2 October 2026 edition
Space-Based Compute Takes Flight as Massive Power Demands Spur Nuclear and Photonic Innovation
Tech giants and infrastructure operators are targeting alternative environments and energy paradigms to break through terrestrial compute bottlenecks, highlighted by Google's orbital datacenter tests and a 9.4GW SMR project in Utah. Concurrently, major cloud capacity shifts and specialised hardware acceleration continue to redefine high-performance infrastructure.
TOP STORY
Google Launches Orbital Datacenter Test and Weighs Starship Requirements
Google has launched experimental AI processors into orbit via SpaceX and Planet Labs, releasing research on the viability of space-based datacenters. The tech giant projects that operational scale in orbit will require roughly 1,600 SpaceX Starship missions alongside novel solutions for orbital networking and formation flying.
GPU MARKET
Nvidia Blackwell Delivers 8x Acceleration for OpenAI's GPT-6 Astra Ultrafast
OpenAI rolled out GPT-6 Astra Ultrafast, built on Nvidia Blackwell GPUs across API access, ChatGPT Work, and Codex. Bespoke inference optimisations tailored to Blackwell silicon provide up to an eightfold boost in token generation over the base model.
DATA CENTERS
NetApp Unveils Novus Storage Architecture for Zettascale AI Workloads
NetApp has introduced Novus, an architecture engineered to resolve metadata processing constraints and deliver 100 TB/s of aggregate throughput over NFS and parallel NFS protocols to support hundreds of thousands of GPUs concurrently.
Nebius Expands Footprint With 50MW Lease from AIB Data Centers
AI cloud infrastructure provider Nebius signed a commercial lease agreement for 50MW of capacity at an AIB Data Centers facility located in the south-eastern United States.
SEMICONDUCTORS
Volantis Raises $88M Series A for Photonic Memory Architecture
Startup Volantis secured $88 million in a Series A round co-led by Abstract Ventures and Lachy Groom. The capital will advance optical memory designed to run inference on 20+ trillion-parameter models at up to 10,000 tokens per second per user.
DeepSeek and Huawei Release Open-Source Tools for Ascend Silicon
DeepSeek and Huawei have published open-source communication and compute libraries for the Ascend 950 processor series, aiming to facilitate software optimisation and lessen reliance on Nvidia's software stack.
Micron Sues YMTC Alleging Tech Theft and Retaliatory Patent Actions
Micron Technology filed a lawsuit against Yangtze Memory Technologies, alleging the Chinese manufacturer poached technical staff to misappropriate proprietary semiconductor technology and subsequently assert patents against Micron.
ENERGY
Valar Atomics Proposes 9.4GW SMR-Powered Campus in Utah
Valar Atomics submitted plans for a massive 9.4-gigawatt datacenter campus powered by small modular reactors, targeting commercial operation of the first nuclear unit by 2028.
Fervo Energy Completes First Enhanced Geothermal Plant in 23 Months
Fervo Energy finished constructing the first commercial enhanced geothermal power facility in under two years, signalling accelerated development timelines for future grid interconnections.
DEALS & FUNDING
Tencent Secures Access to 100,000 GPUs via Oracle Cloud
Tencent has reportedly contracted with Oracle Cloud for access to 100,000 GPUs, leveraging remote cloud infrastructure to secure advanced compute barred from direct mainland export under trade rules.
GMI Cloud Raises $668M Across Equity and Debt
Specialist AI cloud infrastructure provider GMI Cloud completed a $668 million funding round combining equity capital and debt facilities to expand its compute capacity.
What to watch next
Watch for regulatory and political scrutiny surrounding cross-border GPU leasing arrangements like Tencent's reported Oracle deal as cloud pathways become key vectors around hardware export restrictions. In power, monitor progress toward the 2028 operational target for Valar Atomics' 9.4GW SMR proposal as nuclear integration faces rigorous licensing hurdles. Finally, track early industry testing of optical memory architectures as startups attempt to resolve memory bandwidth limitations for frontier models.