Eighteen months of consolidation headlines have settled into a rhythm: another tool acquired, another team folded into a lab, another platform feature that used to be a startup. The pattern itself is old news. What the headlines never answer is the question sitting in your renewal calendar: which layers of this stack are now stable enough to standardize on, and which are still moving too fast to marry? The record of the last year is long enough to answer it layer by layer.
Read the mortality table before the feature table
Buyers compare features. The consolidation record says to compare survival odds first, because in several layers the modal outcome of the past year was a change of owner.
The evaluation and observability layer had the roughest run. Humanloop, a prompt-management and evals vendor, told customers it was shutting down before Anthropic hired its team in August 2025. Weights & Biases went to CoreWeave that May. Observe, an AI observability platform, closed into Snowflake in February 2026. Three exits in ten months, and in each case customers held a contract with a company whose product was no longer the buyer’s priority.
Coding tools churned just as hard at the vendor level. Windsurf was carved up inside one week in July 2025: OpenAI’s reported $3 billion acquisition collapsed, Google paid a reported $2.4 billion to license the technology and hire the founders, and Cognition bought the product, brand, and remaining team three days later. Cursor, the category’s darling, switched its Pro plan from request limits to usage credits mid-subscription in June 2025, then apologized and issued refunds. The tools survived. The assumptions buyers signed under did not.
Even the infrastructure layer is mid-transition: Qualcomm announced in June 2026 that it will acquire Modular, the write-once-run-anywhere inference stack, with the close expected in the second half of the year. Modular’s customers will spend those months inside someone else’s ownership change.
Commit to the model tier, not the model vendor
The model layer looks like the obvious place to standardize. It is, with one correction: the stable unit is the top tier, not any single lab inside it.
Menlo Ventures’ November 2025 survey of 495 US enterprise AI decision-makers found three vendors (Anthropic, OpenAI, Google) carrying 88 percent of enterprise LLM usage. That tier is durable. The ranking inside it is not: in two years, by the same survey series, OpenAI’s share fell from 50 to 27 percent while Anthropic’s rose from 12 to 40 and Google’s from 7 to 21. Anyone who standardized on the 2023 leader bet against the trend; anyone who standardizes on the 2026 leader may be doing it again.
The practical posture is the one enterprises are already converging on: a16z’s 2025 CIO survey found 37 percent running five or more models, up from 29 a year earlier, precisely because leadership keeps rotating. The same survey names the tax you avoid by preparing early: prompts tuned for one provider make switching “a task that can take a lot of engineering time.” Commit to the tier, wire at least two of the three, and keep your prompts and your eval set provider-neutral. Whether you should also be running open weights is its own budget question, and the arithmetic gets its own piece.
The other layer that deserves permanence is the one nobody can acquire out from under you: your own data pipeline, your evaluation sets, your documented workflows. Every consolidation event above made someone’s tooling temporary. None of them touched anyone’s test cases.
Rent whatever the platforms are absorbing
The second stable-looking trap is the plumbing between your data and the model, because the platforms are eating it from below.
AWS made vector storage a native S3 feature in December 2025, at up to 2 billion vectors per index and with a claimed cost reduction of up to 90 percent against specialized vector databases (the claim is AWS’s own; the direction is what matters). Google shipped File Search in November 2025, a fully managed RAG pipeline inside the Gemini API: storage, chunking, embeddings, retrieval, all abstracted away. And OpenAI is deleting its own first attempt at this layer, removing the Assistants API on August 26, 2026 in favor of its newer Responses API, a reminder that even platform abstractions churn on twelve-month clocks.
The buyer’s read: for commodity retrieval over your own documents, the standalone tool you commit to today is competing with a feature your cloud will ship for near-free tomorrow. Rent this layer. The exception is a workload with genuinely specialized retrieval needs at scale, and if you have one, you already know.
Orchestration frameworks earn the same caution for a different reason. LangChain, the most-adopted framework, only reached 1.0 in October 2025, moved its legacy patterns into a side package, and made its first no-breaking-changes promise on that day. The promise may hold. But a stability pledge eight months old is a reason to keep the framework at arm’s length behind your own interfaces, not a reason to weld it into everything.
Make the contract carry the churn
Whatever you do adopt, the contract is where the churn risk should live, and the market data says vendors are pushing it the other way.
Tropic’s benchmark of more than $18 billion in analyzed software spend found the average AI-native contract term stretched to 22.4 months by early 2026, while AI vendors open renewal negotiations asking for 20 to 37 percent increases, against a 3 to 9 percent norm for traditional software. Buyers who negotiate cut those asks roughly in half, landing near 12 percent. (Tropic sells procurement services, so the negotiation win is also its pitch; the contract-term drift is the useful reading either way.) Meanwhile Zylo’s 2026 index of $75 billion in managed SaaS spend found 78 percent of IT leaders hit unexpected charges from consumption or AI pricing in the past year. Longer terms, higher asks, less predictable meters: every one of those shifts transfers churn risk from the vendor to you.
Two clauses matter more than any discount. First, term length matched to layer stability: the model tier can carry a year, a tool in the mortality layers should stay month-to-month or quarterly even at a worse rate. Second, exit rights over your artifacts. The labs’ own documentation is blunt here: an OpenAI fine-tune cannot be downloaded and lives only until its base model is deprecated, and base models retire fast (gpt-4-0613 shuts down October 23, 2026; Anthropic retired models released barely a year earlier, on a minimum of 60 days’ notice). Anything you cannot take with you is not an asset. It is a deposit.
The standardization test
Four questions, asked before any AI renewal or new commitment, sort nearly every case:
- Who could own this vendor in eighteen months, and would the product still be their priority? If the honest answer is “a lab or a cloud, and no,” rent it.
- Is a platform shipping this layer as a feature? If your cloud or your model provider announced the commodity version this year, do not sign multi-year for the standalone version.
- Can we leave with our artifacts? Prompts, evals, embeddings, fine-tuned weights, logs. If any answer is no, the price is not the price.
- Does the term match the layer? Tier-level commitments for the stable layers, short paper for the churning ones, and a renewal cap in writing before the 20 to 37 percent ask arrives.
Standardize on the things that survive acquisition: the three-lab tier, your own data and evals, and the discipline of holding everything else at contract’s length. The rest of the stack has not finished consolidating, and your signature should not pretend it has.