Cloud Software Development in 2027: 5 Key Architecture Decisions
Cloud Software Development Services After October 2026
How much should a company depend on one cloud provider, one region or one type of infrastructure, and what does it cost to depend on it less? Most teams answer that once, during their first migration, and rarely come back to it.
Two recent events are good reasons to come back. A power failure in one of Google Cloud’s London zones knocked out banking apps and other services across the UK for several hours. Separately, Gartner warned that more than $3.5 trillion of AI infrastructure spending could be exposed to impairment by the end of 2030. This article turns the question into five design decisions that teams building cloud software in 2027 will face, whether they do the work in-house or buy cloud software development services from a partner. The October events come up throughout as reference points.
What happened on October 2026
On the morning of Tuesday, October 20, a power event hit europe-west2-b, one of the three zones in Google Cloud’s London region. Services in that zone returned errors and lost network connectivity until Google mitigated the problem later that morning.
Revolut took the most visible hit. BleepingComputer reported that the bank’s status page listed 28 services as partially unavailable, among them sign-ins, card payments, transfers, currency exchange and stock trading. Revolut said “all our customers’ money and data are safe” and later confirmed the Google Cloud outage as the cause. Other UK banks, public bodies and retailers reported problems in the same window.
At the time of writing neither Google nor Revolut had published a full incident analysis, so guessing at any one company’s architecture would be unfair. The general lesson needs no guessing. A cloud region is made of several zones so that one can fail without taking applications down with it. When a single-zone failure turns into a wide outage, some piece of the system was living in that zone alone. Often it is a database, a message queue, a cache or a third-party dependency.
Decision 1: how much resilience each system really needs
Resilience is not one setting for the whole company. Different systems need different protection, and each level has a price.
| Deployment pattern | Survives | Typical extra cost | Typical recovery time |
| Single zone | Server or process failures | Baseline | Hours if the zone fails |
| Multi-zone in one region | Loss of one zone | Moderate: duplicate capacity, cross-zone traffic | Seconds to minutes |
| Multi-region, same provider | Loss of a whole region | High: duplicate stacks, data replication | Minutes, if failover is tested |
| Multi-cloud | Loss of a provider or service | Highest: two platforms, two skill sets | Varies widely |
Indicative comparison. Actual costs and recovery times depend on architecture and data volumes.
A multi-zone design covers most business applications, as long as it really is multi-zone. Every stateful component has to be replicated across zones, the load balancer has to route around a dead zone, and someone has to have tested what happens when a zone disappears.
Payment systems, trading platforms and healthcare systems facing customers often justify going multi-region. Multi-cloud fits a narrower set of cases, usually where a regulator demands it or where one provider’s service is a hard dependency with nothing equivalent elsewhere.
Start with an inventory, and include the systems nobody thinks about until they fail, such as internal tools that the support team relies on during an outage. For each system, write down how long the business can live without it (the recovery time objective) and how much data it can afford to lose (the recovery point objective). Those two numbers pick the pattern. The pattern sets the budget.
Decision 2: where the hidden single points of failure are
Plenty of outages come from parts nobody considered critical. A well-replicated application can still go down because of:
- a managed database left in one zone to save money
- a secrets manager, identity service or DNS setup tied to one region
- a container registry or artifact store that the deployment pipeline needs during recovery
- a third-party API, payment gateway or SaaS tool running in the same zone
- a batch job that holds a lock and never lets go
Finding them means working through dependency maps, running failure tests and reading the provider’s documentation on how each managed service behaves when a zone fails. Chaos engineering, switching components off on purpose in a controlled way to see what breaks, is the most reliable method. Most teams keep postponing it. A gentler start is a scheduled exercise, sometimes called a game day, where the team takes one zone or one dependency offline in a staging environment, watches what happens and writes down every surprise. The first one usually finds something, and the list it produces is a better guide to spending than any vendor checklist.
Regulators care more about this than they used to. In the EU, the Digital Operational Resilience Act has required banks, insurers and other financial companies since January 2025 to manage the risk that comes from critical technology providers, cloud platforms included. Outside finance the rules are lighter, but customers ask the same questions in their security questionnaires.
Decision 3: how to budget for AI workloads on the cloud
The Gartner forecast from October 14 is about the providers’ side of the market. Buyers will still feel it. According to Gartner, AI providers are expected to spend nearly $1 trillion on AI infrastructure in 2026 and another $6 trillion between 2027 and 2030. Gartner expects AI revenue from those data centers to fall well short of that over the same years, which is where the $3.5 trillion impairment estimate comes from.
Gartner calls this its most likely scenario under current trends, and things may turn out differently. For a company running AI features in the cloud, two practical points follow.
AI capacity prices could move either way. If providers have built more than the market uses, GPU and inference prices may drop. If consolidation thins out competition, some services may get more expensive. Long commitments at today’s rates are a bet in both cases.
AI workloads also need the cost discipline any other cloud spend gets. Inference, the cost of running a model every time a user or agent calls it, grows with usage. A feature that costs next to nothing in a pilot can turn into one of the biggest lines on the bill once thousands of people rely on it. Track cost per completed task from the first release. It tells you more than price per token or per GPU hour.
FinOps is the practice of making engineering teams answerable for cloud costs through shared data and regular reviews. It now has to cover AI usage by name: tag AI resources, set budgets and alerts per feature, cache repeated requests, and use smaller models wherever they are good enough.
Decision 4: how much to depend on one provider
The same Gartner release predicted that more than 80% of neocloud providers, the newer firms renting out GPU capacity outside the big hyperscalers, could stop operating independently by 2030. Some will be bought, some will merge, some will close.
If you use one of them for training or inference, that is a supplier risk to plan for now. The October outage made a related point about the big providers. They have bad days too, and the damage depends on how tightly an application is tied to one provider’s region or services.
For most companies, a proportionate response is a handful of habits that keep the option of moving open:
- Define infrastructure as code with Terraform, OpenTofu or similar tools, so environments can be rebuilt quickly in another region or account.
- Package applications in containers and run them on Kubernetes or another portable runtime where that makes sense.
- Keep data in open formats, with tested backups outside the primary region.
- Put provider-specific AI services behind an internal interface, so you can swap a model or inference provider without rewriting the application.
- Write down which managed services have no equivalent elsewhere, and decide on purpose whether you accept that dependency.
Each one adds a bit of work up front, and each makes the next migration, price negotiation or outage cheaper.
Decision 5: who builds and runs it
The last decision is about people. Cloud systems need engineers who understand both the application and the infrastructure under it, and many companies are short of them. That is where outside cloud specialists come in, for one project or for ongoing support.
The scope of that work keeps widening. A typical engagement in 2026 goes well past migration and deployment into architecture reviews, resilience design, disaster recovery, cost management and observability (seeing from logs, metrics and traces what a system is doing and why). For enterprise cloud software development it also means compliance work and evidence for auditors.
Sombra describes its cloud practice in three stages. First an assessment of current infrastructure against business requirements, then action on what it finds, such as cutting cost through better resource use and autoscaling, then ongoing maintenance with monitoring and observability. Its cloud services cover infrastructure management across public, hybrid and multi-cloud environments, migration, disaster recovery, infrastructure as code and FinOps. Sombra holds AWS Advanced Tier status and is a Microsoft Solutions Partner for digital and app innovation.
Application results often depend on this infrastructure work. In a modernization project for Wineshipping, a US wine logistics company, Sombra’s team took the client portal’s availability from 70% to 99.9%. Getting there meant rebuilding how the system was deployed and monitored, along with how it was coded.
What to ask a cloud partner
Before you pick a partner, ask questions that show how the team thinks about failure and cost:
- How would our main systems behave if one zone or one region failed tomorrow?
- How do you test disaster recovery, and how often?
- How do you track and report cloud costs per feature or per team?
- How do AI workloads fit into your cost planning?
- Which parts of your proposed design tie us to one provider, and why?
- Who is on call after go-live, and how fast do you respond to urgent incidents?
A strong cloud software development company answers with specifics from past projects, including what went wrong and what it changed afterward. Vague talk of “best practices” usually means nobody has asked these questions before.
Build or use managed platforms?
There is also the question of how much to build. Managed platforms from the major providers cover a lot of common needs, from databases to AI services, and cut operational work. Custom cloud software development makes sense where the business logic is specific to your company, where performance or cost requirements are unusual, or where portability matters more than convenience.
Most real systems use both. The design work lies in choosing which parts to buy as managed services and which to build, and in keeping the boundary clean enough that you can change one side without disturbing the other. Good cloud software development solutions spell those boundaries out in the architecture documents, so the next team understands them.
Example: costing resilience for one system
Numbers make these decisions easier. The example below is invented, but it follows the way such an assessment usually goes.
A company runs a customer portal on one cloud provider, in a single region. During business hours the portal takes orders worth about $40,000 an hour. The business decides it can live with 15 minutes of downtime and lose no more than one minute of data.
| Option | What changes | Indicative extra cost per month | Meets the target? |
| Stay in one zone | Nothing | $0 | No: a zone failure means hours of downtime |
| Multi-zone | Replicated database, app servers in three zones, tested failover | $3,000–5,000 | Yes, for zone failures |
| Multi-region warm standby | Second region with replicated data and scaled-down servers | $9,000–14,000 | Yes, for region failures, if failover is rehearsed |
Illustrative figures for a mid-sized web application.
One four-hour outage in business hours would cost about $160,000 in lost orders, before support costs and lost customer trust. A multi-zone setup runs roughly $36,000 to $60,000 a year. For this portal, multi-zone is an easy call. Multi-region is worth discussing only if the business decides a full regional outage is a real risk it cannot accept.
Repeat the calculation for each important system and resilience stops being a vague goal and becomes a budget line finance and engineering can agree on. Sombra’s cloud assessments work this way: recovery targets first, then the architecture and cost that meet them.
Summary
October gave two reminders about cloud dependence. The London zone failure showed how fast one hidden dependency can turn a local fault into a national one. Gartner’s forecast showed that the economics of AI infrastructure are still unsettled, with knock-on effects for prices, suppliers and long-term commitments.
For teams planning cloud work in 2027, the five decisions come down to this:
- set resilience requirements per system, from recovery time and data loss tolerance
- find and remove hidden single points of failure, and test failover regularly
- budget AI workloads by cost per completed task, with FinOps from the first release
- reduce dependence on any one provider where the risk justifies the effort
- choose partners who can show how they design, test and run systems under failure
A practical first step is to assess the current setup against these questions, which is where Sombra and other cloud partners usually start. It takes a few weeks and produces a list of specific risks with costs attached. That list is far easier to act on than a general intention to be more resilient.
