Peter Depaepe

Peter Depaepe's thinking centers on a single organizing metaphor: public broadcasters as large, slow ships that need external startups as speedboats to stay maneuverable. "Het klassieke beeld is: wij zijn een groot schip, log; de kleine speedboten zijn de start-ups die uiteindelijk die grote logge boot toch voor een stuk in de juiste richting en sneller in de juiste richting brengen" ¹². This is not abstract innovation-speak for him but a structural argument about why VRT Sandbox exists at all: a company of over a thousand people cannot build everything in-house fast enough, so it sources speed and new formats from outside, testing startups like Pimp My Company and Scribewave directly in live contexts such as Tour de France interactivity ¹². He is explicit that this shift came with a hard-won lesson about the limits of pure experimentation: "innoveren is mooi, maar bedrijven bouwen, daar heb je geld voor nodig" ¹² — innovation without capital to actually build companies goes nowhere.

On the geopolitics of AI, Depaepe is skeptical of half-measures dressed up as sovereignty. The Nvidia-Mistral deal excites him as a jump start for European language models — "wij denken dat we zelf ook in Europa een aantal hele sterke modellen moeten bouwen die veel meer onze eigen taal en ook onze cultuur begrijpen" ¹² — but he flags the contradiction immediately: training those models on American infrastructure means Europe hasn't actually solved sovereignty, only the model layer, while the infrastructure layer remains a separate and unresolved problem that companies around Ghent's Wintercircus are now working on ¹. This is a consistent pattern in how he evaluates announcements: he takes the headline achievement seriously, then immediately asks what layer of the stack it doesn't actually fix.

He is equally attentive to the gap between model capability and user capability. Citing OpenAI's own CFO, he notes the industry's real bottleneck isn't quality but adoption: "op dit moment worden er nieuwe modellen ongeveer om de drie maanden op de markt losgelaten. Die zijn elke keer beter en sterker en kunnen veel meer. En eigenlijk kunnen wij zelf als gebruikers niet volgen" ¹². For him this reframes 2026 not as a year of racing to release more models but of coaching and educating users to extract value from what already exists — a point he treats as more consequential than any single new release ¹².

On discovery and distribution, Depaepe describes a broadcaster preparing for a world where audiences no longer come through apps but through AI assistants. VRT's response is to treat metadata quality as the new SEO, tightening it so language models can find and correctly represent its content, and adopting the BBC/Adobe C2PA standard to stamp content as trustworthy ¹². He is pointed about the strategic choice this implies: VRT has decided neither to pay LLM providers for placement nor to be paid by them for training data, deliberately diverging from Associated Press's paid arrangement with OpenAI, betting instead that credibility and clean metadata will keep it discoverable ¹². He applies the same discovery-fatigue logic to explain Netflix's move into linear TV via TF1, observing that "sommige mensen zijn meer dan een half uur aan het zoeken voor ze effectief op play duwen" ¹ — choice overload, not content scarcity, is what linear scheduling and Netflix's push of its own cheaper IP are actually solving.

Underneath the technology discussion sits a blunt financial argument about public media's fragility. Roughly half of VRT's budget comes from government dotation, the other half from the market, where radio advertising revenue effectively cross-subsidizes television production; as radio listening and radio ad revenue decline, that cross-subsidy — and with it the funding base for video output — comes under direct threat ¹². This is why he treats experiments like AI-personalized audio and Flemish-specific transcription not as novelties but as defensive necessities: generic models routinely mishandle Flemish audio, whereas Scribewave's models, trained specifically on Flemish, correctly transcribe and snippet VRT radio content, which he frames as a language-specialization moat that local AI startups can hold against much larger generic players ¹². He extends the same build-versus-hand-craft logic to personalized media production, noting that generating 80 brand-asset variants instead of two is only feasible by training AI models on a brand's own style guide rather than designing each variant manually.

Despite naming these pressures candidly, Depaepe does not frame VRT as embattled. "Er zijn veel publieke omroepen, zelfs in Europa, die onder druk staan. Wij staan gelukkig niet onder druk. Ik denk dat wij het zeer goed doen" ¹ — a statement that reads less as complacency than as confidence earned by treating each threat (funding, discovery, language, personalization) as a solvable engineering and partnership problem rather than an existential one. Taken together, his recurring move across every topic — infrastructure sovereignty, model adoption, discovery strategy, funding, language specialization — is to locate the layer beneath the obvious headline and ask whether it has actually been addressed, then to answer that question through concrete, testable partnerships with startups rather than through internal build-outs or wait-and-see caution.