Honeywell and Ecolab executives said deploying AI models in buildings and industrial operations is limited by real-world needs for accuracy and cost. Ecolab cut token costs by about 70% to 80% after running its best-available model on high-volume work, where token “tokenomics” became too expensive. They shift toward semi-autonomous, human-led operation and more optimized model setups, rather than fully autonomous agents at scale.
CrowdStrike and CoreWeave announced a partnership to deliver machine-speed AI security by combining CrowdStrike’s adversary expertise and security data with CoreWeave’s training and inference infrastructure. The article cites an eCrime attacker breakout time of 27 seconds in 2025. As a result, they plan to integrate defensive AI and automation into always-on infrastructure and continuous integration/delivery pipelines to speed up security response and reduce misconfiguration windows.
NetApp described how storage operations can be handled by AI agents under a shared control setup while humans still set the limits. In a keynote demo, agents detected a nighttime performance anomaly without waking an engineer. The approach shifts governance to explicit policy boundaries and RACI-style accountability so exceptions and oversight remain human-controlled.
IBM and CoreWeave co-designed workload controls for agent execution, extending IBM’s research infrastructure from model training into isolated code-running and testing. They built a large H100 cluster for IBM Research’s Granite workflows. The work resulted in joint identity integration plus CoreWeave Sandboxes that let researchers choose where agent code runs and what resources it can access.
NVIDIA announced a 64GB version of its DGX Spark desktop AI system aimed at running local models and always-on agents, with optional clustering for more capacity. The DGX Spark 64GB is rated up to 1 petaFLOP of FP4 AI compute. As a result, developers can keep token use local (avoiding per-token API billing) and scale by clustering two 64GB units for 128GB memory and higher throughput, instead of moving workloads to metered cloud services.
Road to KubeCon argues that AI agent harnesses should move from laptop-style, local setups to cloud-native designs that separate the agent loop from supporting infrastructure and services. Koordinator was used by Zhuoyu Technology to push on-Kubernetes GPU allocation above 95% and GPU utilization above 55%, addressing inefficiencies from the default Kubernetes scheduler. As a result, agent and workload management becomes more distributed and scalable, with better scheduling outcomes for AI and other microservice workloads.
Allen Institute for AI released Olmo-core 3, a development framework intended to make mixture-of-experts large language model training more efficient. It reported 52,000 tokens per second on Nvidia B3000 GPUs for a 47-billion-parameter model, about 2.7× higher than Megatron-core at roughly 19,400 tokens per second. The update supports growing expert pools from 8 to 128 while selecting 4 experts per token, enabling scaling to over 1 trillion parameters with lower memory and compute overhead.
Enterprise AI is transitioning from tool use to an “operating model” that connects people, processes, and data in real time to avoid siloed intelligence. Global AI investment is set to reach $2.5 trillion in 2026, up 44% from the prior year. Enterprises now need process redesign, composable architectures, and sovereign data/model control rather than just better models or faster infrastructure.
TechCrunch Disrupt 2026 will host Jas Khaira of Blackstone N1 on the Builders Stage to discuss how AI companies should scale capital and separate early momentum from lasting businesses. The event runs October 13–15 at Moscone West in San Francisco. The focus shifts to investors’ criteria for funding AI growth needs like compute and infrastructure rather than simply raising more money.
Nvidia announced the Shield TV Pro would rise in price effective immediately. Starting October 2, the Shield Pro will cost $299.99 instead of $199.99. Nvidia says higher component costs, including memory, are driving the change, attributing it to industry-wide effects from generative AI supply chains.
Headline closed a $400M European fund, Oura delayed its $15B US IPO, and Mykhailo Fedorov launched Army of Robots. The fund closing was for $400M. The funding and launches signal fresh backing for European AI and robot-related efforts while Oura’s planned listing is pushed out.
NVIDIA will release a DGX Spark configuration with 64GB of unified memory through partner PC makers to run local AI agents and models with its software stack. The 64GB SKU starts at $4,999 and is available from Oct. 23. Developers can scale by clustering two units to pool memory and run larger models with NVIDIA Sync without reconfiguring the software environment.
Volantis raised $88 million to develop its A-1 photonic memory architecture for AI inference. The company targets models over 20 trillion parameters at up to 10,000 tokens per second per user and plans to deliver its first integrated inference engines in 2027. The funding supports expanding its engineering team and moving toward commercial customer deployments of the photonic inference platform.
NetApp and Iterate.ai announced a partnership to run private generative AI using enterprise storage with the AIPod Mini appliance. The setup includes more than 200 agent templates, more than 200 skills, and access to more than 800 tools. As a result, more enterprises can move from pilots to production deployments that keep model and data within their own environment while agents deliver measurable business outcomes.
A San Antonio city councilmember says data centers in his district have expanded from smaller buildings into hyperscale sites tied to AI infrastructure, concentrating local nuisances in specific neighborhoods. He expects the count of data centers in his roughly 55-square-mile district to rise to around 20 over the next few years. The focus shifts from general growth to how these larger facilities affect traffic and where they’re clustered as new projects are proposed.
A tutorial implements Google Research’s Kauldron JAX training library by demonstrating konfig (plain dict configs), kontext (string-key wiring that prevents components from importing each other), ktyping (runtime named-axis shape checks), and kd.train (a readable Trainer), then runs a synthetic CPU training run and a 5-variant config sweep with checkpointing and resume. It installs Kauldron pinned to version 1.4.2 and applies a small compatibility patch because a Trainer can otherwise raise AttributeError before completing a single step. The result is a workflow where changing one config line reshapes dependent values via references and where training components and tensor shapes are validated and rewired through data and strings instead of code edits.
Every AI story that matters,
in your inbox by 8am.
TLDRocket reads all relevant sources, removes duplicate coverage, and summarises the
day in two minutes. Follow companies and topics for alerts, or get the
briefing in Slack. Free, no spam, unsubscribe anytime.
Reading TLDRocket needs no cookies, and the readership counts we rely on come from
our own cookieless analytics. Google Analytics is the exception: it sets cookies and
reports to Google, so it stays switched off until you allow it. You can change your
mind any time from “Cookie settings” in the footer.
Strictly necessary
Session security and form protection (tldrocket-session,
XSRF-TOKEN, 2 hours). The site cannot work without them,
so they need no consent.
Always on
Google Analytics 4 (_ga,
_ga_<id>, up to 2 years). Measures which
stories and sections readers use. Google acts as a third-party processor and may
store the data outside the EU. No advertising, no profiling, no data sold.
