Data centre operators and researchers are shifting from copper wiring toward photonics to cut the energy used by power distribution, cooling, and heat from electronic networking. A typical 100MW data centre uses about 400 tonnes of copper, including up to 70 tonnes in servers and up to 20 tonnes in network wiring. This changes data-centre infrastructure by replacing some copper interconnects with light-based optical links while requiring new manufacturing, installation, and thermal-reliability practices before wider deployment.
Firecrawl announced it closed a $75 million funding round for its AI-focused web scraping platform. The Series B round was led by Los Angeles venture firm Smash Ventures. The funding will be used to enhance its new Alexandria cloud service by adding more third-party data providers and launching a self-service licensing system soon.
Hundreds of workers are building new data centres in rural Inner Mongolia to supply the computing power behind China’s AI platforms. The article says China is building “hundreds” of data centres across provinces including Ulanqab and aims to embed AI across 90% of industries and society by 2030. As a result, infrastructure and power generation plans expand while some rural communities report village relocations and limited local benefits, even as others see job opportunities and faster AI innovation.
BBC reporter Laura Bicker visited Inner Mongolia to describe how Chinese data centres are being built to support the country’s artificial intelligence ambitions. Dozens of data centres are under construction there. The reporting highlights differing attitudes toward the data centre industry in China versus the US ahead of Xi Jinping and Donald Trump meeting in Washington later this week.
GPT-6 improves how prompt caching works with features aimed at raising cache hit rates and adding diagnostics, breakpoints, and controls.
It targets higher cache hit rates and reduces prompt latency and costs.
This changes prompt processing by making caching more effective and more observable, which should cut repeated prompt overhead.
Qualcomm announced the Snapdragon 8 Elite Gen 6 and Snapdragon 8 Elite Extreme Gen 6 smartphone chips with an emphasis on on-device AI features. The Extreme model can run a 30-billion-parameter mixture-of-experts model locally while activating only part of the parameters for each task. As a result, phones can support locally run personal voice agents and sensing hubs that enable speaker differentiation, task automation suggestions, and higher-end camera/video capabilities.
OpenClaw chief architect Vincent Koc described how the project is shifting from single-agent, single-task demos to a system where many agents share sessions, context, and control across machines. In his own setup, he reported running about 60 to 70 agents across 6 to 7 machines and averaging 5B to 10B tokens per day (peaking at 54B in a single day) while deliberately stress-testing the architecture. The result is an agent “multiplayer” approach with a shared control plane, persistent intent and memory, interchangeable remote compute via Crabbox, and a focus on security and model routing to manage cost and risk.
Rabbit is rolling out OS3, a standalone AI agent that runs without needing its R1 hardware. The cloud OS3 lets a single account add up to 5 devices. OS3 will run locally on Windows, Mac, and Linux and can be accessed via a desktop site, messaging apps, or the R1 device, with automatic selection of devices, files, apps, and AI models.
Verda raised fresh funding as it expands its AI cloud business. The Helsinki-based startup brought in $189m in a round led by Emergence Capital. The extra capital supports further growth of its AI infrastructure offering.
Motorola announced the Signature 27 as its first phone to use Qualcomm’s Snapdragon 8 Elite Extreme Gen 6 chipset. The Extreme Gen 6 is Qualcomm’s top-end model. The launch positions Motorola’s next flagship as its most advanced in years, though exact specs, price, and release date are still not provided.
Reactiv built an AI Scheduler for Shopify merchants that automates native iOS and Android app updates using Amazon Bedrock AgentCore. It reduced merchant configuration time by 80 percent and cut time to production by 33 percent. By unifying interactive and scheduled agents on one AgentCore stack with shared memory, tools, and UI protocol, Reactiv enabled bidirectional learning so scheduled updates get better as merchants use the dashboard.
Amazon SageMaker AI Inference Recommendations adds concurrency sweeps to benchmark generative AI endpoints by sending controlled simultaneous requests and measuring throughput and latency to find the saturation point. In the walkthrough, the endpoint’s throughput-to-latency “knee” appears at 256 concurrent requests. Teams can use the results to right-size instance type and serving configuration (and optionally run an SLA-based search) to automate capacity planning and avoid over- or under-provisioning.
Trane Technologies built an AI-powered, agentic diagnostic assistant on Amazon Bedrock AgentCore to turn multi-screen HVAC troubleshooting into a conversational workflow. The team reduced a 20-minute diagnostic process to a 20-second natural language interaction in 3–4 weeks, based on internal technician benchmarking. This changes operations by speeding time-to-insight 60x and shifting staff from reactive work toward more proactive, data-driven optimization.
Tata Elxsi built IRIS, a real-time industrial safety platform on AWS that analyzes camera feeds at the edge, runs computer-vision models in the cloud, correlates detections, and triggers automated alerts. Its measured event-ingestion latency is under 200 milliseconds at p95. This shifts safety monitoring from recorded footage and 15–45 minute detections to continuous, near-real-time alerting without streaming raw video to the cloud.
Salesforce Agentforce for Public Sector is being extended to process unstructured evidence by storing files in Amazon S3, running Amazon Bedrock Data Automation, and exposing extracted results back to users through MCP in the Salesforce console. The integration triggers processing via S3 event notifications and uses DynamoDB and Lambda, including a second Lambda step that saves outputs to a dedicated S3 output bucket. This shifts evidence handling from manual review and extraction to automated, event-driven multimodal processing with natural-language retrieval from Salesforce.
AstroForge is developing an autonomous spacecraft control stack, “Solo,” to reduce reliance on Earth-based flight controllers. The company plans to fly its first autonomous spacecraft in 2027 with its launch partner Stoke Space. It is combining traditional control algorithms with transformer models trained on test data and sensor inputs, aiming for onboard anomaly resolution while also limiting early missions (e.g., “shadow mode” and no radios for Autonomy-1).
Data & Society’s report says opposition to new data center construction in Pennsylvania is driven by how residents experience industry messaging and local impacts rather than by a single political cause. It is based on 18 months of ethnographic fieldwork with 44 Pennsylvanians conducted between 2024 and 2026. The findings shift the debate from measuring construction effects to addressing trust, lobbying and local concerns that the AI industry struggles to reconcile with demand.
Nscale will test Wall Street’s demand for its IPO as its revenue depends mainly on two AI computing customers, Microsoft and Anthropic. Anthropic’s deal is worth $44.6 billion and is contingent on Nscale securing financing, with the customer able to cancel if Nscale misses milestones. The listing on the NYSE and its expected $35 billion valuation plus plans to raise $3 billion shift investor focus toward the risk of concentrated AI infrastructure revenue.
CoolIT says that above 250 kW per rack, liquid-air hybrid cooling becomes impractical and it shifts to near-total heat capture to enable fanless liquid-cooled AI servers. It claims air can fall below 1% of the load, with a 70/30 liquid-air split leaving 75 kW to be handled by air. This changes server design toward integrating all components into a single liquid loop built from modular coldplate blocks rather than relying on air cooling for non-CPU parts.
NVIDIA released Isaac ROS 5.0, an open-source set of GPU-accelerated ROS packages for building robotics apps with AI agents. The update adds an agent-ready inference library for FoundationPose that speeds up object pose tracking by up to 5.5x. It expands support to ROS Lyrical and Ubuntu 24.04 and adds new skills/workflows plus Jetson platform coverage to move faster from development to on-robot deployment.
Linear’s team redesigned its CI to remove bottlenecks created by faster AI-assisted code shipping, by optimizing runner time and PR wait time. Pull request wait time dropped from more than 6 minutes to just over 5, despite test suites nearly quadrupling since the start of the year. CI became faster and more efficient through changes to infrastructure, checkout, gating, setup, and sharding, cutting various CI costs and reducing developer/agent feedback latency.
Verda, a Helsinki-based AI cloud provider, closed a Series B funding round and reached unicorn valuation as it builds out GPU capacity for AI training and inference. The round raised 189 million dollars (about 161 million euros) and was led by Emergence Capital. The company will use the money to expand data centers, hardware, and its software platform, aiming for more than 250 megawatts in operation by 2027.
Verda raised $189 million to scale its full-stack AI cloud and expand compute capacity. The round was led by Emergence Capital and pushed Verda’s total funding to more than $450 million, valuing it at over $1 billion. Verda will use the money to accelerate product development across its AI cloud, increase inference and multiply its compute capacity over the next year, and expand internationally.
Morphotonics raised more than €40 million in Series B to scale its nanoimprint lithography manufacturing technology for waveguides used in AI glasses and other photonic components. The funding round totals over €40 million and includes Invest-NL as a major new investor plus participation from the EIC Fund, Ernij Next, and the EIB. The company will use the money to expand production capacity for its Cypris platform, grow its teams, and build supplier, after-sales, and customer-support infrastructure to enable higher-volume, lower-cost manufacturing.
Zeliq raised a €7 million seed extension to grow its sales prospecting platform and AI-native “Agentic Revenue Workspace.” The round brings Zeliq’s total funding to €21 million in less than three years. It will accelerate development of its AI infrastructure, expand its suite of AI sales agents, and double product and technology headcount while continuing European growth and CRM partnerships.
Taiwanese chip and AI experts describe how Taiwan’s semiconductor dominance makes advanced AI chips widely available while Taiwan is largely absent from training the AI models that run on them under pressure from Beijing.
Lightspring closed a seed funding round to develop 3D-lithography optical interfaces that reduce the photonic integrated circuit packaging bottleneck. The round includes €3.1 million led by Atlantic and OVNI Capital, and Lightspring also secured over €3.5 million total financing after a €350,000 i-Lab grant. The company will use the funding to move its 3D optical coupling process from lab to a production-ready manufacturing line and expand its PDK and design tools for industrial volume.
Koreshield launches authentication and identity and AI infrastructure tooling that screens a support agent’s customer messages, retrieved documents, and proposed tool calls before they can influence model behavior or execution. It says it can be integrated with one call. As a result, it records and checks for issues like data leaks, hidden instructions, policy drift, and unsafe agent actions.
CtrlOps 1.0 was launched as a follow-up to CtrlOps, offering server auditing and management with AI. It provides a security hardening score across 25 checks plus a PDF audit report. It changes server access and deployment workflows by letting users review logs and revoke or assign roles without using SSH or handling credentials off-device.
Rollouts automate traffic migration between two model deployments by running canary, blue-green, or rolling upgrades with built-in health checks and rollback controls. The default canary traffic ladder is 5% → 25% → 50% → 100% with a wait window of at least 390 seconds per step when using metric gates. This replaces manual hard swaps or DIY staged cutovers by letting the platform enforce step ordering, pause on failed gates, and complete or cancel without leaving traffic in an ambiguous state.
Hugging Face added GGUF support to Transformers so GGUF checkpoints can be loaded with from_pretrained and run locally on Apple Silicon using GGML/Metal kernels. The initial focus is on Qwen3.5 architecture and the benchmark run is on a MacBook Pro M2 Max with 32 GB unified memory. This changes Transformers to become an efficient GGUF local-inference and serving option, while GGUF files now work through the standard Transformers API instead of requiring llama.cpp as the only inference path.
Jun Kim, creator and maintainer of oMLX, has joined Hugging Face to support the MLX local-AI community. The article says oMLX will remain under the Apache 2.0 license. This funding and full maintenance shift is expected to stabilize oMLX and speed up long-term development, while Hugging Face works to help transformers model definitions transition into reference MLX implementations.
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.
