Every quarter, some version of this meeting happens: a CIO has a Databricks or Snowflake footprint growing faster than anyone forecast, a backlog that is now measured in quarters rather than sprints, and two slides in front of her. Slide one is a global integrator’s statement of work. Slide two is a hiring plan for nine permanent engineers. She is being asked to pick. The spending data says neither slide, on its own, is a plan — because the thing both are meant to absorb is growing faster than either instrument can move.
The platform side is outrunning both of your options
Start with the demand signal. As-a-service capacity spending jumped 65 percent year over year to a record $31.5 billion in a single quarter. That is the platform you already bought, or are about to. It is contracted, it is metered, and it starts accruing value — or cost — the day it turns on.
Now look at what you are producing on top of it. Organizations are putting 11x more models into production than a year ago while getting 3x more efficient at deploying them. Read those two numbers together: efficiency tripled and volume still went up an order of magnitude. Tooling absorbed part of the shock. It did not absorb most of it.
An eleven-fold change in delivery volume breaks any fixed-shape resourcing decision. A single large SOW is fixed in shape — it was scoped against last year’s backlog, it has change-order friction built into it, and it cannot flex down when a workstream finishes early. A permanent squad is fixed in a different way: it takes two quarters to assemble and it is politically very hard to unwind. Both assume the work has a stable size. It does not.
The mega-program is flat while the platform sprints
Here is the part that should settle the debate. Large-contract managed services — the classic integrator mega-program — grew 2.7 percent year over year to $10.9 billion in quarterly ACV, and actually declined sequentially from the prior quarter. Set that against 65 percent growth in the capacity it was supposed to support.
That gap is not a story about integrators being bad at their jobs. It is a story about instrument fit. The multi-year managed-services contract was designed for work with a knowable five-year shape: a data centre migration, an ERP rollout, a steady-state support tower. Lakehouse platform work in 2026 does not have that shape. The scope changes when a model gets promoted, when a governance finding lands, when a business unit shows up with a use case nobody had on the roadmap. Enterprises are not buying fewer mega-programs because they have stopped needing help. They are buying fewer because the help they need arrives in smaller, faster, more specific units than that contract vehicle can issue.
The integrator did not get worse. The work got faster than the contract.
The hiring plan cannot absorb it either
The natural reaction is to insource everything. Then you look at the budget. 57 percent of teams report increased warehouse and compute spend, against just 36 percent reporting increased team budgets. Compute is winning the internal fight for money by roughly twenty points. The practical consequence: the headcount line is the one your CFO is already squeezing to pay for the platform line. You are being asked to staff a bigger backlog out of a budget that grew slower than the infrastructure bill.
Even if the money appeared, the market would not cooperate. Data scientist roles are projected to grow 34.6 percent over the decade, placing them among the top ten fastest-growing occupations in the US, with computer and information research scientists at 21.8 percent. You are not competing for this labour in a loose market; you are competing in the tightest segment of it, with a comp band set by companies whose entire product is the thing you are trying to staff. A nine-person permanent build-out is a two-to-three quarter recruiting program before it is an engineering one — and the platform meter started running in month one.
And leadership knows the internal bench is not ready. Only 27 percent of executives report a comprehensive AI strategy and just 20 percent believe their workforce is truly AI-ready. Twenty percent is the honest argument for buying capability from outside. It is also the argument for making sure that capability transfers rather than leaving when the invoice stops.
The blocker is ownership, which is why full hand-off backfires
The failure mode of outsourcing used to be technical. It isn’t any more. Technical integration challenges fell from 35 percent to 27 percent year over year, while ambiguous data ownership now sits at 41 percent as a persistent obstacle, alongside poor data quality. Plumbing got easier. Deciding who is accountable for a table, a metric definition, or a model’s drift got harder.
A full delivery hand-off makes the 41 percent worse, almost mechanically. When an external team owns the pipelines, the semantic layer and the model registry, ownership ambiguity is not a bug in the arrangement — it is the arrangement. Nobody inside can answer why a number changed without filing a ticket. Decisions that should take an afternoon take a sprint. And the knowledge that would let you decide faster next time is accumulating on someone else’s balance sheet.
That is the real reason “GSI or DIY” is the wrong frame. The question is not who does the work. It is who owns the decisions — and then, separately, who supplies the hands.
The operating model that actually fits: thin core, embedded specialists, scoped integrator
Split the problem into three layers and staff each with the instrument that fits it.
A thin internal core that owns. Small — often four to eight people. Platform architecture, data contracts and ownership mapping, security and governance posture, the build-versus-buy calls. These are permanent hires, and they are the only roles where a two-quarter recruiting cycle is worth paying. This layer does not need to be big enough to deliver the backlog. It needs to be senior enough that nobody outside the company gets to decide what “correct” means.
Embedded specialists who deliver. Currently-employed practitioners who have done your exact migration, your exact Unity Catalog rollout, your exact reverse-ETL build, brought in against a named outcome and a named end date. They sit inside your sprints and report into your core — not into a delivery manager with a utilisation target. This is where the 11x volume gets absorbed, because this layer scales up and down in weeks instead of quarters. If you are sizing this layer, our guide to hiring Databricks developers covers what the specific skill signals look like.
An integrator kept to what it is genuinely good at. Not everything, and not the core. Specifically: the work that is wide rather than deep, that needs twenty bodies on a predictable task for six months, that spans geographies and time zones, or that carries regulatory sign-off requirements a boutique cannot indemnify.
Which layer gets which work
| Work | Thin internal core | Embedded specialists | Integrator |
|---|---|---|---|
| Data contracts, ownership mapping | Owns | Advises | No |
| Platform architecture decisions | Owns | Advises | No |
| Net-new pipeline and model build | Reviews | Delivers | Overflow only |
| Migration of 400 legacy reports | Scopes | Hard cases | Delivers |
| Multi-region, multi-language rollout | Sets standards | No | Delivers |
| Governance and quality remediation | Owns | Delivers | No |
| 24/7 steady-state support tower | Sets SLOs | No | Delivers |
| Regulatory attestation work | Owns | No | Delivers |
What each instrument actually costs you in time
| Instrument | Time to first useful work | Flex down | Where knowledge ends up |
|---|---|---|---|
| Permanent hire | 2–3 quarters | Very hard | Inside — if they stay |
| Embedded specialist | Weeks | Contract end | Inside, if transfer is scoped |
| Large managed-services SOW | 1–2 quarters | Change order | With the vendor |
Read the middle row. Weeks to start, clean exit, knowledge retained — but only if you scoped the transfer. That conditional is the whole job.
Making the transfer a deliverable, not a hope
With 20 percent workforce readiness, every quarter you buy delivery capacity is also a quarter you could be buying capability. Most organizations only get the first one, because nobody wrote the second into the engagement.
Four things to put in writing before anyone starts:
- A named internal owner per workstream. Not a steering committee. One person in the thin core whose name is on the artifact after the specialist leaves. This is your direct countermeasure to the 41 percent ownership problem.
- Review rights, not just delivery. Your core reviews every merge. That is slower in week two and dramatically faster in month six, because the people who will maintain the thing have read it.
- Documentation as an acceptance criterion. Runbook, lineage, decision log. If it is not written down, the work is not complete — and that applies to the integrator’s towers as much as to individual specialists.
- A declared end date on every augmented seat. Not because you will always honour it, but because an engagement with no end date quietly becomes a dependency with no owner.
The pattern is the same whether the specialist work sits in data engineering or in GTM engineering, where the same squeeze shows up in revenue systems rather than pipelines.
How to sequence the next quarter
If you are holding those two slides, do this instead.
First, take the backlog and sort every item into the three columns above. Most teams discover that only 20 to 30 percent of it genuinely belongs to an integrator — and that a large slice of what they were about to put in the SOW is specialist depth work that an integrator would staff with generalists anyway.
Second, hire the core before you hire anything else, and keep it small. Senior, opinionated, permanent, and few. Every seat you add here beyond the minimum is a seat you will struggle to fill in a market growing 34.6 percent, and a seat you cannot flex when the backlog shape changes.
Third, bring embedded specialists against the deep work with named outcomes and end dates. Speed matters here precisely because compute spend is already accruing while the hiring plan runs.
Fourth, negotiate the integrator down to the wide, predictable, geographically-dispersed and attestation-bearing scope. That conversation is easier than it used to be — the market is telling you the same thing your backlog is. Worth knowing too that the best boutiques are frequently full; if you have run into that, we wrote about why specialist delivery shops turn down work.
The choice was never GSI versus DIY. It is: who owns the decisions, who supplies the depth, and who absorbs the volume — three questions with three different right answers.