The full breakdown

1 · The platform is being rebuilt around skills, not courses

The structural story of the week, and bigger than the AI story. Axim named it as the growth bet, with two funded initiatives for the next 12–18 months: Learning Pathways and competency tracking. Both require breaking the platform's founding assumption — as OpenCraft put it, "the core platform is built around one assumption, that a course is the biggest unit of learning." Pathways is being built on the new Open edX core, deliberately not on the legacy platform. Talk 07

2 · Everyone is talking about AI; almost nobody has evidence yet

Roughly a third of the talks touched AI, but only two presented real numbers from real learners. The most careful of those — a 20,000-learner randomised test of AI-presented video — found no measurable difference at all. The other, WGU's embedded AI practice activities, found a genuine improvement but on a narrow slice of courses. Everything else was architecture, prototypes and intent.

3 · Assessment is the acknowledged crisis

WGU's position: standardized assessments are "increasingly easily gamed by AI", and you cannot bolt AI onto the old model. Four separate talks converged on the same conclusion — stop measuring recall in a snapshot, start gathering evidence continuously.

4 · Government is the fastest-growing vertical, arriving at national scale

Ethiopia (50 universities), Portugal (1.5M K-12 users), India (13M registrations), Israel (1.2M), Uruguay (~1 in 14 citizens). Government is 17% of detected sites and, combined with nonprofits, would be the largest single segment.

5 · Growth has flattened — and the community finally has data honest enough to say so

The most under-discussed finding of the week, and the subject of the next section.

Axim reported growth. eduNEXT reported contraction. Both are correct — they measure different things, and reconciling them is the single most useful thing in the data.

The rest of the footprint picture

• Course volume grew ~50% — roughly 40,000 advertised courses in 2024 to ~68,000 in 2025, led by higher ed and government.
• The ecosystem is concentrating. 25 large multi-tenant instances account for 35% of the footprint and 50% of all courses.
• ~60% of sites didn't upgrade at all last year. Teak is now the main migration target; the long-standing Nutmeg bottleneck finally cleared.
• Geography: Western Europe and North America ≈ 50% of sites; Middle East (250) and East Asia (179) another 30%.

Axim's 145M learner figure is scraped from public data on roughly the 20 largest instances, and they're candid it's a floor. The voluntary telemetry registry has run two years, but ~80% of instances are on versions too old to include it.

The gaps that kept coming up

Six things the platform structurally lacks. These made the list because they surfaced independently across different talks — they're not one person's wishlist.

Notice Salesforce appears twice. That's the tell — customers are rebuilding core platform functions in a sales CRM because the platform doesn't offer them. Blueprints and the approval workflow are the two that fast-growing clients hit first.

Releases: Ulmo (Dec 2025) → Verawood (June 2026) → Willow (Dec 2026). An LTS on a two-year cadence is under discussion.

Axim's official position is deliberately restrained. The argument for the LMS in an AI world is accountability — without the harness, "is there any learning actually happening? Or was it all a series of hallucinations that made the person feel good about what they already know?" Talk 05

The best production evidence

WGU's in-flow AI — a guardrailed prompt embedded in the learning activity, not a general chatbot. Live in 84 courses of ~1,200, ~17,000 learners. Because the programme is competency-based, using the material is optional — and 80–100% opted in. First-attempt pass rate improved 18% on a gen-ed course with ~20k enrolment. Their own caveat is the useful one: "It's not AI plopped into a course. It's how you plop it in."

The honest limits, from the person who built the framework

"We have no idea how much this costs. If you have 1,000 users, how much is my token count going to run? The first one who gets 1,000 users here can tell me."

Felipe Montoya on AI Extensions — performance at scale untested, no token or quota management, real provider inconsistencies around tool-calling. His ask was blunt: "The largest pilot so far was yesterday at the tutorials." Talk 14

Where AI is doing genuinely novel work

• Synthetic students — decomposing a learner into ~7 small models rather than one LLM roleplaying. "A generalist model pretending to be four specialists is not four specialists." The structure agent caught an ontology bug: module 2 asks students to read vitamin labels; module 3 teaches what vitamins are. Talk 53
• Continuous assessment — cognitive task analysis → domain model → learner cognition model → just-in-time activities and rubrics, with automated cut scores. Deliberately offloads to small models and plain Python for predictability. Talk 58
• Agentic course production — Santander's ~100 small agents turning market signals into courses and clips. On cost: "I don't think we should put costs before exploration." Talk 21
• AI grading is harder than it looks — sycophancy is structural: when uncertain, the model completes toward the most likely token, and the most likely answer is the correct one, so it silently upgrades wrong student answers. Talk 35

Unicon interviewed 10 institutions running CBE on Canvas, Brightspace and Blackboard. The verbatim from Calbright is the thesis statement:

"There's no clean system. Institutions are just MacGyvering things together." Talk 12 · 7:13

Every institution named the same #1 gap — progress tracking, done in manual spreadsheets. Skills get tagged but can't drive what happens next. CBE roles (coach, mentor, advisor) don't exist in any LMS role model.

The credentials stack has stabilized: W3C Verifiable Credentials → Open Badges 3.0. Good metaphor from DCC: "if a verifiable credential is a car, an Open Badge is a school bus." Blockchain was quietly dropped — it was the wrong layer.

But the employer loop is not closed, and they said so. Employers are interested but "have no way to really consume verifiable credentials" — ATS and HRIS vendors haven't adopted the schemas. The reframe offered: the near-term value of learning records is learner confidence and visibility, not employer verification. Talk 10

The one hard outcome number surfaced in a hallway an hour before the talk: veterans holding Education Design Lab durable-skills credentials entered jobs earning $10,000 more on average than their same-cohort peers. The usual gender gap inverted — women gained substantially more.

What these programmes cost, and who won them

None of this was stated on stage — the talks describe scale, never money. These figures come from public funding records, and the confidence varies a lot by country.

Read the Portugal number carefully. €29.7M is the whole PRR investment line, whose own objective text also covers unrelated items such as fitness-monitoring tools for PE classes. The learning platform is a sub-project inside that envelope, not its price tag. Two further notes: EduQA is an institute created by Decree-Law 105/2025, not a product — the conference talk presented it as the platform. And Skilup, who gave the talk, appears in the award records as Kautak Learning LDA, its Portuguese entity — a corporate link the award listing itself does not show, confirmed by the presenter after publication. The record pairs Kautak Learning with NTT DATA Portugal, so it establishes participation in the supplier consortium, not prime status. There is also active Portuguese reporting on ~€7.1M of separate exam-digitalisation contracts that failed; that appears to be a different system under the same ministry, and shouldn't be conflated with this one.

The pattern is worth noting on its own: the philanthropic programme publishes its number; neither government one does. Portugal is traceable because EU recovery funds carry a transparency obligation. India's is only reachable by RTI.

Operations at that scale is its own discipline. Eshi has 3M+ enrolments across 12 open-enrolment courses and a 74% completion rate. One week produced 149,000 signups and thousands of Monday-morning quiz resets. The fix wasn't engineering — it was collapsing to one final exam per course, an intake form, and bulk processing. "We didn't scale our team, we scaled our system." Talk 63

Flexibility is an asset in higher ed and a liability in corporate training

From hundreds of sales meetings: professional-training buyers like Open edX, get excited, then stall — never on capability. The recurring phrase was "powerful but risky." Discovery, which higher ed reads as diligence, reads to them as unbounded scope. Same product, different framing: "highly customizable open-source LMS" loses; "enterprise learning platform with open-source flexibility" wins. And this market is larger and growing faster than higher ed. Talk 32

Contribute vs. fork, with receipts

Contributors get 2× the productivity benefit of free riders; a private fork costs the average company ~5,000 hours per release cycle, roughly $250k. Read that one carefully: it comes from a 560-respondent Linux Foundation survey across open source generally, not Open edX — and it is per release cycle, not per year. Open edX ships twice a year, and the figure scales with company size and fork count. OpenCraft's cautionary tales: McKinsey Academy died at 6–9 months of reconciliation work per release; FUN, France's first government instance, has been frozen for ~10 years. But they concede that for every quote they made, two or three providers quoted a tenth of the price.

Upgrade debt is the biggest silent liability

68% of tracked sites — 795 of them — didn't move a single release in a year. That's the diagonal of a year-on-year migration matrix, not one shared version: the biggest clusters are Nutmeg (139 sites) and Juniper/Koa (102), with some Hawthorn instances six or seven years old still in production. Those stranded sites hold roughly half of all course content, and 90% are still operational. For contrast, 236 sites moved Lilac → Redwood and 162 moved Redwood → Teak. The real forcing function is that security patches only ship for the latest release. Advice: upgrade once a year, skipping one release; "latest stable" means a few months after release, not release week. Talk 44

Accessibility: intent is no longer a defense

The European Accessibility Act wants VPATs and evidence, not diversity statements. There is no official Open edX VPAT — eduNEXT published an unofficial first draft. The live screen-reader demo found real bugs: first heading is h3 not h1; a graph whose alt text reads "graphic"; drag-and-drop that throws focus to the top of the page — brutal in a timed exam; and in Studio, edit fields announced as buttons with no confirmation that Save worked. Talk 30

1. Synthetic Students at Scale — the most original idea of the conference; the failing-sentence examples will change how you review content.
2. Using Content Experiments — the only randomized AI data anyone presented, plus a method you could run next week.
3. Global Footprint Report + State of Open edX — read together, the honest picture of ecosystem health.
4. The Professional Training Opportunity — the most commercially useful talk; explains why deals die.
5. How to Build Plugin-Friendly Django Apps — architecture wisdom that generalizes far beyond Open edX. "Ugly is not unmaintainable."
6. Collaborate on the Core, Compete on the Edges — the economic case for contributing, with a usable decision framework.
7. Open edX Through a Screenreader — 20 minutes that will change what you ship.
8. One Platform, Three Models — the best framework anywhere for "how much Open edX do you actually need."
9. Continuous Mastery + Embedding AI in Courses — WGU's theory, then WGU's shipped data.
10. From Overwhelmed to Optimized — unglamorous and the most immediately actionable talk of the week.

Speaker directory

Every person who appears on stage across the 57 recorded sessions — their role, their organisation, the talk they were in, and a button that jumps straight to the second they start speaking. Filter by role, organisation or topic.