papers/hanzo-ai-cloud/sections/09-provider-marketplace.tex

\section{Provider Marketplace}

\label{sec:marketplace}
% ============================================================================

Hanzo AI Cloud is not a single-operator cloud. It is a marketplace of compute
providers, each staking HANZO tokens as collateral for their service quality
commitments.

\subsection{Stake-Weighted Selection}

Providers register on-chain with:

\begin{itemize}[leftmargin=1.1em]
    \item \textbf{Hardware attestation}: Verified GPU inventory (model, VRAM,
    count) via TPM-based attestation.
    \item \textbf{HANZO stake}: Minimum 10,000 HANZO per GPU. Higher stake
    increases scheduling priority.
    \item \textbf{SLA declaration}: Committed latency (p99), uptime (monthly),
    and throughput (tokens/sec).
    \item \textbf{Pricing}: Per-token, per-image, or per-second pricing in
    HANZO.
\end{itemize}

The scheduler selects providers using a score function:

\begin{equation}
\text{score}(p) = \alpha \cdot \frac{S_p}{\sum_j S_j} +
\beta \cdot \frac{1}{L_p} + \gamma \cdot \frac{U_p}{T_p} +
\delta \cdot \frac{1}{C_p}
\end{equation}

where $S_p$ is stake, $L_p$ is historical latency, $U_p/T_p$ is uptime ratio,
$C_p$ is price, and $\alpha + \beta + \gamma + \delta = 1$ are
user-configurable weights. Users who prioritize cost set $\delta$ high; users
who prioritize reliability set $\gamma$ high.

\subsection{SLA Enforcement}

SLA violations are detected automatically and enforced on-chain:

\begin{itemize}[leftmargin=1.1em]
    \item \textbf{Latency violation}: If measured p99 exceeds declared p99 by
    $>$20\% over a 1-hour window, the provider is penalized 0.1\% of stake.
    \item \textbf{Uptime violation}: If monthly uptime falls below declared
    SLA, the provider is penalized proportionally (up to 5\% of stake per
    month).
    \item \textbf{Incorrect output}: If a verifier detects an incorrect
    inference result (via PoAI spot-checking), the provider is slashed up
    to 25\% of stake.
    \item \textbf{Compensation}: Slashed HANZO is distributed to the affected
    users as compensation, not burned.
\end{itemize}

\subsection{Compute Bidding}

For batch workloads (training, fine-tuning, bulk inference), providers submit
bids in a sealed-bid auction:

\begin{enumerate}[leftmargin=1.4em]
    \item The user submits a job specification (model, dataset size, deadline).
    \item Providers submit sealed bids (price, estimated completion time).
    \item The scheduler selects the bid with the best score (price $\times$
    time, weighted by stake and historical performance).
    \item Payment is escrowed in HANZO. Released on verified job completion.
\end{enumerate}

% ============================================================================