Broadcom Secures $63 Billion for Anthropic AI Compute; Samsung SDS Breaks Ground on Gumi Data Center
The AI compute build-out accelerates with significant financing for chip deployments and new hyperscale datacenter construction, driven by surging demand for processing power.
The story
The financial and physical layers of AI infrastructure saw major movement this week, headlined by a $63 billion transaction to deploy AI compute capacity for Anthropic using Broadcom technology. Announced on October 26, the deal involves a special purpose vehicle purchasing Broadcom XPU accelerators and networking equipment, then leasing it to Anthropic, with deliveries slated for 2027.
This financing, arranged by Bank of America, Blackstone, Citi, and Morgan Stanley, addresses the equipment side of the AI compute equation, though site and power contracts remain separate. Meanwhile, Samsung SDS began construction on its Gumi AI data center in North Gyeongsang Province, South Korea.
The facility is set to initially provide 60MW of power, with plans to expand to 120MW, targeting completion by the end of 2028 and operations starting in March 2029. This project represents a concrete step in Samsung Group's broader 60 trillion won investment plan for the Yeongnam region. These developments underscore the intensifying capital investment in AI, with major hyperscalers projected to deploy roughly $900 billion in combined AI-related capital expenditure in 2026 alone.
Silicon
Instinct MI450 GPUs & Helios rack-scale solutions
Maker: AMD
What: Powering AI workloads for major clients including Anthropic, Meta, Microsoft, OpenAI, and Oracle Cloud Infrastructure.
For Whom: Hyperscalers and cloud providers for AI data center operations.
Regulus (integrating SensPro AI DSP and NeuPro-Nano NPN32 NPU)
Maker: Mobilint (with Ceva IP)
What: Delivers 10 TOPS of AI throughput at less than 3W, with up to 8GB memory, for edge AI applications.
For Whom: Developers and system integrators for local processing in edge AI devices and USB accelerators.
Core Ultra processors (with Intel AI Boost NPU and Intel Arc GPU)
Maker: Intel
What: Provides 11 TOPS (INT8) NPU, integrated GPU with XMX engines, and up to 24GB LPDDR5 unified memory for on-device inference.
For Whom: Compact AI mini PCs for local LLMs, image generation, and background AI tasks within a 28W–45W power envelope.
The build-out
Gumi AI Data Center
Who: Samsung SDS
Scale: Initially 60MW, expandable to 120MW; 427.3 billion won investment (~$320M USD)
Where: Gumi, North Gyeongsang Province, South Korea
AI Data Center at Drax Power Station
Who: Drax, Brookfield, Radiant
Scale: Targets over 1GW across three phases; initial £2.5 billion for Phase 1 (up to 100MW)
Where: North Yorkshire, UK
Supply & policy signals
China's MOFCOM and MOST released a draft amendment to the Catalogue of Technology Prohibited or Restricted for Export, covering 40 AI, biopharma, and solar cell technologies.
Implication: Companies exporting affected technologies will need to apply for licensing or face prohibitions, potentially impacting international collaboration and sourcing, though exemptions are noted for public domain AI technologies.
The technology distribution market in North America grew nearly 30% year-over-year to $28.9 billion in Q3 2026, with enterprise infrastructure, networking gear, and semiconductors accounting for two-thirds of a $6.5 billion surge.
Implication: AI infrastructure demand is expanding beyond hyperscalers into a broader ecosystem of partners, indicating strong channel activity for servers, storage, and high-speed networking.
Leading-edge GPUs are projected to exceed 4,000W thermal design power by 2029.
Implication: This will push air cooling beyond practical limits, accelerating the adoption of two-phase direct liquid cooling as a structural requirement for data center design.
What we'll be watching
- Public consultation for the Drax Power Station AI data center planning application is expected to begin in November 2026.
- China's Ministry of Commerce is soliciting public feedback on its draft technology export controls until November 15, 2026.
- Microsoft plans to begin deploying its Maia 300 AI chip in 2027, expecting it to become a central platform for its AI models.
- Samsung SDS targets completion of its Gumi AI data center by the end of 2028, with operations scheduled to begin in March 2029.
Reporting + analyst voices: grounded via Google Search at publish time.