**Three moves this morning each hand a founder a different job to do. A personal-agent startup called Instinct raised to a reported ~$2.5B valuation in weeks — and the story turned out to be its data-license terms, a free lesson in what your own agent's ToS should never say. Amazon is shutting Mechanical Turk and SageMaker Ground Truth on Sept 30 — a hard migration deadline if you buy human labeling or run anything human-in-the-loop. And OpenAI's own Broadcom-built inference chip beat an Nvidia system on throughput-per-watt — a leading indicator that your token bill keeps sliding down.** Here's the whole edition in one screen, and the one thing to do about each:

The through-line: your product's legal exposure, your data-pipeline vendors, and the cost of every token you serve all moved the same morning. Two of the three you can act on before lunch.

1. Instinct's $2.5B markup — and the ToS lesson hiding inside it#

On Aug 26, 2026, Instinct — an always-on personal AI assistant from San Francisco's Spear Street Technology, led by founder Noah Shinn — reached roughly $350M in total funding at a reported ~$2.5B valuation, in a round co-led by Index Ventures and Benchmark. The markup is the headline number: Instinct went from a ~$500M valuation to ~$2.5B in a matter of weeks, roughly 5x, and it's still invite-only. You connect your apps and devices — email, messaging, calendar, and in testing even screen, audio, and location — and it acts for you by text and call.

What it means: The valuation tells you where the money is going — personal agents — but the durable lesson is the controversy that landed two days earlier. TechCrunch reported that Instinct's terms of service grant a "perpetual and irrevocable" license to access, store, reproduce, and modify user materials — including screen captures, keystrokes, audio, and location — and to use them to train and fine-tune its models, with little carved out. Testers also flagged security gaps, including an agent that kept summarizing a user's Gmail after access was revoked. For any founder shipping an agent, the takeaway is concrete and free: your product touches far more private data than a normal app, so your data-license, retention, and revocation terms are product-defining, not boilerplate. Write them to the narrowest grant that makes the product work, exclude training by default, honor revocation instantly, and get them reviewed before you scale. The startups racing into this category that get sued or shamed first will be the ones whose terms outran their trust. If you're deciding how much autonomy to hand your own agent, the tool-vs-companion line regulators are now drawing is the adjacent read.

2. Amazon is closing Mechanical Turk — a hard Sept 30 deadline#

Amazon confirmed it will shut down Mechanical Turk on Sept 30, 2026, 21 years after Jeff Bezos launched the crowdwork marketplace he once called "artificial artificial intelligence." Amazon stopped accepting new customers on July 30, and the sister annotation services SageMaker Ground Truth and Amazon Augmented AI close on the same date — so Amazon is exiting human-data-collection infrastructure entirely. The premise cracked as the models MTurk helped train got good enough to do the tasks themselves; a widely-cited 2023 study estimated a third to nearly half of MTurk workers were already using LLMs for text tasks, quietly poisoning the human-signal the platform sold.

What it means: If you're a solo founder or small team that leans on MTurk, Ground Truth, or Augmented AI for data labeling, model evaluations, survey recruitment, or human-in-the-loop review, this is a hard operational deadline, not industry trivia. Do two things this week. First, export what you can't rebuild — task templates, worker-qualification records, and any collected data — because access ends when the platform does. Second, pick a replacement now rather than in the Sept-30 crush: Prolific for research and surveys, Mercor and Scale AI for labeling and expert data, and LLM-assisted or synthetic labeling with a human spot-check for high-volume, lower-stakes annotation. If evals are the piece you're moving, note that the cheapest migration is often to convert a slice of human review into an automated LLM-judge pipeline and reserve humans for the disputed cases.

3. OpenAI's Jalapeño chip beat an Nvidia system on efficiency#

At Hot Chips on Aug 25, 2026, OpenAI published the first benchmarks for Jalapeño, an inference chip it co-developed with Broadcom. On SemiAnalysis's InferenceX benchmark, across OpenAI's own GPT-OSS-120B, DeepSeek R1, and Moonshot's Kimi K2.5, Jalapeño delivered 1.5-1.9x more throughput per kilowatt than an Nvidia Blackwell (GB300) system and materially lower latency on ultra-low-latency workloads — a 700W part measured against Nvidia's 1,400W flagship. OpenAI, which began design work in mid-2024 and went from team-hiring to tape-out in about 16 months, plans small-volume deployment by the end of 2026 and a larger ramp in 2027 — and says it will keep buying Nvidia.

What it means: You can't rent a Jalapeño, so this is context, not a cost lever — but it's the right kind of context. Every additional serious competitor for serving tokens pushes inference price and latency down, and that input cost is what most directly bounds what a team of one can afford to run at scale. The move it implies is defensive and cheap: keep your model and inference layer portable — behind an abstraction or a router — so that when cheaper capacity shows up, switching is a config change, not a rewrite. Read the direction, not the exact multiple: the benchmark is against today's Blackwell, and Nvidia's next generation will narrow the gap. For where the underlying economics sit right now, our GPU rental price map tracks the H100/H200/B200 rates that set today's API floors — and if you'd rather run models yourself, this morning's companion piece on local LLMs for coding covers the other end of that trade.

Also on the wire#

Every figure in this edition is dated and linked to at least two independent outlets where available. Instinct's valuation is investor-reported, not a filing, and is marked "reported"; the data-license and security details are as reported by TechCrunch. Jalapeño's benchmark figures are OpenAI's own, measured on SemiAnalysis's InferenceX against a current Nvidia Blackwell system, and are stated as claims.