Infrastructure
-
86 minutes
Claude’s two 5.1 models threw errors for 86 minutes on Tuesday morning
Anthropic’s status page opened an incident at 09:48 UTC on 15 September for “intermittent error spikes” on Claude Mythos 5.1 and Claude Fable 5.1.
If retries fired in that window, it was them. There is nothing to change.
-
Nearly cut
Two PaperCut bugs, three days old, reached 395 organisations
Kept although it predates this window, because it is the attack the three fixes above are defending against, and this paper has never printed it.
Upgrade to 24.1.10, 25.0.13 or 26.0.5; the fix is a version, not a setting. Then look through your own estate for the shape rather than the product: a bug that only writes configuration is a privilege escalation whenever some other component reads that configuration as a class name, a path or a URL.
-
Two weeks
Azure retires the NVv3 and NVv4 GPU families on 30 September
Microsoft retires six NVv3 sizes and eight NVv4 sizes on 30 September, pointing customers at NVads-A10 v5 and NVads-v710 v5. UAE North runs to 31 January 2027. Reserved-instance sales for both families ended on 2 November 2025.
Migrate before 30 September, and register the feature flag first or the resize errors.
-
The seller’s terms
Oracle’s customers prepaid a sum equal to two fifths of what it spent building in the quarter
Oracle’s customers prepaid it $11.4bn in the three months to 31 August, against $28.5bn it spent on capital projects. Its buyers are financing its build. The cash-flow line reads “increase in deferred revenues from customer prepayments with significant financing component,” and that accounting label means the money arrives far enough ahead of the service that it works as a loan from the customer to the builder of the capacity.
If you are negotiating OCI capacity and are asked to prepay, put delivery dates and refund terms in the same contract. A seller whose buyers finance the build has no reason to discount it.
-
The default moved
vLLM switched every model to a new runner and deleted ten architectures on its way out
Version 0.29.0 makes Model Runner V2 the default for all models, turns FlashInfer all-reduce on by default, removes ten model architectures and deprecates the `api_server` entry point that sits in a great many Dockerfiles. NVIDIA’s TensorRT-LLM removed its own sampler the same morning.
Before you bump vLLM, grep your serving config for the ten removed architectures and your Dockerfiles for vllm.entrypoints.openai.api_server. If you are on TensorRT-LLM’s TRTLLM sampler, migrate before rc26 rather than after, and if you serve Qwen3 MoE with LoRA adapters, verify the adapter is actually being applied.
-
Who owns the compute
Google is paying to keep a Finnish reactor running for 22 years to power four data centres
A 13 billion euro commitment across 2027 and 2028 arrives with 629 MW of new wind, a 94 MW battery, and a power purchase agreement tied to extending the life of Loviisa.
Nothing here adds European capacity you can buy this year. If you are modelling EU-region pricing, assume supply loosens after 2027 rather than before, and do not read the generation figures as capacity numbers.
-
Export controls
The Blackwell servers that reached ByteDance came through a 33 per cent stake
Aivres, one-third owned by blacklisted Inspur, shipped over $3 billion of Nvidia Blackwell systems that reached ByteDance and Alibaba, the New York Times reported. The Sunday investigation traced $5.6 billion of technology from the California server maker into Southeast Asia between April 2024 and February 2026, the Blackwell systems more than half of it. The route is the rule’s arithmetic: Inspur has been on the Entity List since 2023 and the 50 per cent Affiliates Rule BIS launched in September 2025 does not reach a 33 per cent holding. Nvidia spokesman John Rizzo said the company does not support diversion, and on the Times’s account it did not need to. The paper’s verdict: nothing changes for a builder until BIS names Aivres, which it did not do when it listed six other Inspur subsidiaries in March 2025 without naming it.
Nothing to do this week. The number to watch is whether BIS names Aivres before the mid-September US-China talks; until it does, the Blackwell capacity some Chinese labs train on arrived through a rule’s arithmetic.
-
Provenance
A quarter of vLLM’s last three hundred commits carry an agent’s name
One of them is signed off by codex@openai.com. The Developer Certificate of Origin that signature satisfies was written to be signed by a person, and vLLM’s contributing rules do not mention agents at all.
If your compliance process treats a DCO sign-off as provenance for a dependency, that assumption is now partly false for vLLM and probably for others. Check what your own contribution rules say about agent-authored commits before someone asks you.
-
Serving
SGLang ships day-zero serving for a 2.4 trillion parameter mixture of experts
SGLang v0.5.19 landed on 5 September with day-zero support for Qwen3.8, 2.4 trillion parameters in total against 95 billion active. IBM Granite 4.2 and Ling 3.0 flash and tiny arrive in the same release. Day-zero means the checkpoint serves on the release that ships beside it, rather than after a follow-up patch.
-
Vector stores
Rebalancing a Qdrant cluster got a third faster, and the fix was to stop parsing
Qdrant 1.19.1 moves raw payloads between shards instead of deserialising them, which its changelog says makes transfers “1.5x faster” than before. The same release makes replica-state changes and transfer restarts crash-safe, to stop consensus desynchronising.
-
Vertical integration
Nvidia is putting $2B into the cloud that sells Anthropic its GPUs
Nvidia is putting $2 billion directly into Nscale, TechCrunch reported on Friday, and Nscale spends its capital on Nvidia accelerators. The rest of the $3.5 billion pre-listing raise is $1.5 billion in convertible notes, ahead of a possible flotation this month.
If you are not inside one of these financed chains, price 2027 capacity as scarce and do not assume a published rate reflects what large buyers pay.
-
Infrastructure
Neocloud security is now part of model supply
SemiAnalysis reports serious weaknesses found while testing GPU-cloud providers: container escape paths, weak cluster policies, exposed monitoring, and outdated software. Its broader data does not yet show a universal AI-driven explosion in vulnerabilities. The practical conclusion is narrower and more urgent: every new compute vendor is another security boundary your model stack inherits.
Ask for the provider’s patch cadence, tenant isolation, admission controls, and incident process with the same seriousness as price and accelerator availability.