Skip to content
Palantir Eases Model Switching, Yet Top 3 Vendors Take 69% of Enterprise AI Spend
Industry Insights 19 min read

Palantir Eases Model Switching, Yet Top 3 Vendors Take 69% of Enterprise AI Spend

Palantir's new batch model migration feature reduces switching costs, but Menlo Ventures data shows enterprise AI spending increasingly concentrates on Anthropic, OpenAI, and Google, which now capture 69% of deployment spend, while open-source model usage grows but revenue share shrinks, prompting enterprises to evaluate total task costs rather than per-token prices.

DataFunTalk
DataFunTalk
DataFunTalk

Palantir Lowers Model Migration Effort with Batch Replacement and Evaluation

On November 17, Palantir released a bulk model migration capability in Pipeline Builder. Developers can select a single pipeline, multiple projects, or an entire enrollment and map existing model references to a new model. The feature currently supports Use LLM nodes and Text to embeddings expressions. Pipeline Builder scans for pipelines using the source model, generates a Proposal for each qualifying pipeline, and skips those that do not reference the model. Developers must review each Proposal and merge it to the main branch before changes take effect. This preserves change governance while reducing repetitive configuration work.

However, interface compatibility does not guarantee identical business outcomes. A single workflow may involve text extraction, classification, vectorization, and tool calls, and different models handle prompts, output formats, and complex instructions differently. Post-migration issues such as inconsistent results, latency shifts, or cost increases can still arise, requiring task-level validation.

Palantir's earlier AIP Evolve combines model tuning with evaluation. Enterprises set optimization targets (cost, latency, evaluation scores), configure test cases and operational constraints, and AI FDE agents propose model and prompt adjustments. Users can inspect evaluation results, agent execution traces, and change proposals, enabling continuous model selection as task requirements evolve.

Menlo Ventures: Top Three Vendors Capture 69% of Enterprise AI Deployment Spend

Menlo Ventures' "2026: The State of Enterprise AI" report (based on a survey of nearly 400 U.S. enterprise decision-makers, public financials, and third-party data) estimates annualized enterprise generative AI deployment spend at $176.8 billion as of September 2026, up ~5.4x from an adjusted $33 billion in 2025. Anthropic, OpenAI, and Google together account for approximately $122.7 billion, or 69% of the total deployment market, up from 50% a year earlier. This 69% includes both model API services and first-party AI applications sold directly by these vendors.

Menlo splits deployment spend into infrastructure (~$99.5B) and applications (~$77.3B). The three vendors' first-party AI applications are estimated at ~$40 billion annualized, with coding tools (Claude Code, Codex) contributing ~$24B and enterprise chat assistants (Claude, ChatGPT, Gemini) ~$15.5B.

Looking solely at the enterprise LLM API market, concentration is even starker: Anthropic and OpenAI's combined spend share rose from 67% to 86%. Anthropic grew from 40% to 55%, OpenAI from 27% to 31%, while Google fell from 21% to 9%. The market is rewarding models that enterprises actually adopt for production workloads.

Open-Source Usage Rises but Revenue Share Falls

From September 2025 to September 2026, open-weight models' token usage share increased from 11% to 19%, yet their share of model spend dropped from 6% to 4% . Menlo estimates the average token price for open models fell 3.3x over the same period. Conversely, frontier models' token share grew from 18% to 42%, and their spend share from 35% to 63%. Enterprises are processing more tokens with open models but allocating a larger budget proportion to frontier models, indicating that task performance outweighs per-token price in procurement decisions.

Enterprises Shift to Task-Level Cost Calculation

For simple tasks like classification or extraction, per-token price, accuracy, and latency can be compared directly. Agent-driven workflows are more complex: a task may require multi-step reasoning, repeated tool calls, retrieval, error handling, and retries. Two models with identical APIs can differ in the number of calls and time needed to complete the task.

Example: An agent that queries orders, correlates customer data, and generates recommendations must identify intent, select tools, and structure results. A cheaper model that produces more retries may erase per-inference savings. A more capable model with higher per-call cost might reduce total steps and lower overall task cost. Both scenarios require empirical testing.

Therefore, enterprises must calculate model cost within the full task chain: model invocation fees, tool execution, retries, result verification, and any human review costs. Dividing total test-period cost by successful task completions yields an average cost per successful task metric. Comparisons require consistent task scope, success criteria, and cost measurement.

Menlo's data aligns with this engineering need: frontier models' rising token and spend shares show enterprises increasing use of high-capability models. The report also notes that ~80% of tokens on OpenRouter (a developer/startup-heavy platform) come from agent-style calls, signaling that multi-step tasks drive sustained model usage.

Palantir's AIP Evolve places migration, evaluation, and cost optimization in one flow, supporting such comparisons. Technically feasible substitution does not eliminate the preference for models that meet task requirements; complex reasoning and continuous tool-use workloads may continue to command a premium for leading models, while well-defined, easily verified tasks may shift to lower-cost alternatives.

Model Vendors Expand into Applications; Remaining 31% Market Grows Rapidly

Anthropic, OpenAI, and Google now sell both model APIs and enterprise AI applications. Menlo estimates their first-party applications total ~$40B (coding tools ~$24B, enterprise assistants ~$15.5B). Model upgrades directly fuel adoption of these products.

The AI application market reached ~$77.3B in 2026, with coding tools and enterprise chat assistants comprising ~76%. These categories benefit from standardized operating environments (code repositories, enterprise apps) where agents can access context and verify progress via tests or query results. Business processes lacking standardized interfaces and validation methods require more engineering to achieve comparable automation.

Despite the top three's share growth, the non-top-three segment still represents ~$54.1B, up ~3.3x year-over-year. Menlo finds ~55% of new AI spend comes from net-new budgets and ~22% from existing software budgets. The expanding pie allows both incumbents and challengers to grow simultaneously.

This raises the bar for independent AI applications and enterprise platforms. Model vendors provide general-purpose assistants; third-party products must solve business-specific problems: connecting proprietary data, calling system APIs, controlling execution permissions, and verifying outcomes. Customers already paying for base models need additional value from a platform layer. If new software merely forwards model requests, it cannot justify sustained procurement spend.

Palantir, Microsoft, Databricks Compete for the Enterprise Business Entry Point

At Ignite 2026 (November 17), Microsoft introduced Microsoft IQ, organizing Work IQ (collaboration/work info), Fabric IQ (governed business data & semantics), Foundry IQ (enterprise knowledge), and Web IQ (external information) into an enterprise context management system, plus 350+ third-party connectors. Business plugins enable agents to reuse existing business process knowledge, skills, and data sources. Capabilities are at various preview stages.

Days earlier (November 11), Databricks announced general availability of Genie and Genie Ontology, which organizes relationships among enterprise data, metrics, people, and processes, with Unity Catalog and Unity Gateway handling governance and access control. Upcoming persistent Genie and Genie Search aim to let agents continuously execute work across enterprise data and applications.

Palantir centers on its Ontology to model business objects, relationships, and operational rules, with AIP connecting models into existing workflows. Recent Pipeline Builder bulk migration and AIP Evolve let enterprises adjust model configurations per task requirements while reusing established business structures.

Though their foundations differ, all three platforms expand the enterprise context, tools, and executable workflows accessible to agents. Models handle reasoning; platforms must connect reasoning results into live business systems.

This explains how model spend concentration and platform revenue growth can coexist. Model vendors leverage leading capabilities and first-party apps for higher share; data and enterprise software platforms leverage business system integration, permission management, and executable workflows. An enterprise may use Palantir for process management, Databricks for data organization, Microsoft for context access, and still call Anthropic, OpenAI, or Google models. A single business task's cost may be split across multiple vendors.

Ultimately, procurement bills must tie to concrete work: whether the model meets task requirements, how many invocation dollars per successful execution, and whether the platform reduces integration and maintenance effort. Palantir now automates part of model migration but still requires human review and business re-validation. Menlo's report shows enterprises directing more spend to leading model vendors while continuing to expand purchases of other AI products. Replaceable models solve only part of the technical selection problem; enterprises ultimately pay for the model's ability to complete tasks and the engineering work needed to run those capabilities continuously in business systems.

Original Source

Signed-in readers can open the original source through BestHub's protected redirect.

Sign in to view source
Republication Notice

This article has been distilled and summarized from source material, then republished for learning and reference. If you believe it infringes your rights, please contactadmin@besthub.devand we will review it promptly.

Enterprise AIOpen Source ModelsModel MigrationPalantirAI SpendingAgent WorkflowsMenlo VenturesPlatform Competition
Reader feedback

How this landed with the community

Sign in to like

Rate this article

Was this worth your time?

Sign in to rate
Discussion

0 Comments

Thoughtful readers leave field notes, pushback, and hard-won operational detail here.