US Non-Dominance
Within three years the best model available to an Australian enterprise for a given task will, for a meaningful share of tasks, not be American — and AU procurement, which assumes US-default, is not ready for that.
A validation run. Researched position, no experiment.
Confidence
50%human-committedExpiry
35duntil review · 8 Oct 2026Lead time
—not yet mainstream · opened 10 Mar 2026Ownership
AWAdam Witanowskimonthly cadenceWhere it is
A conviction bet with a named person attached. The lab director opened this on judgement in March with no supporting run, and the record will score it. What has accumulated since: Chinese open-weight releases (DeepSeek, Alibaba's Qwen line) reached parity with US open weights on our task evals in Q2 and lead on cost; Mistral and two Gulf-funded labs now ship models AU regulators would have to form a view on; and US export controls on accelerators are reshaping where inference capacity sits, including in Southeast Asia. The strongest counter-evidence is also clear: frontier capability is still US-default, the gap at the top has not closed, and every client procurement framework we have read names US providers or hyperscalers by default. The gate is adoption. The models exist and are cheap; the question is whether an AU bank's risk function will run a Chinese open-weight model on customer data, and the current answer is no regardless of the evals.
Why a Quantium decision hinges on it
Quantium's model recommendations are read as neutral. If the cheapest parity model for extraction is a Chinese open-weight and the firm cannot recommend it because of a procurement reflex, the firm is leaving money on the table and saying so to clients. If the firm recommends it and a regulator or a board objects, that is a reputational event. Either way the firm needs a position — on provenance, on weights-hosting, on what 'sovereign' means when the weights are Chinese and the hosting is Australian — before a client asks in a way that cannot be deferred. The sponsor holds the question of whether the firm wants a public position at all.
Field attributes
Position
What is demonstrated, what is hype, what would have to be true.
The shape every position request answers. Signal-tier fields carry a draft; assessed and tested fields carry a validated one.
- 01Two Chinese open-weight models matched the best US open-weight on our extraction and classification evals in Q2 at 0.4–0.6× the hosted price (x-open-weight-parity, cross-referenced).
- 02A Gulf-funded lab's model is now hosted in an AU region by a domestic provider with residency terms; the weights are not American and the hosting is Australian.
- 03Three of five client procurement frameworks reviewed name US hyperscalers or providers as default-approved and have no path for a non-US model to be assessed.
- 01'China has caught up' at the frontier. On agentic and long-horizon tasks the gap is real and has not closed on any eval we run.
- 02Sovereign AI meaning a national model. No AU-trained model is competitive and none is planned at scale; sovereignty here means hosting and control, not origin.
- 03Export controls as a permanent US advantage. Every control so far has been followed within two quarters by a workaround in capacity or architecture.
- 01An AU regulator or a major bank's risk function would have to approve a non-US open-weight model on customer data, in writing, at least once.
- 02Provenance assurance for open weights — who trained it, on what, with what backdoor risk — would have to exist as a service a client can buy.
- 03The parity on extraction and classification would have to extend to at least one agentic task class, or the field stays confined to the cheap tier.
- 01Publish pos-sovereign-inference as the firm's position: recommend by eval and cost, disclose provenance, and treat hosting as the sovereignty control.
- 02Build a provenance checklist for open-weight models and run every recommended model through it; the checklist is worth more than the position.
- 03Score the conviction in March 2027: was the director right, and by how much?
Signals · 9 in this cluster
What the cluster is made of.
Every item carries its source, tier and sightings. Detector-found signal sits beside human drops; downstream they are indistinguishable except by provenance.
Cross-reference from x-open-weight-parity: two Chinese open weights match US open weights on extraction and classification
Eval run across the firm's extraction and classification tasks. Qwen and DeepSeek releases match the best US open-weight within noise and trail the frontier by the same margin. Agentic task evals show a persistent gap. Borrowed from the open-weight field with its tier.
extracted claimNon-US open weights are at parity on the cheap tier and behind on the agentic tier.
Agentic long-horizon benchmark: US frontier leads non-US by 14–22 points
Multi-step tool-use benchmark. US frontier models lead every non-US entry by a wide margin and the gap has been stable for three releases. The counter-evidence the field carries deliberately.
extracted claimThe frontier gap on agentic tasks has not closed.
Panel: 'Sovereign AI for the Indo-Pacific'
Demand-band signal. Every panellist said sovereignty; none said whose weights. The absence is the signal.
Sleeper Weights: Detecting Training-Time Backdoors in Open Language Models Is Currently Infeasible
Shows that trigger-conditioned behaviours can be trained into open weights and survive fine-tuning, and that current detection methods do not find them. Applies to any open weight regardless of origin; makes provenance an assurance question rather than a political one.
Slack drop: three of five client procurement frameworks name US providers as default-approved
A sector owner's review of procurement documents across engagements. No assessment path for a non-US model in three of five. The procurement-gate claim rests on this.
Alibaba releases Qwen open weights with an Apache licence and an AU-region hosted endpoint
Open weights with a permissive licence, hosted in Sydney by a domestic provider at about half the price of the US-open equivalent. The hosting is the news: a non-US model with AU residency.
extracted claimNon-US open weights are available hosted in AU regions at a cost advantage.
'Follow the megawatts: where inference is actually being built in 2026'
Maps announced data-centre capacity by jurisdiction. Non-US capacity share rising; AU share small but growing. Used as the demand-side echo of the export-control signal.
'Would you ever recommend a Chinese model to us? Be honest.'
Asked by a deputy secretary in a sovereignty workshop. The honest answer was 'for some tasks, with provenance disclosed, hosted here' and the room went quiet. This is the adoption gate in one question.
US Commerce expands accelerator export controls; Singapore and UAE inference capacity announcements follow within a quarter
Control expansion followed by capacity build-outs in jurisdictions AU data could be processed in. Geopolitical signal; what matters for the field is where cheap inference ends up.
Claims · 4 supporting, 1 refuting
The atoms.
A document cannot go stale; an assertion can. Claims are immutable and stamped with the extractor that produced them, so staleness, diffs and the graveyard operate at claim level.
Chinese open-weight models are at parity with US open weights on extraction and classification tasks and cost less to run hosted.
AU enterprise procurement frameworks default to US providers and have no assessment path for a non-US model; the gate is procurement, not capability.
Provenance risk in open weights — training data, backdoors, licence terms — is unassessable with current tooling and is the real objection behind 'not Chinese'.
Export controls on accelerators are moving inference capacity toward Southeast Asia and the Gulf, which changes where AU data can cheaply be processed.
The frontier capability gap on agentic and long-horizon tasks has not closed and non-US models are not competitive there.
Position history · the diff is the product
3 validation runs against a fixed brief. Confidence 40% → 50%.
Parity confirmed on extraction and classification; frontier gap confirmed on agentic tasks. Provenance is the real objection. Position drafted for the sponsor's decision on whether to publish.
- Chinese open-weight models are at parity with US open weights on extraction and classification tasks and cost less to run hosted.
- The frontier capability gap on agentic and long-horizon tasks has not closed and non-US models are not competitive there.
- Provenance risk in open weights — training data, backdoors, licence terms — is unassessable with current tooling and is the real objection behind 'not Chinese'.
- c-us-non-dominance-4 ↓ 0.62 → 0.55
Scoring · ordinal bands
Agents propose. A named human commits.
Uncommitted scores are visibly marked and never leave the building. Bands, not point estimates — false precision is the tell that a number was generated rather than derived.
Impact
committed · RMTouches every model recommendation and the firm's public neutrality.
Timeline
committed · AWThe models are here; the procurement change is years. Conviction that it comes sooner.
Cost of being wrong
committed · LFA wrong public position either loses the cheap tier or invites a provenance incident.
Demand
committed · ABGovernment asks about sovereignty constantly and about non-US models never; the question is latent.
TAM
agent-estimatedNot a market; a condition on every market. Agent-estimated as unmeasurable.
Cost of entry
agent-estimatedA position paper and a provenance checklist. Agent-estimated.
Relevance · per vertical
Why it matters here, or explicitly does not.
Ranking is per vertical, not global. Sector owners commit notes against agent drafts.
Sovereignty is the stated priority and the frameworks default to US hyperscalers; the contradiction will surface in the next procurement cycle.
Mechanism · Provenance-disclosed model selection with AU hosting as the sovereignty control; position shared with the DTA on request.
No bank will run a Chinese open-weight on customer data this year; the question is whether one will run it on non-customer workloads and what APRA says when asked.
Mechanism · Non-customer workloads only, with the provenance checklist, until a regulator view exists.
Every model recommendation the firm makes is exposed to the provenance question.
Mechanism · Provenance line on every recommendation in the library; agent draft, awaiting a human.
Red team · the strongest case against
The strongest case against: this is a conviction dressed as a trend. Parity on cheap tasks is not non-dominance; the tasks that decide platform choice are agentic and the US lead there is stable or widening. Procurement defaults to US providers for reasons — legal recourse, alliance alignment, supply-chain assurance — that are not reflexes and will not move because a Qwen model is cheaper. And the AU government's own sovereignty posture is explicitly aligned with the US, which makes the political cost of a non-US recommendation higher than the field admits.
- —The parity evidence is on the tier where model choice matters least. On the tier that decides platforms, the field's own claim says the gap has not closed.
- —Procurement defaults encode alliance and legal-recourse judgements that a cost eval does not address. Calling it a reflex is the field's weakest move.
- —A provenance incident — a backdoor, a licence change, a data-training scandal — in one Chinese open-weight would set the field back years, and the base rate for such incidents is not zero.
Source diversity
- ML research20%
- Lab / vendor releases20%
- Geopolitics / policy20%
- Infra / market10%
- Internal / Engel30%
A field supported by one epistemic community is a flag, not a finding.
Cross-pollination · typed joins
Connected, not merely similar.
Enabling, compounding, substituting, blocking. A satisfied dependency trigger is a far stronger signal than semantic proximity.
The field exists because open weights exist; the parity evidence comes from that field's experiment.
Non-US weights hosted on AU infrastructure is the configuration that makes 'sovereign' mean control rather than origin.
Provenance risk in weights is a supply-chain security question; the two fields share the backdoor threat model.
A gateway the firm controls is what lets a non-US model be swapped in or out without a procurement event.
Share graph
Provenance running forward.
Discovery, not accountability. No counts, no rankings, no rollups to managers.
Convergence · who else is here
- PRPriya Raman · Research engineer · inference1 drop
- AWAdam Witanowski · Lab Director (acting)1 drop
- MTMei Tanaka · Research lead · evals1 drop
- LFLena Fischer · Red team & assurance1 drop
- ABAisha Bello · Sector owner · Government1 drop
- RMRohan Mehta · Exec sponsor1 drop
- CDClaire Dubois · Sector owner · Banking1 drop
Several people’s drops meet here. An informal working group already exists and probably does not know it.
Lineage
What this field produced, and what it killed.
Experiments, recommendations and graveyard entries stay attached. The reasoning that killed a claim is the reusable asset.
Open questions · return to the pile
Every run leaves a record. Separately, its question either closes or returns to the pile with notes — which is what the next person proposing the same thing will see.
- 01What would a bank's risk function need to see to approve a non-US open weight on non-customer data, and who writes it?
- 02Does the cheap-tier parity extend to any agentic task class in the next two releases?
- 03Is provenance assurance a service the firm could sell, or a liability it should avoid?