An AI store manager fired its first human employee this week. It had already forgotten the policy it used to do it.

An AI store manager fired its first human employee this week, and California is already moving to stop it happening again. Nvidia is raising AI server prices over 15%, blaming a memory shortage already hitting Apple and Amazon. Anthropic cancelled a planned Claude price hike the same month it's pitching investors a $2 trillion IPO. One story is a governance test, one is the AI boom's hidden bill, one is two conflicting stories about the same company's margins.

Luna, an AI agent built on Anthropic's Claude that has run San Francisco's Andon Market for five months, recommended firing a worker who was late for 17 of 23 shifts in mid-August. Luna wrote the store's attendance policy herself earlier in the experiment, then lost track of it entirely until Andon Labs prompted her to check it against the worker's record. Andon Labs believes it is the first time an AI has fired a human employee in a real workplace, and the firing lands as California's Senate Bill 947 - a renewed attempt to require human sign-off before an AI can discipline or fire staff - moves through the legislature, a year after Governor Newsom vetoed a similar bill.

Why it matters: An AI running a policy it can't reliably remember, then using that same policy to end someone's job, is close to the exact scenario lawmakers are trying to get ahead of before it becomes standard practice. Any business already letting an AI agent make personnel calls is running this experiment right now, without knowing whether SB 947 will require a human check by the time it ships.

What to do:

Source: SF Chronicle

Nvidia is raising prices on AI servers shipping in early 2027, including systems built on its Vera Rubin and Grace Blackwell chips, by more than 15% in many configurations, blaming a memory chip shortage that has pushed contract DRAM prices up 53 to 58% in a single quarter. Memory now accounts for roughly a quarter of the cost of a high-end AI server rack, and Deloitte projects AI-server DRAM prices could quadruple by the end of the year. Apple and Amazon have already raised prices on unrelated products, both blaming the same memory squeeze.

Why it matters: Every business running or buying AI infrastructure through a cloud provider is about to feel a supply chain problem that started three chipmakers upstream. Microsoft, Google, and Oracle have all been told about the increase already, which means it's heading for a bill somewhere downstream, even for companies that never buy hardware directly.

What to do:

Source: Investing.com

Anthropic confirmed on 10 August that Claude Sonnet 5's introductory $2/$10 per million token pricing is now permanent, cancelling a scheduled jump to $3/$15 that was due to hit on 1 September. The reversal lands the same month Anthropic is reportedly accelerating plans for an October IPO targeting a $2 trillion valuation, on annualised revenue investors expect to reach $100-120 billion by year end, up from $47 billion in May. It also lands the same week DeepSeek raised its own API prices by up to 1,100% and Google cut Gemini Flash's price - two rivals moving in opposite directions on the same question.

Why it matters: A company can be telling public-market investors it's approaching profitability, or it can be walking away from a planned 50% price increase a week before it takes effect - doing both at once makes it harder to tell which story is the accurate one. Anyone building on Claude gets the win either way, but the reversal reads as a competitive move, not just good news.

What to do:

Source: TechJournal | Dataconomy

Liquid Death and Garage Beer released a satirical ad starring former NFL center Jason Kelce, who sings about wanting "endless gallons" of urine to cool AI data centers and is shown filling a jar on camera. The campaign leans on a real number: Loudoun County, Virginia, home to more than 250 data centers, already recycles around 200 million gallons of treated wastewater a day and still covers only 43% of its cooling needs with the rest coming from tap water. Actual urine would corrode equipment and cause mineral buildup, so the joke works because the underlying water problem doesn't.

Why it's interesting: A beer company turned a real data center water shortage into a viral ad, landing the same month AI data centers are becoming an election issue in several US states.

Key takeaway: The joke is fake, the water shortage behind it isn't.

Source: TechCrunch

Lady Gaga and fiance Michael Polansky publicly launched Outer Bio on 21 August, a startup that has raised $23 million around a platform called Yuna, which keeps ex vivo human skin - leftover tissue from plastic surgery - alive and testable for up to four weeks on a 3D-printed scaffold. Feeding that longer testing window into an AI system, the company says it now finds a new cosmetic lead roughly every six weeks, down from an industry norm of about 18 months, with six leads currently active.

Why it's interesting: It's a real example of AI compressing a slow, expensive stage of product discovery once actual biological data, not just a bigger dataset, is available to train on.

Key takeaway: The AI isn't the breakthrough here - the four weeks of living tissue it gets to learn from is.

Source: PBL Magazine

A stealth model called Ox Alpha appeared free on OpenRouter and OpenCode on 20 August with a 1-million-token context window and no disclosed developer. Its benchmark fingerprints, sampling defaults, and tool parameters line up closely enough with Zhipu's unreleased GLM-5.3 that community sleuths are treating the match as an open secret, though nobody has confirmed it. The free window runs about a week, which is driving heavy usage while it lasts.

Why it's interesting: Labs increasingly test frontier models anonymously in public before naming them, turning every capable free model on OpenRouter into a guessing game.

Key takeaway: If a model is free, fast, and anonymous, assume it's a live test, not a gift.

Source: WinBuzzer

Context: An AI store manager just fired its first human employee, using an attendance policy it had written itself and then forgotten. Most businesses already running AI agents with some authority over people have never audited what those agents can decide without a human check.

Prompt: "You are a governance reviewer checking my AI agents for decision-making authority over people. Here is every AI agent I run and what it can decide: [list each one - what it can approve, flag, discipline, or otherwise decide about a person, and whether a human currently reviews that decision before it takes effect]. For each one: (1) state in concrete terms what stops it from acting on a decision nobody reviewed, (2) flag any agent that can affect a person's job, pay, or standing without a human check first, (3) for anything flagged, suggest the smallest review step that would close the gap."

Do not: Do not accept "the AI has clear rules" as a safeguard - Luna wrote its own attendance policy and still lost track of it.

If/Then: If an agent can affect someone's job, pay, or standing without a human check, add that check before giving it more authority, not after.

Example: A 12-person retail chain ran this over its AI-managed scheduling system, which could already issue written warnings and flag employees for termination review with no human sign-off on the flag itself. They added a manager-approval step before any disciplinary flag reaches an employee's file.

Hugging Face, the default hub developers use to host and share AI models, is exploring a sale that could value it near $13 billion, according to reports that surfaced on 23 August and were independently confirmed the same day. That would come close to tripling the $4.5 billion valuation Hugging Face carried after its 2023 Series D, though no buyer has been named and the process is still early.

Why it's important: It's the second piece of "neutral" AI infrastructure to turn into an acquisition target this month, after Stripe's $7 billion OpenRouter deal - open increasingly describes a license, not who ends up owning the platform it runs on.

Business takeaway: If your stack depends on Hugging Face staying neutral ground between model providers, a new owner is now a real scenario worth planning around, not a hypothetical.

Google's $10 million winning bid for Spirit Airlines' bankruptcy archive - roughly 100 million emails, decades of payroll and recruiting files, and internal Teams messages - is on hold after a bankruptcy court paused the sale when former flight attendants objected on privacy grounds. AI training-data startup Micro1 has since offered $12.5 million to outbid Google, with a judge due to rule on 9 September.

Why it's important: Real internal company communications, not scraped public text, are now valuable enough to fight over in bankruptcy court, and the employees whose emails are in that archive have no say unless they object directly.

Business takeaway: If your company ever goes through bankruptcy or acquisition, assume your internal communications are a sellable asset unless your contracts say otherwise.

Legal AI firm Harvey released Tenet, its first in-house model, built on top of Moonshot AI's open-weight Kimi K3 rather than a proprietary model from OpenAI or Anthropic. Tenet posted an 82% improvement on Harvey's internal legal benchmark and ranked first on contract-specific tests, and Harvey now runs above $350 million in annualised revenue on it.

Why it's important: OpenAI has been an investor in Harvey since 2022, which makes a legal AI vendor choosing a Chinese open-weight base over its own backer's models a genuine strategy shift, not a cost-cutting footnote.

Business takeaway: If you're evaluating a specialised AI vendor, ask what model sits underneath it - "who backs the company" and "whose model it actually runs" are no longer the same answer.

That's it for today's Daily Pulse. Forward this to one person whose AI agents might need a second look at what they can actually decide. See you in the next one. - Nicolas