#ai-infrastructure

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.

TensorX and Solstice line up $1bn to finance Europe’s sovereign AI buildout

TensorX and Solstice have announced a $1bn financing facility to fund AI hardware and data-centre capacity across the EU, targeting the growing demand for sovereign compute that stays on European soil. Alongside this, Solstice is launching aiUSX, a yield-bearing asset that lets companies put idle AI-earmarked capital to work as infrastructure lending. The product is capped at $5m at launch and is designed to generate yield that can later offset inference costs. Both companies operate within the Deus X Capital ecosystem, which positions itself as the connective tissue enabling the partnership.

China wants its green power wired straight into the data centre

China is pushing its data centre industry to connect directly to renewable energy sources rather than drawing from a coal-heavy public grid. The flagship project is a 500MW solar farm in Ningxia's Zhongwei desert, commissioned by China Datang Corp, which uses dedicated 110kV lines to supply a data centre cluster — described as the country's first large-scale direct green-power supply for computing. The full first phase pairs the solar array with a 1.5GW wind farm and storage, targeting 4.3 TWh of annual generation against a projected 2.29 TWh data centre demand. The initiative is part of Beijing's 'east data, west computing' strategy and a national goal to source roughly 80% of AI data centre power from renewables by 2030, up from about 10% in 2023. Significant challenges remain, including curtailment, grid bottlenecks, and intermittency, with the wind component still under construction and due in September.

The CPU's growing role in agentic AI infrastructure

Arm-sponsored content arguing that CPUs play a critical but underappreciated role in agentic AI infrastructure. While accelerators handle model performance, CPUs act as the control plane — managing data movement, workload scheduling, and secure isolation. Arm's Neoverse platform underpins custom silicon from AWS (Graviton), Google (Axion), Microsoft (Azure Cobalt), and NVIDIA (Grace Hopper/Blackwell), all reflecting a shift toward purpose-built Arm-based processors in cloud and AI datacenters. The piece introduces the Arm AGI CPU, built with Meta, targeting rack-level density for agentic AI deployments.

Slack Outlines Four-Phase Journey to a Multi-Cloud AI Serving Platform

Slack's engineering blog details how its AI serving infrastructure evolved through four phases: from self-managed Amazon SageMaker in an escrow VPC, to Amazon Bedrock for reduced operational overhead, then a hybrid Bedrock Provisioned Throughput and On-Demand model to handle 10× traffic swings, and finally a multi-cloud architecture adding Google Cloud Vertex AI. The final setup introduced a provider-agnostic serving layer with secretless authentication, API normalization, unified observability, and intelligent routing based on latency and error metrics. Results include roughly 10% quality improvement on complex reasoning tasks and ~67% latency reduction for short prompts, alongside improved geographic failover and reduced single-provider dependency.

OpenAI wants to claim more of the AI stack with Jalapeño, its first custom chip

OpenAI has announced Jalapeño, its first custom inference accelerator chip, co-developed with Broadcom and supported by Celestica. Designed to run all LLMs, it was built in just nine months — claimed to be the fastest ASIC development cycle ever for high-performance semiconductors. OpenAI says the chip will make AI faster, more reliable, and cheaper, and plans to deploy it at gigawatt scale in Microsoft and partner data centers by year-end. However, the announcement lacks technical benchmarks or specs, with a detailed report promised 'in the coming months.' The move is part of a broader multi-generation roadmap and reflects OpenAI's ambition to own the full AI stack — from models to underlying hardware — raising questions about future developer flexibility and ecosystem lock-in.

Runpod hits $1bn valuation on $100M raise

Runpod, a cloud startup renting GPU compute to AI developers, has raised $100M led by Summit Partners at a $1bn valuation — a tenfold increase from its ~$100M seed valuation in 2024. The company reports $240M in annualized revenue, over 1 million developers on the platform, and 20 billion inference requests processed. Unlike rivals focused solely on inference, Runpod offers the full ML lifecycle: experimentation, training, fine-tuning, and deployment. It uses an asset-light model, renting capacity rather than owning data centers, and supports both Nvidia and AMD chips. The company says it turned down buyout offers exceeding $500M to remain independent. Key risks include reliance on third-party hardware, margin pressure when chips are scarce, and well-funded competitors like CoreWeave.

China Telecom’s $1.7bn server deal hands Huawei a big win

China Telecom awarded a $1.7bn tender for 40,000 servers, with roughly 71% of the value flowing to Huawei's Kunpeng ecosystem despite Huawei not bidding directly. The larger package of 28,000 servers runs on Arm-based Kunpeng chips, replacing Intel and AMD x86 processors. This mirrors similar orders from China Mobile and China Unicom, all part of Beijing's Xinchuang campaign to purge foreign technology from critical infrastructure. The market for Xinchuang hardware is projected to reach nearly 789 billion yuan in 2026. Analysts note the shift is now AI-driven, as mixed CPU-GPU systems for agent-style AI and reinforcement learning favor Arm server chips. Caveats remain: Kunpeng still relies on a British-licensed Arm instruction set, and Chinese chips lag in the most demanding AI training workloads, keeping RISC-V on Beijing's radar as a fully sovereign alternative.

ByteDance seeks a record $20bn offshore loan to fund its AI build-out

ByteDance is in early talks with banks for a roughly $20bn offshore loan, which would nearly double its 2024 record of $9.5bn. The funds are intended to support an estimated $22.7bn capital expenditure plan for 2026, primarily directed at AI infrastructure. The company faces unique procurement challenges due to US export controls blocking access to Nvidia's top accelerators, pushing it to develop custom CPUs on Arm and RISC-V tracks, partner with Qualcomm for inference chips, and source from smaller Chinese chip suppliers. Simultaneously, ByteDance remains Microsoft's largest AI customer, spending over $1bn annually on OpenAI models via Azure. The loan would help finance both its domestic hardware buildout and its continued reliance on Western AI models.

Blackstone plans $30B for Japan AI data centres, and is unbothered by bubble talk

Blackstone plans to invest $30 billion in AI data centres in Japan over three to five years, targeting facilities with combined capacity exceeding one gigawatt. President Jonathan Gray dismissed AI bubble concerns, arguing the risk of underbuilding compute capacity outweighs overbuilding. The move extends Blackstone's existing data-centre portfolio — anchored by its 2021 QTS acquisition — into a market attractive for its stable power grid and government support. Gray also signalled accelerated private-equity activity in Asia, following a record $13.1 billion regional fund raise. The announcement comes amid a global race for AI infrastructure, with hyperscaler capex approaching $690 billion annually and rivals like China pursuing sovereign-funded national compute grids.

SpaceX Reflection AI deal: $6.3bn for compute

SpaceX has signed a $6.3bn compute deal with Reflection AI, an open-source AI startup founded in 2024 by former Google DeepMind researchers. Reflection will pay $150mn per month to rent Nvidia GB300 chips at SpaceX's Colossus 2 data center in Memphis through 2029. Notably, Nvidia is on both sides of the deal — supplying the hardware and having invested ~$800mn in Reflection. SpaceX is rapidly becoming a compute landlord, with existing rental contracts with Anthropic (~$1.25bn/month) and Google (~$920mn/month), giving it over $80bn in committed compute revenue through 2029. Reflection, valued at $8bn after a $2bn raise last October, is now raising again at a $25bn valuation despite not yet shipping a public frontier model.

Masayoshi Son dismisses Musk’s space data centres as an AI-race bet

SoftBank founder Masayoshi Son publicly dismissed the concept of orbital data centres at a shareholder meeting, arguing the economics don't work. While space offers cheaper solar electricity, the costs of launching hardware, maintenance, and communication latency far outweigh the savings. Son also stressed that the AI race will be decided in the next few years, not a decade out — making long-horizon space projects irrelevant to the current competition. His comments implicitly target Elon Musk and Jeff Bezos, who have both championed orbital data centre concepts. SpaceX has itself acknowledged viability risks in filings. Son frames rivals pursuing space infrastructure as distracted from the contest that actually matters now.

NVIDIA Powers Over 400 of the World’s 500 Fastest Supercomputers

NVIDIA now powers over 400 of the TOP500 fastest supercomputers (81%), with nearly 90% of newly added systems built on NVIDIA technology. NVIDIA GPUs accelerate a record 238 systems and NVIDIA networking connects a record 376. On the Green500 efficiency list, the top 8 systems all run NVIDIA GPUs, led by KAIROS at France's University of Toulouse using a Grace Hopper Superchip at 73.3 gigaflops per watt. Grace CPU adoption reached 26 systems, and new Blackwell architecture (B200/GB200) systems are entering rankings globally. Europe has a record 35 NVIDIA AI HPC supercomputers in development.

Oracle’s workforce shrank by about 13% as it bankrolls its AI buildout

Oracle ended fiscal year 2026 with roughly 141,000 employees, down from 162,000 — a reduction of about 21,000 people or 13%, one of the deepest cuts in the company's history. The layoffs were framed as a reallocation of resources from legacy operations into AI and cloud infrastructure. Oracle's capital expenditure for the year reached approximately $50bn, with a contracted revenue backlog surpassing half a trillion dollars driven by large AI deals. Analysts estimate the workforce reductions free up $8–10bn annually in cash flow, which is being funneled into data-center construction. The cuts hit some teams by 30% or more, and a $16.3bn data-center financing deal required bond manager PIMCO to anchor $10bn after US banks stepped back. Oracle joins a broader tech industry trend of converting payroll into AI capital expenditure.