Meta contractor contaminated Cheyenne's water, city says

Micron and Anthropic sign a multi-year AI memory supply deal

Micron Technology ký thỏa thuận nhiều năm cung cấp HBM, DRAM, SSD cho trung tâm dữ liệu của Anthropic, cùng hợp tác tối ưu kiến trúc bộ nhớ cho AI, và đầu tư chiến lược vào vòng Series H của Anthropic. Thỏa thuận này phản ánh xu hướng ngành khi các nhà sản xuất chip và đám mây vừa là nhà cung cấp vừa là cổ đông của các phòng thí nghiệm AI.

Lập trình viên nên đọc bài này để hiểu cách các công nghệ xử lý bộ nhớ (HBM, DRAM) và kiến trúc lưu trữ mới đang định hình hiệu suất, tiết kiệm năng lượng cho các mô hình AI lớn, từ đó tìm hiểu cách tối ưu hóa ứng dụng của mình với những tiến bộ này.

Open, convenient and predictable: Introducing Provisioned Throughput

Together AI is launching Provisioned Throughput, a new inference tier that sits between serverless and dedicated GPU deployments. It offers reserved token capacity via Provisioned Throughput Units (PTUs) at $0.05 per PTU per minute, a 99% uptime SLA, and token-based pricing — without requiring customers to manage GPU infrastructure. Initially available for MiniMax M3 and GLM-5.2, it targets production workloads migrating from proprietary closed-model APIs, with costs claimed to be up to 90% lower than Claude Opus 4.8 at list price. Capacity is available in North America and EMEA with a one-month minimum commitment.

Amazon returns to the bond market for at least $25bn to fund its AI build-out

Amazon has launched its largest bond sale of 2026, raising at least $25bn across eight tranches with maturities extending to 2066. The funds are earmarked for AI infrastructure including data centres and custom Trainium silicon, part of a broader $200bn capital spending plan for the year. Amazon's total borrowing since early 2025 has now exceeded $70bn across multiple currencies. Investor demand reached $62bn in orders before spreads tightened, leaving a final book of ~$41bn. The sale reflects a wider trend among hyperscalers — collectively guiding to over $650bn in AI capex in 2026 — that increasingly outpaces their operating cash flows. Analysts noted cooling demand compared to earlier offerings, signalling bond buyers are growing more selective as AI-linked debt accumulates in the market.

Samsung plans a floating AI data centre by 2028

Samsung Heavy Industries plans to launch a purpose-built 50MW floating data centre barge by 2028, parked nearshore with onboard power generation and LNG fuel tanks. The move is driven by growing land constraints, water scarcity, and community opposition to onshore data centres. Samsung has signed partners including Capital Clean Energy Carriers, Lloyd's Register, and Supermicro to test AI servers at sea. Rivals Mitsui OSK Lines and Hitachi are pursuing similar ship-based approaches, while China has already deployed an undersea data centre. The economics remain unproven due to saltwater corrosion, storm risks, and connectivity costs, but the initiative signals how far the industry will go to meet AI infrastructure demand.

A £35bn plan to build 14 mini nuclear reactors in the UK

Polish billionaire Michał Sołowow's firm SGE has announced plans to build 14 GE Vernova Hitachi BWRX-300 small modular reactors across three UK sites at an estimated cost of £35bn. The fleet would generate 4.2GW — enough for roughly 8 million homes or 11% of UK electricity demand — with a target of first power in 2034. The project is privately financed but seeks government-backed Contracts for Difference and National Wealth Fund engagement. The push is driven by surging power demand from AI data centres, EVs, and heat pumps. Key risks include the fact that no BWRX-300 reactor is commercially operational yet, and nuclear projects have a long history of delays and cost overruns.

Why AI Infrastructure must evolve for Agent Experience — Akshat Bubna, Modal CTO

Modal CTO Akshat Bubna discusses how AI infrastructure must evolve from developer experience (DX) to agent experience (AX), following Modal's $355M Series C. The conversation covers why Kubernetes was never designed for bursty AI workloads, Modal's shift to elastic inference for custom models, GPU snapshotting for faster cold starts, the DeFlash speculative decoding system, Auto Endpoints for optimized inference deployment, sandboxes for RL rollouts requiring up to 100,000 concurrent environments, networked containers with private IPv6 and RDMA for distributed training, and Modal's supercloud strategy spanning 17 cloud providers. Key insight: agents need tight feedback loops, fast iteration, and observability-first infrastructure rather than human-readable YAML configs.

Hot French startup ZML releases free product to speed inference across lots of AI chips

ZML, a Paris-based AI startup backed by Yann LeCun and $20M in VC funding, has launched ZML/LLMD, a free LLM inference server designed to run open-source models across a wide range of AI chips — including Nvidia, AMD, Google TPU, Apple Metal, and Intel Arc. The goal is to eliminate vendor lock-in and help enterprises mix chip types for cost and energy efficiency. Unlike ZML's earlier open-source ML framework, LLMD is free but not open source, with monetization plans deferred until adoption is better understood. The startup competes with vLLM-backed Inferact, SGLang-backed RadixArk, and Baseten, but differentiates by targeting cross-chip compatibility and co-designing silicon with chip partners. ZML's 20-person team is notable for its angel investors including Docker/Dagger founder Solomon Hykes and Hugging Face co-founders.