The United Kingdom’s Met Office has spent the past year running its operational weather and climate workloads on a purpose-built supercomputing cluster hosted by Microsoft Azure, and the early results are striking: a step-change in raw compute capacity, measurable gains in operational resilience and observability, and a scientific model upgrade that the Met Office says would not have been possible without the new platform.
The Met Office is Britain’s national weather service and a major global centre for meteorology and climate science. Its forecasts underpin everyday decisions—from commuter travel and aviation operations to national emergency planning and defence logistics—and the organisation operates a long lineage of supercomputers dating back to the 1950s. That history includes early experiments on the EDSAC at Cambridge and the procurement of the Ferranti Mercury (nicknamed “Meteor”) in 1959, milestones that mark how numerical weather prediction has been tightly coupled to computational capability since the discipline’s inception.
In 2021 the Met Office contracted Microsoft and Hewlett Packard Enterprise to design and deliver a next‑generation supercomputing capability. The intention was explicit: move beyond a single, on‑premises centre to a managed, cloud‑hosted HPC service that could scale, evolve and deliver continuous upgrades over a decade. After a phased transfer and a period of parallel operations, the Met Office completed the transition to the Azure‑hosted supercomputing cluster and has since been producing operational forecasts and research outputs from the new environment.
The Met Office’s new platform addresses both. The initial configuration of the Azure-hosted system was designed to deliver more than 60 petaflops of aggregate peak computing capacity and a processor core count at the scale of millions. The architecture is assembled in multiple quadrants of dedicated HPC hardware integrated into Azure’s fabric, combining HPE Cray EX systems with high‑performance storage and an “active data archive” sized in the exabyte range. That combination is intended to free scientists to run higher‑resolution models, longer ensembles, and more realistic physics without being constrained by an inflexible on‑premises refresh cadence.
Because different sources have used different rounding and terminology (aggregate peak FLOPS versus sustained application FLOPS, for example), readers should treat single‑figure core counts and petaflop numbers as indicative of order‑of‑magnitude capability rather than immutable specifications. The important point is the step change: the new platform multiplies the Met Office’s previous operational capacity by several times and provides an upgrade path.
Several availability metrics have been reported by Met Office executives in media briefings: very high availability for integrated services (approaching 99.95%), sub‑percent downtime for the supercomputing service itself and reports of critical workload availability at or near 100% since the migration. These figures, while encouraging, originate in Met Office communications during the first year of operations and should be seen as early indicators rather than long‑term guarantees—high operational availability over short windows is different to demonstrable, multi‑year reliability under diverse stress conditions.
Key operational benefits claimed and observable:
Where AI is likely to add value:
However, the public sector nature of the Met Office’s remit means that accountability and transparency must remain front and centre. The public funds the service, and the outputs—forecasts, warnings and climate projections—are public goods. That requires clear cost transparency, demonstrable reproducibility and strong contractual guardrails to protect national interests.
This first year should be seen as a successful proof point for an architecture that pairs purpose‑built HPC hardware with cloud operational tooling. Yet it is also a cautionary case: long‑term success will depend on rigorous governance, contractual clarity, preserved scientific reproducibility and active management of vendor dependency risks.
If the Met Office can sustain the operational availability and scientific uplift it has reported, while preserving public control over data, costs and research integrity, this partnership could stand as a blueprint for how national science agencies modernise critical infrastructure without losing sight of public accountability. The next phase—further capacity upgrades, deeper AI experimentation and continued monitoring of costs, resilience and reproducibility—will determine whether this platform becomes a durable national asset or a transient technical success.
Source: ITPro Met Office hails huge efficiency gains in first year of cloud supercomputing with Microsoft Azure
Background
The Met Office is Britain’s national weather service and a major global centre for meteorology and climate science. Its forecasts underpin everyday decisions—from commuter travel and aviation operations to national emergency planning and defence logistics—and the organisation operates a long lineage of supercomputers dating back to the 1950s. That history includes early experiments on the EDSAC at Cambridge and the procurement of the Ferranti Mercury (nicknamed “Meteor”) in 1959, milestones that mark how numerical weather prediction has been tightly coupled to computational capability since the discipline’s inception.In 2021 the Met Office contracted Microsoft and Hewlett Packard Enterprise to design and deliver a next‑generation supercomputing capability. The intention was explicit: move beyond a single, on‑premises centre to a managed, cloud‑hosted HPC service that could scale, evolve and deliver continuous upgrades over a decade. After a phased transfer and a period of parallel operations, the Met Office completed the transition to the Azure‑hosted supercomputing cluster and has since been producing operational forecasts and research outputs from the new environment.
Why this matters: compute, data and science at scale
Modern weather and climate science is computationally hungry. Numerical weather prediction (NWP) models solve complex equations across three dimensions and through time, assimilating vast observational streams—satellites, buoys, aircraft, radar—and producing ensemble forecasts that quantify uncertainty. Two interlinked bottlenecks have historically constrained forecast accuracy and lead time: compute capacity (how fine and physically realistic models can be) and data management (how observations and model output are stored, searched and reused).The Met Office’s new platform addresses both. The initial configuration of the Azure-hosted system was designed to deliver more than 60 petaflops of aggregate peak computing capacity and a processor core count at the scale of millions. The architecture is assembled in multiple quadrants of dedicated HPC hardware integrated into Azure’s fabric, combining HPE Cray EX systems with high‑performance storage and an “active data archive” sized in the exabyte range. That combination is intended to free scientists to run higher‑resolution models, longer ensembles, and more realistic physics without being constrained by an inflexible on‑premises refresh cadence.
- Compute scale: the platform delivers tens of petaflops of peak performance, enabling higher‑resolution global runs and more ensemble members.
- Data scale: an active archive and high‑performance storage provide the throughput and capacity to store and retrieve petabytes of model output and observations.
- Operational model: Microsoft provides a managed HPCS (high‑performance computing as a service) offering, combining dedicated hardware halls with cloud orchestration, telemetry and automation.
The move to Azure: technical anatomy and what’s new
Dedicated HPC in a cloud envelope
This is not the same as taking a commodity virtual machine fleet and running HPC jobs on it. The Met Office system sits in dedicated halls within Microsoft data centres, where racks, power, cooling and network are optimised for tightly coupled HPC workloads. The initial hardware mix includes HPE Cray EX systems and AMD EPYC CPUs, tied to a high‑performance interconnect and a tailored storage solution. The top‑level design choices that matter most:- Four quadrants of HPE Cray EX supercomputers provisioned and integrated into the Azure operational model.
- A compute footprint designed to be upgraded iteratively over the contract term, enabling periodic increases in raw FLOPS and core counts.
- An active data archive intended to hold and serve multiple exabytes of historical and operational model output.
- Co‑located telemetry, observability tooling and managed services to provide an “HPC‑as‑a‑service” operational model.
Performance numbers and discrepancies to note
Public statements about raw numbers vary by source and must be read carefully. Microsoft and Met Office programme briefings described the first generation as exceeding 60 petaflops of aggregate peak computing capacity and cited a processor core count in the order of 1.5 million. Journalistic reporting has occasionally quoted a higher figure (for example, 1.8 million cores), but the authoritative programme briefing and vendor materials align on approximately 1.5 million CPU cores and ~60 petaflops for the first phase. The system was explicitly designed to grow over time—as hardware generations advance and the Met Office’s needs evolve, additional capacity and upgrades will be delivered across the multiyear contract.Because different sources have used different rounding and terminology (aggregate peak FLOPS versus sustained application FLOPS, for example), readers should treat single‑figure core counts and petaflop numbers as indicative of order‑of‑magnitude capability rather than immutable specifications. The important point is the step change: the new platform multiplies the Met Office’s previous operational capacity by several times and provides an upgrade path.
Data throughput and I/O
Operational forecasting and research produce and consume enormous daily volumes. Met Office engineering estimates for the initial operational period noted the system handles between 200 and 300 terabytes of model and observation data per day in active workflows. The active archive’s design—paired with high‑throughput cloud storage services—was explicitly chosen to support such throughput numbers while keeping data accessible for on‑the‑fly analysis, ensemble re‑runs and long‑term research.Operational gains: resilience, observability and availability
One of the largest operational benefits the Met Office has emphasised is improved resilience and faster incident detection and resolution. The managed service model brings built‑in telemetry, automation and observability capabilities that, according to Met Office leadership, have allowed the organisation to detect and remediate issues faster than with their prior on‑premises setups.Several availability metrics have been reported by Met Office executives in media briefings: very high availability for integrated services (approaching 99.95%), sub‑percent downtime for the supercomputing service itself and reports of critical workload availability at or near 100% since the migration. These figures, while encouraging, originate in Met Office communications during the first year of operations and should be seen as early indicators rather than long‑term guarantees—high operational availability over short windows is different to demonstrable, multi‑year reliability under diverse stress conditions.
Key operational benefits claimed and observable:
- Improved observability: telemetry and monitoring at scale provide faster mean time to detection and mean time to repair for HPC and data pipelines.
- Minimal impact on long‑running workflows: automation and scheduling resilience reduce the chance that ten‑year scientific experiments are interrupted by platform churn.
- Built‑in scalability: the capacity to grow compute resources for specific research campaigns without procuring new on‑prem hardware.
- Dedicated hosting: an architecture designed to meet data sovereignty and security requirements by keeping the system physically based in the UK within dedicated halls.
Scientific outcomes: the new model and what it delivers
The Met Office has followed the platform change with a major scientific model upgrade that introduces enhanced physical parameterisations, the assimilation of more aircraft observations, and recalibrated ensemble treatment of uncertainty. The headline impacts reported by the Met Office include:- More realistic rainfall forecasts, with improvements across light and heavy precipitation regimes thanks to a modern cloud microphysics scheme.
- Improved cloud cover and cloud base forecasts, which are operationally significant for aviation, renewable energy forecasting and visibility/weather warnings.
- Better UK temperature predictions, aiding energy demand planning and winter resilience planning.
- An extended global ensemble forecast range—moving the practical range from seven to ten days for the global ensemble—which improves medium‑range situational awareness and early warning lead time.
Strengths: what's most compelling about the transition
- Scalability and upgrade cadence: Moving to a managed, upgradeable HPC platform allows the Met Office to adopt newer processor generations and architectures without the capital ramp of building and commissioning on‑prem datacentres.
- Operational observability: The telemetry and automation from the cloud provider improve incident response and reduce unplanned downtime for critical scientific workflows.
- Data accessibility: Migrating decades of observations and model archives into active, searchable storage opens new possibilities for reuse, ensemble reanalysis and reproducible science.
- Energy and sustainability claims: The deployment was designed with energy efficiency and renewable energy sourcing in mind, reducing the per‑FLOP carbon intensity of operations relative to older on‑prem equipment.
- Rapid innovation cycle: The managed service model encourages more iterative science releases and lowers the friction for experimenting with hybrid AI/physics approaches.
Risks and open questions
While the transition offers tangible benefits, it also introduces a set of risks and governance challenges that the Met Office, its government sponsors and the wider user community must manage carefully.Vendor lock‑in and operational dependency
A long, multi‑year contract with a single cloud provider—and deeply integrated managed services—creates a dependency that is difficult and costly to unwind. Even if the hardware sits in dedicated halls, the operational tooling, telemetry, APIs and orchestration layer will be tied to Microsoft’s Azure ecosystem. The Met Office must therefore:- Ensure contractual rights to portability, data extraction and long‑term archiving in open formats.
- Preserve internal capabilities and expertise so that science teams remain platform‑agnostic where possible.
- Design exit and continuity plans in the event of service disruption or contract disputes.
Cost structure and fiscal transparency
Large HPC contracts often shift capital expenditure into predictable operational expenditure, but OPEX models can escalate if not tightly governed—especially when usage spikes for an emergency response or prolonged research campaigns. The Met Office and government stakeholders will need robust governance to keep costs predictable and aligned with public value.Reproducibility and scientific auditability
Long‑running scientific projects and published results depend on being able to reproduce model runs and underlying datasets. Moving to a managed cloud environment requires:- Clear, versioned archival policies for model code, configurations and input data.
- Guarantees about long‑term accessibility of archived runs without hidden egress or access costs.
- Documentation and metadata standards to preserve context for future reanalysis.
Data sovereignty, security and national resilience
While Microsoft provides UK‑based hosting and dedicated operational protections, national stakeholders will rightly scrutinise the implications for digital sovereignty. Important questions include the legal protections around data access, the resilience of cross‑border network links, and the independence of critical services during geopolitical stress.Single‑point upgrades versus diversified research paths
Scientific risk also emerges from centralising capability. If the majority of operational and experimental modelling consolidates on one platform, bugs, systemic architecture constraints or misconfigurations could propagate across many experiments. Diversity of approach—maintaining some on‑prem or partner‑hosted capabilities—can act as a hedge.AI, hybrid workflows and the future of forecasting
Met Office leadership has been careful to temper over‑exuberant promises about AI. Their stated position is pragmatic: AI techniques are promising as accelerants for specific tasks—data assimilation, model emulation, bias correction and downscaling—but they are not substitutes for physics‑based models that remain central to understanding processes and providing physically consistent forecasts.Where AI is likely to add value:
- Emulation of expensive physics to provide rapid approximations for ensemble members.
- Intelligent prioritisation of compute resources for forecast runs that have the highest expected impact.
- Explainable machine learning to identify “forecasts of opportunity” where skill can be exploited.
Practical recommendations and governance for national weather services
For other national meteorological services or research organisations contemplating a similar path, the Met Office case offers practical lessons:- Negotiated service-level metrics: Define measurable availability, throughput and job success SLAs for both operational and long‑running scientific workflows.
- Portability and open formats: Insist on contractual guarantees for code, configuration and data portability in open or standard formats.
- Dual‑track capability: Maintain a small on‑prem or partner diverged capability for disaster recovery, benchmarking and scientific diversity.
- Cost governance: Implement usage caps, burst controls and transparent reporting to avoid runaway OPEX during high‑demand periods.
- Reproducibility guardrails: Archive metadata, code and input datasets with version control and immutable storage primitives to support reproducible science.
- Workforce investment: Train and retain scientific and operational staff in cloud‑native HPC practices, observability tooling and hybrid AI approaches.
- Data sovereignty and legal clarity: Secure explicit contractual terms that define jurisdiction, access rights, and emergency governance.
The bigger picture: public value, research acceleration and national resilience
The Met Office’s move to an Azure‑hosted supercomputing service is emblematic of a broader shift in public sector infrastructure: mission‑critical, high‑value national services are increasingly pairing domain expertise with commercial cloud capability. Done well, this can accelerate research, broaden access to data, and deliver tangible public value—earlier warnings, better aviation safety, faster disaster response and more efficient energy planning.However, the public sector nature of the Met Office’s remit means that accountability and transparency must remain front and centre. The public funds the service, and the outputs—forecasts, warnings and climate projections—are public goods. That requires clear cost transparency, demonstrable reproducibility and strong contractual guardrails to protect national interests.
Conclusion
Twelve months after transitioning its forecasting and research workloads to a dedicated Azure‑hosted supercomputing platform, the Met Office reports large efficiency gains, higher compute capacity and improved operational resilience—outcomes that have already enabled a major scientific model upgrade with practical benefits for rainfall, cloud and temperature forecasts, and an extended ensemble range for medium‑term forecasting.This first year should be seen as a successful proof point for an architecture that pairs purpose‑built HPC hardware with cloud operational tooling. Yet it is also a cautionary case: long‑term success will depend on rigorous governance, contractual clarity, preserved scientific reproducibility and active management of vendor dependency risks.
If the Met Office can sustain the operational availability and scientific uplift it has reported, while preserving public control over data, costs and research integrity, this partnership could stand as a blueprint for how national science agencies modernise critical infrastructure without losing sight of public accountability. The next phase—further capacity upgrades, deeper AI experimentation and continued monitoring of costs, resilience and reproducibility—will determine whether this platform becomes a durable national asset or a transient technical success.
Source: ITPro Met Office hails huge efficiency gains in first year of cloud supercomputing with Microsoft Azure
