Build vs buy: the data platform question for 2026

For a mid-market Australian business in 2026, the honest answer to "build or buy the data platform?" is buy the core, build the opinionated layer on top. About 90% of businesses should live here. The remaining 10% genuinely benefit from building end-to-end, and almost all of them know who they are before they ask the question.

The decision isn't binary, and the industry framing as if it is costs businesses millions. Let's break it into the three layers that actually matter, and tell you which ones are commodity and which ones are the thing worth owning.

The three layers, in order

Layer 1: storage and compute. Object storage, table formats, query engines. Commodity. Buy it. This is the most undifferentiated part of the stack, and the vendors (Snowflake, Databricks, BigQuery, AWS analytics, Azure Synapse) compete on price and features in ways that genuinely benefit you. Building this yourself in 2026 is building undifferentiated heavy lifting. Don't.

Layer 2: ingestion and orchestration. Getting data from source systems in, scheduling transformations, managing dependencies. Mostly buy. Fivetran, Airbyte, dbt, Dagster. The tools work, the integrations exist. Where you might build: proprietary source systems that nobody has a connector for, or regulatory environments where hosted connectors won't fly.

Layer 3: the opinionated layer. Your business's data model. Your semantic definitions. Your domain-specific transforms. Your metrics store. The layer that turns raw tables into "revenue at risk" or "sales-qualified lead" or "customer-lifetime-value." Build this. No vendor sells your business. Buying this layer means letting someone else own your definitions, which in practice means letting them own your ability to answer your own questions.

The failure modes cluster at the boundaries. Businesses that try to "buy" the opinionated layer end up with a vendor's opinion about what a customer is, applied to a business where customers work completely differently. Businesses that try to "build" the storage layer end up two years in, on a bespoke platform no new hire can operate.

When buying wins harder

• Speed to value matters more than TCO. Buying a managed warehouse gets you to first dashboard in six weeks. Building gets you there in six to twelve months. If the business case has a window, buying wins even if the five-year cost is higher.
• Fewer than 50 data users. Custom platforms create an operating burden. At small scale, the ops overhead of a build is a tax you can't amortise.
• Your data team is five or fewer. Every hour your senior data engineers spend operating the platform is an hour they're not spending on the opinionated layer, where the business differentiation lives.
• Cloud data spend is under $200k a year. The economics of building to save money don't bite until the warehouse bill is a real line item on the P&L.

When building (parts of it) wins

• Severe cost pressure at scale. If you're spending $5m+ a year on warehouse compute, the economics of an open lakehouse start to matter. Meta and Netflix built their own for good reasons; none of those reasons apply to a business with a $400k data bill.
• Proprietary workflow that's core to the product. If your data platform is the product (data-as-a-service, analytics-as-a-service companies), buying a generic platform and hoping to differentiate on top is building your business on rented land.
• Regulatory or sovereignty requirements that vendor managed services can't meet. Rare, but real. Australian government work, some health data, specific defence scenarios.
• You have a team of 20+ platform engineers who'd otherwise be idle. The operating cost of a built platform is the full-time attention of a real team.

Buying gets you to adequate fast. Building gets you to excellent slow, and only if you have the team to operate it for a decade afterwards. Most businesses don't need excellent. They need the opinionated layer on top of adequate, shipped this quarter.

The hybrid pattern that works

Here's the default we recommend for 2026 mid-market:

1. Managed warehouse or lakehouse. BigQuery, Snowflake, or Databricks. Pick based on your existing cloud and your team's skills.
2. Managed ingestion. Fivetran for SaaS sources, Airbyte self-hosted for the weird ones, custom ingestion only for the handful of sources with no existing connector.
3. dbt for transformation. Open source, version controlled, testable, standard. The closest thing the industry has to a consensus.
4. A semantic layer you own. Cube, dbt Semantic Layer, or a custom YAML-driven definition registry. This is where "what is a customer" gets written once and everyone uses the same answer.
5. Opinionated domain models in dbt. Your revenue model, your retention model, your inventory model, your usage model. Written by your team, reviewed by the business, tested every night.

That pattern costs somewhere between AUD $15k and $45k per month all-in for a mid-market business (licences plus cloud plus a small platform team), and it gets you to a place where new AI use cases plug in in weeks, not quarters.

The cost comparison honest people don't show you

A fully managed setup for a mid-market business:

• Warehouse: $5k–$20k per month.
• Ingestion: $1k–$8k per month.
• Orchestration + transformation: $500–$3k per month.
• Semantic layer: $0–$3k per month.
• People (one platform engineer + half a senior analytics engineer): $22k–$30k per month fully loaded.
• Total: $28k–$64k per month.

A built-from-scratch open-source setup for the same business:

• Cloud storage and compute: $3k–$15k per month (yes, cheaper).
• People (three platform engineers, one SRE): $45k–$70k per month fully loaded.
• Operational cost (on-call, incident response, upgrade debt): another $5k–$15k per month in team time.
• Total: $53k–$100k per month.

The "build is cheaper" argument only works if you don't count the team. Which is exactly what the build-your-own posts on Hacker News don't count.

The rule we give clients

If you're choosing between build and buy, the answer is almost certainly buy the three infrastructure layers, build the opinionated layer, and let the vendor deal with the boring part. The businesses that should build everything know it before they ask. The ones asking "should we build it?" are the ones who shouldn't.