Four stories from yesterday, and none of them is about a smarter model. The Pentagon asked three men with no military service to describe the next war. A memory company posted the kind of gross margin software companies brag about, and told its customers prices keep climbing until 2028. California decided an algorithm cannot be the one who fires you. And OpenAI revealed that for four weeks this summer, thousands of accounts were quietly trying to read the part of its model nobody is supposed to see. The common thread: the fight moved from who builds the best AI to who controls the hardware, the rules and the secrets underneath it.
The Pentagon put Musk, Luckey and Gingrich in charge of imagining the next war
- Defense Secretary Pete Hegseth announced Project Meridian at Marine Corps Base Quantico: a review to identify which military domains the United States must dominate and which capabilities it has to master, with findings due within 120 days and an unclassified public report at the end. Pentagon technology chief Emil Michael oversees it; Elon Musk, Anduril co-founder Palmer Luckey and former House Speaker Newt Gingrich co-lead it, with a wider team from government, industry, academia and policy.
- The scope runs from artificial intelligence, autonomy, directed energy and robotics to biotechnology, and covers battlefields "from subterranean depths to the cislunar frontier". None of the three co-leads has served in the military. Both Musk and Luckey run companies competing for Pentagon money: SpaceX took a $4.16 billion Space Force contract in May for threat-tracking satellites plus a separate $2.29 billion award, and more than $8 billion in 2026 defense awards in total. Hegseth acknowledged the tension in his own framing: "the best forecasters of future conflict are not all inside the Pentagon".
Why it mattersThe people who will sell the hardware are writing the shopping list. That is the whole story, and it is not hidden: it is the explicit design of the project. The defense of it is real too, because Anduril and SpaceX genuinely do know more about autonomous systems than most of the Pentagon's acquisition staff. What is missing is anyone in the room whose business does not get bigger if the answer is "more autonomy, more satellites, more drones". Watch the unclassified report in late January, and read it as a procurement forecast rather than a strategy document. What it means for Latin America: whatever Meridian recommends becomes the export catalog within a few years, and regional militaries buy from that catalog. If you work in public policy or security in the region, the window to decide your own rules on autonomous systems is now, while the standard is still being drafted and not yet installed.
Micron hit an 86.25% gross margin and says memory stays scarce through 2028
- Micron guided to $61.5 billion in revenue for its first fiscal quarter with a record 86.25% gross margin, which works out to more than $53 billion of gross profit against roughly $2.06 billion of operating expenses, a tax rate near 15.5% and earnings per share of $38.15 plus or minus a dollar on about 1.15 billion shares. An 86% gross margin is software-company territory, not a number memory manufacturing has ever produced.
- CEO Sanjay Mehrotra was blunt about why: "In calendar 2027 as well as 2028, we see demand exceeding supply. In fact, we see greater tightness in the industry in 2027 and in 2028 versus 2026." He added, "We do not have line of sight to when supply and demand will return to balance." NAND and DRAM shipments are expected to grow in the low-to-mid 20s percentage range over the next two years, HBM faster than conventional DRAM, and new capacity including a $100 billion New York megafab is still years from relieving the squeeze.
Why it mattersThis is the AI boom arriving in a place consumers actually feel. Data centers are buying memory at any price, so every gigabyte in a laptop, a phone or a game console now competes with a training cluster that does not check the price tag. A margin that high is not clever management; it is the signal of a market where the seller sets the terms. And Mehrotra's second quote matters more than the first: a company with record profits saying it cannot see the end of the shortage is telling you this is structural, not a quarter of good luck. What it means for Latin America: memory is priced in dollars and imported, so this lands as higher prices on laptops, phones and servers across the region, amplified by local tariffs and currency. If you have a hardware purchase planned for 2027, buying earlier is cheaper than buying on schedule, and quoting a client a device price six months out without a revision clause is a way to lose money.
California made it illegal for an algorithm to fire you by itself
- Governor Gavin Newsom signed SB 947, the No Robo Bosses Act, authored by Senator Jerry McNerney: employers cannot rely solely on an automated system to discipline or terminate an employee, a human has to corroborate the decision, and the worker has to be told that such a system was used. It takes effect July 1, 2027. Business groups objected that the standard of "primarily relies" is left undefined.
- It arrived with a package of workplace AI laws signed the same day: SB 951 requires 60-day notice of AI-caused mass layoffs that hit 25% or more of a workforce; AB 1883 restricts AI surveillance that infers emotional states or collects neural data; AB 1979 bars AI from replacing a licensed clinician's judgment and extends medical confidentiality to health chatbots; AB 1331 bans workplace surveillance in bathrooms. Newsom vetoed two others, and rideshare drivers, independent contractors under Proposition 22, were left out of the protections entirely.
Why it mattersNotice what this law does not do: it does not ban the algorithm, it bans the algorithm acting alone. That is the shape regulation is taking everywhere, and it is cheap to comply with and easy to fake. A human who rubber-stamps a hundred AI recommendations a day satisfies the letter of SB 947 and none of its intent, which is exactly why the undefined "primarily relies" matters. The more interesting clause is SB 951: a company now has to say out loud that AI caused a mass layoff, which creates the first public dataset on AI job losses. And the Prop 22 exclusion is the honest part of the story, because the workers most managed by algorithms today are the ones the law skipped. What it means for Latin America: contact centers, BPO and logistics operations in the region run performance scoring and shift allocation on exactly these systems, with no equivalent law coming. If you lead one of those teams, write the human-review step into your own process now and keep the log, because your US clients will start asking for it as a contract term long before any local regulator does.
OpenAI says 16,000 requests in two days tried to steal its model's hidden reasoning
- OpenAI disclosed a coordinated distillation campaign it attributes to individuals associated with the Chinese startup Moonshot AI, maker of Kimi. It began July 1, 2026 and peaked on July 24 and 25 with roughly 16,000 attempted requests from more than 4,000 users; more than 15,000 users in total showed the same prompt-pattern activity before OpenAI shut it down on July 28.
- The technique: the operators extracted encrypted reasoning traces and replayed them across different sessions to get plaintext output, manipulating interactions "so that protected reasoning could be reproduced in forms visible to the requester". OpenAI banned the fraudulent accounts, closed the replay pathway, added checks to catch streamed output that leaks reasoning, and deployed further anti-distillation mitigations. It published no direct response from Moonshot AI.
Why it mattersThe hidden reasoning is the product now. A model's final answer can be copied cheaply by anyone; the chain of thought that produced it is the expensive part, which is why the labs encrypt it and why someone built a four-week operation to replay it. Two things follow. First, the moat is thinner than the valuations assume, because this attack needed accounts and patience, not a breakthrough. Second, the timeline deserves attention: this was disrupted on July 28 and disclosed on October 1, which is the same voluntary-disclosure instinct the FTC is now treating as evidence. What it means for Latin America: if your business sends prompts and gets back reasoning you consider proprietary, assume the same trick works against you at a smaller scale. Keep reasoning server-side, never stream it to the client, and rate-limit by behavior pattern instead of by request count, because 16,000 requests spread across 4,000 accounts passes every simple limit.
The big picture
Read the four together and the layer that matters this year is not the model. The Pentagon is deciding what autonomous war looks like, and it invited the vendors to help write the definition. Micron is proving that whoever owns the memory sets the price of everything built on top of it, and has no reason to stop. California is writing the rule that a human stays in the loop, and leaving just enough ambiguity for that human to become a signature. And OpenAI is showing that the real asset is not the answer but the reasoning, and that it can be siphoned by patient people with ordinary accounts. For two years the question was which model is smartest. Yesterday every story was about the hardware under it, the law around it, and the secrets inside it.
Go deeper
Sources & further reading
- CNBC — Musk, Luckey and Gingrich to help Pentagon with warfare initiative
- The Hill — Hegseth puts Musk, Luckey, Gingrich in charge of future warfare review
- Tom's Hardware — Micron projects tightening RAM shortages through 2028
- KQED — Newsom signs slate of AI workplace laws, barring 'robo bosses'
- The Hacker News — OpenAI disrupts reasoning extraction campaign linked to Moonshot AI
- Quartz — OpenAI disrupts Moonshot AI model reasoning theft campaign