# growth became optional at gabriel; diageo puts 40% of media in a seal > gabriel reaches break-even and funds 45 cities with subscriptions, bndes picks who allocates r$500 million of ai money, and teresa torres reminds us delivery didn't come free - edition: Tuesday, September 15, 2026 (2026-09-15) - notebook: product & markets - topics: market · brazil · product · media - items: 10 from 9 sources - original: https://tonho.wtf/en/daily/2026-09-15-produto/ - portuguese edition: https://tonho.wtf/diario/2026-09-15-produto/ - authorship: written by an llm pipeline, reviewed and translated by antonio leandro (tonho.wtf) --- Before, if Gabriel didn't grow, it broke: the fixed cost was already contracted and the model depended on scale. "Now, growing became an option," says Erick Coser. It's the best line of the day and it organizes the rest, because the difference between expanding with the customer's money and expanding with other people's money isn't speed — it's who decides when to stop. On the other side of the same question, the whole day is about where the check comes from: the BNDES and Finep AI fund coming off paper with a manager picked, Airbnb working the mezzanine layer, Kroger finding out that ad profit grows faster than shelf profit, Perifa Mídia creating inventory where nobody had priced anything. The defensive half is the more interesting one. Diageo took 40% of the fiscal year's media budget and spent it not on a campaign but on a seal that proves the bottle is theirs. Baymard measured ten B2B medical and pharma sites and none reaches "decent". Teresa Torres devotes an entire episode to dismantling the line that AI made delivery free. Add Scott Young saying that retaining what you read depends on how much you already know, and the second thread shows up: before deciding where the expansion money comes from, someone needs to know how much it costs to finish. ## product and growth **[Delivery Isn't Free](https://www.producttalk.org/delivery-isnt-free-all-things-product-podcast-with-teresa-torres-petra-wille/)** — Teresa Torres and Petra Wille take the most repeated line of the moment and separate two things that travel glued together: building *one* feature got dramatically cheaper, building a production product didn't. The symptom she describes is recognizable to anyone who has run a coding agent without brakes — Frankenstein data model by feature 15, the same snippet duplicated in 17 places, performance on the floor three weeks later. "The first 60 to 70% is easy. It's a prototype… Closing that last 30% is months to years of work." And she points at the hidden second half of the line: "delivery is free" is almost always followed by "…so only taste matters" or "…so only discovery matters", and both conclusions are wrong. **[B2B medical and pharma UX benchmark](https://feeds.baymard.com/link/9825/17462397/b2b-medical-pharma-ux-benchmark-2026)** — Baymard evaluated 10 sites (AbCam, Thermo Fisher, McKesson, Sigma Aldrich and others) on 400+ parameters, generating 3,400+ weighted scores and 2,600+ best-practice examples. None gets past "mediocre"; half are "poor". The problems aren't aesthetic: price frequently obscure, total order cost and delivery time missing from the buying area, search that falls apart when the query is slightly off, and a cart that can't be saved or handed off. That last one teaches the most — in B2B, whoever builds the order is rarely whoever approves it, and none of the ten had modeled that. ## brand and ip **[Selo Drink Seguro](https://braziljournal.com/metanol-no-more-diageo-cria-selo-high-tech-para-garantir-que-o-johnnie-e-walker/)** — Diageo pooled the budgets of its five biggest brands and allocated around 40% of the media investment planned for fiscal 2026/2027 into an anti-fraud system: the same technology as the chip in the Brazilian passport and in euro banknotes, 27 trillion combinations, verification in five seconds by pointing the camera at the seal. Since February that's 32 million bottles sealed across 49 products, 100% of the portfolio in Brazil, with the cost absorbed in the P&L and not passed on to the consumer. The context is the 2026 methanol crisis and the 4.7% of the spirits market that Euromonitor attributes to counterfeiting. Two things matter for anyone building: the category leader's biggest awareness play of the year is a piece of engineering, not a film; and the executive already admits the platform could turn into loyalty, which means a defensive tool is, by construction, a first-party data collector at the exact moment of purchase. **[bioSuc](https://thedieline.com/biosuc-is-juice-with-a-little-character/)** — Teaser: a concept project by Sara Nogaledo that swaps polished fruit photography for retro cartoon characters — grapefruit, peach and pineapple with personalities of their own — with simple black type at the top holding the line. Dieline promises the interview with the process reasoning; only the teaser is available. Even from the cover you can see the decision: on a shelf where everyone licenses the same fruit photo, the character is the only asset the neighbor can't copy without the plagiarism being obvious. ## media and attention **[Kroger Precision Marketing](https://digiday.com/media/krogers-ad-business-sees-most-profit-growth-since-2021/)** — Profit from the ad operation grew 24% in the second quarter, the best advance since 2021, with media monetization up 88 basis points on the year and e-commerce growing 20% in the second consecutive quarter of profitable growth. The detail that counts more than the number: CEO Greg Foran mentioned e-commerce and retail media in the first sentences of the call, before any mention of merchandising. Foran spent six years running Walmart US and helped build Walmart Connect — and the explanation he gives for the new inventory is organizational, not technological: "stronger collaboration between our merchandising and media teams expanded advertising inventory". Add to that Kroger's AI assistant, which this year was born with ads inside, while Walmart and OpenAI launched theirs without and have been fitting them in since. **[Perifa Mídia](https://adnews.com.br/post/clarissa-crisostomo-lanca-perifa-midia-para-expandir-publicidade-ooh-em-favelas-e-periferias)** — Clarissa Crisóstomo launches an OOH operation dedicated to favelas and peripheral neighborhoods, projecting to manage more than 80,000 points and proprietary formats: the Perifa Clock, digital street furniture with the first three units in Rocinha, and the Perifa Praça, which turns courts and communal spaces into a touchpoint, plus deals with residents for walls and gates. Expansion starts with Rio (Rocinha and Vidigal) and São Paulo, after tests with Café Pilão, O Boticário, 99 and Festival Favela Cria. The product play isn't selling cheap media in forgotten territory — it's that this inventory didn't exist as inventory, it existed as a wall. Naming, standardizing and measuring is what turns surface into something sellable, and the proprietary format is what keeps the advertiser from dealing directly with the owner of the wall. ## market and capital **[Gabriel](https://startups.com.br/negocios/exclusivo-gabriel-atinge-breakeven-e-amplia-presenca-para-45-cidades/)** — Operating break-even after three straight months of positive Ebitda, and from 4 to 45 cities since January. The mechanics: subscription with equipment on loan (Erick's analogy is the carrier's router), average ticket of R$800 a month ranging from R$400 to R$20,000, four-year contracts, and the first 2022 cohort renewing now with net dollar retention of 115%. Revenue grew 100% in the last year and 200% the year before with a team the size of the 2022 one, sustained by vertically integrated hardware — its own router costs half what buying one ready-made would. The expansion strategy is density, not spread: cameras installed free in squares and commercial streets to build familiarity before selling into the condo on the corner. In my reading, the asset here is density, not the camera — and the proof by contrast is in the text itself, when a sales leadership mistake set São Paulo back a year and gave facilities companies time to try copying the model, something that didn't happen in Rio because the mesh closed first. The line that settles the capital decision: "Not needing to raise money at any price gives me the peace of mind not to bring in people I wouldn't want as partners" — from someone who has already raised R$66 million in a SoftBank Series A in 2021. Worth noting that the third front for 2027 is expanding the attributes the network reads (vehicle color and model, pedestrian clothing color), explicitly without facial recognition: that's where the setup brushes against the regulatory discussion, and saying it in an interview is already positioning. **[Monashees in the BNDES and Finep AI fund](https://startups.com.br/negocios/venture-capital/monashees-e-escolhida-para-gerir-novo-fundo-de-ia-do-bndes-e-finep/)** — The public call received 17 proposals; Monashees came in first, Bossanova second, Patria High Growth third. The FIP aims to mobilize up to R$500 million, of which BNDESPAR commits up to R$125 million and FNDCT/Finep up to R$80 million, with a 25% cap per anchor unitholder. In other words: public money comes in as an anchor to pull the rest, not as a check. The two constraints that change the thesis for anyone about to raise are the sector list (agri, health, infrastructure, industry, bioeconomy, defense and sovereignty, cybersecurity and public safety) and the rule that 30% of the Finep contribution goes to the North, Northeast and Center-West. Curious that public safety is on the list on the same day the sector's most visible startup announces it has stopped needing a round. **[Airbnb's Mezzanine Financing Strategy](https://www.thediff.co/archive/airbnbs-mezzanine-financing-strategy/)** — Teaser: Byrne Hobart opens the edition of The Diff with Airbnb's mezzanine financing structure — the layer between equity and senior debt — alongside notes on counterparties, sanctions in practice and data moats. Only the headline is outside the paywall; noted here for subscribers. ## reading **[How to stop forgetting what you read](https://www.scotthyoung.com/blog/2026/09/14/stop-forgetting-what-you-read/)** — Scott Young (2026) argues that retention depends far more on what you already know than on your intelligence, because reading is already an act of retrieval: the text only sticks to whatever prior knowledge is there for it to stick to. Two practical consequences follow. The first is that the real trade-off isn't quantity against depth, it's accuracy — starting with biased books produces a self-reinforcing loop, which is why he prefers textbooks and survey courses, with Oxford's Very Short Introductions as the recommendation. The second is the warning against the elaborate note system that becomes an obstacle to simply reading more. For anyone building product, this rhymes directly with the Torres episode: what she says only a domain expert catches in AI-generated outputs is exactly the prior knowledge Young describes. The base isn't ornament — it's what lets you spot the error. ## stalled sources Today's list looks less like a queue and more like a graveyard: First Round Review at 323 days without publishing, Gurwinder at 261, Calculated Risk at 247, Matthew Ball at 211, Elad Gil at 148. Commoncog (22 days) and Granted (18) are still within a plausible cadence.