Two AI infrastructure announcements landed in the same week.
Ours was a model. We released flash-1-mini — a 4-billion-parameter bilingual Canadian legal AI — under Apache 2.0, alongside an open benchmark (CBLRE), an open training corpus, and the methodology behind both.
The model runs on a MacBook.
It runs on a Raspberry Pi.
The weights are downloadable. You own the file.
Perplexity's was an orchestrator.
At Intel's Computex keynote in Taipei, CEO Aravind Srinivas demonstrated what the company calls the first hybrid local-server inference orchestrator:
Software that decides, in real time and mid-task, which parts of an AI workload run locally on your device and which route to frontier models in the cloud.
A local model flags what's sensitive and keeps it on the machine; everything else goes to the cloud.
It's closed-source.
The routing logic is proprietary.
Perplexity is valued at around $20 billion, with annual recurring revenue north of $450 million.
Editorial update, October 11, 2026: These are the market estimates cited in the June essay. A corresponding public company filing or primary financial statement is not supplied here; valuation and annualized revenue estimates should not be treated as audited present-day figures. Intel's Computex record confirms the on-stage appearance; it does not independently verify these financial amounts, the "first" claim or the full privacy architecture described here.
Two announcements. Two different positions on the same emerging stack.
The two layers
For three years the AI conversation has been about the model.
Which lab built the best one, which scored highest, which would be the national champion. That conversation is ending.
A model is a file. Files commoditize.
The frontier model that first did something specific cost tens of millions to train; the open-weight model that matches it on a narrow task can be fine-tuned for a tiny fraction of that, and the fraction shrinks every quarter.
My prediction was that more economies would gain access to capable open-weight models. The timing and breadth were uncertain.
The model layer is becoming infrastructure — important, but not where the durable position sits.
The position is one layer up: orchestration.
It decides which model runs which task. Whether your request stays on your device, runs in your country's cloud region, or routes to a global hyperscaler.
What data crosses what border. What gets logged, verified, retained.
That's the layer Perplexity just planted a flag in.
It's also where the next decade of value in AI accumulates.
The only question that matters is whether that layer is closed or open.
Why closed orchestrators commoditize
Perplexity's orchestrator makes the routing call for you.
A local model decides what's sensitive, the proprietary logic decides what leaves your device, and you accept those decisions — including the privacy model, the compliance posture, and an audit trail you can't see.
For a lot of buyers that's fine. For others it's a problem.
A government that needs to specify which workloads stay inside national borders can't inspect or modify a black box.
A regulated firm that has to show an auditor why a given data classification went to a given compute environment has a compliance gap.
Closed routing logic is structurally harder to audit than open routing logic — and the regulatory direction, from the EU AI Act to Canada's draft strategy to every sovereign-cloud framework, pushes toward auditability.
Then there's the open-source clock.
I expected open implementations to reproduce more of this functionality over time; there is no predictable lag guaranteed here.
Closed Unix became open Linux.
Closed frontier models became a wall of open weights that now match them on most tasks.
The closed product is the R&D advance team for its open equivalent.
Once a credible open orchestrator exists, the marginal cost of routing logic falls toward zero, and the orchestrator stops being the product — it becomes a feature of whatever's around it.
None of this means Perplexity fails.
Closed services can hold a layer for decades on user experience and switching costs alone — Windows and Adobe both prove it, and Perplexity will likely build a real, successful business here.
But holding a layer and owning the durable position underneath it are different things.
Why open orchestrators compound
The open layer is the substrate.
It's what the closed orchestrators, the picked champions, the regulated buyers, the researchers, and the builders who haven't started yet all build on top of.
It sits below the competitive layer and above the commodity layer — which is the only place a small team can hold ground against a $20-billion company.
For orchestration specifically, that substrate is four things:
- a reference architecture for hybrid local-cloud routing,
- an open routing engine anyone can fork,
- an open methodology for evaluating routing decisions, and
- an open compliance framework regulated buyers can adopt without lock-in.
None of it exists comprehensively yet. It will. That's the layer we're building.
What we shipped, and what's coming
flash-1-mini is the proof at the model layer: a specialist Canadian legal AI, built by five people in Toronto, deployable openly, running on consumer hardware.
CBLRE was our early evaluation-method proposal, not an adopted standard or a compliance credential. The public harness and methodology do not by themselves make gated data openly available or establish an independently reproduced result.
In July we plan to ship flash-1 (9B).
In September, we plan to ship flash-1-pro (27B) — and alongside it the Reference Orchestration Layer v1.0: open code showing how to run flash-1-pro in production legal workflows.
Editorial update, October 11, 2026: This was a June roadmap, not a report of a completed September release. The later legal-model pivot and July postmortem describe changed direction. They do not establish that the 27B model or the promised regulated-workflow reference implementation shipped. For CBLRE's public-method and gated-data limits, see the updated evaluation account.
Citation verification, multi-model routing, audit-trail generation, human-in-the-loop checkpoints. Apache 2.0, frozen at release.
Fork it and build your own orchestrator on top.
To be clear about what that is and isn't: it's not a competitor to Perplexity.
We are not running a half-billion-dollar commercial orchestration service.
We're publishing the reference logic anyone can build one from.
So when a Canadian regulated buyer needs auditable open routing for procurement, or a European government needs to pin workloads to EU data centres, the open infrastructure already exists.
The closed orchestrator can win on user experience for years.
The open substrate compounds in the background until it's the thing everything else stands on.
Why this is a Canadian moment
This June account referred to reported policy proposals. The subsequently published AI for All strategy supports sovereign AI infrastructure and scaling Canadian companies. The separate Canada Strong Fund was announced with $25 billion over three years for projects and companies across the economy, not an AI-only allocation.
Editorial update, October 11, 2026: Primary records: the April 27 Canada Strong Fund backgrounder and the June 5 AI strategy release. The latter postdates this June 3 essay. Plans, announcements and operating capacity are different evidence; neither source proves my prediction that open orchestration will compound.
Most of that flows to closed, picked-champion mechanisms.
That's one position. It isn't the only one.
A Canadian institution deploying AI for regulated work in 2027 faces a choice: pay for closed proprietary orchestration that may or may not clear compliance, or deploy open orchestration that's auditable end to end, runs on Canadian infrastructure where required, and routes to global cloud where it doesn't.
The open path needs open infrastructure to exist by then.
That's what we're building toward.
In three weeks we put a public demo of flash-1-mini on Telus H200 capacity in Rimouski, Quebec.
This was our hosted-demo claim: Canadian compute and a model available for inference. It did not establish that a complete production legal-workflow or compliance stack had shipped.
Not a sandbox. Actual inference on Canadian infrastructure.
The thesis
Models commoditize fast.
Orchestration software commoditizes slower, but it commoditizes.
The open infrastructure underneath both compounds over a decade.
Perplexity is building one position, and a real one — they'll probably win the headlines and a good business with them.
We're at a different layer, shipping the open substrate anyone, including Perplexity, can build on.
The closed orchestrators get the coverage. The open infrastructure gets used.
Closed orchestrators will commoditize. Open ones will compound.
A small team in Toronto. Apache 2.0. Not VC-funded. Canadian government R&D backing.
AI you can own.