Three restructurings, one benchmark and two drone incidents this week, and in every case the interesting part was the thing nobody would name out loud.
That is the shape of the week. Etsy's memo says one thing and its org chart says another, Meta's chart hides the variable that explains the result, and Sofia is calling an explosion near a gas compressor a stray decoy. The Bratislava deal is the exception worth studying, because Eldridge said plainly what it was buying and why.
Etsy cut 12% of its staff and told them it wasn't AI
Etsy's Q2 shareholder letter landed on August 5 with a 220-person cut attached, about 12% of the company, mostly Product and Engineering. CEO Kruti Patel Goyal, seven months into the job, told staff that neither cost-cutting nor AI drove the decision, then added that "AI is changing how all of us work, and it will continue to change how we build products and solve problems." Net income swung from $29 million a year ago to a $46.7 million loss. The company's ChatGPT shopping app is live in beta.
The next morning Challenger, Gray & Christmas published its July numbers: 33,429 announced US cuts, the lowest month in two years, with AI given as the reason for 10,970. That is the fifth consecutive month AI has led every stated cause, and 112,713 of the 477,033 cuts this year. The line that matters more than the count: Challenger notes firms are inconsistent about naming AI because saying it "can win over investors while pushing employees away." Visa's 7% cut was confirmed as AI-attributed; Montefiore's is contested by its nurses' union.
So the official attribution number is a floor, set by legal and comms. Read your competitors' cuts by function rather than by press release: when Product and Engineering go first at a company shipping an AI shopping surface, the memo is not the evidence, the org chart is. Keep the counter-signal in view too, because announced hiring plans hit 16,095 in July, the highest July since 2022. The work is moving, not evaporating.
AI Insights. Real Growth. Higher GMV, Better Profits
The difference between growing stores and stagnant ones isn't more effort. It's better insights. StoreClaw analyzes your Shopify and Amazon data, surfaces your biggest growth opportunities, and helps you increase GMV while protecting profit. Start free with bonus tokens. No credit card required.
Bratislava sold the thing everyone else pretends they built
On August 10 Eldridge, the $75 billion holding company chaired by Todd Boehly, took a significant ownership stake in Sudolabs, a Bratislava firm that builds custom agentic-AI systems for clients. Jozef Petro and Pavol Madar founded it in 2019, it has shipped more than 100 client products, and roughly 70% of its book is already American. Terms were not disclosed. Operations stay in Slovakia while the capital funds US expansion.
Notice how different this is from the Etsy memo. Eldridge, whose other positions include Chelsea FC and the LA Dodgers, did not announce an internal AI transformation. It bought a team of people who do the work, in a country most US allocators could not find a Series A in, and said so in plain language. That is the honest version of the same trade: the capability is scarce, the people who have it are the asset, and buying them is faster than claiming you already have them.
For regional founders the read is a pricing one. Sudolabs was not a product company with a multiple to argue about, it was a services shop with American revenue and a Slovak cost base, and that combination is now buyable by institutional money rather than only by strategic acquirers. If you run something similar, the question this week is what share of your revenue is invoiced in dollars, because that ratio is doing more work in a valuation conversation than your headcount does.
Meta's benchmark chart ran a different harness in every bar
Meta Superintelligence Labs shipped Muse Code and Muse Spark 1.2 on August 5, claiming 82.9% on Terminal-Bench 2.1 against Claude Opus 5's 86.7%. Independent analysts pulled the chart apart over the following two days and found that each bar had been run inside a different coding harness: Muse Spark 1.2 in Muse Code, Opus 5 in Claude Code, GPT-5.6 Terra in Codex, Grok 4.5 in Grok Build. Hold the harness constant and the claimed 6.7-point gain over Muse Spark 1.1 falls to roughly two points. As of August 6 the model had no entry on the independently verified leaderboard, and Meta's previous model came in 3.8 points below its own published figure once someone else ran it.
Nothing here was falsified. Every number is real, and the variable that produces most of the gap was left unlabelled, which is a different offence and a harder one to call out. It is the same move as a layoff memo that lists every cause except the operative one.
The practical consequence is that vendor benchmarks are now unusable for procurement decisions, and you should stop treating them as a shortlist. Pick the two coding agents you are choosing between, run both inside the same harness on twenty tickets from your own backlog, and score them on merged pull requests rather than on pass rates. It costs an afternoon. The alternative is buying a two-point improvement at the price of a seven-point one.
Drones over a Patriot depot, and nobody will name the sender
Six unidentified drones flew over the Bundeswehr's Mechernich logistics complex in North Rhine-Westphalia overnight on August 6, the only facility in Europe that repairs Patriot air-defence systems. Germany's Military Counterintelligence Service and Berlin's Joint Drone Defence Centre are investigating, with the early read pointing to Fly-380 VTOL airframes carrying more than 10kg over a 400km range. Two days later a drone crossed from Romania and exploded inside Bulgaria near Kardam, about a kilometre from a Trans-Balkan pipeline compressor station. Sofia's preliminary position is that it was a stray Ukrainian decoy. Kyiv denies targeting Bulgaria at all.
Attribution is the whole story, and it is precisely what both governments have declined to do. Naming a state actor triggers consultation mechanisms and a political decision nobody wants in August, so the incidents get filed as unexplained and the airspace stays open. This is the gap between the counter-drone budgets announced across Europe this year and the coverage that exists over a depot repairing the continent's air defence.
The commercial consequence lands on anyone with physical infrastructure near a border or an energy corridor, which in this region is most industrial operators. Procurement for detection follows incidents rather than programmes, so the next twelve months of counter-UAS spending will be shaped by Mechernich and Kardam more than by any white paper. If you are building in sensing, RF detection or airspace monitoring, the buyers who ignored you in the spring are the ones now writing unbudgeted line items.
Short Signals
Five things to install or test this week, tagged by the seat they help.
Productivity: OpenAI made the good model free. OpenAI made GPT-5.6 Luna the default for Free and Go users on August 6 with unlimited text chats and a Think button for harder questions, and gave Sol an effort slider on paid tiers. OpenAI's own evals put responses containing at least one factual error 62% lower on Luna than on 5.5 Instant. Re-run your three worst prompts from last quarter before you renew any seat-based wrapper.
Dev: Copilot now serves Kimi K3 at a fifth of frontier pricing. GitHub added Kimi K3 to Copilot on August 6 across VS Code, JetBrains, Xcode, the CLI and mobile, hosted on Fireworks at $3 per million input tokens and $15 per million output. It is off by default on Business and Enterprise until an admin enables the policy. Switch your model picker for one routine ticket this week and compare the diff quality against your default.
Agent infra: Cloudflare shipped a browser built for agents, not people. Cloudflare launched Kitesurf on August 6, a stateless browser running on Workers that uses three to seven times less CPU and memory than Chromium for screenshots and HTML extraction. It is free during beta and needs no client changes, just add browser=kitesurf to an existing Browser Run endpoint. If you run scraping or agent workflows that keep hitting concurrency limits, this is a one-line test.
Research: Anthropic reopened biology on Fable 5. Anthropic rewrote the classifier gating biology on August 7, cutting silent fallbacks to weaker models by about 85% so health and education questions now stay on Fable 5. Virology, toxicology and molecular design still route away. Worth retesting if you abandoned a health or lab-data workflow because the answers kept degrading without explanation.
Capital: Romania will cover 60% of a returning founder's loan. The Romanian government approved a €100 million "Diaspora Invests at Home" scheme on August 6, run through the Development and Investment Bank. Grants cover up to 60% of an investment loan capped at €200,000, against loans of €5,000 to €500,000 over ten years. You need twelve months abroad, a three-year hold, and 10% of your own money, or 5% if you are under 35.
Next edition soon,
Çelik