7 min read · 13 sources

Nvidia’s dedicated inference silicon is no longer a paper launch. The Groq 3 LPX accelerator is in full production as an extension of the Vera Rubin data center platform, with Nebius Group as the first committed cloud customer. A full rack-scale deployment bolts up to 256 LP30 accelerators onto Vera Rubin GPUs over high-bandwidth interconnects, and the design splits the work cleanly: Vera Rubin GPUs handle context processing, the LPX boxes handle token generation. That separation is what makes it interesting for agentic AI, where the bottleneck is decode latency on long, multistep chains rather than raw pretraining throughput. If you are sizing inference fleets for agent workloads, this is the architecture to model against.

ClickHouse’s LogHouse ingests 50 million OpenTelemetry events per second and stores 177 PiB uncompressed, and the in-memory queues that handled 99% of cases started losing telemetry the moment backpressure hit.

OpenAI is clawing back corporate share from Anthropic

Source: techcentral.ie ↗

Ramp’s spending data across more than 70,000 US organisations tells a different story than the one most AI Twitter threads are running with. Anthropic led in May (41% to 39%) and pulled further ahead in July (44% to 40%). Q3 trends, though, show OpenAI growing faster, and Ramp economist Ara Kharazian credits new developer-facing products for the rebound. Anthropic’s specialised Fable model is dragging on its own numbers: customers are balking at cost and a 30-day data retention policy. The aggregate number is the more interesting one. Paid AI adoption across Ramp’s customer base went from roughly 50% in March to 56% in July. The pie is growing, and the lead is flipping quarter to quarter. Procurement teams who locked in annual Anthropic contracts in June are now staring at a very different competitive picture.

IBM wants one chip to run your mainframe and your Kubernetes pods

Source: venturebeat.com ↗

IBM is developing a dual-architecture processor that natively executes both Arm and IBM Z instructions on the same cores. The pitch is operational: an enterprise could run modern AI and cloud-native software on the same physical hardware that is currently sitting in a glass-walled room running CICS and DB2. The licensing and capacity-planning implications are bigger than the silicon. If this lands, the “lift and shift” conversation for z/OS workloads changes shape, and so does the cost model for shops that have been keeping two fleets alive with separate power, cooling, and audit regimes. No timeline in the source, but worth tracking the moment IBM starts quoting one.

Stop reaching for Kafka Streams when you actually need a queue

Source: kai-waehner.de ↗

Kai Wähner’s post on when not to use stream processing is the kind of architecture-team reading list material that gets passed around Slack for a reason. He defines stream processing narrowly: data-arrival-initiated continuous computation over unbounded events, Flink, Kafka Streams, Spark Structured Streaming. He then argues it is only the right tool for stateless transformation, stateful streaming, complex event processing, and real-time ML inference. Everywhere else, call it what it is: a request-response API call, a state query, a correction flow, or a lock. Two concrete cases back the argument. Kestra 2.0 ripped out Kafka Streams and rebuilt around plain queueing. Confluent Control Center was rearchitected in May 2025 around Prometheus. If your team has a “we need a stream processor” slide in the next design review, read this first.

Defenders need the same model access as attackers, Stratechery argues

Source: stratechery.com ↗

Ben Thompson’s Autonomy and Innovation essay uses the white-hat/black-hat framing to argue that AI capability is separable from intent. Defensive AI needs the same offensive access to vulnerabilities, and the only viable move is making sure defenders have equivalent access to top models. Thompson criticises the Trump administration’s response to Anthropic’s Fable release and to the earlier Hugging Face attack, noting that Hugging Face repelled the attack with help from open-weight Chinese models. The parallel he draws is bug bounties: build the incentive structure so the defensive side has the same firepower. Operationally, the takeaway is that any enterprise security programme that treats model access as a procurement decision rather than a defensive capability is going to be on the wrong side of the asymmetry.

Atlassian Code Context gives your coding agent a real brain

Source: byteiota.com ↗

Atlassian Code Context launched on August 12 as a capability of the Teamwork Graph. It indexes GitHub and Bitbucket repos alongside Jira, Confluence, Loom, and 50+ other integrations, and exposes the result to AI agents. Atlassian’s internal benchmarks claim 44% more accurate agent results with 48% fewer tokens, which lines up with what you would expect when semantic retrieval replaces speculative context dumps. Setup is a one-liner curl install of the Teamwork Graph CLI, dropping agent skill files into ~/.agents/skills. That directory is read by Cursor, Codex, Gemini CLI, GitHub Copilot, Claude Code, and Rovo Dev, so the same skill file lights up every agent on your engineer’s machine. Permissions flow from the existing repo access model, which is the part that actually matters for security review.

Anthropic's Slack agent stops waiting to be @-mentioned

Source: venturebeat.com ↗

Claude Tag now reads the full Slack conversation and jumps in unprompted, which Anthropic is calling “multiplayer AI.” The behavioural change is small in the spec and large in the operational consequences. Once an agent is a participant rather than a tool, your incident channel has a new actor, and the audit trail, the runbook, and the on-call escalation all need to know it is there. Worth piloting in a low-stakes channel before letting it loose on production.

Thomson Reuters bets $40M that a domain-tuned model beats a frontier one

Source: siliconangle.com ↗

Thomson is Thomson Reuters’ first proprietary LLM, now the default in CoCounsel Legal’s Tabular Analysis feature. Total spend was about $40M over two years, with the final training run at roughly $450,000, which is a useful data point for anyone arguing about build-versus-fine-tune economics. The base was an open-weight model, augmented with proprietary content and continual learning on legal domain work, with hundreds of subject-matter experts in the loop. The integration is the real moat: it is wired into Westlaw (40,000+ databases, 150+ years of content) and Practical Law. The lesson for enterprise AI teams is that a smaller, well-curated model with deep domain integration can win on workflow without needing to beat GPT-class models on benchmarks.

ClickHouse's LogHouse ingests 50M OTel events per second and is rebuilding around S3

Source: clickhouse.com ↗

ClickHouse’s LogHouse handles 50 million OpenTelemetry events per second and stores 177 PiB uncompressed across ClickHouse servers, Kubernetes, and cloud infrastructure. The interesting part is the failure mode. The original two-layer design used DaemonSet agents feeding scalable gateways, with in-memory queues absorbing backpressure. That fell over at roughly 10 GBps compressed because the queues could not ride out a downstream outage. ClickHouse is previewing a custom S3-backed collector pipeline: same agent-to-gateway topology for the 99% case, with durable overflow into S3 when ClickHouse is unhappy. If you are running OTel at scale, the S3-as-buffer pattern is worth stealing.

Neoclouds crossed $25B in 2025 and want 20% of the AI cloud market by 2030

Source: infoworld.com ↗

The neocloud segment — CoreWeave, Lambda, Nebius, RunPod, Vultr — generated over $25B in 2025 revenue. Gartner forecasts they could capture 20% of the $267B AI cloud market by 2030. Of 100+ neoclouds globally, only 10-15 operate at meaningful US scale, per McKinsey. The pitch is NVIDIA-native, InfiniBand-equipped, bare-metal GPU infrastructure with flexible contracts, aimed directly at AI workloads that hyperscalers still treat as a secondary concern. Enterprise standardisation is still the bottleneck: procurement, security review, and FinOps tooling lag the actual performance story. If you have not yet been asked to evaluate a neocloud for a training or inference workload, you will be by Q1.

Google puts Antigravity on a leash

Source: infoworld.com ↗

Antigravity is moving into Gemini Enterprise, and with it comes the cost controls IT has been asking for. Monthly project-level spend thresholds, pooled token quotas shared across developer teams, opt-in overage with monthly caps, and centralised usage metrics covering tokens, API calls, and developer activity. Some configurations will require extra Gemini Enterprise licences. The visibility closes a gap that has been forcing per-project billing workarounds for the last year.

Quick hits for the tail

Source: oneuptime.com ↗

OneUptime’s design for a canonical MTTR timeline uses an append-only canonical_event table with idempotency on (source_system, source_event_id) and corrections handled by superseding rather than mutating. PagerDuty, Jira Cloud v3 changelogs, Slack, and observability events are all mapped, and MTTR facts are only derived from approved mappings of a selected policy version. The Jira detail is worth a look: it uses the history created timestamp, not the updated timestamp, which is a common subtle bug in homemade incident timelines.

Cursor explained how it scaled beyond Git: S3 is the source of truth, local NVMe repositories handle latency-sensitive operations. If your monorepo has outgrown Git and you have been pretending otherwise, this is a useful reference architecture.