Enterprise AI • Compute • Sovereign AI • Deployment Infrastructure
Published by GreyRadius Consulting | 28 September–4 October 2026
Executive Highlights
Asia / IndiaAWS launches in-country Claude inference across India
Global / Europe, Asia and GCCAlibaba Cloud announces new regions in Türkiye, Finland and the Netherlands — disclosure / context update.
GCC / United Arab EmiratesMubadala and Together AI establish UAE sovereign-AI infrastructure partnership
Global / United StatesAnthropic discloses at least $518bn of long-term AI infrastructure commitments
GlobalTrane and LiquidStack launch 2.5MW coolant distribution unit for AI infrastructure
Global / United States, Taiwan and Asia-PacificGMI Cloud raises $668m for AI infrastructure expansion
Asia-Pacific / Australia and GlobalSharon AI secures $356m GPU-backed debt facility
North America / United StatesVolato subsidiary signs $1.17bn of take-or-pay AMD GPU infrastructure orders — disclosure / context update.
North America / United StatesGimlet Cloud and Cerebras plan 100MW wafer-scale inference deployment
North America / United StatesGeneral Compute signs multi-year Cerebras inference infrastructure agreement
North America / United StatesCoreWeave makes NVIDIA Vera Rubin NVL72 commercially available, with Cognition as early customer
North America / United StatesSynopsys and Amazon sign $1bn custom-chip infrastructure agreement
North America / United StatesCScale raises $145m for AI optical interconnect technology
North America / United StatesNebius signs 50MW, 12-year US data-centre lease
North America / United StatesGoogle confirmed as customer for Crusoe's Amarillo AI campus
North America / United StatesSkybox and Prologis file $550m fourth Hutto data-centre building
North America / United StatesZone Frontier secures 4,077-acre Texas site with potential for more than 500MW of data centres
North America / United StatesValar proposes multi-GW SMR-powered data-centre concept in Utah
North America / United StatesAmazon signs 20-year Calvert Cliffs agreement supporting 190MW nuclear uprate
North America / United StatesFERC pauses PJM backstop plan addressing 6.8GW power shortfall
North America / United StatesUS Senate blocks data-centre electricity-cost legislation 57–43
North America / United StatesAccelevation and shareholders raise $540m in AI-infrastructure IPO
North America / United StatesVirginia introduces Data Center Accountability Framework and bans project NDAs — disclosure / context update.
North America / United StatesOracle reports 2.8m construction hours at New Mexico Project Jupiter
North America / United StatesOracle's Michigan AI project passes one million union craft hours
North America / United StatesSpectrum brings AI compute to a network with more than 1,000 edge sites
Europe / United Kingdom£16.5bn (~US$21.90bn), 1GW Chapelcross data-centre campus enters Scottish planning process
Europe / United KingdomCivo starts 8MW UK sovereign-AI edge facility with 38MW expansion target
Europe / IrelandGoogle receives planning permission for 72,400 sq m Dublin data-centre expansion
Europe / United KingdomAWS receives planning permission for Iver data-centre campus
Europe / PortugalNxN acquires operational 1MW Portuguese facility and plans up to €100m (~US$113.58m) expansion
Europe / PolandPoland's Konin signs LOI around 100MW first phase of potential 260MW campus
Europe / United KingdomAmazon signs 36MW Shetland wind PPA
Europe / FinlandFinland proposes replacing first-come grid allocation for large loads
Europe / SpainDigital Realty and Blackfuel launch multi-MW liquid-cooled AI deployment in Barcelona
Europe / FranceGlobal Switch selected for large liquid-cooled NVIDIA Blackwell deployment in Paris
Europe / United KingdomAWS adds in-region Claude inference in London
Europe / LuxembourgMicrosoft launches Azure Extended Zone in Luxembourg
Europe / France, Germany, Italy and PolandESA awards 30-month sovereign cloud contract to OVHcloud and CGI
Europe / PolandPoland receives PLN1.56bn (~US$401m) for three-site sovereign data-processing network
Europe / European UnionEU opens €80m (~US$90.86m) of AI hardware and compute-infrastructure funding — disclosure / context update.
Europe / SpainBalearic Islands launches redundant data centre for critical government systems
Europe / United Kingdom and continental EuropeVirtus secures £2.45bn (~US$3.25bn) financing for UK and European data-centre build-out — disclosure / context update.
Europe / FinlandEnetron's claimed 250MW Finnish project approval contradicted by municipality
Asia-Pacific / Southeast Asia and AustraliaMeta contracts Firmus GPU capacity across Southeast Asia
Asia / China and Southeast AsiaTencent reportedly signs five-year Oracle deal for access to 100,000 GPUs
Asia / South Korea and United StatesSemifive signs $52m contract for next-generation AI inference accelerator
Asia-Pacific / AustraliaHyperDC files plans for 400MW IT campus outside Melbourne
Asia / Macau, ChinaBrightRay plans 40MW liquid-cooling-capable AI data centre in Macau
Asia / Hong Kong, ChinaAmazon agrees $278m–$312m acquisition of two Hong Kong data centres
Asia / MalaysiaAmazon agrees RM316.84m (~US$77.66m) purchase of 45.46-acre Malaysian data-centre site
Asia / South Korea and SingaporeAWS adds in-region Claude inference in South Korea and Singapore
Asia / IndonesiaIndonesia halts construction at BDx project over permitting
Middle East / EgyptEgypt's Intro Group plans $270m first phase of 80MW Kemet Data Center
Filter
Sovereign AI / Cloud
AWS launches in-country Claude inference across India
29 September 2026 | Asia / India
AWS expanded Amazon Bedrock so Claude Opus 5, Claude Sonnet 5 and Claude Haiku 4.5 can process inference within India. AWS uses geographic cross-region inference across its Mumbai and Hyderabad Regions, allowing workloads to scale between Indian regions without inference processing leaving the country. The launch is particularly relevant for financial services, healthcare and public-sector customers with local-processing requirements. It does not represent a new AWS Region or disclosed incremental MW, but materially expands the AI workloads that can be deployed while meeting Indian data-residency requirements.
Test geographic routing between Mumbai and Hyderabad, service quotas and failover. Check logs, backups and support access separately from the stated in-country inference processing commitment.
GreyRadius Insight
Local model availability can unlock regulated demand sooner than a new campus build. Enter with a workload-specific compliance and integration offer, backed by measured throughput, total cost and customer approval to move into production.
Sovereign AI / Cloud
Alibaba Cloud announces new regions in Türkiye, Finland and the Netherlands
29 September 2026 | Global / Europe, Asia and GCC
Alibaba Cloud announced plans to establish new cloud regions in Türkiye, Finland and the Netherlands over the next 12 months, while expanding existing infrastructure in Malaysia, Germany, the UAE, France and Hong Kong. The programme extends Alibaba's international cloud footprint as enterprise AI workloads increase. No aggregate investment, MW or server count was disclosed. The announcement is therefore a geographic cloud-infrastructure expansion commitment rather than delivered operating capacity, with individual region commissioning dates still to be established. This report revisits the regional-expansion programme covered in the previous issue.
Track each new region’s service-launch date, local approvals and enterprise customer commitments. The regional expansion repeats an earlier disclosed programme; the 29 September report should not be counted as incremental capacity beyond that programme.
GreyRadius Insight
Geographic expansion becomes commercially useful when customers can deploy supported workloads locally. Entrants should identify gaps in migration, integration and local support, then validate paid demand before committing teams or infrastructure across all three markets.
Sovereign AI / Cloud
Mubadala and Together AI establish UAE sovereign-AI infrastructure partnership
29 September 2026 | GCC / United Arab Emirates
Mubadala and Together AI announced a strategic partnership to explore AI infrastructure and ecosystem opportunities in the UAE. Together AI will establish an Abu Dhabi presence, bringing its AI-native cloud, inference and research capabilities closer to regional customers. The companies specifically position the partnership around sovereign AI capability. However, no investment amount, GPU count, MW capacity, physical site or delivery timetable was disclosed. The agreement is therefore exploratory rather than funded or commissioned sovereign compute capacity.
Watch for funded projects, anchor workloads and a defined operating model for the Abu Dhabi presence. Clarify data control, local support and infrastructure responsibilities before assigning capacity to the partnership.
GreyRadius Insight
A regional partnership can open customer access before physical supply is committed. A useful UAE entry strategy starts with a specific sovereign workload and buyer, then builds local infrastructure around an evidenced commercial requirement.
Capital / Investment
Anthropic discloses at least $518bn of long-term AI infrastructure commitments
29 September 2026 | Global / United States
Anthropic's confidential IPO filing shows at least $518bn of AI infrastructure commitments over roughly a decade, according to Reuters. Around 80% is described as non-cancelable regardless of usage. Reported obligations include approximately $111.1bn with Google, $110bn with Amazon, $31.4bn with Microsoft and $161.2bn of Broadcom-related lease obligations. A separate xAI arrangement could reach $84.5bn but has more flexible terms. These are long-term contractual and lease obligations rather than commissioned infrastructure or annual capex, but they provide unusually strong evidence of forward demand for compute capacity.
Examine payment schedules, termination rights and counterparty concentration. Supplier forecasts should reconcile obligations with cash receipts and delivered capacity, rather than allocate the full decade-long total to near-term demand.
GreyRadius Insight
Long contracts can finance supply while concentrating demand risk in a few AI labs. Stage commitments against tenant credit quality and cash collection; a non-cancelable contract remains exposed to counterparty failure.
Infrastructure / Power
Trane and LiquidStack launch 2.5MW coolant distribution unit for AI infrastructure
30 September 2026 | Global
Trane Technologies and LiquidStack launched the CDU 2.X, an architecture-agnostic coolant distribution unit designed for high-density AI facilities. It provides up to 2.5MW of cooling capacity, with 3,750 litres per minute at 3.5 bar, supports facility inlet temperatures up to 45°C, and offers dual-feed and automatic-transfer configurations. Orders are open, with shipments scheduled to begin in Q2 2027. The product is relevant to future GPU deployment because multi-MW liquid-cooling distribution is increasingly required at rack densities beyond conventional air-cooling limits.
Check Q2 2027 shipment availability, hydraulic compatibility and performance at the facility’s actual inlet temperature. Validate redundancy and maintenance access across the complete cooling loop before placing orders.
GreyRadius Insight
A larger CDU expands design options, while the weakest link in the thermal chain still limits uptime. Evaluate pumps, heat rejection, controls and service response as one system, especially for mixed generations of GPU racks.
Capital / Investment
GMI Cloud raises $668m for AI infrastructure expansion
1 October 2026 | Global / United States, Taiwan and Asia-Pacific
GMI Cloud raised $668m, comprising $223m of Series B equity and a $445m credit facility. NVIDIA participated in the equity financing. GMI says the capital will expand GPU capacity in the US, Taiwan and broader Asia-Pacific and reported more than $600m in contracted annualized revenue. The financing is platform-level rather than tied to one specific data-centre development, so it should not be translated directly into MW or GPU capacity. Contracted revenue nevertheless provides important evidence behind the company's infrastructure financing.
Separate equity received, credit available and debt actually drawn. Reconcile company-reported contracted ARR with live revenue, customer concentration and the delivery timetable for new GPU clusters.
GreyRadius Insight
Equity and credit support different risks: early expansion needs flexibility, while debt needs dependable cash flows. Assess capacity investments against payment schedules and utilization, with credit drawn only when delivery and customer activation are sufficiently certain.
Capital / Investment
Sharon AI secures $356m GPU-backed debt facility
1 October 2026 | Asia-Pacific / Australia and Global
Sharon AI secured a $356m committed senior secured GPU-backed SPV debt facility at a fixed 9.95% interest rate excluding fees. The facility is secured against GPUs and associated cash flows and is intended to finance AI-factory deployments linked to customer contracts. Sharon AI has separately stated a goal of scaling toward 68,000 GPUs by mid-2027, but that target predates this financing and should not be treated as the number of GPUs funded by this particular facility. The near-10% rate provides a useful indicator of current financing economics for specialist GPU infrastructure.
Review drawdown conditions, fees, covenants and debt repayment against customer cash flows. The company’s stated 9.95% rate and broader GPU target require contract-level validation; this facility does not fund the entire target automatically.
GreyRadius Insight
GPU-backed debt magnifies both utilization gains and hardware obsolescence risk. Stress-test price declines, contract renewals and residual GPU value, and ensure fixed financing costs remain serviceable under a slower customer ramp.
Capital / Investment
Volato subsidiary signs $1.17bn of take-or-pay AMD GPU infrastructure orders
28 September 2026 | North America / United States
Volato Group disclosed that subsidiary Alignment Engine signed two four-year take-or-pay orders worth $1.17bn with a large AI customer. The contracts include approximately $133m of contractual prepayments, with roughly $40m due following execution. Deployment in Ohio is scheduled in phases, using AMD MI355X GPUs from 31 December 2026 and MI455X GPUs from 30 June 2027. The combination of take-or-pay terms and prepayments gives the planned infrastructure a contracted revenue base, although the GPU systems and supporting facility capacity have not yet been delivered. The underlying contracts were entered on 22 September; this is the 28 September public-release update.
Verify prepayment collection, delivery acceptance and the two GPU deployment dates. The contracts were entered in the prior week; the 28 September release is a disclosure update, with execution dependent on power, facility readiness and hardware delivery.
GreyRadius Insight
Take-or-pay terms improve financing visibility but cannot remove delivery risk. Model whether upfront cash covers procurement and whether delay remedies protect customer and lender cash flows. Compare debt repayment with each GPU generation’s expected commercial life.
Compute / Accelerators
Gimlet Cloud and Cerebras plan 100MW wafer-scale inference deployment
28 September 2026 | North America / United States
Gimlet Cloud partnered with Cerebras to deploy approximately 100MW of wafer-scale AI compute, creating one of the week's largest disclosed non-GPU accelerator projects. The first facility is targeted to come online before the end of 2026, while Gimlet is expected to become an early customer for Cerebras' next-generation CS-4 systems in 2027. Gimlet recently raised $300m in Series B funding. The 100MW represents planned compute infrastructure and should not be translated into an equivalent GPU count because Cerebras uses wafer-scale processors rather than conventional GPU architecture.
Watch for the first site going live, metered power and workload benchmarks. Test software compatibility, model coverage and effective cost per output before treating the 100MW plan as broadly substitutable compute.
GreyRadius Insight
Alternative accelerators can improve inference economics where the workload fits. A diversified procurement strategy should compare migration effort and sustained throughput alongside chip price, with capacity commitments tied to validated production workloads.
Compute / Accelerators
General Compute signs multi-year Cerebras inference infrastructure agreement
29 September 2026 | North America / United States
General Compute signed a multi-year agreement to deploy Cerebras wafer-scale systems, with customer capacity expected to become available in Q1 2027. General Compute will finance the hardware and sell dedicated inference capacity, initially targeting agentic coding workloads. The company describes it as its largest single hardware commitment and the first deployment drawing on the $400m debt facility it secured with Upper90 in July. Neither the number of Cerebras systems, MW capacity nor transaction value was disclosed, so the deployment cannot yet be quantitatively compared with Gimlet's 100MW Cerebras project.
Track Q1 2027 delivery, customer reservations and drawdowns under the debt facility. Establish who bears commissioning delays and how hardware repayment aligns with the duration of customer contracts.
GreyRadius Insight
Dedicated inference can offer predictable performance, but agentic workloads have uneven utilization. Financing works best when minimum customer payments cover fixed costs and capacity can be reassigned if an anchor workload changes.
Compute / Accelerators
CoreWeave makes NVIDIA Vera Rubin NVL72 commercially available, with Cognition as early customer
30 September 2026 | North America / United States
CoreWeave announced commercial availability of NVIDIA Vera Rubin NVL72 infrastructure after previously installing a working rack. Cognition became the first disclosed customer and began using the systems in early September. Each NVL72 rack combines 36 CPUs and 72 GPUs and is 100% liquid cooled. CoreWeave also plans standalone Vera CPU infrastructure, configured at 128 Vera CPUs and 11,264 cores per rack, connected through BlueField-4 DPUs and Spectrum-X Ethernet. This is operating accelerator availability rather than a future hardware roadmap, although CoreWeave did not disclose the total number of NVL72 racks available.
Validate reservable rack capacity, regional availability and production pricing. Benchmark long-running agent workloads, including CPU demand, network throughput and cooling constraints, rather than extrapolate from one available rack.
GreyRadius Insight
New hardware matters when it reduces the cost of a completed business task. Buyers should compare end-to-end throughput and reliability, and avoid premium capacity commitments before testing their own workload mix.
Compute / Accelerators
Synopsys and Amazon sign $1bn custom-chip infrastructure agreement
30 September 2026 | North America / United States
Synopsys and Amazon signed a multi-year $1bn silicon IP agreement supporting future Amazon custom chips and AWS infrastructure, with Amazon becoming the lead customer for Synopsys' expanded silicon-IP business. AWS' custom silicon portfolio includes Trainium accelerators for AI training/inference and Graviton CPUs, built around its Nitro infrastructure architecture. Amazon reported in July that its custom-chip business had exceeded a $25bn annual revenue run rate. The agreement is relevant to AI infrastructure because it supports continued vertical integration of hyperscale compute rather than reliance exclusively on merchant accelerators.
Follow the specific IP licences, design milestones and eventual production programmes. The agreement value does not establish when new chips become available or how much merchant-GPU demand will be displaced.
GreyRadius Insight
Hyperscaler custom silicon creates opportunities across design tools, IP and integration. Suppliers need a customer-specific qualification strategy; access to a large chip programme depends on technical fit and delivery capability as well as price.
Compute / Accelerators
CScale raises $145m for AI optical interconnect technology
30 September 2026 | North America / United States
Palo Alto-based CScale emerged from stealth after raising $145m in Series C funding, taking total funding to $188m. NVIDIA and Intel Capital participated, while Atreides Management, Valor Equity Partners and Premji Invest co-led the round. Founded in 2023, CScale is developing optical interconnect technology for AI scale-up networking, with an architecture intended to contain optical failures without interrupting compute. The funding is relevant to the increasingly important networking layer between accelerators, although the company has not yet disclosed production volumes, bandwidth specifications or customer deployments.
Track disclosed bandwidth, error recovery, thermal performance and customer qualification. Optical-networking capital is an early demand indicator; volume orders and field reliability will determine its commercial relevance.
GreyRadius Insight
As accelerator clusters grow, interconnect reliability can constrain productive compute. Procurement should evaluate degraded-mode performance and repair time, because a faster link that interrupts an expensive cluster may worsen overall economics.
Infrastructure / Power
Nebius signs 50MW, 12-year US data-centre lease
1 October 2026 | North America / United States
Nebius signed a lease with AIB Data Centers covering 50MW at a southeastern US facility. The initial term is 12 years, with two five-year renewal options. The underlying site reportedly has a pre-existing 65MW electric-service agreement. Two halls are planned for Nebius, with delivery targeted in approximately 10 and 14 months. Nebius will provide prepayments, while project-level debt and equity will help fund development. This is contracted future capacity rather than operational capacity, but the existing electrical-service agreement materially differentiates it from projects where power remains speculative.
Verify the conditions attached to the 65MW service agreement, phased hall acceptance and prepayments. Model the gap between lease commitments, construction funding and the 10- and 14-month delivery targets.
GreyRadius Insight
A named tenant and an existing power agreement improve bankability. They still require coordinated delivery: lenders should release capital against power, mechanical completion and tenant acceptance milestones rather than a single campus date.
Infrastructure / Power
Google confirmed as customer for Crusoe's Amarillo AI campus
30 September 2026 | North America / United States
Crusoe confirmed that Google is the end user for its Goodnight Campus under construction near Amarillo, Texas. The campus in Claude, Armstrong County, is planned to span at least six single-storey data-centre buildings. Crusoe describes the development as part of its strategy to build AI infrastructure near abundant, lower-cost energy sources. The identification of Google is commercially important because it converts an anonymous project pipeline into infrastructure associated with a hyperscale customer, improving visibility into future utilization, although the report does not provide a new aggregate MW figure.
Track Google’s contracted scope, energized halls and network readiness at Goodnight Campus. Named customer confirmation strengthens visibility, but the number of buildings cannot be converted into undisclosed MW.
GreyRadius Insight
An identified hyperscaler makes a supplier pipeline more actionable. Equipment and service entrants should pursue the funded phase and qualification requirements, while allowing for customer-specific designs and concentration risk.
Infrastructure / Power
Skybox and Prologis file $550m fourth Hutto data-centre building
2 October 2026 | North America / United States
Skybox and Prologis filed plans for Hutto 4, an 846,300 sq ft / 78,625 sq m data-centre building near Austin, Texas. The project carries an estimated $550m investment, with construction scheduled from January 2027 to October 2028. It would become the largest building at their Hutto campus. The wider PowerCampus is planned to provide 600MW of IT capacity across 3.88m sq ft. Hutto 4 is filed/planned rather than under construction, and the 600MW number refers to the wider campus rather than this individual building.
Verify final permits, funding and construction commencement in January 2027. Keep the US$550m building budget separate from the wider campus’s 600MW IT target when assessing procurement opportunities.
GreyRadius Insight
Large buildings offer a visible construction pipeline, but revenue arrives through accepted capacity. Suppliers should link inventory and staffing to phased orders and delivery windows rather than the campus’s ultimate potential.
Infrastructure / Power
Zone Frontier secures 4,077-acre Texas site with potential for more than 500MW of data centres
29 September 2026 | North America / United States
Zone Frontier announced a long-term ground lease covering approximately 4,077 acres in Potter County, Texas. It can designate 200–800 acres for data centres and at least 1,800 acres for solar generation and battery storage, with the company estimating potential data-centre capacity above 500MW. The lease has an initial 30-year term plus two 10-year extension options and is supported by a long-term water agreement. Crucially, base rent for individual site areas begins only when construction starts, and Zone plans to secure a tenant before beginning construction.
Watch tenant commitments and the conditions activating rent on individual areas. Confirm water rights, generation approvals and a deliverable first site before valuing the full 500MW-plus development potential.
GreyRadius Insight
Construction-linked rent and a tenant-first approach can reduce early holding costs. The project remains an option until demand and power converge; use staged commitments to preserve flexibility while utility and customer negotiations proceed.
Infrastructure / Power
Valar proposes multi-GW SMR-powered data-centre concept in Utah
1 October 2026 | North America / United States
Valar Atomics proposed Project Beehive on more than 9,000 acres of federal land near Price, Utah. According to DCD, the concept could combine data centres with approximately 456 small modular reactors totaling potentially 9.6GW of electrical generation capacity, plus nuclear-fuel and waste-storage facilities. Valar aims for its first reactor in 2028. The project is extremely early-stage and faces substantial regulatory, licensing, financing and construction requirements; the multi-GW number is therefore a proposed generation concept, not secured data-centre power or operating compute capacity.
Track land authorization, reactor licensing, fuel supply and financing. Treat the proposed nuclear generation scale separately from data-centre IT load; the first operating reactor is the essential proof point.
GreyRadius Insight
Advanced nuclear may address long-term firm-power needs, but licensing and deployment uncertainty remain central. Keep speculative reactor capacity outside base-case compute supply forecasts and evaluate nearer-term power options for customer delivery.
Amazon and Constellation signed a 20-year power agreement covering 690MW, including support for a planned 190MW increase in generation capacity at Maryland's 1,790MW Calvert Cliffs nuclear plant. The uprate is expected between 2030 and 2032 and forms part of more than $3bn of infrastructure investment. Electricity will continue entering the PJM grid, so the agreement should not be described as 690MW of dedicated Amazon data-centre IT capacity. The 190MW represents additional generation capacity.
Monitor uprate approvals, the 2030–2032 delivery schedule and retail-supply terms. The 690MW agreement includes existing generation; only 190MW is proposed incremental capacity.
GreyRadius Insight
Long-term offtake can unlock investment in existing generation faster than a completely new nuclear plant. Data-centre operators should still model interim electricity exposure, because the agreement’s future expansion cannot meet immediate load growth.
Infrastructure / Power
FERC pauses PJM backstop plan addressing 6.8GW power shortfall
30 September 2026 | North America / United States
US federal regulators ordered PJM to revise and temporarily delay its proposed electricity-procurement backstop addressing a projected 6,800MW power shortfall, driven in significant part by rapidly increasing data-centre demand. FERC suspended the programme for five months while examining cost allocation and related provisions, emphasizing that large new electricity users should bear appropriate costs rather than shifting them to existing ratepayers. The action is important for data-centre developers because grid cost allocation is increasingly becoming part of project economics rather than a generic utility expense.
Follow the revised PJM proposal and the final allocation of backstop costs. Update development models for potential large-load charges, forecast revisions and the timing of new generation procurement.
GreyRadius Insight
Grid access is increasingly a negotiated financial obligation. Developers should stress-test returns against connection delays and allocated generation costs, and reflect those exposures in tenant pricing before signing long contracts.
Infrastructure / Power
US Senate blocks data-centre electricity-cost legislation 57–43
30 September 2026 | North America / United States
The US Senate blocked the Ratepayer Protection Act in a 57–43 vote, short of the 60 votes required to advance. The House had previously passed the measure 417–3. The legislation sought to require regulators to assess whether large electricity users such as data centres should pay incremental grid-infrastructure costs rather than shifting those expenses to households and other customers. The failed vote leaves the underlying cost-allocation dispute unresolved and signals continued federal political scrutiny of electricity costs associated with rapid data-centre expansion.
Track utility and state action even though the federal procedural vote failed. Confirm which incremental costs current tariffs already allocate to large users; the bill’s failure does not eliminate project-level charges.
GreyRadius Insight
Political resistance to cost shifting can persist without new federal legislation. A resilient entry case should cover the project’s full grid impact and explain community benefits rather than assume households will subsidize expansion.
Capital / Investment
Accelevation and shareholders raise $540m in AI-infrastructure IPO
30 September 2026 | North America / United States
Data-centre infrastructure supplier Accelevation and existing shareholders raised $540m through a US IPO, selling 30 million shares at $18 each. Accelevation itself sold 10 million shares while Olympus Partners-linked shareholders sold 20 million. The company's revenue increased from less than $3m in 2021 to nearly $448m in 2025, illustrating the scale-up occurring in the data-centre equipment and services supply chain. The IPO priced below the marketed $20–$24 range, offering a useful signal on public-market appetite for AI infrastructure businesses.
Separate primary proceeds to Accelevation from shares sold by existing owners. Compare the discounted offer price with backlog quality, cash conversion and the working capital required to fulfill equipment orders.
GreyRadius Insight
AI infrastructure demand can grow while public investors become more selective. Suppliers should fund expansion against profitable, collectible orders; IPO proceeds and shareholder liquidity have different implications for operating capacity.
Infrastructure / Power
Virginia introduces Data Center Accountability Framework and bans project NDAs
1 October 2026 | North America / United States
Virginia Governor Abigail Spanberger introduced a Data Center Accountability Framework that includes a ban on NDAs for data-centre projects, accelerated work on noise regulation and review of environmental impacts from backup generators, particularly diesel systems. The framework also calls for stronger planning and community-engagement tools for local authorities. Because Northern Virginia contains the world's largest concentration of data-centre infrastructure, changes to disclosure and development requirements could influence site-development timelines and community engagement across one of the industry's most important markets. The executive order was signed on 18 September; this is a 1 October news report on the earlier framework.
Review the executive order’s actual scope, implementation timetable and local planning requirements. The framework predates the 1 October report; check obligations governing disclosure, generator impacts and noise rather than assuming every provision is already implemented.
GreyRadius Insight
Community transparency now affects the route from land acquisition to construction. Developers should make power, water, noise and employment commitments reviewable early, with community engagement built into the critical path. Budget for design changes before tenant delivery dates become fixed.
Infrastructure / Power
Oracle reports 2.8m construction hours at New Mexico Project Jupiter
29 September 2026 | North America / United States
In a construction update published on 29 September, Oracle said more than 1,500 union workers had contributed to Project Jupiter in New Mexico by 31 August, with the workforce accumulating more than 2.8m labour hours. Oracle described training and apprenticeship opportunities connected with the build. These are cumulative, company-reported construction indicators rather than new MW, energized halls or operating compute. The update provides evidence of labour intensity, but does not establish a commissioning milestone.
Track electrical, mechanical and commissioning milestones alongside reported labour hours. Separate cumulative activity through August from changes announced this week; hours worked do not establish completed or energized capacity.
GreyRadius Insight
Construction intensity helps explain delivery effort but offers limited evidence of remaining risk. Customers and financiers should monitor critical-path completion and acceptance testing, with workforce data used to assess execution constraints rather than capacity availability.
Infrastructure / Power
Oracle's Michigan AI project passes one million union craft hours
28 September 2026 | North America / United States
Oracle announced that construction of its The Barn AI data-centre development in Saline Township, Michigan, passed one million union craft hours. The project is drawing on electricians, welders, equipment operators, carpenters, ironworkers and plumbers. Oracle did not disclose incremental MW, capex or a commissioning date in this update, so the milestone is evidence of construction activity rather than newly available compute capacity. It is relevant to delivery economics because labour availability is becoming a material dependency alongside power, cooling equipment and accelerators.
Watch critical-trade staffing, equipment delivery and energization schedules. The one-million-hour milestone demonstrates activity, while completion forecasts still require evidence from the project’s electrical and mechanical programme.
GreyRadius Insight
Skilled labour can constrain delivery even when capital and chips are available. Contractors and equipment suppliers should coordinate installation windows early, because a missed trade sequence can delay an entire hall’s revenue start.
Compute / Accelerators
Spectrum brings AI compute to a network with more than 1,000 edge sites
28 September 2026 | North America / United States
West Java’s governor ordered BDx Data Centers to halt construction until required permits were issued and the environmental assessment completed. The company had announced a 640MW Jatiluhur AI Campus 2 on 22 September. The current development is the permitting order, rather than newly announced capacity. Reporting said it remained unclear whether construction had physically stopped, and BDx had not responded to a request for comment. The effect on the project’s delivery schedule remains unresolved.
Measure the latency achieved for actual customers, available GPU capacity and orchestration reliability. A 1,000-site network footprint is a deployment platform; it does not mean AI services are active at every site.
GreyRadius Insight
Edge inference is attractive where response time, bandwidth or local processing creates measurable value. Start with those workloads and compare total operating cost; distributed sites add support complexity that centralized compute can avoid.
Infrastructure / Power
£16.5bn (~US$21.90bn), 1GW Chapelcross data-centre campus enters Scottish planning process
2 October 2026 | Europe / United Kingdom
CX Tech filed a Planning Permission in Principle application for a 1GW technology and data-centre campus next to the former Chapelcross nuclear site in Scotland. The developer estimates total investment at £16.5bn (~US$21.90bn). Plans cover eight buildings totaling 162,000 sq m, plus an on-site substation, with proposed direct-wire connections to renewable generation. Subject to approvals, construction is envisaged between 2027 and 2035. The 1GW figure is proposed project scale; it should not yet be considered secured grid capacity or delivered IT load.
Follow detailed consents, direct-wire agreements and the first funded building. Test the 2027–2035 sequence against utility delivery and customer demand instead of treating all eight buildings as one committed project.
GreyRadius Insight
Long phased developments preserve expansion options but expose investors to changing technology and power economics. Underwrite each phase on its own demand, density and connection assumptions before committing to the entire campus.
Infrastructure / Power
Civo starts 8MW UK sovereign-AI edge facility with 38MW expansion target
1 October 2026 | Europe / United Kingdom
Civo is fitting out an 8MW AI edge data centre in Hertfordshire, intended to scale to 38MW and support high-density sovereign AI workloads including NVIDIA Vera Rubin NVL72 systems. It is planned as the first of 40 UK edge sites. Civo says it has secured five additional locations in Lancashire, Greater Manchester, London and Oxfordshire representing a combined 150MW. The facilities are designed for direct-to-chip liquid cooling. The initial Hertfordshire 8MW is under fit-out, while the 38MW expansion and wider 40-site network remain planned.
Distinguish the 8MW fit-out from the 38MW expansion and wider site programme. Monitor initial customer commitments, available models and liquid-cooling acceptance before counting the planned edge network as usable capacity.
GreyRadius Insight
Sovereign edge capacity needs enough local demand to support distributed operations. Prioritize workloads with measurable residency or latency requirements, and scale site count only when recurring customer revenue supports the operating model.
Infrastructure / Power
Google receives planning permission for 72,400 sq m Dublin data-centre expansion
28 September 2026 | Europe / Ireland
Ireland's planning authority approved Google's third-phase expansion at its Grange Castle campus after an earlier local refusal. The proposed building covers approximately 72,400 sq m on a 50-acre site. Conditions require evidence of engagement around power purchase agreements before development begins and renewable fuel for backup generators. The approval removes a major planning barrier but does not mean construction has started or incremental grid capacity is secured. The project is therefore permitted rather than operational.
Check compliance with pre-construction energy conditions and renewable backup-fuel requirements. Planning permission removes one barrier; power procurement and network delivery remain separate conditions for operation.
GreyRadius Insight
Conditional approvals should feed directly into the investment model. Price the required energy and backup arrangements before committing to a build, and align tenant service dates with the slowest remaining approval or utility dependency.
Infrastructure / Power
AWS receives planning permission for Iver data-centre campus
1 October 2026 | Europe / United Kingdom
Buckinghamshire Council approved AWS' proposed data-centre development at the Ridgeway Distribution Centre in Iver. The campus will replace 17 industrial and warehouse units and includes plans for 27 backup generators. Amazon must contribute £600,382 (~US$796,972) toward projects addressing air pollution associated with the development and monitor generator emissions. Planning approval establishes development rights but does not demonstrate that construction, grid energization or commissioning has started.
Monitor the approval conditions, generator emissions controls and utility connection. Include the air-quality contribution and ongoing monitoring costs when comparing the site with alternative UK developments.
GreyRadius Insight
Local environmental requirements can affect both upfront cost and operating flexibility. A market-entry assessment should include generator permissions and community obligations before the developer locks in the facility’s electrical design.
Capital / Investment
NxN acquires operational 1MW Portuguese facility and plans up to €100m (~US$113.58m) expansion
30 September 2026 | Europe / Portugal
NxN Data Centers entered Portugal through the acquisition of an operational 1MW data centre in Évora. The company simultaneously announced plans to invest up to €100m (~US$113.58m) over three years in Portuguese expansion. The existing 1MW is operating capacity; the €100m (~US$113.58m) is a forward investment plan and should not be treated as already commissioned additional capacity. The transaction gives NxN an immediate operating foothold from which it can build a larger Iberian portfolio.
Compare utilization and recurring revenue at the acquired 1MW site with the proposed expansion pipeline. Require site-level budgets, power availability and customers before treating the three-year investment plan as executable.
GreyRadius Insight
An operating foothold can accelerate entry by providing local processes and customer references. Expansion should follow demonstrated demand, with the acquired facility used to test service quality and commercial economics in Portugal.
Infrastructure / Power
Poland's Konin signs LOI around 100MW first phase of potential 260MW campus
29 September 2026 | Europe / Poland
Energa Obrót, Central Energy Group and Konin municipality signed a letter of intent for a proposed data-centre development. The first phase targets 100MW of IT compute capacity, with potential expansion to 260MW, while total investment has been estimated at just under €3bn (~US$3.41bn). Construction is targeted from early 2027. Because the week's milestone is an LOI rather than a binding construction, financing or power agreement, the 260MW expansion and investment value should be treated as proposed rather than committed capacity.
Follow the LOI into binding power, financing and tenant contracts. Validate the 100MW first-phase schedule independently of the potential 260MW expansion and the total investment estimate.
GreyRadius Insight
An LOI aligns stakeholders but offers limited protection against delivery risk. Suppliers should preserve access to the opportunity through qualification and design work while deferring major exposure until funded orders are signed.
Infrastructure / Power
Amazon signs 36MW Shetland wind PPA
30 September 2026 | Europe / United Kingdom
Amazon signed a PPA with Statkraft for output from the Mossy Hill wind farm near Lerwick in Shetland, Scotland. The wind project is expected to become operational in 2030, with Amazon taking approximately 36MW of capacity. Amazon now has more than 50 carbon-free energy agreements across the UK representing over 1GW, including 41 solar installations on logistics properties plus off-site wind and solar projects. The 36MW figure represents contracted renewable-generation capacity rather than data-centre IT load.
Track Mossy Hill’s construction and 2030 operation date. Assess output variability and balancing obligations; the contracted 36MW nameplate capacity does not provide continuous 36MW supply to a data centre.
GreyRadius Insight
Renewable procurement supports energy strategy but needs a separate reliability solution. Evaluate hourly matching, storage and firming costs so sustainability commitments and round-the-clock compute economics are assessed together.
Infrastructure / Power
Finland proposes replacing first-come grid allocation for large loads
30 September 2026 | Europe / Finland
Finland proposed amendments to its Electricity Market Act that would replace first-come, first-served grid connections with priority groups. Connections with intake capacity of 3MW or less would receive the highest priority, potentially pushing large electricity consumers such as hyperscale data centres further down connection queues. The measure remained a draft during the reporting week rather than enacted legislation. If implemented, it could materially change the value of Finnish development sites that have land but do not yet possess firm grid allocations.
Monitor the draft’s passage, exemptions and treatment of existing connection applications. Projects without firm allocations should reassess their place in the queue and the cost of delayed energization.
GreyRadius Insight
Connection priority can change the economics of land already acquired. Secure contractual power rights and an alternative development sequence before committing to sites whose value depends on an assumed first-come position.
Infrastructure / Power
Digital Realty and Blackfuel launch multi-MW liquid-cooled AI deployment in Barcelona
30 September 2026 | Europe / Spain
Blackfuel selected Digital Realty's BCN1 Barcelona facility for a multi-megawatt AI inference deployment. The infrastructure combines liquid-cooled Dell AI Factory systems and AMD GPUs with Digital Realty's ServiceFabric private-connectivity platform. The companies did not disclose the exact MW figure, GPU count or customer capacity. Nevertheless, this is a physical multi-MW AI deployment rather than a software-only product announcement and demonstrates growing use of existing European colocation facilities for high-density inference workloads.
Validate installed versus customer-ready capacity, production workloads and the service’s pricing. Test private connectivity, regional data controls and support arrangements alongside the AMD hardware’s inference performance.
GreyRadius Insight
Existing colocation can offer a faster route to local AI services when power and cooling are ready. Market opportunity depends on repeat production demand; sell measured performance and service assurance rather than accelerator access alone.
Infrastructure / Power
Global Switch selected for large liquid-cooled NVIDIA Blackwell deployment in Paris
29 September 2026 | Europe / France
Global Switch's Paris facility was selected to host what the company describes as one of Europe's largest NVIDIA Blackwell deployments. The cluster will operate in a predominantly liquid-cooled, high-density environment at an NVIDIA DGX-Ready facility. Neither the customer, GPU count nor MW load was disclosed. The Paris campus is powered with renewable electricity and is scheduled to begin supplying waste heat to Levallois-Perret in 2027. The project highlights how liquid-cooling capability is becoming a competitive requirement for existing colocation assets.
Seek the customer identity, contracted load and accepted cooling capacity. Track the heat-offtake launch separately; a planned district-heating agreement does not demonstrate current heat recovery.
GreyRadius Insight
Established connected facilities can win high-density AI demand through targeted upgrades. Operators should compare retrofit cost and disruption against the revenue from committed tenants, with thermal reliability treated as a core service requirement.
Sovereign AI / Cloud
AWS adds in-region Claude inference in London
30 September 2026 | Europe / United Kingdom
AWS launched in-region support for Claude Opus 5.5 and Claude Sonnet 5 on Amazon Bedrock in its London region. Inference requests and data remain within the United Kingdom, supporting regulated workloads that require local processing. The announcement materially expands the set of frontier AI models that enterprises can deploy while maintaining UK data residency, although AWS did not disclose additional physical infrastructure, accelerator counts or incremental MW supporting the launch.
Benchmark London-region throughput and model pricing with UK production workloads. Confirm the scope of residency across inference, observability and recovery before procurement teams treat the launch as end-to-end compliance.
GreyRadius Insight
UK-local frontier models can reduce one barrier to adoption. The stronger commercial offer combines deployment controls, workflow integration and predictable service capacity; access to a model is only the start of customer value.
Sovereign AI / Cloud
Microsoft launches Azure Extended Zone in Luxembourg
30 September 2026 | Europe / Luxembourg
Microsoft and Post Luxembourg launched an Azure Extended Zone in Luxembourg, placing a small-footprint extension of Azure inside the country for workloads requiring low latency, local data residency and compliance. Extended Zones provide selected compute, storage and networking services rather than the full service catalogue of a standard Azure Region. Luxembourg's government says the infrastructure will allow businesses and public-sector organizations to develop and run AI services locally under European data-protection requirements. No physical MW or server count was disclosed.
Verify the local service catalogue, dependencies on the parent Azure region and recovery locations. Test whether the available compute and network services meet each workload’s residency and resilience needs.
GreyRadius Insight
A local extension can serve targeted demand with a smaller footprint, but feature gaps may limit migration. Match entry strategy to supported applications and customer requirements rather than assume the capabilities of a full cloud region.
Sovereign AI / Cloud
ESA awards 30-month sovereign cloud contract to OVHcloud and CGI
28 September 2026 | Europe / France, Germany, Italy and Poland
The European Space Agency selected OVHcloud and CGI for a 30-month contract to build Digital EO, combining compute, storage, networking and AI tooling for European Earth-observation data. ESA expects its Earth-observation holdings to exceed 500PB by 2035. Infrastructure will be deployed in Italy, France and Germany, with support operations in Poland. The contract is described as multi-million-euro but its precise value was not disclosed. The award represents a concrete European sovereign-cloud procurement rather than a general policy commitment.
Follow delivery milestones, sovereign-operation requirements and migration from existing data holdings. With a 500PB-plus future archive, assess storage tiers, egress cost and interoperability early in the programme.
GreyRadius Insight
Sovereign demand becomes actionable through funded procurement and integration work. Suppliers should target specific storage, data-management and processing roles, with evidence that their services fit the buyer’s operational and sovereignty requirements.
Sovereign AI / Cloud
Poland receives PLN1.56bn (~US$401m) for three-site sovereign data-processing network
2 October 2026 | Europe / Poland
Poland’s National Data Processing Center received PLN1.56bn (~US$401m) in EU-backed funding for three interconnected data centres, fibre links and supporting IT infrastructure. The programme is overseen by the Ministry of Digital Affairs and NASK-PIB. Earlier procurement was cancelled, and reporting now estimates completion by 2029. The first-phase cost ceiling is PLN1.6bn (~US$411m). This is funding for developing sovereign infrastructure, with no disclosed MW or server count; it is not commissioned national compute capacity.
Track renewed procurement, the three-site delivery plan and connectivity contracts. Separate the grant from signed build orders and monitor schedule recovery after the earlier procurement cancellation.
GreyRadius Insight
Public funding creates a route to market but does not guarantee timely execution. Entrants should align bidding, partnerships and staffing with procurement milestones, and price operational resilience across all three sites rather than each facility in isolation.
Sovereign AI / Cloud
EU opens €80m (~US$90.86m) of AI hardware and compute-infrastructure funding
29 September 2026 | Europe / European Union
The EU Chips Joint Undertaking highlighted two existing funding calls with a combined €80m (~US$90.86m) budget targeting AI hardware development and deployment. Eligible participants include AI-chip designers, system developers, AI data centres, HPC/AI infrastructure operators and cloud providers. One call covers AI-chip demonstrators for European compute infrastructure, while the other focuses on AI-compute evaluation and procurement. Applications close 27 October 2026. This is programme funding rather than installed compute; no accelerator count or MW capacity has yet been procured. The calls opened on 7 July; the 29 September report is an application-opportunity update.
Confirm eligibility, consortium requirements and the 27 October submission deadline. The funding calls opened earlier in the year; this week’s report highlights an existing application opportunity rather than a new award or commissioned compute.
GreyRadius Insight
Grant access can reduce technology-development exposure but does not validate a market. Build proposals around an identifiable operator, workload and procurement path, and reserve commercial funding for the steps needed to convert a demonstrator into repeat customer orders.
Sovereign AI / Cloud
Balearic Islands launches redundant data centre for critical government systems
30 September 2026 | Europe / Spain
The Balearic Islands government launched a new data centre in Palma, Mallorca, integrated with its existing Sant Pere facility to create a distributed, redundant architecture for government and health-service systems. Authorities say either facility can support 100% of technological operations if the other fails. The new facility is designed to a Tier III-equivalent standard with stated availability above 99.982%. This is operational sovereign/government infrastructure rather than proposed capacity, although the government did not disclose its MW or server count.
Test failover under full production load, recovery time and the independence of power and fibre paths. Claimed site availability is less useful than evidence that critical public services survive a facility outage.
GreyRadius Insight
Government infrastructure creates value through continuity of services. Procurement should use tested recovery objectives and common-mode failure analysis, ensuring that geographic redundancy also means operational independence.
Capital / Investment
Virtus secures £2.45bn (~US$3.25bn) financing for UK and European data-centre build-out
28 September 2026 | Europe / United Kingdom and continental Europe
Virtus Data Centres completed a £2.45bn (~US$3.25bn) financing package involving 13 banks. The package includes a £1.2bn (~US$1.59bn) green capex facility and will support existing and future infrastructure in the UK and continental Europe, including the 78MW Saunderton campus and future London19 development in Slough. Virtus operates 14 UK facilities and is developing additional projects in Germany and Italy. The £2.45bn (~US$3.25bn) is portfolio-level financing rather than the capex assigned solely to Saunderton or any single project. This adviser announcement revisits the financing covered in the previous issue.
Reconcile the £2.45bn (~US$3.25bn) facility with actual drawdowns and site-level construction commitments. This adviser release revisits financing covered in the previous issue; monitor Saunderton’s power, leasing and commissioning milestones for new execution evidence.
GreyRadius Insight
Portfolio financing can spread capital across operating and developing assets, improving flexibility while obscuring individual project returns. Require asset-level cash-flow coverage and staged capex; operational portfolio income should not conceal an unfunded or unleased development phase.
Infrastructure / Power
Enetron's claimed 250MW Finnish project approval contradicted by municipality
2 October 2026 | Europe / Finland
Enetron Park announced that zoning had been approved for a 250MW data-centre and energy campus in Pyhäjoki, with construction targeted for early 2028. However, municipal meeting records showed that local officials had actually rejected planning for the project in a 5–2 vote because no applicable general or local plan existed for the site. Local reporting subsequently suggested the project may not proceed in Pyhäjoki. This item should therefore be treated as an unresolved/disputed development rather than approved capacity, and the 250MW must not be added to Finland's committed pipeline.
Verify municipal decisions directly and seek a revised planning route. Exclude the disputed 250MW from committed supply until enforceable approvals and a feasible construction programme are documented.
GreyRadius Insight
Developer announcements should be tested against the authority that controls delivery. Site investment decisions need a clear chain from land rights to planning permission and utility access, especially where public claims conflict with local records.
Compute / Accelerators
Meta contracts Firmus GPU capacity across Southeast Asia
29 September 2026 | Asia-Pacific / Southeast Asia and Australia
Meta entered agreements with Firmus covering contracted GPU compute capacity plus expansion options at AI factories in Southeast Asia built around NVIDIA infrastructure. Meta is already consuming GB300 NVL72 capacity from Firmus' Melbourne facility. Firmus and DayOne are separately developing a Batam, Indonesia AI campus previously described as capable of accommodating 170,000 GPUs, but that figure is the broader project target and is not Meta's contracted GPU quantity. Financial terms and Meta's committed GPU count were not disclosed, so expansion options should not be interpreted as firm contracted capacity.
Separate Meta’s firm reservations from expansion options. Follow each Asian site’s energized capacity, deployment dates and export permissions; the Batam project target does not reveal Meta’s contracted GPU quantity.
GreyRadius Insight
An anchor customer can help finance regional AI factories, while options leave expansion risk with the operator. Suppliers should size their entry around funded deployment phases and customer concentration, rather than the entire campus ambition.
Compute / Accelerators
Tencent reportedly signs five-year Oracle deal for access to 100,000 GPUs
1 October 2026 | Asia / China and Southeast Asia
Tencent reportedly signed a five-year Oracle Cloud lease giving it access to 100,000 GPUs across multiple Oracle data centres in Southeast Asia. DCD, citing the Financial Times, puts the agreement at approximately $7bn, with Tencent reportedly paying 30% upfront. The arrangement would allow Tencent to use accelerators unavailable inside China because of US export controls. Neither Tencent nor Oracle had publicly confirmed the transaction in DCD's report, making this a significant but secondary-source-only reported contract, rather than a company-verified commitment.
Seek confirmation from Oracle and Tencent, equipment eligibility and deployment locations. Verify whether the reported upfront payment secures delivered systems or future capacity before using the deal in regional demand forecasts.
GreyRadius Insight
Offshore compute can change where demand is served without eliminating compliance exposure. Market-entry cases should account for export rules, remote-access controls, data movement and the consequences of a future policy change.
Compute / Accelerators
Semifive signs $52m contract for next-generation AI inference accelerator
30 September 2026 | Asia / South Korea and United States
South Korean ASIC company Semifive signed a $52m contract with an unnamed US AI fabless company to develop a next-generation inference accelerator targeting hyperscalers and cloud providers. Semifive said the agreement is its largest contract and represents about 60% of the company's $88m in H1 2026 new-order revenue. The project covers design, software, packaging, testing and mass production through Semifive's Spec Hand-off model. Commercial production is targeted for 2028, making this an AI-chip supply-chain development rather than currently available compute capacity.
Watch tape-out, foundry allocation, packaging qualification and the 2028 production target. Confirm customer concentration and milestone payments before translating this design order into recurring manufacturing revenue.
GreyRadius Insight
A large accelerator design contract opens a route into US demand, but execution spans several suppliers. Build contingency into foundry and packaging plans and tie commercialization spending to verified technical milestones.
Infrastructure / Power
CDC proposes 1.4GW Wagga Wagga data-centre campus
30 September 2026 | Asia-Pacific / Australia
CDC Data Centres is exploring a major campus near Wagga Wagga, New South Wales, reported at approximately 1.4GW and A$15bn (~US$10.46bn) of investment. The development would comprise four buildings and would place major compute capacity well outside Australia's established Sydney and Canberra clusters. The project remains at the planning and consultation stage. Consequently, 1.4GW is a proposed facility-scale figure, not commissioned IT load or proof of a secured grid connection. The project's location makes future transmission, grid allocation and permitting milestones particularly important.
Track planning consent, transmission studies, connection offers and a funded first phase. Confirm whether the 1.4GW figure measures facility demand or IT load before comparing it with other Australian projects.
GreyRadius Insight
A regional campus can diversify development beyond established hubs, but remote locations add fibre, staffing and grid dependencies. Evaluate a commercially viable first tranche independently of the full multibillion-dollar campus vision.
Infrastructure / Power
HyperDC files plans for 400MW IT campus outside Melbourne
29 September 2026 | Asia-Pacific / Australia
HyperDC filed plans for a 39-hectare data-centre campus targeting 400MW of IT capacity at Moorabool, approximately 75km west of Melbourne. The proposed site is close to AusNet's Moorabool Terminal Station and the 300MW Victorian Big Battery, although physical proximity does not establish that the developer has secured a 400MW grid connection. No commissioning schedule or financing package was disclosed. The project therefore remains proposed capacity, with power allocation and planning approval representing major execution dependencies.
Require written grid allocations and permitting milestones. Proximity to a terminal station and battery is useful site context, but it does not establish access to the project’s planned 400MW.
GreyRadius Insight
Site attractiveness should be priced through the cost and timing of energization. Secure connection rights and network-upgrade obligations before paying a premium for land marketed as close to power infrastructure.
Infrastructure / Power
BrightRay plans 40MW liquid-cooling-capable AI data centre in Macau
30 September 2026 | Asia / Macau, China
BrightRay signed a prefabricated-module framework agreement for a planned AI data centre in Coloane, Macau. The five-storey facility could provide up to 40MW of IT capacity across 9,000 sq m, with delivery targeted for Q3 2027. The proposed Tier III-quality facility would support air cooling, full liquid cooling and hybrid deployments, with rack densities ranging from 10kW to 500kW per rack. The project is planned rather than operational, but the upper rack-density figure illustrates the mechanical design requirements emerging around future AI infrastructure.
Validate Q3 2027 delivery, utility availability and the density supported simultaneously across racks. A 500kW maximum rack specification does not mean the entire 40MW facility can operate at that density.
GreyRadius Insight
Flexible cooling can widen the customer pool, but unused high-density capability adds cost. Align electrical and thermal design with contracted workloads, with modular expansion for higher densities as demand becomes firm.
Capital / Investment
Amazon agrees $278m–$312m acquisition of two Hong Kong data centres
29 September 2026 | Asia / Hong Kong, China
Amazon Data Services Hong Kong agreed to acquire iTech Tower 3.1 and iTech Tower 3.2 in Fanling from Grand Ming Group. The minimum gross consideration is HK$2.18bn ($277.9m) and could reach HK$2.445bn ($311.7m) with additional payments. The first phase of Tower 3.1 was delivered in December 2025 and is already leased to Amazon, with fit-out continuing; Tower 3.2 remains under construction. The transaction converts leased/developing infrastructure into Amazon-controlled property and expands AWS' physical position in Hong Kong.
Follow transaction completion, fit-out of Tower 3.1 and construction of Tower 3.2. Separate operating space, developing space and contingent consideration when assessing the capacity Amazon gains.
GreyRadius Insight
Owning infrastructure can improve control over a strategic location and remove lease renewal exposure. The investment case still rests on usable power and workload demand; property acquisition alone does not produce additional cloud capacity.
Capital / Investment
Amazon agrees RM316.84m (~US$77.66m) purchase of 45.46-acre Malaysian data-centre site
28 September 2026 | Asia / Malaysia
Sunsuria agreed to sell Amazon Data Services Malaysia a 45.46-acre / 183,973 sq m site in Sepang for RM316.84m (~US$77.66m), for proposed data-processing services and infrastructure. Completion is targeted for Q2 2027, subject to shareholder and other approvals. No MW capacity, construction budget or commissioning schedule has yet been disclosed. The transaction therefore represents land control and future development optionality rather than additional operating Malaysian cloud capacity.
Monitor shareholder approvals and the Q2 2027 closing target. Watch for a disclosed power allocation, building programme and construction budget before assigning MW or cloud-launch dates to this land transaction.
GreyRadius Insight
Land control preserves a route to future growth in Malaysia. Investors should value that option separately from a commissioned asset, with development spending conditional on utility capacity and a credible customer-demand case.
Sovereign AI / Cloud
AWS adds in-region Claude inference in South Korea and Singapore
29 September 2026 | Asia / South Korea and Singapore
Amazon Bedrock introduced Claude Opus 5 and Claude Sonnet 5 in Seoul, while Claude Sonnet 5 became available with in-region inference in Singapore. AWS states that inference requests and associated data remain within the region called by the customer. The capability targets regulated workloads in financial services, healthcare and the public sector. Like the India launch, this is a material cloud-residency development rather than construction of a new cloud region or disclosure of physical compute capacity.
Validate single-region quotas and peak-load behavior in Seoul and Singapore. Confirm the exact model available in each market and test how the application responds when regional capacity is exhausted.
GreyRadius Insight
Residency choices trade routing flexibility for local processing assurance. Buyers should size reserved capacity and fallback behavior around business-critical peaks, while providers differentiate through reliable local service rather than model availability alone.
Infrastructure / Power
Indonesia halts construction at BDx project over permitting
30 September 2026 | Asia / Indonesia
West Java authorities ordered construction halted on a BDx Data Centers project because required permits had not been secured. BDx had previously announced its 640MW Jatiluhur AI Campus 2, with an initial 120MW building designed for direct-to-chip liquid cooling and very high rack densities. The material development this week is the permitting halt, not the older capacity announcement. The impact on the project's previously targeted early-2027 delivery remains unresolved, making it a strong example of announced MW diverging from deliverable capacity because of local approvals.
Require evidence of permit issuance and environmental clearance, then verify whether work has resumed. Do not assume the halt order has already become a confirmed schedule delay; the physical construction status remains uncertain.
GreyRadius Insight
Breaking ground before approvals creates a gap between publicity and deliverability. Market-entry execution should treat local permits and community requirements as spending gates, with tenant commitments and construction exposure matched to clearance status.
Infrastructure / Power
Egypt's Intro Group plans $270m first phase of 80MW Kemet Data Center
2 October 2026 | Middle East / Egypt
Intro Group plans to invest $270m in the first phase of the Kemet Data Center in Egypt's Suez Canal Economic Zone. The full project is designed for 80MW across 80,000 sq m and will be developed in two phases, with part of its electricity expected to come from solar generation. Egypt's communications minister and Intro Group reviewed completed designs, equipment and systems contracts, implementation timing and the planned service launch. The $270m relates specifically to phase one rather than the entire project.
Track the phase-one service date, firm power mix and committed customer demand. Clarify how much of the eventual 80MW is funded by the initial US$270m investment.
GreyRadius Insight
Egypt’s connectivity position can support regional infrastructure, but customers will buy resilience and service economics. Validate international routes, energy reliability and payment risk alongside the initial construction budget.
Executive Dashboard
Signal
Key Data
Date
Region
Executive Implication
AWS launches in-country Claude inference across India
Claude Opus 5, Claude Sonnet 5 and Claude Haiku 4.5; Mumbai and Hyderabad Regions
29 September 2026
Asia / India
Local model availability can unlock regulated demand sooner than a new campus build.
Alibaba Cloud announces new regions in Türkiye, Finland and the Netherlands
Türkiye, Finland and the Netherlands over the next 12 months; Malaysia, Germany, the UAE, France and Hong Kong
29 September 2026
Global / Europe, Asia and GCC
Geographic expansion becomes commercially useful when customers can deploy supported workloads locally.
Mubadala and Together AI establish UAE sovereign-AI infrastructure partnership
Abu Dhabi presence
29 September 2026
GCC / United Arab Emirates
A regional partnership can open customer access before physical supply is committed.
Anthropic discloses at least $518bn of long-term AI infrastructure commitments
$518bn of AI infrastructure commitments over roughly a decade; 80% is described as non-cancelable regardless of usage; $111.1bn with Google, $110bn with Amazon, $31.4bn with Microsoft and $161.2bn of Broadcom-related lease obligations
29 September 2026
Global / United States
Long contracts can finance supply while concentrating demand risk in a few AI labs.
Trane and LiquidStack launch 2.5MW coolant distribution unit for AI infrastructure
CDU 2.X; 2.5MW of cooling capacity; 3,750 litres per minute at 3.5 bar; 45°C
30 September 2026
Global
A larger CDU expands design options, while the weakest link in the thermal chain still limits uptime.
GMI Cloud raises $668m for AI infrastructure expansion
$668m; $223m of Series B equity; $445m credit facility; more than $600m in contracted annualized revenue
1 October 2026
Global / United States, Taiwan and Asia-Pacific
Equity and credit support different risks: early expansion needs flexibility, while debt needs dependable cash flows.
20-year power agreement covering 690MW; 190MW increase in generation capacity; 2030 and 2032; $3bn of infrastructure investment
30 September 2026
North America / United States
Long-term offtake can unlock investment in existing generation faster than a completely new nuclear plant.
FERC pauses PJM backstop plan addressing 6.8GW power shortfall
6,800MW power shortfall; five months
30 September 2026
North America / United States
Grid access is increasingly a negotiated financial obligation.
US Senate blocks data-centre electricity-cost legislation 57–43
57–43 vote; 417–3
30 September 2026
North America / United States
Political resistance to cost shifting can persist without new federal legislation.
Accelevation and shareholders raise $540m in AI-infrastructure IPO
$540m through a US IPO; 30 million shares at $18 each; $3m in 2021 to nearly $448m in 2025
30 September 2026
North America / United States
AI infrastructure demand can grow while public investors become more selective.
Virginia introduces Data Center Accountability Framework and bans project NDAs
ban on NDAs for data-centre projects
1 October 2026
North America / United States
Community transparency now affects the route from land acquisition to construction.
Oracle reports 2.8m construction hours at New Mexico Project Jupiter
1,500 union workers; more than 2.8m labour hours
29 September 2026
North America / United States
Construction intensity helps explain delivery effort but offers limited evidence of remaining risk.
Oracle's Michigan AI project passes one million union craft hours
The Barn; one million union craft hours
28 September 2026
North America / United States
Skilled labour can constrain delivery even when capital and chips are available.
Spectrum brings AI compute to a network with more than 1,000 edge sites
640MW Jatiluhur AI Campus 2
28 September 2026
North America / United States
Edge inference is attractive where response time, bandwidth or local processing creates measurable value.
£16.5bn (~US$21.90bn), 1GW Chapelcross data-centre campus enters Scottish planning process
1GW; £16.5bn (~US$21.90bn); eight buildings totaling 162,000 sq m; 2027 and 2035
2 October 2026
Europe / United Kingdom
Long phased developments preserve expansion options but expose investors to changing technology and power economics.
Civo starts 8MW UK sovereign-AI edge facility with 38MW expansion target
8MW AI edge data centre in Hertfordshire; 38MW; 40 UK edge sites; 150MW
1 October 2026
Europe / United Kingdom
Sovereign edge capacity needs enough local demand to support distributed operations.
Google receives planning permission for 72,400 sq m Dublin data-centre expansion
72,400 sq m on a 50-acre site; power purchase agreements before development begins
28 September 2026
Europe / Ireland
Conditional approvals should feed directly into the investment model.
AWS receives planning permission for Iver data-centre campus
17 industrial and warehouse units; 27 backup generators; £600,382 (~US$796,972) toward projects addressing air pollution associated with the development
1 October 2026
Europe / United Kingdom
Local environmental requirements can affect both upfront cost and operating flexibility.
NxN acquires operational 1MW Portuguese facility and plans up to €100m (~US$113.58m) expansion
operational 1MW data centre in Évora; up to €100m (~US$113.58m) over three years
30 September 2026
Europe / Portugal
An operating foothold can accelerate entry by providing local processes and customer references.
Poland's Konin signs LOI around 100MW first phase of potential 260MW campus
letter of intent; 100MW of IT compute capacity; 260MW; €3bn (~US$3.41bn)
29 September 2026
Europe / Poland
An LOI aligns stakeholders but offers limited protection against delivery risk.
Amazon signs 36MW Shetland wind PPA
2030; 36MW of capacity; 50 carbon-free energy agreements across the UK representing over 1GW
30 September 2026
Europe / United Kingdom
Renewable procurement supports energy strategy but needs a separate reliability solution.
Finland proposes replacing first-come grid allocation for large loads
3MW or less
30 September 2026
Europe / Finland
Connection priority can change the economics of land already acquired.
Digital Realty and Blackfuel launch multi-MW liquid-cooled AI deployment in Barcelona
BCN1 Barcelona facility; multi-megawatt AI inference deployment; liquid-cooled Dell AI Factory systems and AMD GPUs
30 September 2026
Europe / Spain
Existing colocation can offer a faster route to local AI services when power and cooling are ready.
Global Switch selected for large liquid-cooled NVIDIA Blackwell deployment in Paris
one of Europe's largest NVIDIA Blackwell deployments; liquid-cooled, high-density
29 September 2026
Europe / France
Established connected facilities can win high-density AI demand through targeted upgrades.
AWS adds in-region Claude inference in London
Claude Opus 5.5 and Claude Sonnet 5
30 September 2026
Europe / United Kingdom
UK-local frontier models can reduce one barrier to adoption.
Microsoft launches Azure Extended Zone in Luxembourg
Azure Extended Zone in Luxembourg; compute, storage and networking
30 September 2026
Europe / Luxembourg
A local extension can serve targeted demand with a smaller footprint, but feature gaps may limit migration.
ESA awards 30-month sovereign cloud contract to OVHcloud and CGI
30-month contract; 500PB by 2035; Italy, France and Germany
28 September 2026
Europe / France, Germany, Italy and Poland
Sovereign demand becomes actionable through funded procurement and integration work.
Poland receives PLN1.56bn (~US$401m) for three-site sovereign data-processing network
PLN1.56bn (~US$401m); three interconnected data centres; 2029; PLN1.6bn (~US$411m)
2 October 2026
Europe / Poland
Public funding creates a route to market but does not guarantee timely execution.
EU opens €80m (~US$90.86m) of AI hardware and compute-infrastructure funding
€80m (~US$90.86m) budget; AI-chip demonstrators for European compute infrastructure; AI-compute evaluation and procurement; 27 October 2026
29 September 2026
Europe / European Union
Grant access can reduce technology-development exposure but does not validate a market.
Balearic Islands launches redundant data centre for critical government systems
either facility can support 100% of technological operations if the other fails; Tier III-equivalent standard with stated availability above 99.982%
30 September 2026
Europe / Spain
Government infrastructure creates value through continuity of services.
Virtus secures £2.45bn (~US$3.25bn) financing for UK and European data-centre build-out
Portfolio financing can spread capital across operating and developing assets, improving flexibility while obscuring individual project returns.
Enetron's claimed 250MW Finnish project approval contradicted by municipality
250MW data-centre and energy campus in Pyhäjoki; rejected planning for the project in a 5–2 vote; unresolved/disputed development rather than approved capacity
2 October 2026
Europe / Finland
Developer announcements should be tested against the authority that controls delivery.
Meta contracts Firmus GPU capacity across Southeast Asia
GreyRadius helps leadership teams translate AI adoption, infrastructure, sovereignty and market shifts into investment priorities, GTM choices and execution-ready growth plans.
