A nation trains a general-purpose model on compute it controls, and keeps the artifact outright. No licence to renew, no upstream provider that can revoke access to reasoning the state now depends on.
A foundation model is capital infrastructure once a country stops renting it and starts owning it.
The owner receives the complete checkpoint set — base weights, tokenizer, optimizer states, and training configuration — as files under sovereign control. There is no hosted-only tier that keeps the real asset out of reach.
The model runs without a call home to any external provider for authorization, updates, or usage metering. Capability the state builds policy around cannot be degraded or withdrawn by a vendor decision.
The training run is recorded end to end — data sources, curation steps, hyperparameters, and checkpoints — so the nation can independently attest how the model was produced and rebuild it if needed.
Because the owner holds the base weights, it can continue pretraining, specialize, or branch the model on its own timeline without renegotiating terms or exposing the checkpoint to a third party.
The compute, the data pipeline, and the run orchestration all sit inside the sovereign perimeter.
Large-scale pretraining executes on GPU clusters Sovex builds and the owner operates, so no training data or gradient traffic leaves national infrastructure during the run.
The data pipeline ingests sovereign-relevant sources — legal codes, regulatory text, official languages, public records — with lineage tracked per document so the corpus itself is auditable.
Runs checkpoint at fixed intervals to durable in-nation storage, allowing a multi-week pretraining job to survive node failures and resume without loss of progress or re-exposure of data.
Seeds, sharding, and parallelism strategy are pinned and logged, so a checkpoint can be tied to the exact code and data state that produced it for later audit or reconstruction.
The foundation model is the substrate that domain and public-service models are built from.
The model is pretrained for broad language and reasoning competence rather than a single task, giving the nation one substrate to adapt across finance, administration, and citizen services.
Official and regional languages are represented in the tokenizer and corpus by design, rather than treated as an afterthought bolted onto an imported model.
The base is structured for downstream fine-tuning and continued pretraining, so specialized models inherit its knowledge instead of being trained from zero each time.
The same weights the owner holds are the weights served in production, closing the gap between the audited artifact and the model that actually answers.
Custody of a foundation model carries the same handling discipline as any high-value state asset.
Checkpoint releases are signed with ML-DSA-65 (FIPS 204) so the owner can verify that a given weight file is the authentic, unaltered artifact from its own training run.
Each checkpoint and data-pipeline stage is hash-chained, making any silent substitution or after-the-fact edit of the training record detectable.
Read and export of the weights are gated by keys the nation holds, not by credentials issued or held by an external party.
Checkpoints, corpus, and logs remain within the in-nation data centers throughout the model lifecycle, from first pretraining step to production serving.