How a fractional CTO helps a startup choose its first cloud architecture
A fractional CTO helps a startup choose its first cloud architecture by settling the constraints first, sorting reversible decisions from one-way doors, comparing three realistic options on the same criteria, and writing the choice down with a date to revisit it. Most first cloud architecture decisions are made in a fortnight, by a founder, under a funding clock. This is a question we hear regularly. The question is rarely which cloud. It is almost always what we commit to now that we will not have to unpick in eighteen months. What follows is the seven-step sequence we run. It starts with what the business has to prove, moves through the choices that are cheap to reverse and the ones that are not, and ends with how the call is documented so the next engineer, the next investor and the next auditor can read it. Scope: startups from pre-seed to Series A, UK and EU context, written October 2026.
1. Set the constraints before you look at providers
The first pass is a short, bounded exercise, because a fractional CTO works on a part-time, flexible basis. Six constraints come first: expected load twelve months out, data residency and GDPR posture, the latency budget from where your users actually are, the skills your team really has, a budget ceiling, and the funding runway behind it. Architecture follows those constraints. Skip them and you buy a stack sized for a load profile you never reach. If your early customers include UK public bodies, the route through the Digital Marketplace G-Cloud framework narrows which providers you can sell on, so it belongs on the same list.
2. Separate one-way doors from reversible choices
The split that governs the whole engagement is between decisions that are expensive to reverse and decisions that are cheap to change. One-way doors: the data model, the tenancy model, the region your data sits in, the identity provider, and the evidence trail you keep for regulators and investors. Reversible choices: compute runtime, message queue, managed database flavour, and most third-party SaaS.
Founders routinely over-debate the reversible ones. A fortnight can go on whether to run containers or a serverless runtime, when either can be swapped behind an interface in a week or two of engineering. Meanwhile the tenancy model gets agreed in a ten-minute call, and it shapes pricing, onboarding, data deletion and every security questionnaire an enterprise buyer will send you. Moving from a shared schema to one database per customer later means migrating live customer data, so the cheap-looking decision turns out to be the expensive one.
Identity is the same trap. Picking a provider that cannot federate with the systems your buyers use creates an integration project rather than a configuration change. Sort the list before you argue about anything.
3. Compare the three realistic first architectures on the same criteria
| Option | Who it suits at this stage | What the team must own | Main risk | Reversibility |
|---|---|---|---|---|
| Managed application platform | Small teams shipping one product with no platform engineer | Application code, data model, vendor relationship | A ceiling on customisation and a cost curve that climbs with usage | Moderate: the app moves, the operating habits do not |
| Raw cloud primitives | Teams with at least one engineer who has run production infrastructure | Networking, identity, backups, patching, cost control | A slow start and a real chance of misconfiguration | High on components, low on the data and identity decisions underneath |
| Hybrid | Teams with one workload that needs control and another that does not | The boundary between the two, plus consistent identity and logging | Two operating models and two sets of skills to maintain | Good, provided the boundary is documented |
Read the table column by column rather than row by row. The first question is who the option suits at this stage, because the wrong answer there wastes the most time. A two-person team with no infrastructure experience and a managed application platform is a sensible pairing. The same team on raw primitives spends its first quarter on networking and identity instead of on the product. The second question is what the team must own, since ownership is where the hidden cost sits.
The lowest monthly bill is rarely the lowest total cost. Engineering days are usually the largest line in a startup’s infrastructure budget, so a cheap primitive stack that quietly consumes senior engineering time every month can cost more than a managed platform with a higher invoice. That is why we price the two together rather than comparing provider invoices. Our cloud pricing analysis of AWS, Azure and Google Cloud covers the billing mechanics and the traps inside them, so the numbers do not need repeating here.
Reversibility is the column founders skip. Compute runtime can change later without a customer noticing. The data and identity choices underneath it cannot, which is why the hybrid option earns its place for teams with one workload that needs control and another that does not. The price of hybrid is a boundary to document and two operating habits to maintain.
The same method works outside cloud. Our guide to choosing a transport management system applies weighted requirements, integration tests and exit terms to a different category, and the reasoning transfers.
4. Decide data, tenancy and residency before anything is provisioned
Provisioning commits you. Before anyone clicks create, a fractional CTO walks through the data questions in order. Where does personal data live, and where is it processed? Which region holds the primary copy and the backups, and does that match what your contracts and your users expect? Which subprocessors touch it, from the database vendor to the error tracker to the email provider? How long is it retained, and what does deletion actually do across backups, logs and analytics?
Then the evidence question. If an investor or an auditor asks in eighteen months, what will you be able to show? A written record of region choices, subprocessor lists and retention rules is cheap to keep now and expensive to reconstruct later. Residency is the one-way door most often walked through by accident, because the default region in a signup form is rarely questioned. UK and EU teams tend to have a view on where data should sit, and the technical answer has to match the contractual one. That discussion is general information rather than legal advice; you should check the specific requirements with a qualified professional. The same trail feeds the due diligence pack a buyer will ask for.
Settle the residency question in writing before the first environment exists, and treat the region default in any signup form as a decision rather than a setting. This is general information; speak to a qualified adviser about your specific situation.
5. Set the cost ceiling and the exit terms at the same time
Cost posture and exit terms get agreed in one conversation, while the startup still has bargaining power. Reserved capacity against on-demand is a posture rather than a permanent state: commit where usage is predictable, stay flexible where it is not, and write down what would trigger a re-plan. Then the exit. Portability of data, the format it leaves in, egress charges, notice periods and lock-in clauses are all cheaper to negotiate before the workload is live. Once production traffic runs on a platform, the terms you would like are the terms you no longer have the standing to ask for.
| Option | Cost | Commitment | Best for |
|---|---|---|---|
| On-demand | Highest unit rate, no upfront spend | None, cancel or stop any time | Pre-seed workloads with unproven or spiky usage |
| Reserved or savings plan | Lower unit rate against an upfront or term commitment | One to three years, steady spend required | Predictable baseline traffic after product-market fit |
| Committed spend agreement with exit terms | Negotiated rate, includes egress and notice terms | Multi-year, terms fixed at signature | Series A teams moving production traffic and needing portability |
6. Write the architecture decision record and the trigger to revisit it
The artefact is one page. It names the constraints, the options considered, the choice, the risks accepted with it, and the trigger that would reopen the question. That is the architecture decision record, and it is the difference between a decision and a preference. An outside fractional CTO can write it, having run similar trade-offs before, and can remain neutral. A founder-only decision often lacks this external perspective.
Signals that the record is missing are easy to spot:
- nobody in the room has run this trade-off before
- two founders disagree and the disagreement is settled by seniority
- a one-way door is involved and nobody has named it
- the reasoning exists only in conversation
Triggers keep the choice honest. Revisit when monthly infrastructure spend crosses the ceiling you set, when the first enterprise security questionnaire arrives, when engineering headcount doubles, or when a new region becomes a customer base. Write the trigger into the record, so the review happens on evidence rather than on memory.
7. Run it for 30 days, then decide whether to keep the CTO
Once the architecture is chosen, the next thirty days prove whether the choice holds: the first deploy, the first backup restore, the first cost review, the first on-call rota. Our first 30 days onboarding plan sets out how that period runs. Later, when a live workload has to move, the cloud migration playbook covers the sequence and the risks. None of it holds unless someone owns the drill that proves the architecture still works. If you are making this call now, book a CTO as a Service review and bring your constraints, your current spend and the decisions you have already locked in.
A first cloud architecture is a decision about commitment, not about vendors. It rests on a small number of one-way doors, taken under a funding clock, that a startup must later be able to explain. Run the seven steps: constraints first, reversibility second, options compared on identical criteria, data and residency settled before provisioning, cost and exit terms agreed while there is bargaining power, the whole thing written down, and a date set to revisit it. Founders who skip the sequence rarely fail loudly. They pay quietly, in migration cost and engineering time, eighteen months later.

FAQ
Does a fractional CTO design the first cloud architecture or just review it?
On a first architecture the useful split is that the CTO owns the decision and the written record, while the team owns the build. The CTO sets constraints, runs the option comparison, settles data and residency, and signs off the architecture decision record. Engineers then implement against it. A CTO who only reviews without owning a decision tends to leave the same open question behind.
How long should choosing a first cloud architecture take?
The decision itself is a matter of weeks rather than months for a pre-seed to Series A team, provided the constraints are agreed first. Most of the elapsed time goes on confirming data, residency and contract terms. Fix the review trigger at the same time as the decision, so the choice has a stated expiry rather than an assumed permanence.
Can a founder make this decision without a fractional CTO?
Often yes. The failure mode is choosing without a written record of why, which makes the decision hard to defend to a new engineer, an investor or a buyer during due diligence. A fractional CTO earns the fee where the decision is contested, where a one-way door is involved, or where nobody on the team has run the trade-off before.
How does this differ from the cloud migration work a fractional CTO later does?
Choosing a first architecture commits to as little as possible while the team learns quickly. A migration moves an existing, live workload with uptime and customer commitments in play, which is a different risk profile and a different sequence. Make the first decision with the eventual migration in mind, so exit terms and portability are agreed at the start.