Enterprise AI • Compute • Sovereign AI • Deployment Infrastructure
Published by GreyRadius Consulting | 19–25 October 2026
Executive Highlights
IndiaSTTGDC and Neysa plan deployment of up to 9,000 GPUs in Noida. Up to 9,000 GPUs | Noida, India | Target Q1 2027.
IndiaVultr announces 20MW AMD AI compute cluster in Chennai. 20MW announced cluster scale | AMD MI355X | Chennai, India.
GCC — Saudi ArabiaNebius commits to up to 50MW at Saudi Arabia's Oxagon AI data centre. Up to 50MW customer allocation | 100MW under construction | 360MW first-phase target | 1.5GW ultimate ambition | Q4 2027 availability target.
GCC — United Arab Emirates / United StatesAleria outlines NVIDIA GPU infrastructure expansion in UAE and United States. 8,640 planned Blackwell Ultra GPUs | 16,000-GPU target | 28 planned Vera Rubin NVL72 racks in UAE.
United StatesMeta's Prometheus AI campus reaches first operational milestone in Ohio. First phase operational | Hundreds of thousands of GPUs reportedly installed | Approximately 1 million GPUs targeted at full scale.
United StatesNebius secures $3.5 billion financing facility for AI infrastructure expansion. US$3.5 billion financing | One-year maturity | SOFR + 2.75 percentage points initially | More than US$40 billion in referenced customer commitments.
United StatesGoogle and Voltus announce 1GW virtual power plant arrangement. 1GW virtual power plant target | Previous 100MW agreement | Approximately three-year duration | PJM market.
United StatesVantage signs agreement for more than 1GW of solid-state transformer equipment. More than 1GW equipment agreement | 5MW modules | 800V DC architecture | Greater than 98.5% claimed efficiency.
United StatesTerraWatt proposes 1GW data centre redevelopment in Illinois. Up to 1GW proposed campus capacity | 200MW development blocks | Illinois, United States.
United StatesEndeavour and Trevi propose desalination using data centre waste heat. Up to 200,000 gallons/day/MW claimed potential | Waste-heat utilisation | Desalination application.
SwitzerlandSwitzerland advances sovereign-cloud procurement worth up to CHF7 billion. CHF7 billion procurement ceiling | Reported equivalent of US$8.5 billion | Swiss Government Cloud programme.
NorwayCrusoe commits to another 40MW AI data centre at Norway's Tørdal campus. 40MW proposed DRA02 | 12MW existing operational IT capacity | NVIDIA Vera Rubin NVL72 planned.
Germany / EuropeNVIDIA introduces modular AI factories targeting installation in as little as 14 days. As little as 14 days claimed installation | Approximately six-month comparison | Early 2027 deployment target | Up to 30% claimed power reduction under specified conditions.
United KingdomMicrosoft withdraws from delayed Nscale data centre in Essex. Original completion target: late 2026 | Grid connection reportedly delayed until the 2030s | Microsoft anchor tenancy withdrawn.
IcelandEmerald AI and NVIDIA demonstrate AI workload power flexibility. Approximately 25% reported rack-power reduction | 30-minute test duration | Iceland pilot.
SpainKrolmap proposes €1.5 billion, 225MW off-grid data centre in Zaragoza. €1.5 billion proposed investment | 225MW planned capacity | Zaragoza, Spain | Off-grid power concept.
South KoreaKT Cloud begins construction of 50MW AI data centre in Ansan. 50MW planned capacity | 36,880m² floor area | 2029 completion target | Ansan, South Korea.
ChinaZTE introduces liquid-cooled AI data centre module supporting up to 250kW per rack. Up to 250kW claimed rack support | Liquid-cooling architecture | Modular AI data centre design.
KazakhstanAkashi secures $31.2 million financing for phased AI data centre development. US$31.2 million reported financing | 10MW initial phase | 2027 initial target | 100MW ultimate campus ambition.
United States / GlobalOracle introduces GPU cluster-health validation for AI cloud workloads. Released 22 October 2026 | Passive and active cluster validation | Slurm and Kubernetes support | No additional GPU capacity disclosed.
Filter
AI Compute Infrastructure | GPU Cloud
STTGDC and Neysa plan deployment of up to 9,000 GPUs in Noida
23 October 2026 | India
STTGDC India and Indian AI cloud provider Neysa announced a deployment at the STT Noida 2 data centre in Uttar Pradesh. The initial phase is designed to accommodate up to 9,000 GPUs, supporting Neysa's Velocis AI cloud platform for training and inference workloads. Commercial operations are targeted for Q1 2027. The agreement establishes a prospective deployment location and customer, but does not confirm GPU installation, commissioning or customer availability. The accelerator models and precise power requirements were not disclosed.
Track the path from hosting agreement to installed accelerators, tested cooling and commercially available service. The Q1 2027 target depends on equipment delivery and commissioning; up to 9,000 GPUs describes a prospective deployment ceiling, not bookable supply. The next commercial signal is the proportion of the planned fleet reserved by paying customers and the terms under which those reservations convert into service. A large deployment with weak demand visibility can create pricing pressure before it creates attractive returns.
GreyRadius Insight
Enterprise buyers should request capacity reservations tied to a defined accelerator configuration, acceptance test and service date. Benchmark their own training or inference workload and include storage, network, data movement and support in the comparison. Commit in tranches so delayed commissioning does not leave a critical workflow without computing capacity. Set a minimum acceptable throughput and maximum job-completion time for the actual workload. Include remedies for late service commencement and an alternative-provider plan. Review cost per successful training run or per accepted inference output; inexpensive GPU-hours can be poor value when networking, storage or failures extend completion time.
AI Infrastructure | GPU Cloud
Vultr announces 20MW AMD AI compute cluster in Chennai
21 October 2026 | India
Vultr announced plans for a 20MW AI computing cluster in Chennai using AMD Instinct MI355X accelerators. The proposed deployment targets enterprise AI training, fine-tuning and inference workloads, potentially expanding access to AMD-based GPU computing within India. The announcement identifies the accelerator architecture and planned deployment scale, but does not confirm the GPU count, energisation date or commercial service launch. The 20MW figure should not be treated as verified operational IT load.
Confirm whether 20MW refers to total facility power or IT load, then track energisation, accelerator deliveries and service launch. The commercial test is reliable workload performance and customer uptake on the AMD platform; announced power alone does not establish computing throughput. A broader accelerator supplier base could improve purchasing flexibility, but hardware diversification creates value only where the software stack can support it efficiently. Watch customer migration effort and repeat usage after initial trials, since these reveal whether demand extends beyond discounted access.
GreyRadius Insight
Evaluate AMD-based supply through cost per completed workload at an agreed quality and service level. Include software compatibility, engineering effort and migration costs in the pilot. For providers, stage capital against validated customer demand and operational readiness; a differentiated hardware option creates value when customers can use it reliably. Separate portable workloads from those requiring substantial software changes. Include engineering time and performance tuning in switching costs, then compare these with the expected savings over the contract term. Diversify where the business case is positive, while retaining a common evaluation method so supplier comparisons remain credible.
AI Data Centres | Hyperscale Investment
Nebius commits to up to 50MW at Saudi Arabia's Oxagon AI data centre
22 October 2026 | GCC — Saudi Arabia
Nebius signed an agreement with Saudi AI company Humain and developer DataVolt to become an anchor customer at the Oxagon data centre development within NEOM. Nebius will initially take up to 50MW of the 100MW currently under construction, with phased capacity availability targeted from Q4 2027. The development has a proposed first-phase scale of 360MW and an ultimate campus ambition of 1.5GW. Nebius intends to deploy NVIDIA-powered AI infrastructure. The customer agreement represents a material commercial commitment, but the ultimate campus target does not constitute operational or fully financed capacity.
Track delivery of the customer-backed allocation separately from the 100MW under construction and larger campus ambitions. The critical dependencies are power availability, phased acceptance, equipment delivery and when the agreement begins generating revenue. An anchor customer can improve financing prospects while increasing dependence on a single counterparty. Watch minimum purchase obligations, delivery conditions and termination rights where disclosed. Commercial commitment is strongest when the tenant and developer carry clearly defined obligations across each phase.
GreyRadius Insight
Underwrite the first contracted tranche against its own capital, power and delivery obligations. Tie expansion to accepted capacity and customer uptake, while clarifying responsibility for delays across developer, operator and equipment suppliers. Use the larger campus figures as scenario inputs until each phase has financing and a viable critical path. Run a downside case in which the first tranche is delayed and later campus phases do not proceed. Confirm that shared power, cooling and network investments remain viable under that scenario. Treat expansion as a sequence of separately justified commitments, with customer concentration and cash collection included in the investment review.
Sovereign AI | GPU Infrastructure
Aleria outlines NVIDIA GPU infrastructure expansion in UAE and United States
23 October 2026, reported | GCC — United Arab Emirates / United States
Aleria announced plans to expand AI computing infrastructure across the UAE and United States. The proposed US deployment includes 8,640 NVIDIA Blackwell Ultra GPUs, with a longer-term target of 16,000 GPUs. The company also outlined plans for 28 NVIDIA Vera Rubin NVL72 racks in the UAE. The announcement indicates potential expansion of high-density AI compute infrastructure, but equipment delivery, installation, energisation and customer availability require further confirmation. The US GPU quantities and UAE rack quantities describe different deployment commitments and should not be combined into a single installed-capacity figure. Direct source URL was not established in the supplied research; source verification remains pending.
Source: DatacenterDynamics
Strategic Watch
Confirm the original announcement and delivery timetable before treating the expansion as available supply. US GPU quantities and UAE rack quantities describe different geographies, units and technology generations; adding them would obscure readiness and procurement risk. The proposed technology generations may involve different delivery schedules and infrastructure requirements. Watch whether power and cooling designs can accommodate the intended systems without expensive redesign, and whether the two locations have distinct customers and service propositions.
GreyRadius Insight
Build separate UAE and US deployment cases covering power, cooling, delivery dates and applicable equipment requirements. Avoid translating planned racks into usable service capacity without a validated configuration. Buyers should preserve alternative supply until allocation, installation and service acceptance are documented. Assign each location its own demand forecast, unit economics and commissioning plan. Require an upgrade path and clarify who pays if equipment specifications change before delivery. Frame sovereignty around actual data handling and operational control, then validate those requirements contractually rather than inferring them from the hosting country.
Operational AI Infrastructure | Hyperscale
Meta's Prometheus AI campus reaches first operational milestone in Ohio
23 October 2026 | United States
Meta reported that the first phase of its Prometheus AI data centre campus in Ohio had become operational. The development reportedly already accommodates hundreds of thousands of GPUs, with an ultimate deployment ambition of approximately 1 million GPUs. Unlike projects that remain at the announcement or construction stage, the first-phase milestone indicates that some AI computing infrastructure has entered operation. However, the reported GPU counts should be distinguished from independently verified customer-accessible capacity, and the ultimate GPU target should not be interpreted as installed infrastructure.
The first operational phase is a stronger delivery signal than a campus proposal, but installed GPU counts still reveal little about productive utilisation. Watch workload availability, failure rates, networking performance and the relationship between delivered computing output and energy use. At this scale, small losses in reliability or utilisation can consume substantial installed resources. Watch whether reported operating progress translates into stable training throughput and shorter completion times; neither energisation nor a GPU inventory reveals the rate at which useful computing is delivered.
GreyRadius Insight
For infrastructure strategy, distinguish internal hyperscale capacity from commercially accessible cloud supply. Benchmark execution through accepted workload output, utilisation and reliability rather than GPU totals alone. Operators can apply the same discipline to their own ramp: energised capacity must be converted into productive computing before further expansion is justified. Use a capacity waterfall in investment reviews: installed equipment, healthy equipment, schedulable capacity and productive workload time. Record the reason for each loss, including maintenance, network constraints and idle demand. Direct the next investment toward the largest recoverable constraint before assuming that more accelerators will improve output.
AI Infrastructure Financing
Nebius secures $3.5 billion financing facility for AI infrastructure expansion
23 October 2026 | United States
Nebius secured a US$3.5 billion senior secured financing facility arranged by Citi to support data centre development and GPU procurement. The one-year facility initially carries interest at Term SOFR plus 2.75 percentage points, with the spread potentially increasing to 3.50 percentage points. Nebius also referenced more than US$40 billion in long-term customer commitments as part of its broader financing strategy. The facility provides capital for infrastructure development but does not establish that individual data centres have been commissioned or that purchased GPUs are operational.
The one-year facility creates a timing mismatch with long-lived infrastructure assets. Track drawdowns, commissioning dates, cash generation and the route to refinancing; the potential spread increase makes delayed delivery more expensive. Long-term customer commitments are not equivalent to cash available for debt service. Distinguish committed facility size from drawn debt, and contracted customer value from collections. The main risk is the interval between paying for infrastructure and receiving customer cash; a strong order book can coexist with a material near-term funding requirement.
GreyRadius Insight
Match financing duration to the cash-generating milestone of each project. Stress-test delayed commissioning, lower utilisation and refinancing at higher rates, including required collateral and liquidity headroom. Prioritise phases where customer obligations and revenue commencement align with debt servicing rather than relying on the aggregate value of future commitments. Map monthly cash requirements through construction, GPU delivery and customer acceptance, including interest and contingency. Identify a refinancing route before maturity approaches and test a scenario where it is delayed. Link new expansion approvals to liquidity remaining after committed phases, rather than assuming future customer commitments can immediately fund current spending.
Grid Flexibility | Data Centre Energy
Google and Voltus announce 1GW virtual power plant arrangement
22 October 2026 | United States
Google expanded its partnership with Voltus to develop a 1GW virtual power plant across the PJM electricity market. The agreement builds on an earlier 100MW arrangement and is expected to run for approximately three years. Voltus aggregates flexible demand and distributed energy resources, including battery systems and controllable commercial electricity loads. The initiative could help manage grid constraints affecting data centre development. However, the 1GW figure represents a virtual power plant target, not newly generated electricity, installed data centre capacity or a confirmed grid connection.
Watch how much flexible capacity is enrolled, available and delivered during grid events. A 1GW virtual power plant target depends on aggregated resources and operating conditions; it should not be counted as new generation or a guaranteed data-centre connection. The system value depends on delivery at the time and location the grid needs support. Aggregated resources may be constrained during the same peak event, so nominal enrolled capacity can overstate dependable response. Watch availability testing and the rules governing settlement and non-performance.
GreyRadius Insight
Evaluate flexibility as a contracted operating capability with a baseline, response time, duration and verification method. Identify workloads that can pause or shift without breaching service obligations, then compare compensation with disruption and recovery costs. Any benefit to connection planning should be confirmed with the relevant utility or grid operator. Separate the value of grid-service revenue from any potential improvement in access to power. Test both cases independently, using contracted availability and verified response. Protect customer service levels through workload scheduling and reserve capacity; only include connection benefits once the utility accepts the relevant operating arrangement.
Data Centre Electrical Engineering | AI Infrastructure
Vantage signs agreement for more than 1GW of solid-state transformer equipment
21 October 2026 | United States
Vantage Data Centers signed an equipment supply agreement with Heron Power covering more than 1GW of solid-state transformer capacity for its hyperscale infrastructure portfolio. The equipment is designed to support high-density AI facilities using 800V DC electrical architectures. Heron's modular Link converter is rated at 5MW, with a manufacturer-stated efficiency target exceeding 98.5% from medium voltage to 800V DC. These figures describe equipment specifications and procurement commitments rather than commissioned data centre capacity or independently measured operating performance.
Track manufacturing delivery, electrical certification, installation and measured performance under representative loads. The agreement can improve equipment visibility, but the claimed conversion efficiency must be assessed across operating conditions and alongside redundancy and maintenance requirements. Electrical architecture choices can create dependencies extending beyond the converter itself. Watch compatibility with backup power, protection systems, downstream distribution and field service. A component efficiency advantage can be offset by added conversion stages or reduced availability elsewhere in the power chain.
GreyRadius Insight
Compare the complete electrical system rather than one component’s peak efficiency. Include partial-load losses, redundancy, spare parts, protection design and maintenance skills. Qualify equipment against the intended rack architecture and agree acceptance criteria before deployment; procurement scale should follow validated performance and a credible service plan. Evaluate lifetime energy savings at the expected load profile alongside downtime exposure and replacement lead times. Require a supported design for failure isolation and safe maintenance. Preserve staged deployment where feasible so operational evidence from the first installation informs later procurement, rather than making the entire portfolio dependent on unproven performance.
Data Centre Development | Industrial Redevelopment
TerraWatt proposes 1GW data centre redevelopment in Illinois
23 October 2026, reported | United States
TerraWatt proposed redeveloping a former steel mill site in Illinois into a large-scale data centre campus with potential capacity of up to 1GW. The development concept involves 200MW blocks, allowing capacity to be delivered in phases. Reusing an industrial location could offer land and infrastructure advantages, although the actual benefits depend on grid access, environmental requirements and site engineering. The project remains a proposal; the reported 1GW is an ultimate development ambition rather than permitted, financed, energised or operational capacity.
The first 200MW block will test whether the former industrial site offers usable infrastructure advantages. Watch environmental remediation, utility studies, planning permissions and anchor demand; historical industrial power access does not establish a new 1GW connection entitlement. Brownfield redevelopment can move costs into less visible categories such as remediation, demolition and utility reinforcement. Track which costs are fixed before customer commitment and which can be deferred. The credibility of the first block matters more to near-term supply than the ultimate campus headline.
GreyRadius Insight
Complete a site-specific assessment of remediation, grid entitlement and infrastructure reuse before pricing a brownfield advantage. Secure phased customer commitments and release development spending against permits and power milestones. A smaller deliverable phase can be commercially stronger than a larger campus concept whose utility and environmental dependencies remain unresolved. Compare the site with realistic alternatives on the cost and date of the first usable capacity. Carry explicit allowances for environmental findings and infrastructure that cannot be reused. Secure development rights in stages where possible, preserving the ability to stop or resize the project if utility or customer assumptions fail.
Cooling | Water Sustainability
Endeavour and Trevi propose desalination using data centre waste heat
23 October 2026, reported | United States
Endeavour and Trevi Systems announced a partnership to investigate using data centre waste heat to support desalination and water recovery. The proposed approach could produce up to 200,000 gallons of clean water per day per MW under specified operating assumptions. The technology is relevant to data centre sustainability because cooling and water availability can constrain facility development in water-stressed locations. However, the stated production figure is a technology potential rather than a verified, continuously measured commercial operating result. Project-specific energy consumption, water quality, operating costs and deployment schedules remain important validation requirements.
Validate the water-production claim against feedwater salinity, waste-heat temperature and the full system energy balance. A useful commercial result must include auxiliary electricity, water quality, brine disposal and availability when computing loads fluctuate. A viable heat-to-water arrangement needs stable demand for both computing and recovered water. Watch whether heat availability matches the desalination process and whether integration introduces a new dependency into data-centre cooling. Water production should remain a separate verified outcome from computing reliability.
GreyRadius Insight
Pilot the technology using local feedwater and actual waste heat, measuring cost per unit of compliant water and net energy use. Establish ownership of water infrastructure and responsibility for brine and downtime. Investment should follow repeatable operating evidence and a dependable water offtake arrangement, rather than the maximum production claim. Evaluate the combined facility and the water process as separate businesses before assessing integration benefits. Specify a safe bypass if the desalination equipment is unavailable. Base investment on delivered water that meets the buyer’s quality standard, with disposal, maintenance and auxiliary energy included in the net benefit.
Sovereign Cloud | Government Procurement
Switzerland advances sovereign-cloud procurement worth up to CHF7 billion
23 October 2026 | Switzerland
Switzerland's federal government published a major procurement tender for its Swiss Government Cloud programme, with a reported ceiling of CHF7 billion (approximately US$8.5 billion, as reported). The programme aims to modernise federal government cloud infrastructure through public and private cloud services, replacing ageing systems and strengthening government digital capabilities. The procurement creates a potentially significant opportunity for cloud infrastructure providers. However, the tender ceiling does not represent awarded contracts, committed capital expenditure or operational cloud capacity. Supplier selection, sovereignty requirements and implementation milestones remain outstanding.
Track tender lots, eligibility, sovereignty requirements and award schedules before estimating supplier revenue. The CHF7 billion ceiling establishes procurement scope, while actual demand depends on awards, workload migration and services consumed over the contract term. The opportunity will favour suppliers able to demonstrate compliance and execute migration, not simply offer infrastructure. Watch whether tender structure allows smaller specialist providers to participate through lots or partnerships, and whether funding ceilings translate into committed minimum consumption.
GreyRadius Insight
Cloud suppliers should qualify the opportunity through procurement eligibility, operating control, data access and support requirements before bidding. Build migration and legacy-system integration costs into the proposal. Government buyers should evaluate exit arrangements and portability alongside sovereignty, so a compliant hosting model does not create avoidable long-term switching costs. Build a bid-or-partner decision around capability gaps, qualification cost and probable contract scope. Define a commercially viable migration sequence and price the transition period when old and new systems run together. For buyers, test service portability and recovery using representative applications before accepting a long-term operating dependency.
AI Colocation | Data Centre Expansion
Crusoe commits to another 40MW AI data centre at Norway's Tørdal campus
22 October 2026 | Norway
Polar DC and Crusoe announced plans for a second AI data centre, DRA02, at the Tørdal campus in Drangedal, Norway. The proposed facility has a reported capacity of 40MW and is intended to support Crusoe's AI cloud infrastructure, including planned NVIDIA Vera Rubin NVL72 systems. The development builds on DRA01, where Crusoe already operates 12MW of IT capacity. The expansion strengthens the site's prospective AI infrastructure offering, although construction timing, grid readiness and accelerator installation remain unconfirmed. The 40MW proposed facility figure and existing 12MW IT load should not be treated as directly comparable without confirming their definitions.
Confirm the definition of the proposed 40MW and its power-delivery schedule before comparing it with the existing 12MW IT load. Tenant commitment reduces demand uncertainty, but construction, high-density cooling and accelerator commissioning remain distinct delivery risks. The existing campus provides useful operating evidence, but additional capacity can introduce different grid, construction and cooling constraints. Watch whether common infrastructure has spare headroom or requires upgrades that become the critical path for the new building.
GreyRadius Insight
Use the operating facility’s performance as evidence for expansion assumptions without assuming identical conditions. Contract for measurable power, cooling and service milestones at DRA02, and align accelerator delivery with tested site readiness. Keep financing and customer reservations phased to reduce the cost of equipment arriving before infrastructure is usable. Use separate readiness checks for the building, electrical supply, cooling and computing platform. Align rent commencement and equipment acceptance with the service each party actually controls. A shared campus should create measurable delivery or operating advantages; avoid paying for those advantages before the supporting infrastructure has been demonstrated.
AI Infrastructure Engineering | Modular Data Centres
NVIDIA introduces modular AI factories targeting installation in as little as 14 days
21 October 2026 | Germany / Europe
NVIDIA introduced its DSX POD modular AI factory architecture at GTC Berlin. The approach enables partners to integrate and validate AI computing systems inside factory-built modules before delivery to prepared data centre sites. NVIDIA claims the architecture could reduce on-site installation from approximately six months to as little as 14 days, with partner deployments targeted for early 2027. The platform supports NVIDIA Vera Rubin infrastructure and involves technology partners including Dell, HPE, Lenovo, Schneider Electric and Vertiv. NVIDIA also described potential efficiency improvements, including up to 30% less power for equivalent token output under specified assumptions. These remain design and prospective deployment claims rather than independently measured commercial results.
The 14-day claim concerns on-site installation at prepared locations, not the entire development timeline. Watch factory lead times, site preparation, utility readiness and customer acceptance; a faster final installation stage may leave the overall power bottleneck unchanged. Factory integration may transfer risk from the site to the manufacturing and logistics chain. Watch module availability, transport constraints, interface standards and fault resolution after delivery. A compressed installation window has limited commercial benefit if acceptance testing or missing utility infrastructure delays revenue.
GreyRadius Insight
Compare total time to service across design, procurement, factory integration, site works and commissioning. Modular construction may improve repeatability and reduce on-site work, but its value depends on readiness at the receiving site. Release orders against a coordinated delivery schedule and verify claimed power savings on the intended workload before including them in returns. Allocate ownership of every interface between module supplier, civil contractor, utility and operator. Compare the cost of schedule certainty with potential vendor dependence and future upgrade restrictions. Approve the modular approach where the full delivery plan, including factory slots and receiving-site readiness, improves the date and cost of usable service.
Grid Constraints | Hyperscale Development
Microsoft withdraws from delayed Nscale data centre in Essex
19 October 2026 | United Kingdom
Microsoft withdrew from Nscale's proposed data centre development in Loughton, Essex, where it had been expected to serve as the anchor tenant. The facility had previously targeted completion in late 2026, but grid access is now reportedly not expected until the 2030s. Nscale stated that it remains committed to the project and is considering alternative power arrangements and development locations. The withdrawal represents a significant commercial setback because the project loses an anticipated hyperscale customer while facing prolonged electricity-access uncertainty. Microsoft's earlier UK infrastructure investment announcements should not be counted as new commitments during this reporting period.
Track the tenant exit’s effect on financing, equipment commitments and alternative-site plans. A grid date slipping into the 2030s can invalidate the commercial timetable for a late-2026 target, even if the land and development rights retain long-term value. Loss of the anticipated anchor tenant can weaken the financing case even if the developer retains a technically feasible site. Watch whether replacement customers accept the revised schedule and whether alternative power can meet reliability, cost and permitting requirements without creating a different delay.
GreyRadius Insight
Make a credible power-delivery date an investment gate before committing to site-specific computing equipment. Align customer contracts and supplier deposits with utility milestones, and assess the cost of relocation or alternative generation. Treat grid schedule uncertainty as a commercial variable affecting revenue, financing and equipment obsolescence. Calculate the cost of waiting against relocation, redesign and orderly suspension. Include deposit recovery, holding costs and the changing value of equipment over time. Set explicit dates for reviewing the site decision and align stakeholder commitments with those dates, so sunk development spending does not determine the next capital allocation.
AI Infrastructure Energy | Grid Flexibility
Emerald AI and NVIDIA demonstrate AI workload power flexibility
23 October 2026, reported | Iceland
Emerald AI and NVIDIA reported results from a power-flexibility pilot in Iceland involving AI computing infrastructure. The test reportedly demonstrated an approximately 25% reduction in AI rack power consumption sustained for 30 minutes, illustrating the potential for dynamically managing AI workloads in response to electricity constraints. Such flexibility could help data centres participate in demand-response programmes or reduce peak grid demand. However, the reported result relates to a limited-duration pilot. It does not establish that the same reduction can be sustained continuously or that the approach increases grid connection capacity. Workload performance effects and commercial deployment conditions require further verification. Direct source URL was not established in the supplied research; source verification remains pending.
Source: Emerald AI / NVIDIA
Strategic Watch
Confirm the original pilot evidence and quantify effects on job completion, latency and subsequent power demand. A 25% rack-power reduction for 30 minutes is a limited flexibility result, not evidence of continuous energy savings or a larger grid allocation. Power reduction and energy reduction are different outcomes. Deferred work may return after an event and raise subsequent demand; commercial value depends on the complete operating cycle. Watch repeat events, recovery periods and whether flexibility is available when the grid requests it.
GreyRadius Insight
Define flexibility by workload class and service obligation. Test response speed, performance degradation, job recovery and the energy consumed after an event, then estimate the net commercial benefit. Expand participation only where delivered grid services exceed lost computing value and measurement demonstrates repeatable performance. Build a workload-level flexibility budget covering maximum delay, checkpointing cost and service impact. Compare grid compensation or avoided peak costs with lost computing output and recovery overhead. Base participation on net value across the event and rebound period, with separate limits for training, batch inference and latency-sensitive services.
Data Centre Development | Dedicated Power
Krolmap proposes €1.5 billion, 225MW off-grid data centre in Zaragoza
23 October 2026, reported | Spain
Krolmap proposed developing a large data centre project in Zaragoza, Spain, with an estimated investment of €1.5 billion and a planned scale of 225MW. The concept incorporates renewable electricity generation, energy storage and cogeneration to support an off-grid operating model. The development is relevant because access to grid power is increasingly affecting data centre siting and construction schedules. However, the project remains a proposal. Its financing, permitting, generation arrangements, storage duration and ability to maintain continuous power supply require validation. The proposed 225MW should not be treated as operational capacity.
Assess whether the off-grid design can maintain reliable supply during low renewable output, equipment outages and fuel interruptions. The proposed 225MW requires a credible hourly power balance, reserve margin, permitting route and operating-cost model before it supports a delivery promise. An off-grid model transfers electricity-system responsibilities to the project and its partners. Watch the redundancy of fuel, generation and storage, plus maintenance coverage and realistic outage assumptions. Independence from a grid queue can come with greater exposure to operating complexity and fuel economics.
GreyRadius Insight
Compare the dedicated-power concept with feasible grid-connected alternatives using full lifecycle cost and reliability. Model storage duration, fuel supply, maintenance outages, emissions and water needs. Phase the data-centre load against commissioned power assets and require independent engineering validation before committing to continuous-service guarantees. Require hourly and outage-scenario modelling before treating dedicated power as a schedule advantage. Separate the generation investment from the data-centre return so cross-subsidies are visible. Contract for fuel, maintenance and reserve capacity with the same discipline as the tenant service commitment; expand computing load only against proven dependable supply.
Data Centre Construction | AI Infrastructure
KT Cloud begins construction of 50MW AI data centre in Ansan
23 October 2026, reported | South Korea
KT Cloud reportedly commenced construction of an AI data centre in Ansan, South Korea. The facility is planned to have 50MW of capacity and approximately 36,880m² of floor area, with completion targeted for 2029. The construction milestone is more significant than an early-stage investment announcement because it indicates progress toward physical infrastructure delivery. Nevertheless, the project still requires construction completion, power energisation, computing equipment installation and commissioning before it can provide usable capacity. The precise definition of the 50MW capacity figure and the proportion allocated to AI workloads require confirmation. Direct source URL was not established in the supplied research; source verification remains pending.
Source: DatacenterDynamics
Strategic Watch
Confirm the original construction report, power-capacity definition and scope of work against the 2029 target. Construction start reduces one form of development uncertainty, while grid energisation, cooling readiness and customer contracting still determine service commencement. A 2029 delivery target exposes the project to changes in computing density, cooling specifications and customer demand during construction. Watch whether the design preserves flexibility without paying for unnecessary capacity too early, and whether prospective tenants are committing to a defined delivery configuration.
GreyRadius Insight
For tenants, align reservations with explicit delivery milestones and confirm who bears construction and power-delay risk. For developers, test financing and demand assumptions across the full construction period. Equipment selection should remain coordinated with the expected service date so a 2029 building is not locked prematurely into an unsuitable computing configuration. Protect long-lead infrastructure decisions with scenarios for rack density and cooling demand. Defer replaceable computing choices where possible while securing genuinely critical equipment. Review tenant requirements at agreed design stages, and price changes explicitly so late specification shifts do not erode project returns or destabilise commissioning.
Liquid Cooling | High-Density Infrastructure
ZTE introduces liquid-cooled AI data centre module supporting up to 250kW per rack
23 October 2026, reported | China
ZTE introduced a modular data centre design intended to support high-density AI computing workloads. The system reportedly supports rack power densities of up to 250kW, using liquid-cooling infrastructure to manage the thermal demands of advanced accelerator systems. The announcement is relevant to operators planning AI facilities because rack density affects cooling architecture, electrical distribution and floor-space requirements. However, the 250kW figure is a stated design capability rather than independently verified sustained operating performance. The announcement does not establish that a commercial facility has deployed the system at that density.
Validate 250kW rack support against sustained thermal loads and the required site infrastructure. Cooling-fluid temperatures, pumping power, redundancy and maintenance access can affect whether the design rating is usable in a commercial deployment. Higher rack density can concentrate the operational impact of a cooling or electrical failure. Watch redundancy boundaries and what happens to adjacent systems during maintenance or an outage. The most useful evidence is sustained accepted computing output with realistic failures and service operations included.
GreyRadius Insight
Assess compatibility at facility level: power distribution, heat rejection, coolant quality, leak detection and field maintenance must support the rack design. Require sustained-load and failure-mode testing before relying on the density claim. Compare usable computing output per site constraint, including energy and cooling overhead, rather than maximising rack density alone. Compare designs using usable output per constrained megawatt and the cost of maintaining the target service level. Run failure and maintenance tests at the intended density, including coolant loss and component replacement. Expansion should depend on demonstrated thermal stability and recovery procedures, with trained staff and spares available before customer workloads begin.
Data Centre Financing | Emerging AI Infrastructure Markets
Akashi secures $31.2 million financing for phased AI data centre development
23 October 2026, reported | Kazakhstan
Akashi reportedly secured US$31.2 million in development-bank financing for a planned data centre project in Kazakhstan. The development targets an initial 10MW phase in 2027, with an ultimate campus ambition of 100MW. The financing represents a potential step toward project execution, but the reported funding amount should not be interpreted as sufficient to finance the full proposed campus. Construction, grid access, equipment procurement and commercial demand remain subject to verification. The distinction between the initial 10MW development and ultimate 100MW target is important when assessing delivery risk and investment requirements. Direct source URL was not established in the supplied research; source verification remains pending.
Source: Development-bank financing announcement / industry reporting, 23 October 2026. **Direct source hyperlink not independently established; verification pending.**
Strategic Watch
Confirm financing terms, disbursement conditions and what the US$31.2 million covers. The initial 10MW phase and ultimate 100MW ambition represent different funding requirements; the full campus should remain outside committed-supply estimates until separately supported. Early financing can help establish execution credibility, but subsequent capital may depend on construction progress, customer contracts or sponsor contributions. Watch the conditions attached to each disbursement and distinguish financing secured from funds available for immediate spending.
GreyRadius Insight
Underwrite the initial phase as a standalone project with a funding plan, power agreement and credible customer demand. Clarify whether financing covers construction, equipment or only part of development. Reserve the larger campus for staged options; release subsequent capital only after the first phase demonstrates delivery and commercial acceptance. Prepare a complete sources-and-uses plan for the first phase, including contingency, connectivity and commissioning. Identify the funding gap under a delayed tenant ramp and confirm who must cover it. Preserve expansion land and options economically, while excluding later phases from the base return until their funding and demand are independently supported.
GPU Cloud Operations | Infrastructure Reliability
Oracle introduces GPU cluster-health validation for AI cloud workloads
22 October 2026 | United States / Global
Oracle Cloud Infrastructure introduced Cluster Health Checks to help customers assess GPU node health and validate computing clusters before running AI and high-performance computing workloads. The service includes passive checks covering GPU, driver and networking conditions, alongside optional active tests using Slurm or Kubernetes. Supported Oracle Kubernetes Engine environments can also use GPU host-repair functionality. The release improves infrastructure diagnostics and may help customers identify operational issues before expensive workloads begin. However, it does not represent additional physical GPU capacity, a new cloud region or a disclosed improvement in utilisation.
Track whether health checks identify faults early enough to reduce failed jobs and wasted GPU-hours. Diagnostic coverage is useful only when operations teams can remediate findings promptly; check supported configurations, test overhead and repair responsibilities. A cluster can pass a point-in-time test and still fail during demanding distributed workloads. Watch diagnostic recurrence, false positives and whether problems sit in GPUs, networking or configuration. Operational improvement should be assessed against failures during real jobs, not simply the number of checks run.
GreyRadius Insight
Integrate checks into cluster handover and pre-job procedures, with clear thresholds for quarantining unhealthy nodes. Measure failed-job frequency, time to repair and recovered productive GPU-hours against test overhead. Prioritise fixes that improve accepted workload completion; additional diagnostics create business value when they change operational decisions. Assign one team ownership of fault triage and define when a node returns to service. Analyse job failure and repair data by hardware and software configuration, then use the findings in capacity scheduling and supplier discussions. Calculate the value of recovered GPU-hours after test time and reserved spare capacity, making reliability part of procurement economics.
Executive Dashboard
Signal
Key Data
Date
Region
Executive Implication
STTGDC and Neysa plan deployment of up to 9,000 GPUs in Noida
Up to 9,000 GPUs | Noida, India | Target Q1 2027.
23 October 2026
India
Track the path from hosting agreement to installed accelerators, tested cooling and commercially available service. The Q1 2027 target depends on equipment delivery and commissioning; up to 9,000 GPUs describes a prospective deployment ceiling, not bookable supply.
Vultr announces 20MW AMD AI compute cluster in Chennai
20MW announced cluster scale | AMD MI355X | Chennai, India.
21 October 2026
India
Confirm whether 20MW refers to total facility power or IT load, then track energisation, accelerator deliveries and service launch. The commercial test is reliable workload performance and customer uptake on the AMD platform; announced power alone does not establish computing throughput.
Nebius commits to up to 50MW at Saudi Arabia's Oxagon AI data centre
Up to 50MW customer allocation | 100MW under construction | 360MW first-phase target | 1.5GW ultimate ambition | Q4 2027 availability target.
22 October 2026
GCC — Saudi Arabia
Track delivery of the customer-backed allocation separately from the 100MW under construction and larger campus ambitions. The critical dependencies are power availability, phased acceptance, equipment delivery and when the agreement begins generating revenue.
Aleria outlines NVIDIA GPU infrastructure expansion in UAE and United States
8,640 planned Blackwell Ultra GPUs | 16,000-GPU target | 28 planned Vera Rubin NVL72 racks in UAE.
23 October 2026, reported
GCC — United Arab Emirates / United States
Confirm the original announcement and delivery timetable before treating the expansion as available supply. US GPU quantities and UAE rack quantities describe different geographies, units and technology generations; adding them would obscure readiness and procurement risk.
Meta's Prometheus AI campus reaches first operational milestone in Ohio
First phase operational | Hundreds of thousands of GPUs reportedly installed | Approximately 1 million GPUs targeted at full scale.
23 October 2026
United States
The first operational phase is a stronger delivery signal than a campus proposal, but installed GPU counts still reveal little about productive utilisation. Watch workload availability, failure rates, networking performance and the relationship between delivered computing output and energy use.
Nebius secures $3.5 billion financing facility for AI infrastructure expansion
US$3.5 billion financing | One-year maturity | SOFR + 2.75 percentage points initially | More than US$40 billion in referenced customer commitments.
23 October 2026
United States
The one-year facility creates a timing mismatch with long-lived infrastructure assets. Track drawdowns, commissioning dates, cash generation and the route to refinancing; the potential spread increase makes delayed delivery more expensive. Long-term customer commitments are not equivalent to cash available for debt service.
Google and Voltus announce 1GW virtual power plant arrangement
1GW virtual power plant target | Previous 100MW agreement | Approximately three-year duration | PJM market.
22 October 2026
United States
Watch how much flexible capacity is enrolled, available and delivered during grid events. A 1GW virtual power plant target depends on aggregated resources and operating conditions; it should not be counted as new generation or a guaranteed data-centre connection.
Vantage signs agreement for more than 1GW of solid-state transformer equipment
More than 1GW equipment agreement | 5MW modules | 800V DC architecture | Greater than 98.5% claimed efficiency.
21 October 2026
United States
Track manufacturing delivery, electrical certification, installation and measured performance under representative loads. The agreement can improve equipment visibility, but the claimed conversion efficiency must be assessed across operating conditions and alongside redundancy and maintenance requirements.
TerraWatt proposes 1GW data centre redevelopment in Illinois
Up to 1GW proposed campus capacity | 200MW development blocks | Illinois, United States.
23 October 2026, reported
United States
The first 200MW block will test whether the former industrial site offers usable infrastructure advantages. Watch environmental remediation, utility studies, planning permissions and anchor demand; historical industrial power access does not establish a new 1GW connection entitlement.
Endeavour and Trevi propose desalination using data centre waste heat
Up to 200,000 gallons/day/MW claimed potential | Waste-heat utilisation | Desalination application.
23 October 2026, reported
United States
Validate the water-production claim against feedwater salinity, waste-heat temperature and the full system energy balance. A useful commercial result must include auxiliary electricity, water quality, brine disposal and availability when computing loads fluctuate.
Switzerland advances sovereign-cloud procurement worth up to CHF7 billion
CHF7 billion procurement ceiling | Reported equivalent of US$8.5 billion | Swiss Government Cloud programme.
23 October 2026
Switzerland
Track tender lots, eligibility, sovereignty requirements and award schedules before estimating supplier revenue. The CHF7 billion ceiling establishes procurement scope, while actual demand depends on awards, workload migration and services consumed over the contract term.
Crusoe commits to another 40MW AI data centre at Norway's Tørdal campus
40MW proposed DRA02 | 12MW existing operational IT capacity | NVIDIA Vera Rubin NVL72 planned.
22 October 2026
Norway
Confirm the definition of the proposed 40MW and its power-delivery schedule before comparing it with the existing 12MW IT load. Tenant commitment reduces demand uncertainty, but construction, high-density cooling and accelerator commissioning remain distinct delivery risks.
NVIDIA introduces modular AI factories targeting installation in as little as 14 days
As little as 14 days claimed installation | Approximately six-month comparison | Early 2027 deployment target | Up to 30% claimed power reduction under specified conditions.
21 October 2026
Germany / Europe
The 14-day claim concerns on-site installation at prepared locations, not the entire development timeline. Watch factory lead times, site preparation, utility readiness and customer acceptance; a faster final installation stage may leave the overall power bottleneck unchanged.
Microsoft withdraws from delayed Nscale data centre in Essex
Original completion target: late 2026 | Grid connection reportedly delayed until the 2030s | Microsoft anchor tenancy withdrawn.
19 October 2026
United Kingdom
Track the tenant exit’s effect on financing, equipment commitments and alternative-site plans. A grid date slipping into the 2030s can invalidate the commercial timetable for a late-2026 target, even if the land and development rights retain long-term value.
Emerald AI and NVIDIA demonstrate AI workload power flexibility
Approximately 25% reported rack-power reduction | 30-minute test duration | Iceland pilot.
23 October 2026, reported
Iceland
Confirm the original pilot evidence and quantify effects on job completion, latency and subsequent power demand. A 25% rack-power reduction for 30 minutes is a limited flexibility result, not evidence of continuous energy savings or a larger grid allocation.
Krolmap proposes €1.5 billion, 225MW off-grid data centre in Zaragoza
Assess whether the off-grid design can maintain reliable supply during low renewable output, equipment outages and fuel interruptions. The proposed 225MW requires a credible hourly power balance, reserve margin, permitting route and operating-cost model before it supports a delivery promise.
KT Cloud begins construction of 50MW AI data centre in Ansan
50MW planned capacity | 36,880m² floor area | 2029 completion target | Ansan, South Korea.
23 October 2026, reported
South Korea
Confirm the original construction report, power-capacity definition and scope of work against the 2029 target. Construction start reduces one form of development uncertainty, while grid energisation, cooling readiness and customer contracting still determine service commencement.
ZTE introduces liquid-cooled AI data centre module supporting up to 250kW per rack
Up to 250kW claimed rack support | Liquid-cooling architecture | Modular AI data centre design.
23 October 2026, reported
China
Validate 250kW rack support against sustained thermal loads and the required site infrastructure. Cooling-fluid temperatures, pumping power, redundancy and maintenance access can affect whether the design rating is usable in a commercial deployment.
Akashi secures $31.2 million financing for phased AI data centre development
Confirm financing terms, disbursement conditions and what the US$31.2 million covers. The initial 10MW phase and ultimate 100MW ambition represent different funding requirements; the full campus should remain outside committed-supply estimates until separately supported.
Oracle introduces GPU cluster-health validation for AI cloud workloads
Released 22 October 2026 | Passive and active cluster validation | Slurm and Kubernetes support | No additional GPU capacity disclosed.
22 October 2026
United States / Global
Track whether health checks identify faults early enough to reduce failed jobs and wasted GPU-hours. Diagnostic coverage is useful only when operations teams can remediate findings promptly; check supported configurations, test overhead and repair responsibilities.
Turn Enterprise AI Signals Into Execution
GreyRadius helps leadership teams translate AI adoption, infrastructure, sovereignty and market shifts into investment priorities, GTM choices and execution-ready growth plans.
