# the cut cost reappears: nadir buys globo, uber cuts 3,300

> nadir cuts 90% of film costs and puts the savings into broadcast tv, uber drops 3,300 people to shorten decisions, and cutler names the tax ai still charges from within

- edition: Thursday, September 3, 2026 (2026-09-03)
- notebook: product & markets
- topics: market · media · product · brazil
- items: 10 from 10 sources
- original: https://tonho.wtf/en/daily/2026-09-03-produto/
- portuguese edition: https://tonho.wtf/diario/2026-09-03-produto/
- authorship: written by an llm pipeline, reviewed and translated by antonio leandro (tonho.wtf)

---

Four of today's items describe the same operation with different numbers: someone removes a cost from one place and it reappears in another. Nadir took 90% out of the production cost of an ad film and used the leftover to buy Globo — the money didn't disappear, it migrated from production to media. Uber removed 3,300 people and layers of management to shorten the distance between the CEO and the people who do the work, and the bill reappears in the managers who stayed, with 12.1 direct reports on average and engagement falling. John Cutler names the invisible version of this inside the company: the *recontextualization tax*, the cost of translating a team's reality into the format the executive can read. Nobody was counting that cost because it never showed up on a budget line.

The day's contradiction comes from Benedict Evans, who argues that making tool-building cheap doesn't attack what's expensive: what's expensive is finding out the tool is needed and then getting 500 people across five departments to use it. And the creator economy shows the third act of the same story — the channel that was cheap got expensive, and now automation is being used to push the price back down, with the loss falling on whoever sells, not on whoever buys. Byrne Hobart promises, behind the paywall, the accounting version of that math.

## product and growth

**[TBM 437: AI and the Recontextualization Tax](https://cutlefish.substack.com/p/tbm-437-ai-and-the-recontextualization)** — Cutler names a cost everyone pays and nobody measures: the work of flattening a team's reality to fit the level of detail the most powerful stakeholder can take. The practical argument is that a roadmap tool never travels through the organization because every audience wants the same data a particular way, and the pre-AI alternative was either to convince everyone to look at it the same way, or to make ten versions of every presentation. The way out he describes is to write one skill per audience segment and let the team work however it wants underneath. The text's own caveat holds: it only works if the raw context exists and is up to date — which, in most teams, is exactly what doesn't exist.

**[Get My Eval Test Harness and Run Your First Experiment](https://www.producttalk.org/ai-evals-test-harness/)** — Teresa Torres published the harness she uses to run prompt evals, with a complete example: 15 synthetic interview transcripts, code assertion evals, LLM-as-a-judge with calibration and a golden dataset. The experiment's numbers are what matter for people building: fabricated citations falling from 12.5% to 1.7%, claims with no basis in the transcript from 8.4% to 3.2%, and status label error to zero — just by changing the prompt to spell out the grounding rules. The bar she sets is the right one and it's rare: if a change fixes one error and worsens another, it's not an improvement. The repository sits behind the membership, which is a curious pricing decision for a tool whose value is in becoming a standard.

## brand and ip

**[Roberta's Pizza Carefully Crafts a 'Type Collision' at the Core of Its Brand System](https://thedieline.com/robertas-pizza-carefully-crafts-a-type-collision-at-the-core-of-its-brand-system/)** — The Dieline opens up Roberta's identity system with cofounder Brandon Hoy and designer Drew Price: what looks like typographic chaos is a controlled choice, a deliberate collision between the wordmark's script and the heavy retro display. It's the kind of brand reasoning that almost never leaves the agency and makes it onto the page — the decision to keep the punk spirit of the origin as a design constraint, and not as a moodboard reference. For anyone with a product on a shelf: expensive "sloppiness" is always a system, and the system is the asset.

## media and attention

**[Nadir lança campanha criada 100% com IA após cortar 90% dos custos de produção](https://adnews.com.br/post/nadir-lanca-campanha-criada-100-com-ia-e-viabiliza-estreia-em-tv-aberta-apos-cortar-90-dos-custos-de-producao)** — The Brazilian case of the day comes with the math in the open: in the traditional model, the film would have required 7 locations, 30 actors and R$1.5m to R$2m over as much as 6 months; with AI-generated assets, it came in between R$100k and R$150k in 3 weeks. The part that matters isn't the savings, it's where they went — the Head of Marketing says explicitly that without the cut there would be no budget to run on Globo. Director Erika Dorta makes the honest caveat that the film didn't take the place of a set, because it wouldn't have existed any other way. For a mid-sized brand, AI here didn't replace production: it bought access to mass media, which was what was out of reach.

**[Future of TV Briefing: How strong are the fundamentals of the creator economy, really?](https://digiday.com/future-of-tv/future-of-tv-briefing-how-strong-are-the-fundamentals-of-the-creator-economy-really/)** — In a survey of 1,000 marketing and procurement leaders by Billion Dollar Boy, half admit they get creator pricing wrong and 40% think they paid too much. Tim Peterson adds the reading that was missing: the people publishing these numbers are influencer marketing agencies whose client is the brand, not the creator — complaining about high prices is a commercial tactic as much as a diagnosis. The mechanism he describes is the usual one in digital: automation (Dentsu, L'Oréal, LTK) expands the supply of available creators, and more supply without proportional demand drags price and individual leverage down. A $43.9bn market repeating the trajectory of programmatic, with the IAB's measurement guidelines promised for October.

**[Why news outlets are suddenly launching live shows](https://simonowens.substack.com/p/why-news-outlets-are-suddenly-launching)** — The title essay is behind the paywall, but the open part carries the most useful cost reasoning of the week: Simon Owens reads Netflix's podcast slate refresh not as a bet on new subscribers, but as a churn tool — content that costs a fraction of scripted per hour and gives one more reason to open the app. The phrase he uses, *ambient TV*, is too good for what the format is. Worth it as a reminder that retention tolerates weak acquisition metrics when marginal cost is low.

## market and capital

**[The 'Great Flattening' rolls on as Uber lays off middle managers](https://www.fastcompany.com/91601567/the-great-flattening-rolls-on-as-uber-lays-off-middle-managers)** — Uber cuts 10% of headcount, about 3,300 people, with an explicit thesis in the memo: fewer layers between the CEO and the people executing, faster decisions. The numbers on the redesign matter more than the one on the cut — people seven or more layers from the CEO dropped 20%, and "micro-teams" of one or two reports were cut almost in half. The aggregate result of the move is ambiguous: Bayer projects €2bn in annual savings after cutting up to six layers, Citi went from 13 to 8 layers with financial improvement, and Meta harvested margin in 2024 and internal revolt in the restructuring that followed. The data point nobody is looking at is the manager who stayed: average reports went from 10.9 to 12.1 in a year, 97% pile on individual contributor work, and manager engagement fell from 27% to 22% between 2024 and 2025. Flattening is cheap on the org chart and expensive in the layer that absorbs the rest.

**[SPX e Embraer articulam fusão e criam empresa de cibersegurança de R$ 700 milhões](https://neofeed.com.br/negocios/o-ma-da-ciberseguranca-spx-e-embraer-articulam-fusao-e-criam-empresa-de-r-700-milhoes/)** — Vision Cybersecurity and Tempest Security are joining through a share swap, with SPX as the largest shareholder and Embraer entering the cap table as a relevant shareholder — and staying a client. The new company, still unnamed, is born with more than R$700m in annual revenue, around 900 professionals and more than 600 corporate clients, subject to Cade approval. The deal structure is what there is to learn from: Vision (heir to ISH, from Vitória, founded in 1996) serves mid-sized companies, Tempest (born out of academic research in Recife, 26 years old) serves banks, energy and oil and gas, and Tempest's CEO says client overlap is "practically zero". A merger with zero overlap isn't market consolidation, it's assembling coverage — and the stated thesis is to sell the whole package against niche competitors, first in Brazil, then in Latin and North America.

## who wrote

**[AI, tools and transformation](https://www.ben-evans.com/benedictevans/2026/9/3/ai-tools-and-transformation)** — Benedict Evans attacks the comfortable Valley thesis that, if everyone can generate software, the company reorganizes itself. The argument has two legs: most people are not tool builders and don't spend the day thinking about how their own work could be done another way; and even those who see the problem run into a process that crosses 500 people, five departments and three systems of record, which turns into procurement, a decision and an 18-month sales cycle. The frame that sticks is the spectrum between institutionalized and improvised — SAP on one side, the 10-meg spreadsheet on the other — with tasks migrating in both directions, and the chatbot arriving as one more freeform space next to Excel and email. "AI doesn't change the question: it creates new choices and moves the thresholds." It's the right counter-argument for anyone building an AI tool believing the customer's bottleneck is writing code.

**[AI Doom as an Accounting Problem](https://www.thediff.co/archive/ai-doom-as-an-accounting-problem/)** — Byrne Hobart announces today the edition that treats AI risk as an accounting problem, with capex and cheap tokens on the same agenda. The piece is behind a subscription, but the angle is already the thesis: discussing the bubble through depreciation and investment schedules is the only way out of prophecy and into mechanism. Noted for when I have access.

## stalled sources

First Round Review (311 days), Calculated Risk (235), Gurwinder (249), Elad Gil (136), Y Combinator Blog (79), Investment Idiocy (63), Thinking in Bets (57), Anti-Mimetic (50), SVPG (25), Adjacent Possible (25), Elena's Growth Scoop (21), Kyla Scanlon (21), The Generalist (17), Sherwood News (17), Collab Fund (17).