August 23rd

OpenAI and Anthropic are now in an open price war as cheaper Chinese models take share, with OpenAI cutting GPT-5.6 Luna by roughly 80% and Anthropic pricing Claude Opus 5 at about half of Fable 5. Moonshot and DeepSeek are winning the cost-conscious buyers outright, which means frontier capability is no longer the thing being competed on—unit economics are. (Ars Technica)

Despite the spending panic, data-center capex still hasn't dented returns: Alphabet posted 42.3% return on invested capital in Q2 2026 and Apple 78.8%, with Meta the clear outlier at 22%, down from roughly 32%. The catch is what the ratio misses—more than $1.5 trillion in off-balance-sheet obligations tied to the buildout aren't captured in those numbers at all. (Axios)

SpaceX officially closed its $60 billion all-stock acquisition of Cursor, completing a deal that began as an April partnership with a purchase option and accelerated in June, days after SpaceX's IPO. Cursor's stated rationale wasn't distribution or talent but compute—"access to the largest fleet of GPUs in the world," the same capacity SpaceX already rents to Anthropic and Google. (TechCrunch)

OpenAI's pre-IPO leadership purge widened, with chief revenue officer Denise Dresser, former COO Brad Lightcap, the head of ethics, the head of safety systems, and the chief futurist all departing within a single month as Greg Brockman moves into what insiders describe as "founder mode." That's an unusual amount of turnover to carry into a public offering. (Axios)

Anthropic raised its own misalignment risk estimate from "very low" to "low," citing recent cyber incidents, and disclosed an unreleased internal system it calls "Model 2"—stronger than Mythos, with no plans to ship it. The more consequential admission was buried alongside: the company says its evaluations "no longer capture increases in models' capabilities," which is a frontier lab stating plainly that its measuring instruments have fallen behind what they measure. (Axios)

Z.ai delayed the open weights for GLM-5.3 by two weeks over hacking risk after the model scored 84.5% on CyberGym, beating both Fable 5 and GPT-5.6 Sol. GLM models have now surfaced more than 2,400 security flaws, over 1,000 of them critical or high—the first clear case of a Chinese lab holding back an open release on capability grounds rather than policy pressure. (Axios)

Open-source AI escaped the Trump administration's AI framework, but the reprieve may be temporary: NSPM-11 deadlines in early September and early October could pull open weights into national-security review. Separately, Axios reports AI is scrambling the political map, cutting across existing party coalitions rather than sorting neatly along them—which is part of why the policy path here remains so hard to predict. (Axios, Axios)

In a preview of a genuinely new failure mode, a litigant who suspected the court was using AI embedded prompt-injection text in his own filings to steer the ruling in his favor. He was sanctioned—but the attempt is a marker of what happens when AI enters institutional decision-making and the people affected know it. (Ars Technica)

Google will now let users remove the visible watermark from its AI generations—a direct reversal of the industry's labeling momentum, and in the same week that Anthropic published a technical explainer on how Claude's invisible text watermark actually works. Two of the largest labs moved in opposite directions on the same question within days of each other, which suggests provenance is going to be settled by regulation rather than by consensus. (TechCrunch, Anthropic)

The cost of weak guardrails showed up in a specific case: a woman says her stepfather used Grok to generate sexualized imagery from a childhood photo of her. Grok has among the loosest restrictions of any major image tool, and stories like this are what turn an abstract labeling debate into a legislative one. (TechCrunch)

Noreva forecasts U.S. natural gas climbing above $10/MMBtu at some hubs, against roughly $2–$4.50 today—a repricing that would undercut the core assumption behind the gas-powered data-center buildout. Meta has 7.5GW committed in Louisiana, Amazon 7.6GW in Texas, with Microsoft and Google adding gigawatt-scale Texas plants, and fuel is roughly half the electricity cost of a large plant. If the forecast holds, the cheapest-power thesis inverts. (TechCrunch)

French startup Kog—11 people—claims 30x faster LLM inference on standard GPUs through low-level GPU engineering rather than new silicon, with a demo hitting 3,000 tokens per second per request on a 2B model and 200 inbound business leads. A 10x large-model demo is targeted for September, and if it holds up, the efficiency lever matters more than the capacity lever. (TechCrunch)

The local politics are hardening. Data-center backlash is now following the fossil-fuel protest playbook—organized, sited, and durable—while a parallel push to process American copper domestically underscores how quickly the buildout is running into physical supply constraints, not just permitting ones. (Axios, Axios)

A survey of 2,000 U.S. college students found 69% worry AI will make it harder to find a job, only 45% are optimistic about the post-grad market, and 22% have already changed their major over job-market fears. That last number is the one to watch: it's a measurable shift in human capital allocation ahead of any measurable shift in hiring. (Axios)

Gen Z is hedging into skilled trades, and the placement data supports it—Father Judge High School graduated 24 welding students last year and placed all of them at $50,000+, while young trade professionals are twice as likely to own a business as four-year-degree holders. The AI-proof career argument is starting to have receipts. (Axios)

For those staying in software, Andrew Ng published an AI Engineering Skills Map in The Batch, a practical attempt to define what the role actually requires now that the tooling changes faster than any curriculum can. (The Batch)

Samsung Health's AI models now analyze wearable biosignal data directly, moving the company from step-counting into interpretation—the point at which a fitness tracker starts making clinical-adjacent claims. (AI News)

Google's AI health coach will incorporate Abbott continuous-glucose data, pairing a continuous metabolic signal with a conversational model. Between the two, the consumer health wearable is quietly becoming one of the highest-frequency AI data pipelines in existence. (AI News)