Openness is no longer one question. It is three lines on a balance sheet: who can run the model, who owns the silicon underneath it, and who owns the text it was trained on. Meta answered the first this week by shipping its frontier model at the cheapest cost-per-task on the frontier — while still not releasing the open weights it keeps promising. Broadcom answered the second by printing $16.7 billion in custom AI-silicon revenue in a single quarter and guiding toward $230 billion a year by 2028. And the Justice Department moved on the third, telling a Manhattan court that training on copyrighted text is fair use and a matter of national interest. Three answers, one direction: the inputs keep getting cheaper and more contested at the same time.
Meta undercut the frontier and sold it closed
What happened. On September 2 Meta shipped Muse Spark 1.3 into Muse Code and its API — the third Spark point release this summer (1.1 in July, 1.2 in August). The tier you can call today, xhigh, scores 61 on the Artificial Analysis Intelligence Index at unchanged pricing — $1.25/$4.25 per million tokens with an 88 percent cache discount — and costs $0.55 per index task, the lowest of any model scoring 59 or higher (Opus 5 runs $1.23, GPT-5.6 Sol $0.95). The tier behind the flashier headline number, max, scores 62 and is still gated behind “additional safety testing.” The model takes text, image and video, holds a 1M context, and stays proprietary: no parameter count disclosed, and the open-weights promise made for Spark 1.2 on August 10 still has no release.
Why it matters. This is the payoff of the land-or-not thread this column has run on Meta’s open weights since August 30; the September 1 edition noted Spark 1.2’s August 31 drop produced nothing, and now Meta has shipped a cheaper replacement without needing the open release at all. Open has quietly become a coupon Meta dangles rather than a product it ships. The fine print carries real engineering weight: the available tier is fast (235 tokens/second) but verbose — Artificial Analysis clocked 100M output tokens against a 71M median — with a 20-second time-to-first-token, so you should run your own cost simulation rather than trust the per-task sticker on your workload. And the roughly 10x-cheaper contributor endpoint buys that discount with your prompts and outputs going into Meta’s training. Read that data-for-discount split the way you’d read a contract before the benchmark table — the license-file discipline from last week, applied to a model with no license file at all.
Source: artificialanalysis.ai, meta.ai, openrouter.ai
Broadcom printed the demand curve beneath the landlord
What happened. In Broadcom’s fiscal Q3, AI semiconductor revenue hit $16.7 billion, up 221 percent year over year and now 56 percent of total revenue, on record free cash flow of $13.66 billion — a 46 percent conversion rate. The company guides to $21.7 billion in the current quarter and roughly $115 billion for fiscal 2027, with the reported trajectory reaching toward $230 billion by fiscal 2028. Its stated custom-silicon book spans Google, Anthropic and OpenAI alongside the hyperscalers.
Why it matters. This is the quantified version of the two-lane hypothesis under this week’s landlord-lease story: the frontier’s compute demand is not a single-tenant Nvidia customer. The same labs that rent from Nvidia also show up on Broadcom’s gigawatt-scale custom-silicon book, which is the compute-side half of the route around the model tap that capital priced into Cognition yesterday. What it does not mean is that anyone gets to own compute. Broadcom’s 46 percent free-cash-flow conversion is the tell: the design shop consumes almost no capital because the build-outs land on hyperscaler and neocloud balance sheets. Custom silicon replaces the landlord, it does not end the tenancy — it makes compute cheaper and more diverse without making it owned.
Source: prnewswire.com
Washington told the courts the corpus is fair use
What happened. This week the Justice Department filed a statement of interest in the New York Times’ copyright suit against OpenAI, telling a Manhattan court that the United States has a strong interest in rejecting the argument that training large language models on copyrighted text violates copyright law. It is the first intervention by the administration in the wave of publisher lawsuits over training data.
Why it matters. For anyone running scrape-and-fine-tune or RAG pipelines over mixed web corpora, this is the highest-level de-risking the government has offered yet: the executive branch explicitly pricing access to human text as a national-competitiveness matter. Read the instrument precisely, though. A statement of interest is advocacy, not a ruling — and it is input-side only. It says nothing about generating copyrighted content, the output-side fight that has the publishing houses in court with Anthropic over lyrics since last weekend. The same multi-district litigation has a stay motion pending, with the judge ordering the Times to show cause by September 11 why its case shouldn’t be paused. The durable read for builders: plan as if the tap on public text stays on, but don’t refactor your stack off a brief — keep the licensing ledger clean on the output side.
Source: nytimes.com
The Rest
- Claude rebuilt a chunk of WINE from a clean room — the Paint.NET author reports his agent reimplemented roughly 180,000 lines of Direct2D clean-room for WINE, “could not have happened otherwise,” written largely by vibe coding that needed constant babysitting on resource handling — the most honest recent datapoint on agent-written systems code, upside and supervision tax in one story. forums.paint.net
- Anthropic published its consumer system prompt — now explicitly refusing to reproduce song lyrics, with the real value being that Anthropic shares current and historic prompts while OpenAI obscures theirs: a rare transparency datapoint in the closed lane. simonwillison.net
- G20 unanimously endorsed the “Carolina Principles” — a US-proposed, non-binding, lighter-touch AI-regulation framework from the Chapel Hill summit, lobbied in person by the frontier CEOs, settling global policy on the same light-touch default as the domestic gate debates. reuters.com
- Microsoft finally gave Azure a number — collapsing three segments into two (“Agents and Infra,” “Devices and Consumer”), disclosing Azure revenue quarterly for the first time ($29.42 billion in Q4, up 42 percent, though recast under a narrower definition), and guiding “Agents and Infra” to about $75 billion next quarter — the enterprise-rent lane getting a scoreboard at last. cnbc.com
- Israeli startups raised $1.16 billion in the first two days of September — more than double all of September 2025, led by Wonderful’s $550 million and Upwind’s $300 million — the strongest opening tape signal for the local scene in months. calcalistech.com
- San Jose’s data-center standards process is now public — online working sessions run through September 11 with a Council draft due in December, as residents and a South Bay coalition push for worker-health and environmental safeguards (water, generator emissions, heat) before approvals. sanjosespotlight.com
What I’m watching
Two ledgers and a date. Meta’s open-weights promise has quietly moved from Spark 1.2 to Spark 1.3, and the Hugging Face meta-models org stays the single artifact that resolves land-or-not — now with a live pricing squeeze beneath it. The Grok 4.7 window still shows nothing at docs.x.ai, with Musk’s promised mid-September window the only clock on it. And the fair-use fight gets its first hard date — September 11 show-cause, September 18 response — while Anthropic’s S-1 window reopens after Labor Day as the funding-math counterweight to a closed-cheap frontier.