that one engineer working through Hanzo's agentic stack
out-produced entire contractor teams by the audit's own unit-cost metrics---the
differentiator being the owned agentic \emph{stack}, not the model.
\end{abstract}

\paragraph{Executive Summary.}
This is not a claim that our model is smarter than Claude. It is a claim about
\emph{where} and \emph{how} you can run it. Claude is an excellent frontier model
delivered as a managed cloud service: you call it over the internet, your data
goes to the vendor, you cannot see or own the weights, and it cannot run on a
disconnected device. That is the right product for many commercial uses and the
wrong one for a workload that cannot leave a controlled environment. Hanzo is the
opposite shape: open, owned-weight models you run on your own hardware---in a
data center, at the edge, or air-gapped---with post-quantum security built in,
your data never leaving, and cost driven down because you are not paying a
per-token markup or a hyperscaler margin. America's frontier AI has narrowed to
two closed labs; the only openly available frontier weights today are largely
foreign. For regulated industries, critical infrastructure, and government
buyers---defense included---who need open, sovereign, controllable
AI, that makes Hanzo the American alternative to both the closed duopoly and
foreign open weights.

\tableofcontents
\newpage

% ============================================================================
\section{The Real Axis of Comparison}
% ============================================================================

Discussions of ``which AI is better'' usually compare benchmark scores. For
commercial chat and coding that is a reasonable proxy. For regulated and
sovereign deployment it is the wrong axis. A model that scores two points higher on a
reasoning benchmark is of no use inside a controlled environment if it can only
be reached by sending the data outside that environment.

The decisive axis is the \emph{deployment and trust model}: who holds the weights,
where inference physically runs, whether the data leaves the enclave, whether the
system works without connectivity, what cryptography protects it, and what it
costs at scale. On this axis a closed cloud service and a sovereign self-hosted
stack are not competitors with different scores---they are different categories
of thing. Claude is the leading example of the first category; Hanzo is built as
the second.

We state plainly what closed cloud AI does well: it is operationally simple
(no infrastructure to run), it tracks the frontier of raw capability, and its
provider invests heavily in alignment and safety research. None of that is in
dispute. The argument here is narrow and structural: several requirements that
are mandatory for regulated, critical-infrastructure, and government workloads
(defense among them) cannot be met by a model you
do not hold and cannot run yourself.

% ============================================================================
\section{Case Study: One Engineer on the Hanzo Stack vs.\ a Team on Claude}
\label{sec:oneeng}
% ============================================================================

The cleanest evidence for this paper's thesis is not a benchmark---it is a
controlled natural experiment that already ran in production, on a real regulated
system, and was measured by that system's own engineering audit.

\subsection{Background: a regulated platform, a sprawling estate}
The setting (anonymized) is a U.S.\ regulated securities venue---an SEC-registered
alternative trading system with broker-dealer and transfer-agent functions, among
the most compliance-heavy software domains that exists. By the time of the
engagement it had grown into a sprawling legacy estate: \textbf{200+ microservices}
across \textbf{334 repositories}, roughly \textbf{\$35{,}000/month} of public-cloud
spend, accumulated security debt, and a hard regulatory bar to clear. The
engineering organization was a set of contractor teams, and every engineer on the
program had access to the same frontier chat model (Anthropic's Claude, via
\texttt{claude.ai}).

\subsection{The intervention: same model, different stack}
The variable that differed was the \emph{tooling around} the model. One engineer
worked through Hanzo's owned, agentic stack (\texttt{hanzo.ai})---an orchestration
layer that turns the model into an autonomous build--test--ship loop across hundreds
of repositories---while the contractor teams used the model the conventional way, as
an interactive assistant alongside hand-driven development. This is \emph{not} a
claim that one group had a smarter model: everyone held the same frontier weights.
What separated the outcomes was whether the model was wrapped in an agentic stack or
merely chatted with.

\subsection{The measured comparison}
The platform's own engineering audit covered 334 repositories and
31{,}058 commits over an 81-day window, attributing every commit from raw version-control
history and normalizing cost against each contributor's monthly spend. We report
the three metrics the audit treats as defensible: total commits shipped, unit cost
per line of code (\$/LoC), and shipping rhythm per dollar of monthly spend
(commits per \$1{,}000). The three contractor teams---\textbf{30 engineers} in
total---are anonymized as Team~A, Team~B, and Team~C.

\begin{table}[H]
\centering
\small
\begin{tabular}{l r r r}
\toprule
\textbf{Who (all on the same model)} & \textbf{Commits} & \textbf{\$ / LoC} & \textbf{Commits / \$1k} \\
\midrule
\textbf{1 engineer} on Hanzo (\texttt{hanzo.ai}) & \textbf{15{,}294} & \textbf{\$0.0006} & \textbf{588} \\
\midrule
Team~A --- 4 engineers (\texttt{claude.ai}) & 2{,}112 & \$0.179 & 105 \\
Team~B --- 6 engineers (\texttt{claude.ai}) & 498   & \$0.560 & 10  \\
Team~C --- 20 engineers (\texttt{claude.ai}) & 1{,}415 & \$0.153 & 32  \\
\bottomrule
\end{tabular}
\caption{The client's own engineering audit, anonymized. Everyone used the same
frontier model (Claude); only the single engineer also had Hanzo's agentic stack.
That one engineer shipped more commits than all \textbf{30 engineers combined}, at a
cost per line of code roughly $250\times$ to $900\times$ lower. One person on
\texttt{hanzo.ai} out-built thirty people on \texttt{claude.ai}.}
\end{table}

\paragraph{Reading the numbers honestly.} The audit is explicit that the raw
lines-of-code figure for the agentic-stack engineer is inflated by framework-wide
refactors, bulk dependency upgrades, and automation identities operating under the
same attribution. We therefore do \emph{not} headline raw lines authored; the
defensible signals are (i)~the \emph{commit count}, which reflects discrete units
of shipped, reviewed work; (ii)~the \emph{cost per line}; and (iii)~the
\emph{relative ratio between groups}, since every group's lines were counted by the
identical method, so whatever inflation exists applies uniformly and the
$250\times$--$900\times$ unit-cost gap survives it. Even read at its most
conservative, one engineer plus an agentic stack delivered a team's worth of
audited, shipped output at a fraction of the cost per line.

\subsection{The outcome: it shipped}
The comparison is not about commit velocity in the abstract---the same engagement
\emph{shipped the platform}, and cleared the bar that matters. Over the window the
estate was consolidated from \textbf{200+ microservices into 3 compiled binaries};
the codebase went from \textbf{3.07M lines to 1.06M active} ($\sim$3.5\,M lines of
dead code deleted); a source audit of \textbf{93 findings} (22 critical) plus
\textbf{$\sim$780 dependency vulnerabilities} was driven to \textbf{zero actionable};
and cloud spend fell \textbf{$\sim$90\%} ($\sim$\$35K/mo $\rightarrow$
$\sim$\$3.5K/mo, $\sim$\$380K/yr). The system cleared \textbf{FINRA/SEC approval} and
\textbf{SOC~2}, in record time---with the bulk of that work carried by the single
agentic-stack engineer; a focused \textbf{10-day remediation sprint} alone produced
\textbf{3{,}694 commits} and \textbf{100\% of the critical- and high-severity
security fixes}. The full cost-and-consolidation breakdown is its own companion case
study; the point here is narrower but decisive: the productivity gap above was
\emph{not} bought by cutting corners---it cleared a regulated, externally audited
production bar that the contractor teams, on the same model, did not.

\begin{table}[H]
\centering
\small
\begin{tabular}{l r r}
\toprule
\textbf{Dimension} & \textbf{Before} & \textbf{After} \\
\midrule
Microservices            & 200+                & 3 compiled binaries \\
Codebase (active)        & 3.07M lines         & 1.06M lines ($-$70\%) \\
Cloud spend              & $\sim$\$35K/mo      & $\sim$\$3.5K/mo ($\sim$90\%) \\
Open critical/high findings & 93 (22 critical) & 0 actionable \\
Dependency vulnerabilities  & $\sim$780         & 0 actionable \\
Regulatory status        & legacy, unapproved  & FINRA/SEC approved, SOC~2 \\
\bottomrule
\end{tabular}
\caption{The engagement outcome, anonymized. The productivity comparison was
delivered \emph{into} this result---a regulated, audited production system---not
alongside a prototype.}
\end{table}

\paragraph{Why the stack wins (the mechanism).} The audit is one measured data
point; the reason to expect it to reproduce is structural, not anecdotal.
Interactive chat is human-paced and serial: a prompt is written, an answer is read,
code is pasted, the loop turns once, and the engineer is the bottleneck on every
turn. An agentic stack inverts the bottleneck---the model is wrapped in an
autonomous build--test--ship loop that reads the repository, makes the change, runs
the tests, reads the failures, and iterates, across many repositories at once,
while the engineer supervises rather than types. Throughput then scales with the
orchestration and the available compute, not with one person's typing speed.
Three structural factors compound:
\begin{itemize}
  \item \textbf{The model is one component, not the product.} The stack supplies the
  memory, tool use, sandboxes, retrieval, identity/secrets, and CI integration that
  turn a good answer into a merged, tested commit. Chat supplies only the answer; a
  human supplies the rest, by hand.
  \item \textbf{Parallelism over serialism.} Many agent loops run concurrently
  across the codebase; a chat session is one conversation at a time. The
  productivity gap is the gap between a fleet and a single operator.
  \item \textbf{Owned and unmetered.} Self-hosted on owned hardware, the loop runs at
  full throttle without a per-token meter deciding how much iteration is
  affordable---and iteration is exactly what produces tested, reviewed work.
\end{itemize}
The same frontier weights, accessed two ways, land in two different productivity
regimes. The model is a commodity input; the agentic stack is the durable
advantage---which is why the gap is a property of the \emph{method}, reproducible by
any team that adopts the stack, not a property of one exceptional engineer. The
measured $250\times$--$900\times$ is the proven floor; the theoretical case says it
travels.

\paragraph{What switching gets you.} For an engineering organization the practical
read is direct: adopting Hanzo's stack force-multiplies the engineers already on
payroll---same people, same frontier model, now wrapped in an owned orchestration
layer---so legacy modernization, regulated rebuilds, and net-new work all ship at a
fraction of the cost per line, on infrastructure the organization owns and can
audit. The switch is low-friction (the stack wraps the model access a team already
has) and the gain is structural (it is the method, not the model). That is the case
for moving from renting a chat window to owning an agentic stack.

\paragraph{What this is, in one line.} This is the \texttt{hanzo.ai}-versus-\texttt{claude.ai}
comparison made concrete. It is not ``our model is smarter''---everyone had
Claude. It is that an \emph{owned agentic stack} turns one engineer into a team's
worth of shipped, audited output, at a fraction of the cost per line. The same
ownership that buys sovereignty in the rest of this paper---self-hostable weights,
post-quantum security, data that never leaves the enclave---is what buys this
productivity edge. You own the stack; you do not rent the chat. A buyer adopting
Hanzo gets both halves of that sentence at once.

% ============================================================================
\section{Dimension-by-Dimension}
% ============================================================================

\begin{table}[h]
\centering
\small
\begin{tabular}{p{0.27\textwidth} p{0.32\textwidth} p{0.32\textwidth}}
\hline
\textbf{Dimension} & \textbf{Closed cloud AI (e.g.\ Claude)} & \textbf{Hanzo (sovereign OSS)} \\
\hline
Model weights & Not disclosed; not transferable & Owned (Zen family, 0.8B--1T+); inspectable \\
Where inference runs & Vendor cloud only & Your hardware: data center, edge, air-gapped \\
Your data & Leaves the enclave to the vendor & Never leaves; stays where it lives \\
Disconnected / offline & Not possible (requires connectivity) & First-class (offline on embedded ARM) \\
Provenance & Vendor-controlled & American-owned, full provenance \\
Post-quantum security & Not a product property & Native (ML-KEM-768 + ML-DSA-65) \\
Auditability & Black box & Open stack; cryptographically signed actions \\
Customization & Prompting / limited fine-tune via API & Full fine-tune on your data, in your enclave \\
Cost model & Per-token metered & Owned + self-hosted; no markup or cloud margin \\
Lock-in & High (API + data gravity) & Low (you hold the stack) \\
\hline
\end{tabular}
\caption{Closed cloud AI versus a sovereign self-hostable stack. The comparison
is of deployment model, not of raw benchmark quality.}
\end{table}

\subsection{Data residency and the enclave boundary}
The single most consequential difference is the enclave boundary. To use a closed
cloud model, the data being reasoned over must cross the boundary to the vendor.
For protected health information, customer financial records, regulated
controlled data (CUI included), engineering files, or operational data this is often
simply prohibited. Hanzo runs the model inside the boundary, so the data never
crosses it. This is not a policy preference that can be waived with a contract
clause; it is a property of where computation physically happens.

\subsection{Disconnected and offline operation}
A cloud service is unavailable the moment the link drops. On a remote industrial
site, aboard a vessel, in an air-gapped facility, or at any field location where
connectivity is intermittent or denied---disconnected defense operations
included---there is no link to call.
Hanzo's models are quantized to run fully offline on hardware ranging from a
laptop to an embedded processor, so capability persists in exactly the
environments where a cloud service cannot operate at all.

\subsection{Provenance and the supply chain}
For an open, controllable alternative to a closed cloud model, today's practical
options are foreign open-weight models or Hanzo. Regulated enterprises and
government buyers (defense among them) care about who
produced the weights and whether they can be trusted and inspected. Hanzo's
position is specific: an American-owned, open-source frontier stack with owned
weights, so an organization can have open, auditable AI without depending on
either the closed domestic duopoly or foreign-origin weights.

\subsection{Post-quantum security}
Security against a future quantum adversary is a property of the whole stack, not
a feature of a chat API. Hanzo's transport, identity, and key management are
post-quantum by construction (all three NIST standards in production, formally
verified). A closed chat service does not expose this as something the customer
controls.

\subsection{Cost at scale}
A metered API charges per token on every inference, forever, with the vendor's
margin baked in. Owning the model and self-hosting removes both the per-token
markup and the hyperscaler margin; combined with an efficient transport layer
(the ZAP protocol's 91\% memory and 96\% energy reductions) this is the mechanism
behind roughly an order-of-magnitude lower infrastructure cost at scale. A
companion case study documents a regulated client whose cloud spend fell
approximately 90\% after migrating to this stack.

\subsection{Agentic throughput, restated as cost}
The productivity comparison in Section~\ref{sec:oneeng} is also a cost argument. Hanzo's founding
CEO/CTO, Zach Kelling, was---by public usage tracking (\texttt{claudecount.com})---the
\textbf{\#1 user of Anthropic by token count worldwide in 2025}; the audited output
behind that throughput---one engineer outproducing 30 engineers on the
same frontier model---is documented above. The point here is what changed
afterward: that same agentic throughput now runs on Hanzo's \emph{owned},
self-hostable stack rather than a metered cloud API. The capability that topped a
closed provider's global usage is exactly the capability Hanzo packages as a
sovereign, owned alternative, so a buyer gets the productivity without the
per-token meter, the data egress, or the dependency---frontier-grade agentic
output, owned instead of rented.

% ============================================================================
\section{When Each Model Fits}
% ============================================================================

This is a decision framework, not a verdict. Closed cloud AI is the right choice
when the workload is non-sensitive, connectivity is reliable, operational
simplicity outweighs control, and there is no requirement to own or inspect the
model. A sovereign self-hosted stack is the right---often the only---choice when
any of the following hold: the data cannot leave a controlled environment; the
system must work disconnected or in degraded conditions; provenance and
auditability are required; post-quantum protection is mandated; or per-token cost
at scale is untenable. Regulated industries, critical infrastructure, and
government workloads---defense included---concentrate in the second
column. That is the gap Hanzo is built to fill.

% ============================================================================
\section{Conclusion}
% ============================================================================

The honest comparison is not ``Hanzo's model beats Claude.'' It is that a closed
cloud model and a sovereign self-hostable stack are different categories, and the
requirements of regulated, critical-infrastructure, and government workloads
(defense among them)---data residency, disconnected
operation, provenance, post-quantum security, auditability, and cost---select the
second category by construction. Within that category the choice narrows to
foreign open weights or an American open-source alternative. Hanzo is built to be
that alternative: own the model, own the cryptography, run it where the data
lives.

\end{document}