The end of the stack

Enterprise IT architecture exists to work around two constraints. Integration is expensive. Compute is centralised. The modern enterprise stack, from the data warehouse at the core to the integration middleware at the edge, is a 20-year response to those two facts. Both are dissolving. What replaces the stack is smaller, more federated, and more responsive than what it replaces.

This thesis is not novel in its components. The death of the data warehouse, the compression of middleware, the displacement of enterprise BI, the obsolescence of target-state architecture: each has been written about in isolation. The argument here is that they are the same event, they reinforce each other, and the compounding changes the timeline.

The central move

Intelligence has stopped being scarce. A model that reads a schema, writes a transformation, handles edge cases, and posts a result costs a few dollars per run and minutes to stand up. The economics of integration shift from capital project to commodity call.

This does not mean every enterprise transaction now flows through an LLM. It means the expensive thinking happens once, at design time, and the resulting deterministic pipeline runs cheap at volume.

The architecture splits into three layers.

• Design-time intelligence. The model reads the source system, understands the target, writes the mapper, identifies the exceptions, produces working code. One-off cost, bounded.
• Runtime determinism. The generated pipeline runs on ordinary infrastructure at ordinary cost. Same as any middleware today, without the consulting invoice that built it.
• Intelligence in the loop. Called only where it adds value: exception handling, interpretation of unstructured content, judgement calls, anomaly commentary. The model touches the 1% of events that need it, not the 99% that do not.

This is the shape. Everything below follows from it.

What I am not claiming

I am not claiming that existing enterprise systems disappear. SAP runs 77% of the Fortune 500. Oracle Financials, Workday, core banking platforms: these installed bases persist for a decade or more regardless of what happens at the architectural layer above them. The systems of record are not at risk. The integration, middleware, and analytical layers built on top of them are.

I am also not claiming this happens uniformly. The pattern is barbell. Experimental and edge workloads (management dashboards, internal tools, federated queries, operational reporting) move fast. Core transactional systems and regulatory pipelines move slowly, and for good reasons. Net new build tilts toward the new pattern; existing build persists.

The timeline question is therefore not "when does the old stack die" but "when does net new enterprise IT spend cross over from old pattern to new." My answer is 2027-2029 for most categories, later for regulated cores.

What breaks

Integration middleware. Gartner sized iPaaS at $9bn in 2024 and forecasts $17bn by 2028 on a consensus growth path. The contrarian position is that 2028 is closer to the peak than the base camp. The top five vendors (Salesforce/MuleSoft, Oracle, Informatica, SAP, Boomi) capture 58% of the market selling a solution to a problem that becomes cheaper to solve in-house with an LLM and a connector catalogue. Expect the consensus forecast to be downgraded by 2027.

Data warehousing as the default. The Snowflake and Databricks thesis was: move all your data here, query it in one place. When queries can reach to the data in situ, through connector protocols and scoped access tokens, centralisation becomes optional. It remains correct for high-volume analytical workloads and model training. It stops being the right answer for operational reporting, management dashboards, and ad-hoc questions, which is most of what actually gets built. Centralisation becomes a choice made for specific workloads, not the default architecture.

Enterprise architecture as drawing exercise. Target-state diagrams, three-year roadmaps, canonical data models, architecture review boards. All artefacts of a world where rewiring was expensive and irreversible. When a new flow costs an afternoon, the correct posture is grow-as-needed, not plan-before-build. Architecture becomes editorial, not constructive: decide what is worth having, not what the topology should look like in 2029.

Management information software. Anaplan, Workday Adaptive, Oracle EPM, Tableau-plus-consultancy. The value proposition was schema plus renderer. LLMs write both. Expect the category to compress hard as CFOs replace six-figure line items with dashboards their finance team maintains in-house.

The transformation project. Three-month consulting engagements existed because three months was the shortest credible delivery unit. When the work takes four days, the billing model breaks. The bill-by-hour advisory function survives only where the constraint is genuinely judgement, not delivery.

What survives and grows

Systems of record. More important, not less. If the query reaches to the source, the source must be correct, current, and well-governed. The incumbent ERP and CRM platforms find their moat deepening as federation becomes the access pattern.

Governance and guardrails. Security, audit, data residency, regulatory compliance, access control. These constraints do not dissolve. They matter more when the rate of new flows accelerates. The Chief Architect role converts to Head of Platform Engineering: someone who runs the guardrails and the observability, not someone who draws diagrams.

Judgement about what is worth building. The bottleneck moves from "can we build this" to "should we." Editorial capacity becomes the scarce resource.

Who profits

If the old stack compresses, somebody collects the newly available value. Three pools.

The model providers. Anthropic, OpenAI, and whoever else clears the frontier bar. The design-time intelligence layer is a direct revenue replacement for consulting and integration work. A meaningful fraction of the $400bn global systems integration market becomes model inference revenue over the next decade.

The systems of record. SAP, Oracle, Salesforce, Workday. Their data becomes the queryable surface that the federated architecture runs against. Clean APIs and well-documented schemas become a competitive moat they already partly have.

The governance layer. Okta, HashiCorp, specialised data governance and observability vendors. When integration is cheap and flows proliferate, knowing what is running, who has access, and where the data sits becomes the scarce capability.

Notably absent from this list: the iPaaS vendors, the data warehouse pure-plays, the BI tool vendors, the integration consultancies.

Federation between companies

Inside a company, data sharing is a political problem dressed as a technical one. Between companies, it has been genuinely technical. No trust model, no standard for ephemeral access, no way to expose a narrow query surface without a bilateral integration project.

The pieces are now present. MCP-class protocols provide the wire format. OAuth-scoped tokens and verifiable credentials provide the trust model. Just-in-time connectors mean the pipe is stood up for the decision, used, and torn down. A supplier's inventory planning agent queries your production forecast at the moment it matters, scoped to the SKUs it buys, with a 24-hour access window. No shared database. No master agreement beyond a standing trust relationship.

Supply chain visibility, ecosystem-level intelligence, federated benchmarking: buildable for the first time, because the transaction cost of narrow, time-bound, policy-bound data sharing drops to near zero.

The clock changes too

The quarterly beat of enterprise IT existed because that was the shortest unit in which anything could happen. Scope, build, test, deploy. A year to know if an idea was worth having.

The new clock is days. Hours at the edge. Teams that internalise this run fifty learning cycles for every one their slower competitors complete. The gap between a fast organisation and a slow organisation stops being 2x and becomes 20x, because compounding. The architecture diagram matters less than the metabolic rate of the organisation using it. The correct response for an operator is not to wait for the architecture to settle. The architecture will not settle for a decade. Run personal experiments at the new clock speed, immediately, against real problems. The scarcest knowledge in any organisation is direct, recent, first-hand experience of what this stack can actually do on a Tuesday afternoon.

The case against

Three serious objections.

Enterprise inertia wins for longer than bulls think. Banks still run COBOL. Large enterprises maintain technology a generation behind frontier for reasons that include risk management, audit requirements, vendor relationships, and institutional capacity for change. The new stack may be strictly better and still fail to displace the old one inside the decade. Counter: this argument predicts the failure of the specific predictions below but not the direction of travel. The barbell holds regardless.

The reliability gap is real and under-acknowledged. LLM-generated integrations have a known failure mode: silent drift. The source schema changes, the generated mapper keeps running, the output is subtly wrong for weeks before anyone notices. This is not a theoretical risk. It is the argument against using this pattern for anything that touches money or regulation. Counter: this argues for keeping deterministic test suites and schema contracts around generated pipelines, not against the pattern itself. But it does mean the technology stays on the edges longer than the architecture optimists claim.

Regulated industries have veto power. GDPR, HIPAA, banking secrecy laws, export controls do not care whether the data moved or was queried in place. A federated query returning personal data is still a data transfer. The legal scaffolding required to make cross-company federation routine is a 5-10 year project at regulatory pace. Counter: accepted. Regulated cross-company federation lands in year 5-10 and only in jurisdictions that explicitly bless the model. Internal federation and non-regulated cross-company federation land much sooner.

Falsifiable predictions

1. By end-2027, Gartner or a comparable analyst downgrades its iPaaS 2028 forecast by 20% or more from the current $17bn consensus.
2. By end-2028, new-build corporate data warehousing projects decline 30% from 2024 peak as federated-query patterns displace them for operational and management-reporting workloads. Snowflake and Databricks pivot successfully into compute-for-inference. Fivetran and dbt do not.
3. By end-2029, the Enterprise Architect role as currently defined shows 40% or more decline in job postings from 2024 peak, replaced by smaller, sharper platform-engineering roles.
4. By end-2027, at least one top-five global systems integrator (Accenture, Capgemini, Infosys, TCS, Wipro) publishes a material restructuring of its integration practice, framed as "AI-accelerated," read as managed decline.
5. By end-2030, federated queries between non-competing companies are a standard pattern in at least three industries: supply chain, financial services correspondent networks, healthcare payer-provider data exchange.

Any one of these failing to land inside a year of the stated window weakens the thesis. Three or more failing invalidates it.

Close

The architecture of the future is being built by the people who stop planning it and start shipping it. Everything else is legacy defending ground.