Enterprise AI • Compute • Sovereign AI • Deployment Infrastructure
Published by GreyRadius Consulting | 14–20 September 2026
Executive Highlights
CanadaBell and Saskatchewan outlined up to 900MW of additional AI capacity, creating a pathway to a 1.2GW sovereign AI hub. The wider development could exceed C$50bn, but the new capacity remains conditional on customers, permits and generation agreements.
TürkiyeDigital Realty and Rönesans launched a Türkiye JV with a power-secured Ankara campus providing more than 22MW of carrier-neutral IT capacity. Land, electricity and permits are secured, construction has begun, and completion is targeted for 2028.
United StatesCrusoe opened a second 400,000 sq ft Tulsa factory, taking its US manufacturing footprint to about 1 million sq ft. The expansion internalises production of switchgear, busbars and controls that can determine how quickly AI campuses are energised.
United StatesOracle clarified that Project Jupiter’s data-centre construction is separate from the proposed microgrid facing a stayed air-permit proceeding. Buildings can continue under county permits, while the long-term power pathway remains exposed to regulatory timing.
JapanRebellions and ai& plan to scale locally deployed AI inference infrastructure to 100+ RebelRack systems in Tokyo. ai& has more than US$2bn of committed infrastructure capital and targets five sites by end-2026 and 40MW by end-2027.
North AmericaServerfarm increased its revolving credit facility by US$895m to US$3.895bn, backed by 22 banks and one institutional lender. The liquidity supports multiple AI-ready projects, but project-level power, permits and delivery dates remain the key deployability tests.
BrazilBrazil enacted ReData, a federal incentive regime covering AI training, inference, cloud and HPC data centres. The government estimates roughly R$5.2bn of 2026 fiscal incentives, tied to R&D, sustainability and domestic-market conditions.
United StatesDelos Data raised more than US$100m to develop its Nonstop AI interconnect architecture across chiplets, optics and accelerator connectivity. It claims 10× lower latency and 10× higher efficiency as inference demand is projected by the company to rise 300× by 2030.
United StatesMicrosoft filed for a 123-acre, two-building Virginia data-centre campus with an on-site substation and supporting utility infrastructure. No IT-load MW or operating date was disclosed, so the project remains permitting-stage rather than customer-available capacity.
AustraliaAnthropic signed its first Australian data-centre agreement for inference capacity at Zerra DC’s proposed Western Downs campus. The wider project is planned at 2.16GW with initial operations targeted for 2027, subject to regulatory approvals.
Globalonsemi introduced an Embedded Power Platform that it says can deliver 3–5× higher power density than current solutions. The architecture targets AI data centres as rack designs move toward higher-voltage distribution and megawatt-class power requirements.
GlobalGoogle, NVIDIA and Emerald AI launched the AI Energy Management Alliance to make AI data centres responsive to grid conditions. The model could help unlock faster or larger interconnections by turning flexible workloads into controllable electrical demand.
United States / GlobalCrusoe closed an initial US$3.9bn Series F at a US$30.9bn post-money valuation and disclosed more than US$140bn in contracted value. The capital will expand AI campuses, modular deployments and cloud capacity across its vertically integrated infrastructure model.
United StatesLoudoun County advanced a proposal to pause some data-centre and substation applications requiring legislative approval for up to 12 months. A formal resolution is expected in October, adding local policy risk to Northern Virginia’s existing power and land constraints.
United StatesAbout US$18bn of debt linked to Oracle’s Project Jupiter was reported trading near 89–91 cents on the dollar. The discount provides a market signal that power, permitting and execution uncertainty can directly affect financing economics for very large AI campuses.
United StatesVirginia introduced a statewide Data Center Accountability Framework through Executive Order 22. The framework addresses approvals, electricity costs, environmental standards, clean energy and workforce issues, potentially reshaping development requirements in a core hyperscale market.
TaiwanZettabyte and Bowrington Capital agreed to assess high-density GPU infrastructure and data-centre development in Taiwan. The concept includes potential investment at both site and GPU-asset level, but no MW, GPU count, capital commitment or delivery date has been disclosed.
North AmericaNERC said its foundational Computational Loads Reliability Standards passed an initial ballot, subject to validation. The standards respond to the grid impact of rapidly growing data-centre and other computational loads and could formalise new operating expectations for large facilities.
GlobalMarvell unveiled 2nm optical-interconnect technologies including 400G/lane PAM4, 1.6T coherent options and a 102.4T co-packaged-optics platform. The portfolio targets AI fabrics scaling from 1.6T toward 3.2T and beyond, where bandwidth density and power per bit become critical constraints.
Filter
Data centres / Power / Sovereign AI
Bell Creates Path to a 1.2GW Sovereign AI Infrastructure Hub in Saskatchewan
14 September 2026 | Canada
Bell Canada and the Government of Saskatchewan signed a non-binding MoU for phased development of up to 900MW of additional AI infrastructure capacity, creating a pathway to a 1.2GW Bell AI Fabric hub in Saskatchewan. The incremental capacity is expected to use partner-developed natural-gas generation rather than provincial grid power, while facilities are planned around closed-loop cooling that requires no municipal water. Development remains conditional on customer commitments, commercial agreements, permits, approvals and environmental assessments, so the 900MW is planned rather than deployable capacity. The Saskatchewan government estimates the wider development could represent more than C$50 billion in investment.
The key conversion milestones are anchor-customer commitments, generation agreements, environmental approvals and phased energisation. Executives should track contracted and permitted MW—not the 1.2GW headline—because dedicated generation reduces grid dependence but introduces fuel, permitting and execution dependencies of its own.
GreyRadius Insight
Bell illustrates a broader shift from “find grid capacity” to “build an integrated energy-and-compute system.” In power-constrained markets, competitive advantage will increasingly sit with developers that can combine generation, cooling, permits and anchor demand into one bankable package; announced MW without those layers will carry a widening credibility discount.
Data centres / Cloud
Digital Realty Enters Türkiye With Power-Secured 22MW Ankara Campus
14 September 2026 | Türkiye
Digital Realty and Rönesans Infrastructure formed a joint venture to develop data-centre infrastructure in Türkiye, beginning with an Ankara campus providing more than 22MW of carrier-neutral IT capacity. Land, electricity and permits have reportedly been secured and early construction has begun, making this a more advanced project than a typical announced campus. Completion is targeted for 2028. The partners also intend to pursue capacity in Istanbul. The Ankara project expands Digital Realty's PlatformDIGITAL footprint into a market positioned between Europe, the Middle East and Asia.
Power, land and permits are already secured, shifting the risk from site formation toward construction, customer ramp and ecosystem creation. Watch whether Ankara attracts cloud on-ramps, carriers and anchor tenants quickly enough to justify expansion into Istanbul and establish Türkiye as a regional interconnection node.
GreyRadius Insight
Türkiye’s strategic value is not simply another 22MW of capacity. A carrier-neutral campus between Europe, the Gulf and Asia can become a regional digital gateway if it accumulates network density and cloud connectivity. The investment case therefore depends on ecosystem depth and latency routes as much as on real estate and power.
Hardware / Supply chain / Power
Crusoe Adds 400,000 Sq Ft of US Manufacturing for AI Electrical Infrastructure
14 September 2026 | United States
Crusoe opened its second manufacturing facility in Tulsa, Oklahoma, adding 400,000 sq ft and taking its US manufacturing footprint to roughly 1 million sq ft. The plants manufacture medium-voltage switchgear, low-voltage switchboards, electrical enclosures, industrial controls and fabricated copper busbars—the electrical equipment required between utility supply and energised AI racks. Crusoe employs more than 300 manufacturing workers across its two Tulsa facilities, versus 100 initially expected when its first plant opened in 2024. The expansion represents vertical integration into equipment whose availability can determine data-centre energisation schedules.
Track whether internal production measurably compresses switchgear and electrical-equipment lead times across Crusoe’s campus pipeline. The strategic KPI is not factory square footage but months removed from energisation schedules and the percentage of critical electrical BOM controlled internally.
GreyRadius Insight
AI infrastructure bottlenecks are moving upstream into transformers, switchgear, busbars and controls. Crusoe’s manufacturing expansion shows why developers are internalising schedule-critical equipment: control of the electrical supply chain can become as important as GPU access when the commercial objective is time-to-energisation.
Data centres / Power / Permitting
Oracle Clarifies Construction Status of Project Jupiter in New Mexico
14 September 2026 | United States
Oracle issued a material status clarification on Project Jupiter in Doña Ana County, New Mexico. Oracle said the data-centre campus and proposed microgrid are separate facilities on adjacent sites operated by separate companies. Construction already underway relates to the data-centre campus and is proceeding under county permits; Oracle said the buildings do not require the air-quality permit involved in litigation. The proposed microgrid remains subject to a separate regulatory process whose air-permit proceeding was stayed by the New Mexico Supreme Court. The distinction is important because a regulatory obstacle affecting proposed generation does not currently stop construction of the data-centre buildings, although it could affect the eventual power pathway.
Treat Project Jupiter as two linked but differently permitted assets: buildings and generation. Watch for an interim or alternative power pathway, because continued building construction can create stranded or delayed capital if the microgrid cannot reach approval on the timetable required for compute deployment.
GreyRadius Insight
Project Jupiter demonstrates why “under construction” and “deployable AI capacity” are no longer equivalent. As campuses pair with bespoke generation, executives need a multi-clock readiness model covering buildings, interconnection, generation, environmental approvals and compute installation; the slowest layer determines revenue start.
Compute / Sovereign AI
Rebellions and ai& Plan Up to 100+ AI Inference Racks in Japan
15 September 2026 | Japan
Rebellions and ai& agreed to deploy RebelRack inference infrastructure in ai&'s Tokyo data centre. ai& will begin with an initial purchase and intends to scale to up to 100 or more RebelRack systems. The deployment forms part of ai&'s broader Japanese data-centre expansion, backed by more than US$2 billion of committed infrastructure capital. ai& plans five operational sites by the end of 2026 and targets 40MW of capacity by the end of 2027. The locally deployed inference infrastructure is aimed at enterprises, government institutions and developers and supports Japan's sovereign-AI objectives.
The important milestones are the initial purchase, utilisation, expansion toward 100+ racks and ai&’s progress to 40MW. Watch workload mix and customer adoption: sovereign inference becomes economically meaningful only when local compute is matched with recurring enterprise and government demand.
GreyRadius Insight
Sovereign AI is evolving from data-residency policy into physical control of inference capacity. Japan’s model suggests a market for locally hosted accelerators serving enterprises and government, but sovereignty alone will not sustain economics; utilisation, software compatibility and workload migration will determine whether local infrastructure scales.
Capital / Data centres
Serverfarm Expands Data-Centre Credit Facility to US$3.895 Billion
15 September 2026 | North America
Serverfarm closed an US$895 million increase to its revolving credit facility, bringing total commitments to US$3.895 billion. The facility is supported by 22 banks and one institutional lender and is intended to fund development and construction of multiple large-scale North American data-centre projects. Serverfarm is developing AI-ready, high-density infrastructure for hyperscale and technology customers. The financing materially increases development liquidity, although corporate financing alone does not establish that every project in Serverfarm's pipeline has secured power, permits or a delivery date.
Follow where the additional US$895m is actually deployed and whether named projects have utility capacity, permits and anchor customers. Credit availability improves development optionality, but lenders and customers should distinguish corporate liquidity from sites capable of reaching commercial operation.
GreyRadius Insight
The financing market is rewarding platforms with large pipelines, but capital is increasingly abundant relative to immediately deployable power. The strategic scarcity is shifting from money to power-secured, permitted sites. Developers that can prove energisation dates should command stronger customer and financing terms than those presenting only funded pipelines.
Regulation / Data centres / Sovereign infrastructure
Brazil Establishes ReData Incentive Regime for AI, Cloud and HPC Data Centres
15 September 2026 | Brazil
Brazil enacted the ReData — Special Taxation Regime for Data Center Services, establishing federal incentives for qualifying data-centre investment. Eligible infrastructure encompasses cloud computing, high-performance computing, AI training and AI inference, with tax benefits covering qualifying ICT equipment while imposing R&D, sustainability and domestic-market conditions. Brazilian government reporting places the estimated fiscal incentive for 2026 at approximately R$5.2 billion. The programme connects data-centre investment with Brazil's digital-sovereignty and domestic technology-development objectives.
Implementation rules will determine whether ReData changes project economics enough to redirect investment toward Brazil. Watch equipment eligibility, local R&D obligations, sustainability requirements and announced MW, because incentive value can be diluted by power constraints, import costs or compliance burdens.
GreyRadius Insight
ReData shows governments are using fiscal policy to shape where AI infrastructure lands and what domestic value it creates. Site selection is therefore becoming a policy-arbitrage exercise across tax incentives, energy, equipment imports, R&D obligations and sovereignty requirements—not simply a comparison of land and electricity prices.
Networking / Hardware
Delos Data Raises More Than US$100 Million for AI Interconnect Technology
15 September 2026 | United States
Delos Data raised more than US$100 million to launch and develop its Nonstop AI architecture, targeting the data-movement bottleneck inside increasingly large AI systems. The company says its Nonstop AI Data Interface can provide 10× lower latency and 10× higher efficiency and is being developed across I/O chiplets, near-packaged optics and cards connecting GPUs, XPUs, CPUs, flash and memory. Delos says the financing will accelerate hardware and software development against an anticipated 300× increase in AI inference demand by 2030. The performance figures are company claims rather than independently benchmarked deployment results.
The next proof points are independent benchmarks, silicon or module milestones and design wins with hyperscalers or accelerator platforms. Interconnect claims matter only if latency and efficiency gains persist at production scale and translate into lower system-level power and higher accelerator utilisation.
GreyRadius Insight
As accelerator counts rise, the economic unit of AI infrastructure is shifting from the chip toward the entire data-movement system. Technologies that reduce I/O latency and energy can raise effective GPU utilisation, meaning interconnect efficiency may unlock more usable AI capacity without proportional increases in accelerator count or power.
Data centres / Cloud
Microsoft Files for 123-Acre Two-Building Virginia Data-Centre Campus
15 September 2026 | United States
Microsoft filed for a Department of the Army permit covering its proposed MNZ07 Wellington Data Center in Prince William County, Virginia. The project occupies approximately 123 acres and comprises two one-storey data-centre buildings, an on-site electrical substation, water-storage infrastructure, sanitary and industrial wastewater systems, fibre and electrical conduit and other supporting infrastructure. The permit notice was published on 15 September with a comment period expiring 15 October 2026. No IT-load MW figure or operational date was disclosed, meaning the project remains permitting-stage capacity rather than operational or customer-available capacity.
The permit process, utility arrangement and eventual IT-load disclosure are the critical next signals. A 123-acre site should not be counted as supply until grid capacity, construction phasing and commissioning dates are visible; acreage measures development scope, not usable compute.
GreyRadius Insight
The Virginia filing reinforces a key capacity-planning discipline: separate land bank, permitted pipeline, power-secured pipeline and commissioned capacity. Treating all announced campuses as equivalent can materially overstate market supply and understate the value of sites with confirmed energisation windows.
Compute / Data centres
Anthropic Signs First Australian Data-Centre Agreement at Proposed 2.16GW Campus
16 September 2026 | Australia
Anthropic signed its first Australian data-centre agreement with Zerra DC for capacity at the proposed Western Downs Digital Park in Queensland, roughly 250km from Brisbane. The wider campus is planned at 2.16GW, with initial operations targeted for 2027. Anthropic's capacity will be used for AI inference rather than training. The project is designed around renewable electricity and a closed-loop, air-cooled architecture intended to reduce water consumption. The development remains subject to regulatory approval, including Australia's Foreign Investment Review Board, so the full 2.16GW represents proposed campus scale rather than Anthropic's committed or operational load.
Decompose the 2.16GW campus headline into Anthropic-contracted capacity, permitted phases, secured power and 2027 commissioning milestones. Regulatory approval and the size of the anchor commitment will determine whether the agreement materially de-risks financing for the wider campus.
GreyRadius Insight
Anthropic’s agreement shows how anchor tenants can become infrastructure-financing instruments. A credible AI customer can de-risk a campus, but multi-gigawatt project headlines should be analysed in layers: customer-contracted MW, financed MW, permitted MW, energised MW and ultimately utilised MW.
Power / Hardware
onsemi Launches Power Architecture Claiming 3–5× Higher Power Density
16 September 2026 | Global
onsemi unveiled its Embedded Power Platform, an integrated architecture that uses the silicon wafer itself as the foundation of the package and jointly optimises electrical, mechanical and thermal characteristics. onsemi says the platform can provide 3–5× higher power density than current solutions and is targeting AI data centres alongside automotive and industrial applications. The technology is relevant as AI infrastructure migrates toward increasingly high-voltage distribution and megawatt-class rack architectures, where conversion efficiency and physical power density directly affect deployable compute. The 3–5× figure is a vendor performance claim rather than an independently validated production deployment result.
Watch qualification with data-centre power architectures, production readiness and real efficiency at rack scale. Higher package-level density is valuable only if it reduces conversion losses, thermal burden and footprint across the full power-delivery chain from medium voltage to accelerator.
GreyRadius Insight
At megawatt-class rack densities, power electronics become part of compute economics. Better conversion density can reduce electrical footprint and thermal overhead, allowing more accelerators inside the same facility envelope. The value should therefore be measured in incremental usable compute per MW, not component performance alone.
Power / Grid / Data centres
Google, NVIDIA and Emerald AI Launch Flexible AI Data-Centre Alliance
16 September 2026 | Global
Emerald AI, Google and NVIDIA launched the AI Energy Management Alliance (AEMA) to develop data centres capable of dynamically modifying electricity consumption in response to grid conditions. The initiative brings together technology companies and participants across generation, utilities and storage. Its goal is to enable faster and potentially larger grid connections for AI infrastructure by allowing workloads to act as flexible electrical demand rather than fixed 24/7 load. No new physical MW was announced, but the programme is strategically important because it introduces workload scheduling and demand flexibility as potential tools for overcoming AI data-centre interconnection constraints.
The decisive evidence will be utility pilots showing that workload flexibility changes interconnection outcomes. Track measurable MW of controllable load, response times and operating rules; if utilities recognise flexible compute in planning, software-defined demand could become a new path to faster power access.
GreyRadius Insight
Flexible compute could create a new infrastructure asset class: interruptible or schedulable AI load. If workloads can respond to grid constraints without damaging service levels, operators may trade some utilisation flexibility for earlier interconnection, lower system costs or access to power that fixed-load campuses cannot obtain.
Capital / Compute / Data centres
Crusoe Raises US$3.9 Billion to Scale Vertically Integrated AI Infrastructure
17 September 2026 | United States / Global
Crusoe announced the initial closing of a US$3.9 billion Series F at a US$30.9 billion post-money valuation. Investors include Atreides Management, Mubadala Capital, Valor Equity Partners, Founders Fund, GIC, NVIDIA, Qatar Investment Authority and TPG. Crusoe also disclosed more than US$140 billion in total contracted value across its vertically integrated platform. The capital will support large AI-factory campuses, modular Crusoe Spark deployments and expansion of Crusoe Cloud. Combined with the company's Tulsa manufacturing expansion during the same reporting week, the raise reinforces Crusoe's strategy of controlling infrastructure across energy, electrical systems, data centres and compute.
Track capital deployment against the disclosed US$140bn contracted value: campuses commissioned, cloud capacity activated and customer obligations met. Vertical integration can shorten schedules, but it also concentrates execution risk across energy, manufacturing, construction and compute within one platform.
GreyRadius Insight
Crusoe’s model tests whether vertical integration can turn infrastructure complexity into schedule advantage. Owning more of the stack may reduce supplier dependencies and capture margin across layers, but it also requires exceptional capital discipline; the winning metric will be contracted revenue converted into operational capacity faster than peers.
Regulation / Data centres
Loudoun County Advances Up-to-12-Month Pause on Some Data-Centre Applications
17 September 2026 | United States
Loudoun County moved toward pausing action for up to 12 months on data-centre and electrical-substation applications requiring legislative approval while the county reviews its planning and zoning framework. The proposal is particularly significant because Loudoun is one of the world's largest concentrations of data-centre infrastructure. The measure is not a universal shutdown of data-centre development: projects proceeding under different approval routes may not be affected. A formal resolution was scheduled for consideration in October 2026, adding another regulatory variable to power and land constraints facing future Northern Virginia capacity.
The October resolution should be read at approval-route level, not as a blanket market shutdown. Developers need to map which pipeline projects require legislative action, which can proceed administratively and whether the pause shifts demand—and grid pressure—to neighbouring Northern Virginia jurisdictions.
GreyRadius Insight
Loudoun’s move signals that mature data-centre hubs can become constrained by political carrying capacity as well as electricity. Future site strategy will need a “social and regulatory headroom” assessment alongside MW availability, because community acceptance and zoning pathways can become binding constraints even where capital and demand remain strong.
Capital / Data centres / Power
Oracle's US$18 Billion Project Jupiter Debt Comes Under Market Pressure
18 September 2026 | United States
Approximately US$18 billion of loans associated with Oracle's Project Jupiter AI data-centre development in New Mexico came under market pressure, according to Reuters reporting on Financial Times information. The debt was reportedly trading at approximately 89–91 cents on the dollar as investors assessed Oracle's expanding debt load and uncertainty surrounding infrastructure execution. Project Jupiter also faces local environmental concerns and regulatory complications around its proposed power infrastructure. The development is significant because it provides a direct financing-market signal that power, permitting and execution risk can be transmitted into the cost and liquidity of debt financing supporting very large AI infrastructure projects.
Watch secondary-market pricing, refinancing terms and resolution of Project Jupiter’s power and permitting issues. If the debt remains discounted, it would show that schedule and energisation risk is being capitalised directly into funding costs, potentially raising hurdle rates for other gigawatt-scale projects.
GreyRadius Insight
Project-level debt trading below par is a useful market signal: infrastructure execution risk is no longer confined to developers—it is being transmitted into credit pricing. This could create a financing premium for projects with secured power and clean permitting, making readiness quality a direct determinant of cost of capital.
Regulation / Power / Data centres
Virginia Introduces Statewide Data Center Accountability Framework
18 September 2026 | United States
Virginia Governor Abigail Spanberger unveiled a statewide Data Center Accountability Framework and signed Executive Order 22 to begin implementing its core elements. The framework addresses community transparency and participation in approvals, electricity costs, environmental standards, clean-energy development and workforce issues. It represents a statewide policy response to the scale of data-centre development in Virginia, rather than a single-county planning action. For operators and investors, the framework potentially changes the regulatory environment surrounding future capacity in the world's most important hyperscale data-centre market.
Executive Order 22 is the starting point; agency implementation will determine commercial impact. Watch changes to approval timelines, grid-cost allocation, environmental requirements and community processes, because statewide rules could alter both project economics and site-selection criteria across Virginia.
GreyRadius Insight
Virginia is moving data-centre policy from local permitting toward a broader social-contract framework covering electricity costs, environmental impact and community participation. Operators that treat these as external affairs issues may be disadvantaged; regulatory engagement and demonstrable local value are becoming core development competencies.
Compute / Capital / Data centres
Zettabyte and Bowrington Explore High-Density GPU Infrastructure in Taiwan
18 September 2026 | Taiwan
Zettabyte and Bowrington Capital signed an MoU to assess high-density GPU-cluster and associated data-centre development in Taiwan. The parties intend to evaluate sites, infrastructure requirements and total cost of ownership, with Bowrington potentially participating as a co-owner, co-developer and investor at both the physical site and GPU-asset level. No committed MW, GPU count, investment value or operational date was disclosed, so the agreement remains feasibility-stage infrastructure, not deployable capacity. Its strategic importance lies in the proposed financing structure connecting long-duration infrastructure capital directly with high-density compute assets.
The MoU needs to progress to site selection, committed GPU volumes, financing terms and an operating date. Watch how ownership is split between real estate and GPU assets: mismatched depreciation cycles, utilisation risk and hardware refresh requirements can materially affect returns.
GreyRadius Insight
GPU infrastructure finance is beginning to blur the boundary between real estate and technology assets. Direct investment at GPU level could accelerate cluster deployment, but underwriting must account for much faster hardware obsolescence than facility depreciation. Capital structures will need to match asset life to workload contracts.
NERC reported that preliminary results showed its foundational Computational Loads Reliability Standards had passed their initial ballot, subject to validation. The proposed requirements address the integration of rapidly expanding large computational loads into the North American bulk power system and draw on NERC's Level 3 Alert, incident reviews and Large Loads Working Group. Development involved NERC and Regional Entity staff, electric-industry participants, data-centre and cryptocurrency operators and technical experts. The standards are significant because very large AI and data-centre loads increasingly exhibit operating characteristics capable of affecting bulk-system reliability, making computational facilities an explicit category of grid-planning concern.
Follow ballot validation and the technical requirements that emerge around telemetry, ride-through, load ramping and emergency response. New standards could make grid behaviour a design input for AI facilities, affecting interconnection studies, controls architecture and the operational flexibility expected from large loads.
GreyRadius Insight
NERC’s initiative indicates that very large compute campuses are becoming power-system participants, not ordinary commercial loads. Grid-aware controls, telemetry and workload management may evolve into prerequisites for scale, creating a new intersection between data-centre software architecture and utility reliability engineering.
Networking / Hardware
Marvell Unveils 2nm Optical Technologies for Next-Generation AI Data Centres
20 September 2026 | Global
Marvell announced a set of 2nm optical-interconnect demonstrations aimed at next-generation AI data-centre connectivity. Technologies include 400G/lane optical PAM4, an 800G ZR/ZR+ pluggable with MACsec, 1.6T ZR and O-band coherent-lite technologies and a 102.4T co-packaged-optics platform among 38 demonstrations planned for ECOC 2026. Marvell positions the portfolio against AI architectures scaling from 1.6T toward 3.2T and beyond, where bandwidth density and power consumption per bit become critical constraints on scale-up, scale-out and scale-across connectivity. This is enabling networking technology rather than deployed customer capacity.
The transition from demonstrations to customer sampling and production is the key milestone. Track power per bit, reach, thermal performance and adoption in 1.6T/3.2T fabrics, because networking efficiency increasingly determines how much useful compute can be assembled within a fixed campus power envelope.
GreyRadius Insight
Optical interconnect is becoming a capacity multiplier for AI factories. As clusters scale, the limiting factor is not only accelerator performance but how efficiently data moves between them. Lower power per bit and higher bandwidth density can preserve more of the campus power budget for compute and reduce the networking penalty of larger clusters.
Executive Dashboard
Signal
Key Data
Date
Region
Executive Implication
Bell Creates Path to a 1.2GW Sovereign AI Infrastructure Hub in Saskatchewan
900MW; 1.2GW; C$50 billion
14 Sep 2026
Canada
Bell’s plan shows that gigawatt-scale AI campuses are increasingly being designed around dedicated generation and water-light cooling rather than relying only on existing grids. For market entrants, the competitive asset is becoming an integrated package of power, permitting, cooling and anchor demand.
Digital Realty Enters Türkiye With Power-Secured 22MW Ankara Campus
22MW; 2028.
14 Sep 2026
Türkiye
Türkiye’s position between Europe, the Middle East and Asia can support regional cloud and interconnection demand, but the strategic value of the campus depends on ecosystem density. Carrier neutrality, cloud on-ramps and customer concentration will matter as much as the initial 22MW.
Crusoe Adds 400,000 Sq Ft of US Manufacturing for AI Electrical Infrastructure
400,000 sq ft; 1 million; 300; 100; 2024.
14 Sep 2026
United States
Electrical equipment is moving from a procurement line item to a schedule-critical infrastructure layer. Vertical integration can protect AI developers from long lead times and gives operators greater control over the path from utility interconnect to live rack.
Oracle Clarifies Construction Status of Project Jupiter in New Mexico
Data centres / Power / Permitting
14 Sep 2026
United States
Large AI projects increasingly have multiple permitting clocks. A campus can advance physically while its long-term power architecture remains unresolved, creating a gap between construction progress and genuinely deployable compute capacity.
Rebellions and ai& Plan Up to 100+ AI Inference Racks in Japan
100; US$2 billion; 2026; 40MW; 2027.
15 Sep 2026
Japan
Sovereign AI is broadening beyond domestic data storage into local accelerator infrastructure. Japan’s opportunity is to pair domestic or regional compute options with trusted facilities and enterprise workloads that require lower latency and greater control over data location.
Serverfarm Expands Data-Centre Credit Facility to US$3.895 Billion
US$895 million; US$3.895 billion; 22 banks
15 Sep 2026
North America
The capital stack for AI infrastructure is deepening, yet liquidity and deployability are not the same thing. Investors and customers should distinguish corporate funding headroom from capacity that has power, approvals and a credible commissioning date.
Brazil Establishes ReData Incentive Regime for AI, Cloud and HPC Data Centres
2026; R$5.2 billion
15 Sep 2026
Brazil
Brazil is using tax policy to compete for AI, cloud and HPC infrastructure while attaching sovereignty and development conditions. This can shift location economics, but operators will need to model incentives together with power availability, import exposure and compliance obligations.
Delos Data Raises More Than US$100 Million for AI Interconnect Technology
US$100 million; 10×; 300×; 2030.
15 Sep 2026
United States
As AI systems scale, data movement is becoming a first-order bottleneck alongside compute. The next infrastructure winners may be companies that reduce the energy and latency cost of moving data between accelerators, memory and storage rather than simply adding more chips.
Microsoft Files for 123-Acre Two-Building Virginia Data-Centre Campus
123 acres; 15; 2026.
15 Sep 2026
United States
Land scale is an incomplete proxy for AI capacity. For pipeline assessment, executives should separate site control and permitting from energised MW, because the latter determines when a campus can actually support customer workloads.
Anthropic Signs First Australian Data-Centre Agreement at Proposed 2.16GW Campus
2.16GW; 2027.
16 Sep 2026
Australia
Anchor AI customers can materially improve the financeability of large campuses, but capacity claims need to be decomposed into contracted, permitted, powered and operational layers. That distinction is critical as multi-gigawatt announcements become more common.
onsemi Launches Power Architecture Claiming 3–5× Higher Power Density
3; 5×
16 Sep 2026
Global
Power conversion is becoming a compute-density constraint. As racks move toward megawatt-class designs, gains in packaging and conversion density can free physical space and thermal headroom that would otherwise limit accelerator deployment.
Google, NVIDIA and Emerald AI Launch Flexible AI Data-Centre Alliance
24; 7
16 Sep 2026
Global
AI workloads may become part of the grid solution rather than only a source of load growth. Operators that can shift non-time-critical compute could gain a new route to power access, turning software-defined demand flexibility into an infrastructure advantage.
Crusoe Raises US$3.9 Billion to Scale Vertically Integrated AI Infrastructure
US$3.9 billion; US$30.9 billion; US$140 billion
17 Sep 2026
United States / Global
Crusoe is building a vertically integrated model spanning energy, electrical equipment, data centres and cloud. The strategic question is whether controlling more layers compresses deployment time enough to justify the capital intensity and execution complexity of the model.
Loudoun County Advances Up-to-12-Month Pause on Some Data-Centre Applications
12; 2026,
17 Sep 2026
United States
Northern Virginia’s constraint set is shifting from power scarcity alone to a combined power, land and policy problem. Capacity planning now requires regulatory-path analysis at county level, not just utility interconnection and real-estate diligence.
Oracle's US$18 Billion Project Jupiter Debt Comes Under Market Pressure
US$18 billion; 89; 91 cents
18 Sep 2026
United States
AI infrastructure risk is moving into credit markets. When power and permitting uncertainty affects debt liquidity, the cost of delayed energisation can propagate beyond construction schedules into financing economics and future capital availability.
Virginia Introduces Statewide Data Center Accountability Framework
22
18 Sep 2026
United States
Virginia is moving toward statewide governance of data-centre externalities. For operators, community engagement, grid-cost allocation and environmental performance are becoming core development capabilities alongside site acquisition and engineering.
Zettabyte and Bowrington Explore High-Density GPU Infrastructure in Taiwan
Compute / Capital / Data centres
18 Sep 2026
Taiwan
Financing structures are beginning to connect infrastructure ownership directly with GPU assets. That can broaden the investor base for high-density compute, but it also requires tighter alignment between hardware depreciation, facility life and customer utilisation.
AI data centres are becoming a distinct grid-planning category. Operators will increasingly need to design not only for uptime but also for predictable bulk-system behaviour, making grid compliance and workload controllability part of infrastructure architecture.
Marvell Unveils 2nm Optical Technologies for Next-Generation AI Data Centres
2nm; 1.6T; 102.4T; 38; 2026.
20 Sep 2026
Global
Networking bandwidth is becoming inseparable from accelerator scaling. Optical innovation that increases bandwidth density while reducing power per bit can determine how efficiently very large clusters scale across racks, rows and data-centre campuses.
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.
