United States / United Kingdom Emerald AI — US$150M Series A; 5 demonstrations; >100GW latent-grid-capacity estimate.
Brazil Alibaba Cloud — first South American cloud region; 2 data centres; 106 AZs across 31 regions.
Global NVIDIA Groq 3 LPX — full production; ~3,400 output tokens/sec in cited long-context benchmark.
India / Global Infineon to acquire C2i — AI data-centre power-delivery technology; ~2,800 Infineon employees in India.
Power & Grid | Contracted Capacity | AI Infrastructure
OpenAI's 3.2GW Georgia Power Contract Clears Regulatory Review
26 August 2026 | United States — Georgia
Georgia Power received regulatory approval for its electricity-service contract supporting OpenAI's Project Camellia in Effingham County, Georgia. The agreement covers 3,200MW (3.2GW) of new electricity demand, with OpenAI responsible for the infrastructure and service costs required to serve the campus. OpenAI has also committed up to 1,000MW (1GW) of flexible demand response, allowing load reductions during periods of grid stress. Power is expected to be delivered in phases between 2028 and 2032, so the 3.2GW should be classified as contracted/approved rather than energized or operational capacity. Georgia Power says its wider large-load portfolio is expected to generate approximately US$950 million in annual customer savings beginning in 2029.
Watch the conversion of regulatory approval into phased, energized capacity. The critical indicators are transmission and generation milestones, the economics of the 1GW flexible-load commitment, and whether demand-response participation becomes a replicable condition for other multi-GW AI campuses.
GreyRadius Insight
OpenAI’s 3.2GW agreement signals a change in how AI capacity is secured: grid flexibility is becoming part of the commercial contract, not an afterthought. Markets able to combine new supply with controllable AI load could approve large campuses faster—and gain an advantage over regions that evaluate data centres only as fixed baseload.
AI Compute | Hyperscale Cloud | GPU Infrastructure
AWS and NVIDIA Plan Another 2 Million GPUs Across Global AI Infrastructure
26 August 2026 | Global
AWS and NVIDIA plan to deploy 2 million additional NVIDIA GPUs during 2027–2028, including Blackwell Ultra, Rubin and Rubin Ultra systems. This is additional to AWS's previously announced plan to deploy more than 1 million NVIDIA GPUs beginning in 2026, potentially creating an accelerator estate measured in several million units. The companies also plan to introduce NVIDIA Vera CPU infrastructure on AWS and deepen networking and NVLink integration. A sovereign/public-sector component includes planned AI factories for the U.S. government incorporating 100,000 GPUs on secure AWS infrastructure. Neither the MW requirement nor capex associated specifically with the additional two-million-GPU deployment was disclosed, so this remains announced compute capacity rather than deployed GPUs.
Track where the 2 million GPUs are actually deployed, the MW and cooling infrastructure attached to them, and the utilization achieved after commissioning. Regional allocation will also reveal where AWS sees sufficient power, network capacity and sovereign demand to support the next wave of AI infrastructure.
GreyRadius Insight
At multi-million-GPU scale, accelerator procurement is no longer the best measure of AI capacity. The strategic metric becomes deployable compute: GPUs multiplied by available power, networking, cooling and utilization. Infrastructure providers that solve those constraints together will capture more value than those supplying any single layer.
Powered Land | AI Factories | Infrastructure Development
NVIDIA Links Its AI Ecosystem to Lancium's 15GW+ Powered-Land Pipeline
24 August 2026 | United States
Lancium announced a strategic collaboration with NVIDIA, accompanied by an investment from NVIDIA into the Blackstone-backed infrastructure developer. Lancium reports 4GW of leased capacity and a development pipeline exceeding 15GW of powered land across the United States. The partnership is intended to turn these locations into deployment sites for NVIDIA accelerated compute, networking and software infrastructure. Lancium will use NVIDIA DSX infrastructure designs and technologies intended to optimize GPU deployments within available power envelopes while supporting grid-responsive operation. The 15GW+ figure represents a development pipeline and should not be treated as operational AI capacity; the 4GW is described as leased capacity. NVIDIA's investment amount was not disclosed.
Watch how much of Lancium’s 15GW+ pipeline advances into interconnected, financed and customer-contracted sites. Project-level disclosures on interconnection status, first-power dates and tenant commitments will show whether NVIDIA’s involvement improves the pipeline-to-operational conversion rate.
GreyRadius Insight
NVIDIA is extending its influence upstream because future accelerator sales depend on infrastructure readiness. The strategic opportunity is shifting toward 'compute-ready power': sites where land, interconnection, electrical design and AI architecture are coordinated before GPU deployment. Raw GW pipeline without those elements has much lower strategic value.
AI Infrastructure Finance | GPU Deployment | Data Centres
Blue Owl Leads US$2.4bn Financing for IREN's NVIDIA AI Factory
28 August 2026 | Canada — British Columbia
Blue Owl managed funds led US$2.4 billion of compute-equipment financing for IREN. The financing comprises a US$1.2 billion senior secured term loan and US$1.2 billion of senior secured notes. Proceeds are intended to finance air-cooled NVIDIA accelerated-computing infrastructure, including Blackwell Ultra GPUs, at IREN's Mackenzie campus in British Columbia. Capital will be drawn in tranches as equipment is delivered and commissioned rather than funded entirely upfront. IREN separately reports a global development pipeline exceeding 5GW. The transaction represents financed AI compute infrastructure, but the US$2.4bn financing should not be interpreted as evidence that the corresponding GPU fleet has already been installed or become operational.
Track financing drawdowns against equipment delivery, commissioning and customer contracts. The key question is whether asset-backed structures can reduce the funding gap between ordering expensive accelerators and generating recurring AI-cloud revenue.
GreyRadius Insight
AI infrastructure is developing its own financing architecture. Linking debt directly to compute equipment and commissioning milestones can make expansion more scalable, but it also raises the importance of utilization and technology-obsolescence risk. The winners will pair access to capital with contracted demand and rapid deployment.
AI Data Centres | Infrastructure Platform | Investment
SK Telecom Creates 318MW AI Data-Centre Infrastructure Platform
27 August 2026 | South Korea
SK Telecom created SK Horizon, separating AI-data-centre and associated infrastructure assets from SK Broadband into a dedicated investment platform. SK Horizon is expected to build a portfolio reaching 318MW, encompassing eight operating data centres in South Korea as well as AI facilities under construction in Ulsan and Guro. The platform is backed by approximately KRW3.08 trillion (about US$2.2 billion) from KKR and the IMM Investment-Stonebridge consortium. SK Telecom retains majority ownership; reported ownership is 51% SK Telecom, 29% KKR and 20% IMM. The 318MW represents targeted platform capacity spanning both operating and developing assets, rather than 318MW of currently operational AI capacity.
Watch how SK Horizon allocates capital between existing facilities and new AI capacity, and whether Ulsan and Guro secure anchor customers before full build-out. The operating-versus-development split inside the 318MW target will be more informative than the headline portfolio number.
GreyRadius Insight
SK Telecom’s structure shows how telecom operators can reposition data centres from internal infrastructure into investable platforms. Separating the asset base can attract long-duration capital while preserving strategic control—potentially giving operators a faster route to fund AI expansion without overloading the telecom balance sheet.
Planning Approval | Grid Allocation | Nordic Data Centres
T1 Energy Wins Approval to Convert Norwegian Industrial Site into 50MW Data Centre
27 August 2026 | Norway — Mo i Rana
Local authorities in Mo i Rana rezoned part of T1 Energy's Giga Arctic industrial facility for data-centre development. T1 expects an initial 50MW data centre could become operational during 2027. The underlying industrial building covers approximately 926,000 square feet, while T1 remains in the grid queue for as much as 396MW. The initial project's power position is more advanced than a conventional speculative development because Norway's grid operator had already assigned T1 50MW of grid capacity in March 2026. The material new development this week is planning approval, complementing the previously secured 50MW grid allocation. It should be classified as approved/grid-allocated, not operational.
Watch whether T1 converts its existing building and secured 50MW grid allocation into contracted capacity on the 2027 timeline. Customer commitments and retrofit economics will determine whether the brownfield advantage translates into materially faster time-to-revenue.
GreyRadius Insight
The scarce asset in AI infrastructure is increasingly not land—it is time-to-power. Brownfield industrial sites with usable buildings and existing grid rights can bypass years of greenfield development risk. This creates a growing opportunity to revalue power-secured industrial assets as potential AI infrastructure platforms.
Flexible Compute | Grid Capacity | AI Infrastructure Software
Emerald AI Raises US$150m to Commercialize Grid-Responsive AI Compute
25 August 2026 | United States / United Kingdom
Emerald AI raised a US$150 million Series A at a reported US$1.05 billion valuation, led by Energize Capital and DCVC. Its technology orchestrates AI workloads so data-centre electricity consumption can respond dynamically to grid conditions instead of operating exclusively as inflexible baseload. The company says it has completed five demonstrations across Arizona, Illinois, Virginia, Oregon and London and is moving into commercial full-data-centre deployments. Emerald estimates that widespread use of flexible AI computing could unlock more than 100GW of latent capacity on the existing U.S. electricity grid. That 100GW figure is a company estimate of potential capacity, not contracted, energized or currently available capacity.
Watch commercial deployments for evidence that flexible AI load can deliver meaningful grid response without compromising workload SLAs. Utility acceptance, measured curtailment capability and customer economics will determine whether flexibility becomes bankable infrastructure rather than a demonstration-stage feature.
GreyRadius Insight
If AI workloads can become genuinely dispatchable, the industry gains a new capacity lever without waiting exclusively for new generation. That could change site selection: access to constrained grids may depend partly on how intelligently compute can shape demand, turning workload orchestration into an infrastructure capability.
Cloud Region | Data Residency | South America
Alibaba Cloud Opens Its First South American Cloud Region in Brazil
28 August 2026 | Brazil
Alibaba Cloud launched its first cloud region in Brazil, establishing its first cloud region in South America. The region uses two data centres and provides local compute, storage, networking, database and security infrastructure designed for workloads requiring lower latency and greater local data governance. Alibaba also plans enterprise agentic-AI services for the market. Following the expansion, Alibaba Cloud reports a global footprint of 106 availability zones across 31 regions. The company did not disclose a Brazil-specific investment figure; its broader global commitment is US$53 billion for cloud and AI infrastructure. Unlike many capacity announcements this week, the Brazil development represents launched/operational cloud infrastructure.
Watch enterprise migration, local AI-service availability and expansion beyond the initial two-data-centre footprint. The stronger signal will be whether local residency and latency requirements translate into sustained workload growth and additional South American capacity.
GreyRadius Insight
Brazil’s importance is moving from being a destination for data-centre investment to becoming a regional cloud control point. Local cloud regions allow providers to compete on sovereignty, latency and AI-service availability simultaneously—making operational cloud presence more strategically valuable than announced investment alone.
Inference Infrastructure | AI Accelerators | Production Compute
NVIDIA Groq 3 LPX Enters Full Production for High-Speed AI Inference
24 August 2026 | Global
NVIDIA announced that Groq 3 LPX had entered full production as an inference accelerator extending its Vera Rubin architecture. The infrastructure targets agentic workloads where long contexts and repeated inference steps make token-generation latency and power efficiency increasingly important. In an Artificial Analysis benchmark cited by NVIDIA, Groq 3 LPX running Gemma 4 31B produced approximately 3,400 output tokens per second on 100,000-token long-context workloads, around four times the nearest alternative tested. Nebius was identified as the first AI-cloud adopter. Unlike announced GPU procurement, this is a production-stage compute infrastructure development, although the benchmark is workload-specific rather than a universal performance measure.
Watch real-world performance across different models, context lengths and cloud deployments rather than relying on a single benchmark. Adoption beyond the first cloud provider, cost per token and power efficiency will determine whether specialized inference materially changes infrastructure economics.
GreyRadius Insight
Agentic AI shifts the infrastructure bottleneck from simply training larger models to serving repeated, latency-sensitive inference economically. As inference volumes scale, tokens per second per watt—and ultimately cost per useful task—could become more important purchasing metrics than raw accelerator specifications.
Power Delivery | Semiconductors | AI Data Centres
Infineon Acquires India's C2i to Strengthen AI Data-Centre Power Delivery
24 August 2026 | India / Global
Infineon agreed to acquire Bengaluru-based C2i Semiconductors, a developer of software-defined multiphase controllers, intelligent power-management technology and system-level architectures for AI data centres. Financial terms were not disclosed, and the transaction is expected to close during Q3 2026. C2i's technology addresses rapidly changing power loads generated by high-density AI processors and includes vertical power-delivery architectures and substrate-integrated voltage-regulator capabilities. The acquisition also expands Infineon's R&D footprint in India, where the company currently employs approximately 2,800 people. The transaction does not create additional MW of data-centre capacity directly, but targets the infrastructure layer that converts facility electricity into stable, usable processor power as rack densities increase.
Watch how quickly C2i’s controller and vertical power-delivery technologies move into qualified AI platforms. Customer design wins and improvements in conversion efficiency, transient response and rack-level density will show whether the acquisition affects deployable compute economics.
GreyRadius Insight
The power constraint is moving inside the rack. Securing facility MW is insufficient if conversion and delivery losses prevent that electricity from feeding increasingly dense processors efficiently. Power electronics therefore becomes a strategic compute multiplier: better delivery can increase usable AI capacity without adding equivalent facility-level MW.
Market Data & Intelligence
Metric
Latest Value
Reporting Period
Region
Executive Implication
OpenAI / Georgia Power
3.2GW approved demand; up to 1GW flexible load; delivery 2028–2032
26 Aug 2026
United States — Georgia
Power approval and load flexibility are becoming core prerequisites for multi-GW AI campuses.
AWS / NVIDIA GPU plan
2M additional GPUs in 2027–2028; >1M previously announced; 100,000 for U.S. government AI factories
26 Aug 2026
Global
Compute scale is moving into millions of accelerators, increasing pressure on power, cooling and network deployment.
Lancium / NVIDIA
>15GW powered-land pipeline; 4GW leased
24 Aug 2026
United States
Powered land is becoming a strategic asset, but development pipeline must be separated from operational capacity.
Flexible compute could become a grid-capacity tool if commercial deployments validate dispatchable AI load.
Alibaba Cloud Brazil
First South America region; 2 data centres; 106 AZs across 31 regions
28 Aug 2026
Brazil
Operational local cloud infrastructure strengthens latency, residency and regional AI-service availability.
NVIDIA Groq 3 LPX
Full production; ~3,400 output tokens/sec on cited 100k-token benchmark; Nebius first adopter
24 Aug 2026
Global
Inference responsiveness and efficiency are emerging as infrastructure competition metrics.
Infineon / C2i
Acquisition expected Q3 2026; terms undisclosed; ~2,800 Infineon employees in India
24 Aug 2026
India / Global
Power-conversion technology is moving higher in strategic importance as AI rack density rises.
Turn Infrastructure Strategy Into Execution
GreyRadius helps leadership teams translate cloud, data centre and AI infrastructure shifts into market priorities, investment choices and execution-ready growth plans.
