Amazon Web Services: cloud infrastructure for trading, data and broker operations

Amazon Web Services (AWS) provides the underlying infrastructure and managed building blocks for financial applications. Its capital-markets offering covers trading infrastructure, risk calculations and data workloads, with partner software available alongside AWS services. The operational scope depends on the services and architecture purchased.[1]

AWS’s financial-services reference architecture illustrates exchange data arriving through Outposts and private connectivity, streaming through MSK and Managed Service for Apache Flink, and reaching applications through WebSocket APIs. S3 and Glue support raw, processed and curated datasets; Lake Formation controls access. This is an architectural pattern for ingestion and distribution, rather than a packaged market-data feed or permission to redistribute exchange content.[2]

Direct Connect uses Ethernet between the customer’s router and an AWS router at a connection location. Its requirements include VLAN tagging and BGP configuration; public virtual interfaces reach public AWS services, while private interfaces reach VPC resources. Colocation and connectivity partners can supply parts of the physical path, so the circuit and AWS-side service must be specified together.[4]

A Direct Connect connection is not encrypted by default. AWS documents IPsec through Site-to-Site VPN and MACsec on supported connections. MACsec protects the relevant link from the customer’s data centre to the Direct Connect location; an application’s complete encryption design can require additional controls.[5]

EC2 placement groups make a deliberate trade-off. Cluster placement puts instances close together within one Availability Zone for tightly coupled workloads. Partition placement separates groups of instances across hardware, while spread placement separates a small set of instances. Proximity and failure isolation are different objectives; a placement choice does not establish end-to-end venue latency.[3]

Data Exchange offers public subscriptions and custom private offers through AWS Marketplace. Some providers verify the subscriber’s identity and intended use. Access can involve files exported to S3 or signed URLs, API Gateway, Redshift data sharing or access to a provider’s S3 data. The subscription agreement remains the source of commercial and redistribution rights; acquiring cloud infrastructure does not extend those rights.[11]

Control Tower establishes a governed multi-account landing zone using AWS Organizations, IAM Identity Center and associated services. Account Factory provides account templates, while preventive, detective and proactive controls address different stages of configuration. This helps separate workloads and account ownership before production applications are deployed.[14]

The Direct Connect Resiliency Toolkit helps provision redundant dedicated connections, including configurations across multiple locations. It generates connection documentation and supports BGP failover tests. Topology, port speeds and independent paths still need to match the chosen resilience design; provisioning two links is not by itself proof that all application dependencies can fail over.[6]

Outposts shifts AWS-managed hardware into the customer’s premises while retaining its relationship with an AWS Region. Site readiness, power and network connectivity therefore remain part of delivery. The published model supports workloads needing local infrastructure while using AWS APIs and management conventions.[7]

A concrete implementation case is Jefferies’ use of RDS Custom for Oracle. AWS describes a six-to-eight-month rollout supporting legacy or customised applications while retaining privileged operating-system and database access. The case spans six Regions and 50 accounts. It demonstrates one institutional deployment, rather than a standard implementation deadline for every brokerage.[28]

Basic support covers account and billing assistance; technical cases require a paid support plan. AWS publishes initial-response targets and says it makes reasonable efforts to meet them. These clocks concern the first response, not repair or restoration.[20]

The Compute SLA provides credits of 10%, 30% or 100% at its specified unavailability thresholds. Claims must arrive by the end of the second billing cycle after the incident and include the required evidence. Exclusions apply, and regional and instance credits cannot be stacked for the same affected instance. This service-specific commitment is not an availability guarantee for an entire brokerage application.[21]

AWS’s shared-responsibility model assigns it the underlying cloud infrastructure. With EC2, the customer manages the guest operating system, applications and security configuration. More abstracted services shift additional platform operations to AWS, while data, access permissions and relevant configurations remain customer responsibilities. Applicable financial-services obligations depend on the institution’s own activities and deployment.[22]

IAM guidance recommends federation and temporary credentials for people, roles for workloads, multifactor authentication and least-privilege permissions. Those controls determine which users or processes can read, change or delete resources. Giving an integration access to cloud resources should be treated as a separate decision from authorising it to initiate a business transaction.[10]

Bedrock Agents can call configured functions or APIs. User confirmation is a developer-enabled control for a particular action: CONFIRM permits execution and DENY blocks it. The documentation does not describe universal human approval as an automatic property of every agent.[13]

Bedrock retention is configurable and model-dependent. The “none” setting prevents durable request/response storage and provider sharing for compatible models; an incompatible model is blocked. “default” follows model-specific retention, while “aws_review” permits AWS safety review, including human review, for up to 30 days. Settings are regional, with project controls applying to the documented endpoint scope. Cross-Region inference can retain content in destination Regions. A request’s store=false option is not equivalent to zero retention.[12]

The Customer Agreement allows the customer to select content Regions, with legal-order and other stated exceptions to that location commitment. Account information and customer content have different contractual treatment. Region selection should therefore be read alongside the particular service and data-processing terms.[23]

AWS is Amazon’s cloud business. Amazon’s results released on 5 February 2026 report AWS segment sales of $128.7 billion for the year ended 31 December 2025. That is a dated measure of the business’s scale, rather than a current-year revenue estimate.[27]

The contracting entity is determined by Account Country under the AWS Customer Agreement. Examples include Amazon Web Services, Inc. and Amazon Web Services EMEA SARL; separately listed countries use other local entities. A brokerage should identify the actual party on its account and agreement instead of treating the global AWS brand as one universal contracting company.[23]

Operator

Applicable AWS entity by account country, including Amazon Web Services, Inc. and Amazon Web Services EMEA SARL.

Service scope

Cloud computing, storage, networking and managed services supporting capital-markets applications and data workloads.

Customer types

Brokerage firms; Banks; Asset managers; Exchanges; Financial technology firms

Delivery

Hosted; API; Managed service; Hybrid

Pricing publication

Public pricing

Platforms

Amazon EC2; Amazon S3; AWS Direct Connect; AWS Data Exchange; Amazon Bedrock

Pricing scope

Published service usage tariffs and support minimums; commitments, region, architecture and third-party services affect the total.

Implementation

Multi-account landing zones can use Control Tower; network resilience and application integration require a deployment-specific design.

Support

Basic account and billing support; paid technical-support plans have severity-based initial-response targets.

Permissions and regulatory scope

Cloud infrastructure provision. The customer remains responsible for its applications, data, configuration and applicable financial-services obligations.

Provider metadata checked 20 Oct 2026Source 1Source 23Source 7Source 15Source 14Source 20Source 22Source 2Source 13

CloudTrail Event history supplies 90 days of management events for each account and Region. It does not include data events, Insights events or network-activity events, and it is not an organisation-wide trading ledger. A firm needing a continuing audit record must configure the appropriate trail and event coverage rather than rely solely on the default history view.[9]

The standard agreement’s post-termination window is 30 days, except where termination is for specified cause. Retrieval requires all amounts due to be paid, and ongoing use remains chargeable. This is the contractual retrieval provision, distinct from the account-management process for reopening a closed account.[23]

Account closure has a separate 90-day reopening period. Access to content during that period requires reopening; after permanent closure, AWS describes deletion of content and resources with an exception for CloudTrail trails. Remaining Reserved Instance and Savings Plans commitments can continue to be billed until their terms expire. Backups and exports therefore belong in the closure plan.[24]

AWS’s current exit-transfer procedure requires a complete request at least two months before switching and approval before transfer. Eligible scope is all AWS data, or all data for a chosen service, across the applicable accounts. Credits cover eligible internet transfer, with Direct Connect, CloudFront, Snow Family and Global Accelerator excluded; ordinary service charges continue. The stated outside-EU switching period is 180 days, with deletion or closure requirements.[25]

For eligible EU accounts, the Data Act Addendum establishes a 30-calendar-day transition after the notice period. If technically infeasible, AWS must notify the customer within 14 business days after receiving the complete request, giving reasons and a revised period of up to seven months. These scoped terms take precedence over conflicting standard provisions.[26]