Understanding Why Enterprise Spending on Open-Source Models Concentrates on a Few Vendors

    An industry survey of enterprise AI leaders reveals how spending on open-source models concentrates on a small number of providers, why usage share and spending share can move in opposite directions, and how companies can build a procurement and cost-governance framework that balances scale, stability, and flexibility.

    1. The Phenomenon: Open-Source Model Spending Is Concentrating on a Few Providers

    An industry survey of enterprise AI leaders found that roughly 89% of spending related to open-source models flows to the top three model providers. The largest single provider accounts for about 61%, the second for about 15%, and the third for about 11%.

    These numbers invite an intuitive conclusion: when enterprises choose open-source models, they are not “spreading their bets evenly” but concentrating heavily on a handful of vendors that are stronger in capability, more competitive on price, and more mature in service. In other words, even though the open-source model market looks abundant on the supply side, real enterprise purchasing behavior still shows a clear tendency toward concentration at the top.

    To understand this phenomenon accurately, though, you cannot look at the “89%” figure in isolation. Three questions really need to be answered:

    1. What is the denominator behind that 89%?
    2. Why does the spending share fall even as more enterprises adopt open-source models?
    3. Why does the budget concentrate on a few vendors rather than spreading across many providers?

    Only by breaking these three questions apart can you avoid misreading a partial data point as an overall trend.

    2. First, Define the Denominator: 89% Is Not Total Enterprise AI Spending

    Many misreadings come from an unclear denominator.

    The 89% figure does not mean that “89% of an enterprise’s entire AI budget is spent on open-source models,” nor does it mean “89% of total AI spending flows to a few vendors.” Its more accurate meaning is:

    Among the spending that enterprises direct toward open-source models, about 89% is paid to a small number of model providers.

    In other words, the denominator is “open-source model spending,” not “total AI spending.”

    A simple formula makes this clear:

    1. Open-source model spending share = Open-source model spending / Total enterprise AI spending
    2. A given vendor's share = That vendor's open-source model spending / Total open-source model spending

    The same survey provides a larger budget context: total enterprise AI spending is expected to reach roughly $176.8 billion, including about $77.3 billion in AI application marketplaces, about $34 billion in coding-oriented AI applications, and about $25 billion in enterprise AI chatbots, among others.

    Open-source model spending is just a small slice of this large pie. Misreading it as “the vast majority of the enterprise AI budget goes to open-source models” would seriously overstate its actual impact.

    More noteworthy, the survey also shows that the usage share of open-source models is rising—from about 11% to about 19%—while the spending share is declining, from about 6% to about 4%. This points to a key trend:

    More enterprises are using open-source models, and usage scenarios are multiplying, but the average procurement unit price has dropped noticeably.

    A more rigorous understanding, then, is:

    Enterprises are using open-source models more frequently, but because of falling prices, cost optimization, and changes in task structure, the overall spending share has actually contracted. And within that contracted pool of spending, most of it has concentrated in the hands of a few leading providers.

    That is where the concentration phenomenon truly deserves study.

    3. Why More Usage, But Less Spending?

    “Rising usage share, falling spending share” is not a contradiction. There are usually four reasons behind it.

    3.1 Unit Token Prices Keep Falling

    The cost of model calls can be abstracted as:

    1. Total cost ≈ call volume × average input tokens × input unit price
    2. + call volume × average output tokens × output unit price

    If call volume grows while input and output unit prices fall, total cost can very well decline.

    For example:

    1. Original state:
    2. Call volume = 1,000,000 calls
    3. Average cost per call = $0.002
    4. Total cost = $2,000
    5. New state:
    6. Call volume = 2,000,000 calls
    7. Average cost per call = $0.0005
    8. Total cost = $1,000

    Call volume doubles, but cost is halved.

    This is exactly what commonly happens in the open-source model and enterprise-grade model service market: as inference infrastructure matures, model compression and quantization techniques become widespread, and batch-processing capacity improves, the unit cost per call keeps falling.

    3.2 Inference Optimization Reduces Actual Consumption

    In real enterprise usage, not every request is billed at maximum context and maximum output length. Many scenarios can lower costs through the following approaches:

    • Streamlining prompts;
    • Shortening the context window;
    • Capping output length;
    • Using structured output to reduce wasted generation;
    • Enabling caching for repeated questions;
    • Moving low-priority tasks to offline batch processing;
    • Routing simple tasks to low-cost models.

    These engineering optimizations significantly reduce the hidden costs beyond the “nominal unit price.”

    3.3 Enterprise Tasks Shift from Complex Generation to Structured Processing

    Early enterprise use of large models often centered on complex tasks such as long-text summarization, open-ended Q&A, and creative generation. As adoption deepens, more tasks take on structured, evaluable, batch-processable forms, such as:

    • Information extraction;
    • Tag classification;
    • Form filling;
    • Ticket routing;
    • Code assistance and completion;
    • Log summarization;
    • Document paragraph rewriting;
    • Knowledge-base Q&A.

    These tasks tend to have shorter inputs and outputs, more controllable results, and fewer retries, so the per-call cost is lower.

    3.4 Market Competition Drives Prices Down

    When multiple model providers offer models with similar capabilities, price competition intensifies. To expand usage, vendors often offer:

    • Tiered pricing;
    • Volume-call discounts;
    • Reserved-capacity pricing;
    • Free tiers or trial quotas;
    • Commercial discounts for long-term customers.

    As a result, enterprises may make more calls while their average cost still falls.

    4. Why Does the Budget Concentrate on a Few Vendors?

    Falling prices are only the backdrop. Budget concentration has deeper causes.

    4.1 Leading Models Are More Reliable on Critical Tasks

    When enterprises procure models, they are not just asking “does it work?” but “can it be used reliably in production?”

    Complex tasks place higher demands on model capability, for example:

    • Long-context understanding;
    • Multi-turn conversation consistency;
    • Complex code generation;
    • Structured JSON output;
    • Tool calling;
    • Long-document summarization;
    • Multilingual processing.

    If a model lacks stability in real business scenarios, enterprises often face more hidden costs:

    • Retry costs;
    • Manual review costs;
    • Costs of correcting wrong results;
    • User complaint costs;
    • Costs of business-chain interruptions.

    So even if a low-price model is cheaper on a single call, an enterprise may be more willing to pay a premium for a stable, high-quality model.

    4.2 Enterprise Procurement Naturally Favors Scaled Suppliers

    Enterprise-grade procurement differs from individual developer trials and places more weight on:

    • Service availability;
    • Response stability;
    • Quota management;
    • Fault recovery;
    • Billing clarity;
    • Contract compliance;
    • Data security;
    • Audit support.

    This naturally pushes purchasing behavior toward a few mature providers. For large enterprises, choosing multiple small or uncertain suppliers often means higher governance costs.

    4.3 Migration Costs Are Higher Than the Surface Cost

    After connecting to a model, many enterprises build an entire engineering system around it, including:

    • Prompt templates;
    • Evaluation sets;
    • Routing strategies;
    • Caching logic;
    • Logging and monitoring;
    • Cost reports;
    • Access controls;
    • Business system integration.

    Switching models is not just changing a call endpoint; it may involve extensive regression testing and effectiveness validation. The higher the migration cost, the more enterprises tend to keep using an already-validated provider.

    4.4 Concentrated Procurement Aids Unified Governance

    For finance and platform teams, concentrated procurement is easier to manage:

    • Fewer contracts;
    • Simpler invoicing and reconciliation;
    • Clearer budget attribution;
    • More unified usage tracking;
    • More centralized security review;
    • Lower supplier management costs.

    So even if the technical department wants to maintain diversity, the finance and procurement systems may still push the budget toward a few vendors.

    5. The Gap Between 87% and 89%: Not Something to Over-Interpret

    The survey mentions that the top three shares are about 61%, 15%, and 11%, which add up to 87%. The total share is 89%, leaving a gap of about 2 percentage points.

    This gap should not be over-interpreted.

    5.1 Rounding Introduces Error

    If each share has minor rounding, for example:

    1. The actual figures might be:
    2. 61.4% + 15.2% + 11.1% = 87.7%

    Combined with other spending, arriving at 89% is entirely normal.

    5.2 Statistical Scope May Include Additional Fees

    What enterprises actually pay may include more than the model call itself, such as:

    • Hosted inference fees;
    • Dedicated resource fees;
    • Technical support fees;
    • Quota expansion fees;
    • Compliance and audit services;
    • Data transfer or storage related fees.

    Different scopes can cause the shares not to align exactly.

    5.3 Model Versions and Product-Line Boundaries May Differ

    A single vendor may offer multiple versions, specifications, and regional services. If the survey only consolidates by “provider,” while some sub-products are counted separately, differences can also arise.

    5.4 Sample Composition Affects the Result

    Survey respondents are typically enterprise AI leaders, and different enterprises differ in industry, size, and stage. Large enterprises often have bigger budgets and higher call volumes, and their model-service choices are more concentrated, which amplifies the share of leading vendors.

    So a 2-percentage-point difference is better used to understand that “the statistical scope is not fully consistent” rather than to speculate about any specific vendor.

    6. What Kind of Model Procurement Evaluation Framework Should Enterprises Build?

    Facing the concentration of open-source model spending, enterprises should not merely ask “who is cheapest” but build a quantifiable evaluation framework.

    6.1 Segment by Business Scenario

    Different tasks place different demands on models and should not be judged by the same standard.

    They can be divided as follows:

    Scenario type Typical tasks Key metrics
    High-value core scenarios Customer decisions, complex code, critical documents Accuracy, stability, latency
    Batch back-office scenarios Tagging, extraction, summarization Cost, throughput, consistency
    Real-time interaction scenarios Customer service, Q&A, assistants Latency, availability, experience
    Experimental validation scenarios Exploring new features Flexibility, iteration speed, cost ceiling

    6.2 Establish Unified Evaluation Metrics

    You cannot rely on subjective impressions alone; you should build repeatable evaluation:

    • Task success rate;
    • Structured-output compliance rate;
    • Hallucination rate;
    • Average latency;
    • P95 latency;
    • Retry rate;
    • Cost per task;
    • Manual correction rate;
    • User satisfaction;
    • Security and compliance hit rate.

    6.3 Use Total Cost Instead of Unit-Price Comparison

    A low unit price does not equal a low total cost.

    It is advisable to adopt a total cost of ownership perspective:

    1. Total cost = call fees + retry fees + manual review fees + integration costs + migration costs + operations costs

    Some models have a low unit price, but if their output quality is unstable, they can cause more manual review, so the total cost ends up higher.

    7. How to Reduce Single-Point Dependency Risk?

    Although concentrated procurement has governance advantages, over-concentration also brings risks.

    7.1 Supply Risk

    If a single provider runs into any of the following:

    • Quota restrictions;
    • Service outages;
    • Price adjustments;
    • Version deprecation;
    • Changes in regional availability;

    it can affect business continuity.

    7.2 Bargaining-Power Risk

    When one vendor holds too high a share, an enterprise’s bargaining room in later renewals, capacity expansions, and commercial negotiations may shrink.

    7.3 Technical Lock-in Risk

    If business logic is deeply bound to a model’s output format, prompt structure, or calling method, migration costs will keep rising.

    7.4 Compliance Risk

    Different enterprises have different requirements for data storage, cross-border transfer, log retention, and content safety. A single provider may not satisfy the compliance requirements of all regions or industries.

    To balance the convenience of concentrated procurement with the flexibility of multiple model sources, enterprises can introduce a unified model access layer.

    8.1 Unified Model Gateway

    A model gateway shields business systems from differences between models so that business systems can be built against a unified interface.

    Typical capabilities include:

    • Unified request format;
    • Unified authentication;
    • Call logging;
    • Quota management;
    • Cost metering;
    • Failure retries;
    • Timeout control;
    • Canary releases;
    • Model switching.

    The schematic logic looks like this:

    1. def route_request(task_type, latency_sla_ms, cost_limit):
    2. if task_type == "structured_extract":
    3. return "low_cost_route"
    4. if latency_sla_ms < 800:
    5. return "low_latency_route"
    6. if cost_limit == "tight":
    7. return "budget_route"
    8. return "balanced_route"

    8.2 Multi-Model Routing

    Enterprises can choose different models based on task type, budget, latency requirements, and quality level.

    For example:

    1. Simple classification task → low-cost model
    2. Real-time customer service task → low-latency model
    3. Complex coding task → high-quality model
    4. Offline batch task → overnight batch-processing model
    5. High-risk content moderation → strict-compliance model

    This avoids using the same expensive model for every task, and also avoids binding all business to a single provider.

    8.3 Caching and Batch Processing

    Many enterprise scenarios involve repeated requests, such as:

    • Fixed product introductions;
    • Standard policy Q&A;
    • Common ticket classification;
    • Similar document summaries;
    • Repeated code-completion prompts.

    Semantic caching or result caching can significantly reduce the actual number of model calls.

    At the same time, for tasks that do not require real-time responses, you can switch to batch processing:

    1. Real-time requests → prioritize low latency
    2. Non-real-time requests → enter a queue, processed during low-price or low-load periods

    8.4 Automatic Degradation and Circuit Breaking

    When the primary model behaves abnormally, the system should degrade automatically:

    1. Primary model timeout → switch to backup model
    2. Output format failure → retry or switch model
    3. Rising service error rate → trigger circuit breaker and alert
    4. Abnormal cost increase → limit calls or route to low-cost path

    This can significantly improve business stability.

    9. From Ad Hoc Procurement to AI Cost Governance

    The concentration of open-source model spending is also a reminder that enterprises need to strengthen their governance of AI spending.

    9.1 Tagging

    Every model call should carry clear tags:

    • Business line;
    • Scenario;
    • Environment;
    • User type;
    • Project;
    • Cost center;
    • Model version;
    • Routing strategy.

    Without tags, fine-grained analysis is impossible.

    9.2 Monitoring Key Metrics

    It is advisable to monitor at least:

    1. Call count
    2. Input tokens
    3. Output tokens
    4. Cache hit rate
    5. Retry rate
    6. Average latency
    7. P95 latency
    8. Error rate
    9. Cost per task
    10. Budget consumption rate

    9.3 Setting Budget Alerts

    For example:

    • Daily consumption exceeds a threshold;
    • Single-scenario cost changes abnormally period over period;
    • One model’s call volume spikes;
    • A certain prompt causes abnormal output length;
    • Retry rate exceeds the baseline.

    9.4 Conducting Regular Vendor Reviews

    Enterprises should not sign a contract once a year and then leave usage completely unchecked. It is advisable to periodically evaluate:

    • Whether current pricing is still competitive;
    • Whether model quality still meets business needs;
    • Whether better alternatives exist;
    • Whether there is over-concentration;
    • Whether a second supplier is needed;
    • Whether the routing strategy needs adjustment.

    10. Conclusion: Concentration Is the Result; Governance Is What Matters

    The heavy concentration of open-source model procurement spending among a few providers is no accident. It results from price competition, model stability, engineering integration costs, enterprise procurement habits, and the need for governance at scale.

    For enterprises, what truly matters is not simply judging “whether concentration is abnormal,” but building a sustainable management approach:

    1. First confirm the statistical denominator to avoid misreading data;
    2. Evaluate models by total cost rather than unit price;
    3. Choose models in layers according to scenario;
    4. Reduce single-point risk through multi-model routing;
    5. Achieve fine-grained cost governance through monitoring, metering, and alerts;
    6. Find a balance between concentrated procurement and flexible supply.

    The value of open-source models lies not in letting enterprises endlessly chase a single low-price vendor, but in giving enterprises more comparable, replaceable, and combinable technical options. A truly mature model procurement strategy should look at both short-term cost and long-term sustainability, both call price and business stability, migration cost, and governance complexity.