❯ Federal judge rules the Pentagon’s supply-chain risk label on Anthropic was unlawful retaliation

[the ruling] A federal judge in California ruled on August 28 that the Pentagon’s designation of Anthropic as a supply-chain risk was unlawful, finding that Defense Secretary Pete Hegseth’s label amounted to unconstitutional retaliation against speech, was “arbitrary and capricious,” and denied Anthropic due process under the Fifth Amendment. The designation had ordered every federal agency to stop working with the company, and was temporarily frozen by the same court in March.

[origin and process defects] Per the ruling, the dispute began with the safety lines Anthropic drew in early 2026: refusing to let its models be used for fully autonomous weapons or mass surveillance of American citizens. The case won a temporary injunction in March and later hit a setback on appeal. The judge found the company’s public criticism was a substantial factor in the government’s actions, citing statements from Trump and Hegseth attacking the company and officials’ intent to “make a public example out of Anthropic.” The opinion also flags inverted process — Hegseth publicly ordered the designation before the supporting analysis was finished, with the factual record assembled afterward, as if “to justify the foreordained conclusion.”

[how far the ruling reaches] The judge wrote that “the empty invocation of national security is not a blank check to punish and retaliate against government critics.” That draws a line under the designation process federal agencies use: order first, justify later does not survive judicial review. The more direct industry effect is that the price of holding a safety line just dropped a notch — refusing certain military uses once meant forfeiting all federal business, and now there is at least a path to court.

▪ SIGNALAn AI company blocked a Defense Department blacklist on constitutional grounds; safety red lines now have a legal foothold.

❯ OpenAI will stop supplying models to Cursor on November 12 after the SpaceX acquisition

[cutting supply] OpenAI is ending its partnership with coding tool Cursor, and under its proposal Cursor’s direct access to OpenAI models will end on November 12. The stated reason is blunt: “we cannot be confident that SpaceX will use our technology within our ToS.” The trigger is a change-of-control clause — SpaceX agreed in June to buy Cursor parent Anysphere for $60 billion, closing the deal earlier this month.

[what led here] OpenAI grounded the call in its track record with Musk’s companies: after the Twitter acquisition the company broke OpenAI’s contract terms, and xAI admitted violating its terms of service. The contract gave OpenAI a limited cancellation window after a change of control, and the company says it is giving the maximum notice the agreement allows so developers have time to migrate. OpenAI models currently serve about 5% of Cursor’s traffic; Anthropic has since said it will keep supplying the tool.

[a two-month window] Development teams built on Cursor have a little over two months to switch models or switch tools. The harder lesson lands on every startup distributing someone else’s models: supply can be severed by a single upstream control provision, and who acquires you now belongs in the technical evaluation. The Altman–Musk conflict has moved from rhetoric into the supply chain.

▪ SIGNALTerms of service became a competitive weapon; acquisition ownership now belongs on page one of the architecture review.

❯ CXMT posts first-half revenue of 150.3 billion yuan, up 873%, with 77.6 billion yuan in profit

[first report since listing] CXMT released its first semiannual report since going public: first-half revenue of 150.31 billion yuan, up 873.64% year over year, and net profit attributable to shareholders of 77.605 billion yuan, swinging from a loss. Excluding non-recurring items, profit was 78.793 billion yuan, also a reversal. Gross margin reached 84.84%.

[where the money came from] Per the filing, the company was still loss-making a year earlier, and attributes the result to fast-growing global compute demand layered on tight DRAM supply and sharply higher prices. Operating cash flow was 131.156 billion yuan, up 2,985.64%, with R&D spending of 6.859 billion yuan, up 87.38%. China’s largest integrated memory maker has been listed for a single month, during which its shares rose nearly sixfold, making it the most valuable company on the A-share market.

[a cycle extreme] An 84.84% gross margin is not normal for chip manufacturing; it records a cycle extreme. The cost increase has already reached end devices, and device makers’ bills of materials are the other side of the rising handset ASPs and falling units SemiAnalysis tracked over the same period. Investors betting on a cycle reversal need to answer one question first: how much of this price surge is share won through domestic substitution, and how much is simply an industry-wide shortage.

▪ SIGNALA memory maker clearing 77.6 billion yuan in six months shows AI investment profits piling up upstream along the supply chain.

❯ Xiaomi 18 Fold arrives in September, debuting CXMT LPDDR6 and the in-house 3nm Xring O3

[a double debut] Xiaomi and CXMT confirmed that the Xiaomi 18 Fold, launching in September, will simultaneously debut CXMT’s LPDDR6 memory and Xiaomi’s in-house 3nm Xring O3 processor. It is the first time a domestically produced LPDDR6 part ships alongside a flagship-class processor. Reuters reports a shipment target of 200,000 to 300,000 units.

[chip specs] Per TrendForce, Xiaomi’s in-house silicon previously topped out at the Xring O1; the Xring O3 is fabbed on TSMC’s 3nm process and is the first mobile processor to support LPDDR6, hitting 10,667Mbps and 113.8GB/s of memory bandwidth, a 48% gain over the O1. It packs 24 billion transistors, 26% more than its predecessor, on a 133mm² die. Xiaomi has two more chips at TSMC: the 6nm XRING O100 NPU for running its MiMo model on device, and the 3nm XRING D100 for autonomous driving.

[a validation run] Two to three hundred thousand units is a validation run, not a volume play. The pressure lands on Qualcomm’s and MediaTek’s share assumptions in China’s flagship tier: both are preparing 2nm parts while Xiaomi already has a 3nm in-house chip in a shipping phone. For CXMT, this pushes the proving ground for domestic memory from the server side into consumer devices. Watch the return and yield data on those 200,000 units.

▪ SIGNALAn in-house SoC paired with domestic memory in one phone — what Xiaomi is really selling is proof that a supply chain works.

❯ Tencent open-sources Hy4 preview: 770B parameters, 1M context, self-scored just past GLM-5.3

[model specs] Per Tencent’s announcement, the company released and open-sourced Hunyuan Hy4 preview, a mixture-of-experts model with 770 billion total parameters and 49 billion active, and a context window above 1 million tokens. Weights are on Hugging Face, and the model is wired into Tencent’s own Yuanbao, ima and CodeBuddy, with API access via Tencent Cloud TokenHub and OpenRouter.

[read the benchmark carefully] Tencent says an internal blind evaluation had 163 experts score 203 engineering tasks, with Hy4 preview averaging 2.99 out of 4.00, edging GLM-5.3 at 2.92 and Kimi K3 at 2.94; against GLM-5.3 the win/tie/loss split was 46.8%, 12.8%, 40.4%. The three scores sit close enough to read as rough parity rather than a win, and Tencent designed and ran the evaluation itself, with no third-party benchmark replicating it on the same terms.

[the field has bunched up] Within a single week, GLM-5.3, Kimi K3 and Hy4 have crowded into the same capability band. When rankings rest on self-evaluation and the gaps land in decimals, an engineering team’s selection criteria fall back to license terms, deployment cost and inference reliability — things you can see. Leaderboards stop carrying information at this range.

▪ SIGNALA 2.99-to-2.92 gap says the contest among Chinese open-weight models has moved from benchmarks to engineering delivery.

❯ Z.ai opens GLM-5.3 weights but makes $10B-revenue firms pass its own security review

[the license changed] Z.ai released GLM-5.3’s weights, swapping the permissive MIT license used for GLM-5.2 for a custom GLM-5.3 license: any company with aggregate revenue above $10 billion over any 12 consecutive months must pass Z.ai’s security review before hosting the model for commercial use.

[who the clause targets] Per The New Stack, GLM-5.2 had run under a permissive license, and the new threshold aims squarely at hyperscalers and neoclouds — individuals and smaller teams keep their rights to run, deploy and fine-tune. GLM-5.3 is a 753-billion-parameter coding model, launched two weeks ago and held back from open release for safety hardening, which the company calls the strongest open-weight model for coding. Weights are on Hugging Face and several third-party inference services already serve it through OpenRouter. Local deployment is not cheap: FP8 needs roughly eight H200s, and an aggressive 2-bit quantization runs 230GB to 250GB.

[where the gate sits] Hyperscaler legal teams now have to clear a Chinese company’s security review before listing a model, a structure the open-weight ecosystem has not seen before. The default assumption that open weights mean free commercial use stops holding with this release, and enterprise inference compliance checklists need another line. Watch whether major cloud platforms delay listing it as a result.

▪ SIGNALAn open-source license is being used to screen who may host a model; Chinese labs are putting gates on their own distribution.

❯ Lambda borrows $1 billion for Nvidia GPUs, with repayment riding on Microsoft’s lease

[the financing structure] Nvidia-backed AI cloud provider Lambda has raised about $1 billion in private short-dated debt arranged by JPMorgan to buy Nvidia chips that will be leased to Microsoft, Bloomberg reported. The short tenor is the point: Lambda is betting it can deploy fast, generate cash fast, and repay from lease income.

[borrowing in sequence] This is not a one-off. Lambda closed a $1 billion secured credit facility in May and this week announced a $926 million loan to fund Nvidia GB300 GPUs, while reportedly negotiating a $3 billion pre-IPO round. The repayment source is effectively Microsoft’s lease payments rather than Lambda’s own operating revenue. A separate arrangement leasing GPUs back to Nvidia is also reported to be in progress, putting two cash flows against the same chips.

[what is being priced] The structure pushes AI compute financing toward structured finance: credit market risk pricers are no longer judging Lambda’s credit but the term of Microsoft’s lease and the residual-value curve of Nvidia silicon. If GPUs depreciate faster than the lease pays down, the risk travels back up the chain to the banks lending to these neoclouds.

▪ SIGNALWhen repayment comes from another company’s lease, AI compute debt is no longer priced on the borrower.

❯ Meituan’s Q2 revenue hits 104.6 billion yuan, ending three losing quarters; Keeta profitable in Saudi in 22 months

[back to profit] Meituan reported second-quarter revenue of 104.6 billion yuan (about $15.62 billion), up 14.4% year over year, with adjusted net profit of roughly $372 million, ending three consecutive losing quarters. The turn came as the food-delivery subsidy war cooled — discounting eased noticeably after regulators publicly criticized the instant-retail price war earlier this year.

[the overseas line] On the earnings call, Wang Xing laid out the unit-economics timeline abroad: Keeta reached profitability in Saudi Arabia 22 months after starting trial operations, and turned unit economics positive in Hong Kong in 29 months, which the company reads as evidence the model travels. In the second half Keeta will focus on operating efficiency in existing markets, and Meituan says it remains confident in Brazil’s long-term potential.

[what decides next quarter] Those two figures, 22 and 29 months, give teams building local services abroad a payback period to benchmark against. At home, the subsidy truce came from regulatory pressure rather than the end of competition, and Douyin’s push into instant retail remains the biggest variable in Meituan’s margin recovery. Watch whether discounting returns once the third-quarter peak season arrives.

▪ SIGNALRetreating subsidies bought one profitable quarter; how long Douyin is willing to lose money decides the next one.

❯ OpenAI confirms it is testing a persistent Codex mode, with no near-term launch plans

[confirmed on the record] An OpenAI spokesperson confirmed the company is testing a persistent mode for its coding agent Codex, letting the agent work proactively across sessions and generate its own follow-up tasks until a user puts it to sleep. The spokesperson said the feature remains in testing, with no near-term plans to launch and no broad availability.

[how it differs] Existing modes stop after minutes or hours even when a task is unfinished. Wired found the code in Codex’s command-line repository in August, where it states that once the mode is selected the agent will “continue working until put to sleep,” drawing on prior interactions to decide what comes next. It is not enabled in any shipped build. That CLI has historically been the testbed, and most features there eventually reached the desktop and web clients.

[who should plan ahead] Enterprise security teams need rules in place first: an agent that runs continuously and assigns itself work needs permission boundaries and audit logs designed for a resident service, not a one-off session. An agent that never clocks off also invalidates per-session cost models, so budgeting has to be rebuilt.

▪ SIGNALGoing from “run a task” to “always on” changes both the billing unit and the permission model for agents.

❯ Fable 5.1 and Opus 5.1 said to slip to next week, though Anthropic never announced either

[what is circulating] Posts on social platforms claim Fable 5.1 and Opus 5.1 were set for release on August 28 and have slipped to next week, with some people said to already have access. The caveat comes first: Anthropic has never officially announced either model, so the “delay” is a community expectation window moving, not a confirmed schedule change.

[how much to trust it] The claim so far traces to a single leak-tier account, with no official statement, no reporter follow-up, and no verifiable product page or documentation. More than one version of the date is circulating, which is itself a sign of an unstable source. By contrast, GLM-5.3’s weights and Tencent’s Hy4 landed the same day as downloadable, checkable artifacts.

[how to use this] Teams scheduling model selection should treat it as a prompt rather than a basis: leave slack for a new version arriving, but do not move a launch plan on an unverified date.

▪ SIGNALA model that was never announced cannot be delayed; what is worth noting is how far release timing is now driven by outside expectations.