California Data Center Intel

Recent California data center news

• Jan 6, 2026

  AMAX Advances Major Site Expansions to Host 150kW Class Liquid Cooled AI Infrastructure

  AMAX has begun a major expansion of its Fremont, California production facility to add liquid cooling and power capacity for high-density AI racks (Phase A adds 2 MW of power capacity).

  • Main announcement: AMAX is upgrading its Fremont, CA facility (Phase A) to add 2 MW of power capacity and liquid cooling infrastructure to support 150 kW class AI racks with optional liquid-cooled configurations up to 300 kW per rack, enabling deployment and short-term operation/validation of high-density, rack-scale AI systems under AMAX’s HostMax™ service.
  • Background and technical details: Expansion is engineered for NVIDIA HGX B200/B300, GB200/GB300 NVL72, and AMD MI355X platforms; includes ASHRAE W45 compliant liquid cooling and ASHRAE A2 compliant air cooling, PUE-aware power/cooling architecture, and a full lifecycle workflow (build, burn-in, system validation, benchmarking, short-term operation) with continuous monitoring of power, temperature, coolant flow, and facility conditions. CTO Dr. Rene Meyer provided a statement emphasizing rapid AI infrastructure readiness.
• Jan 6, 2026

  Data Center Jobs: Engineering, Construction, Commissioning, Sales, Field Service and Facility Tech Jobs Available in Major Data Center Hotspots

  Data Center Frontier, in partnership with Pkaza Critical Facilities Recruiting, published a monthly roundup of current data center job openings on its jobs board.

  • Monthly jobs roundup: The post lists roughly 15–18 open roles (examples: Data Center Facility Technician, Electrical Commissioning Engineer, Construction Project Manager, Senior Electrical Engineer, Production Architect, Strategic Sales Account Manager, Mechanical Engineer, Site Selection Manager/Director/VP, Electrical Project Manager, Electrical Superintendent, Project Executive, MEP Construction Project Manager, Mechanical Commissioning Engineer, Engineering Design Director, Navy Nuke Facility Technician) with locations across the United States including Impact, TX; Ashburn, VA; Dallas, TX; Atlanta, GA; Reading, PA; Allentown, PA; Charlotte, NC; New Albany, OH; Lyndhurst, NJ; Boulder, CO; Richmond, VA; Austin, TX.
  • Role and employer context: Positions are listed with mission-critical data center providers, engineering design and commissioning firms, A/E/C architecture firms, equipment rental providers, electrical contractors and general contractors; listings repeatedly cite energy efficiency, sustainable design, and AI infrastructure support, and several technician roles explicitly note acceptance of Navy Nuke / military veterans.
• Jan 6, 2026

  At CES, Nvidia launches Vera Rubin platform for AI data centers

  Nvidia launched the Vera Rubin NVL72 server rack platform for AI data centers at CES.

  • Platform announcement and specs: The Vera Rubin NVL72 is a rack-scale AI platform composed of six silicon pieces (Vera CPU, Rubin GPU, NVLink 6 Switch, ConnectX-9 SuperNIC, BlueField-4 DPU, Spectrum-6 Ethernet Switch) delivering up to 3.6 exaflops NVFP4 inference and 2.5 exaflops NVFP4 training; key hardware details include 72 Rubin GPUs + 36 Vera CPUs (NVL72 configuration), 54 TB LPDDR5x, 20.7 TB HBM4, 1.6 Pbps HBM4 bandwidth, and a 260 Tbps scale-up bandwidth.
  • Background, networking, storage and rollout details: The platform introduces context memory storage (Nvidia Inference Context Memory Storage Platform) using BlueField-4 and Spectrum-X with Dynamo/Nixl for coordinated context retrieval (claimed up to 5x tokens/sec, performance/TCO, and power efficiency vs traditional storage); Nvidia also announced rack innovations (Third-Gen NVL72 Rack Resiliency, cable-free modular trays, NVLink Intelligent Resiliency for zero-downtime maintenance). Initial formats: Vera Rubin NVL72 (rack-scale) and HGX Rubin NVL8 (8 GPUs for x86 servers).
• Jan 6, 2026

  13 predictions for how facilities management will evolve in 2026

  Facilities Dive invited industry experts and readers to share predictions for 2026; respondents said facilities management will prioritize AI, integrated building controls, and turning building data into actionable insights.

  • Main announcement and specifics: Respondents forecast widespread adoption of AI-driven building controls and integrated tech stacks (CMMS + building automation + IoT + asset data). Example: Kent State University uses AI to monitor 1,000 input variables and make 150 control decisions every 15 minutes, saving $470,000 annually; others expect multi-agent AI and purpose-built AI agents to automate workflows and administrative tasks in 2026.
  • Background and additional details: Experts highlighted occupant experience and safety (real-time engagement; AI-powered security), growth in medical-office retailization (1+ billion sq ft added in top 50 markets claimed), the need for secure, high-density server rooms for AI workloads, emphasis on embodied carbon and adaptive reuse, and operational KPIs such as revenue per technician per labor hour.
• Jan 6, 2026

  AMD Showcases Growing AI Hardware Arsenal at CES2026

  Advanced Micro Devices (AMD) unveiled the Helios rack-scale system and new Ryzen AI processors at CES in Las Vegas on Jan. 6, 2026.

  • Main announcement: AMD introduced Helios, a rack-scale system built around the Instinct MI455 AI chip, paired with EPYC central processors and high-speed networking, and also announced Ryzen AI processors for personal computers; the keynote was delivered by CEO Lisa Su at CES in Las Vegas on Jan. 6, 2026. The company positioned these products as infrastructure to meet a surge in AI demand driven by growth from ~1 zettaflop (2022) to >100 zettaflops (today) and user adoption rising from 1 billion toward ~5 billion users.

  • Partnerships and deployment details: AMD highlighted that OpenAI is using AMD systems to support enterprise demand and AI agents, and Luma AI uses AMD hardware for video generation/editing workloads that process tens of thousands of data tokens per second; AMD said the announcements reflect a shift to continuous AI workloads and to selling integrated, rack-scale platforms and open tools to compete with Nvidia. No specific financial terms or implementation timelines were provided in the article.

• Jan 6, 2026

  AWS hikes prices for EC2 Capacity Blocks amid soaring GPU demand

  Amazon Web Services (AWS) has raised prices for some EC2 Capacity Blocks for machine learning workloads. These Capacity Blocks reserve accelerated GPU/Trainium inventory for future start dates and are used to guarantee cluster access for ML training.

  • Main announcement: AWS increased Capacity Blocks pricing by approximately 15% for certain P5 Capacity Block SKUs; example changes include p5e.48xlarge (US East - Ohio) from $34.608 to $39.799 per effective hourly rate per instance (per accelerator) and p5en.48xlarge from $36.184 to $41.612. Regional exceptions include US West (N. California) where p5e.48xlarge moved to $49.749 (from $43.26) and p5en.48xlarge to $52.015 (from $45.23). The P6e / p4d.24xlarge entry remains at $761.904 for 72 B200 accelerators in Dallas Local Zone.

  • Background and details: AWS Capacity Blocks let customers reserve clusters (1–64 instances, up to 512 GPUs or 1024 Trainium chips) for up to six months with reservations allowed up to eight weeks in advance. Analysts attribute the change to GPU supply/demand dynamics and a scarcity premium on guaranteed inventory; competing providers (Google Cloud, Microsoft/Azure) offer different reservation/scheduling models and trade-offs between pricing segmentation and scheduling flexibility.

• Jan 6, 2026 · Tech Xplore

  Nvidia unveils new AI chip platform amid rising competition

  Nvidia unveiled its latest AI platform, Vera Rubin, at the Consumer Electronics Show in Las Vegas and said Rubin-based products will be available from partners in the second half of 2026.

  • Main announcement: Nvidia announced the new Vera Rubin AI chip platform at CES; the platform is described as comprising “six chips that make one AI supercomputer”, claimed to run five times more efficiently than previous offerings, and Rubin-based products are scheduled to be available from partners in H2 2026. The company currently holds an estimated 80% of the global market for AI data center chips.
  • Background and details: The Rubin architecture is described as a “profound shift” from Nvidia’s prior Blackwell generation; Nvidia faces competitive pressure from AMD and Intel, and from large cloud customers (Google, Amazon, Microsoft) developing in-house chips (Google’s Gemini 3 was trained without Nvidia technology).
• Jan 5, 2026

  NVIDIA Rubin Platform, Open Models, Autonomous Driving: NVIDIA Presents Blueprint for the Future at CES

  NVIDIA unveiled the Rubin platform in full production and introduced the Alpamayo open reasoning model family at CES 2026, positioning Rubin as a six-chip, extreme codesigned AI platform and Alpamayo as an open stack for level-4 capable autonomy.

  • Main announcement:NVIDIA announced the Rubin platform is in full production (successor to Blackwell), a six-chip extreme‑codesigned system including Rubin GPUs (50 petaflops NVFP4 inference), Vera CPUs, NVLink 6, Spectrum‑X Ethernet Photonics, ConnectX‑9 SuperNICs, and BlueField‑4 DPUs; NVIDIA claims token generation costs ~one‑tenth of the prior platform and introduced AI‑native KV‑cache storage with up to 5x gains in tokens/sec, TCO performance and power efficiency. The company also introduced Alpamayo (R1 reasoning VLA, AlpaSim simulation blueprints) and said the first passenger car with Alpamayo on NVIDIA DRIVE will be on U.S. roads this year (the all‑new Mercedes‑Benz CLA).
  • Background and details: The announcement was delivered at CES 2026 in Las Vegas; NVIDIA positioned Rubin as built “from the data center outward” and emphasized open models trained on NVIDIA supercomputers across domains (Clara, Earth‑2, Nemotron, Cosmos, GR00T, Alpamayo). Additional product notes: DGX Spark claims up to 2.6x performance for large models; partnerships and ecosystem mentions include Hugging Face, Siemens (expanded partnership), Palantir, ServiceNow, Snowflake, Boston Dynamics, Franka, Synopsis, Cadence, and others.
• Jan 5, 2026

  SOCAMM memory gains ground as AI data centers proliferate

  Samsung unveiled SOCAMM2, an LPDDR5-based memory module and new CAMM2 industry form factor designed specifically for AI data center platforms.

  • Main announcement: Samsung introduced SOCAMM2 (LPDDR5) as an industry-standard CAMM2 memory form factor offering up to 2× bandwidth vs DDR5 RDIMMs, 1.5–2.0× reported performance, and ~55% power consumption of comparable DDR5; modules are smaller due to stacked memory and can be used with or instead of DDR.
  • Background and timeline:Dell originally co-designed CAMM and handed the spec to JEDEC (which added ECC and enterprise features); SK Hynix has announced support (believed behind Micron and Samsung); full market launch is expected around Q2 2026 aligned with Nvidia’s Vera Rubin platform launch.
• Jan 5, 2026

  Everything you needed to know about FLOPs

  Andy Patrizio explains FLOPs benchmarking, floating-point precisions, and how benchmarks are used in supercomputing and AI.

  • Main explanation: The article describes that the Top 500 list measures supercomputers in FP64 (double-precision, 64-bit) as a proxy for scientific computing; FP64 takes twice as long as FP32 and four times as long as FP16, and uses twice (vs FP32) or four times (vs FP16) the memory. The Top 500 is published every June and November. Addison Snell (CEO, Intersect 360) is quoted emphasizing 64-bit remains the de rigueur standard for scientific workloads.

  • Additional details and vendor/AI context: The article explains precision tiers—FP32 (single precision), FP16 (used for AI inferencing), bfloat16 (Google-originated, licensed to Intel/AMD/Nvidia; described as less precise than FP16 but faster), and FP8/FP4 (FP8 used for inference, some neural network training, and edge cases; FP8 only used on GPUs and not by Intel/AMD according to the article). It warns vendors sometimes advertise peta/exaFLOPs using lower-precision metrics (e.g., FP8) and advises readers to ask the precision behind advertised FLOP numbers.