Whenever a cheaper and more open alternative appears in technology, the conversation often turns into a competition over which side will eventually eliminate the other. Linux was supposed to replace Windows, while Android and iOS were repeatedly framed as rivals fighting for complete dominance. In reality, both sides survived because they serve different users, priorities and business models.
Artificial intelligence appears to be heading in a similar direction. Proprietary models from companies such as OpenAI and Anthropic continue to dominate many enterprise workloads, but businesses are increasingly experimenting with open-weight alternatives as well. Rather than choosing one camp exclusively, organisations are beginning to use several models according to the needs of individual applications.
The bigger risk may therefore not be choosing the "wrong" model. It is becoming so dependent on one provider that switching later becomes technically difficult, operationally painful or simply too expensive.
Enterprises Are Already Using Multiple Models
The idea of relying on one AI provider for everything is becoming less realistic as the market matures. Different models have different strengths, price points, latency characteristics and deployment options.
Andreessen Horowitz has reported that 37% of enterprises were already operating five or more AI models in production, combining both open and proprietary systems.
That suggests model diversification is becoming a practical architecture decision rather than simply an ideological preference for open or closed technology.
One application may use a frontier model for complex reasoning, while another relies on a smaller open-weight model for classification or document processing.
Open-Weight Models Are Closing The Capability Gap
Open-weight models have improved substantially over the past few years. They increasingly offer performance that is good enough for workloads that once required access to the largest proprietary models.
The difference matters because companies are no longer evaluating these models only on benchmark scores. Cost, deployment control and operational flexibility have become equally important.
A model does not necessarily need to outperform the best frontier system overall. It only needs to perform well enough on the specific job the company needs it to do.
That is where open-weight models can become particularly attractive.
The Real Appeal Is Control
Performance is only one reason businesses are interested in open-weight AI.
The bigger attraction is often control.
Organisations want greater control over where their data is processed, how models are customised and how much dependence they have on an external vendor.
If a model can be deployed inside the company's own infrastructure, sensitive information does not necessarily need to leave that environment.
For industries handling highly regulated or confidential information, that can make the difference between whether AI can be used at all.
Data Can Stay Inside The Organisation
Healthcare providers, legal firms and financial institutions frequently operate under strict rules concerning sensitive information.
Sending confidential data to an external AI API may not always be acceptable, even when the provider offers strong security controls.
Deploying an open-weight model inside the organisation's own cloud or data centre allows the company to maintain much tighter control over where that information travels.
The model can process internal data without requiring every document or prompt to be sent to another company's infrastructure.
Customisation Can Stay In-House Too
Open-weight models also allow organisations to perform additional training or fine-tuning using their own data.
This can help shape the model around specific terminology, workflows or business requirements.
More importantly, the resulting customisation remains under the organisation's control rather than being tied entirely to one external AI provider.
If the business later changes infrastructure providers, it can potentially take the model and its customisations with it.
That flexibility can become valuable over the long term.
Owning The Weights Reduces Vendor Dependence
Vendor risk extends beyond pricing.
An AI provider could change its terms, discontinue a model, modify API behaviour, introduce new usage restrictions or even disappear entirely.
When a company owns or controls the model weights, those changes are less disruptive.
The organisation can continue running the same model even if the company that originally released it changes direction.
This does not eliminate every dependency, but it reduces the risk of having a critical application completely tied to one external service.
Open-Weight Is Not The Same As Open Source
One important distinction is that open-weight models are not necessarily open-source models.
The term usually means the trained model weights are available for users to download and run.
That does not automatically mean the training data, full training code or every component of the model-development process is available.
Licensing terms may also impose restrictions on how the model can be used.
Companies therefore need to examine the actual licence rather than assuming that publicly downloadable weights automatically provide all the freedoms associated with traditional open-source software.
More Control Also Means More Responsibility
This is where the trade-off becomes important.
A proprietary AI provider handles much of the difficult operational and safety work behind the scenes. The vendor maintains the infrastructure, updates the model and typically applies safeguards around how the system behaves.
Once an organisation operates an open-weight model itself, many of those responsibilities move internally.
Someone needs to manage deployment, security, monitoring, scaling and model behaviour.
The freedom to control the model therefore comes with an obligation to control it properly.
Safety Guardrails Can Be Easier To Remove Than Expected
Open models can also be modified in ways proprietary platforms generally prevent.
Researchers have shown that safety training can sometimes be weakened or reversed with relatively small amounts of additional fine-tuning.
Guardrails applied to models such as Llama and Gemma have reportedly been removable relatively quickly under certain conditions.
This does not mean organisations should avoid open-weight AI. It simply means safety cannot be assumed to remain intact after modification.
Companies running their own models need to treat model governance as part of the deployment rather than assuming the original model creator has already solved the problem permanently.
Narrow Workloads Are Where Open Models Often Shine
Open-weight models become especially attractive when the problem is clearly defined.
Tasks such as embeddings, classification and personally identifiable information detection often do not need the full capabilities of a frontier model.
A smaller model designed specifically for one of these tasks may perform extremely well while using substantially fewer computing resources.
In some cases, the specialised model may even outperform a much larger general-purpose model because it has been optimised for exactly that problem.
Why Pay For Intelligence You Are Not Using?
A frontier model is designed to perform many different tasks.
It might write code, summarise documents, analyse images, answer questions and perform complex reasoning.
That flexibility is extremely valuable when the workload varies.
But if the application only needs to classify customer support tickets into ten categories, most of that capability is unnecessary.
You are effectively paying for intelligence that the application never uses.
A smaller specialised model can avoid much of that overhead.
Embeddings Are A Good Example
Embedding models convert text, images or other information into numerical representations that applications can compare mathematically.
They are frequently used for semantic search, recommendation systems and retrieval-augmented generation.
These workloads generally do not require an enormous conversational model.
A dedicated embedding model can often handle the task faster and at considerably lower cost.
This makes embeddings one of the clearest examples where using the biggest available AI model would be unnecessary.
Classification Is Another Strong Fit
Classification workloads are similarly straightforward.
A business might need to determine whether an email is a complaint, sales request, support issue or billing question.
A properly tuned smaller model can perform that task effectively without requiring a massive general-purpose system.
When millions of classification requests are processed, even a small difference in inference cost can become significant.
This is where specialised open-weight models can deliver substantial savings.
PII Detection Can Benefit From Specialisation
Detecting personally identifiable information is another workload where narrow models can be useful.
Applications may need to identify names, addresses, identification numbers or other sensitive information before documents are processed elsewhere.
A specialised model can perform that detection without invoking an expensive frontier model for every piece of text.
Reducing the amount of unnecessary compute is particularly valuable when the workload is repetitive and high-volume.
The Hard Part Is Not Downloading The Model
One misconception about open-weight AI is that running it in production is as simple as downloading a model from Hugging Face and launching it with something like Ollama.
For experimentation, that can indeed be remarkably easy.
Production is very different.
An enterprise application needs predictable latency, high availability, scaling, monitoring and efficient GPU utilisation.
Achieving all of that requires considerably more infrastructure work.
Self-Hosting Is Similar To Running Open-Source Software
The situation resembles open-source software.
The software itself may be free to download, but operating it reliably still costs money.
Most businesses do not maintain every database, search engine or infrastructure component from scratch simply because the source code is available.
Instead, they frequently pay managed providers to operate those systems.
Open-weight AI is likely to follow the same pattern.
GPU Operations Require Specialist Knowledge
Running models efficiently at scale requires people who understand GPU infrastructure.
Someone needs to manage batching, model loading, memory allocation and scaling across multiple accelerators.
Inference servers need to remain stable when traffic suddenly increases.
Hardware failures need to be handled without taking down the application.
These problems can become highly specialised very quickly.
The cost of the engineering team may eventually exceed the savings from avoiding commercial API fees.
Latency Optimisation Is Another Challenge
A model that works perfectly on one developer's workstation may behave very differently when thousands of users begin sending requests simultaneously.
Batching requests improves efficiency but can increase latency.
Adding more GPU capacity reduces waiting time but increases cost.
Production AI infrastructure therefore requires constant balancing between performance, utilisation and expense.
Large hosted AI providers already invest heavily in solving these problems.
A company self-hosting models needs to decide whether reproducing that expertise internally is worth it.
Inference Providers Offer A Practical Middle Ground
For many businesses, the most realistic approach is using a specialised inference provider.
Companies such as Fireworks, Together AI and Base10 focus on hosting and scaling open-weight models for customers.
The business can choose from open models without needing to manage the underlying GPU fleet itself.
This provides much of the flexibility associated with open weights while removing a significant portion of the operational burden.
Managed Open Models Can Provide The Best Of Both Worlds
This model offers an interesting compromise.
You gain access to a wider selection of open-weight models and may achieve lower inference costs than using premium frontier APIs.
At the same time, the inference provider handles hardware provisioning, scaling and optimisation.
The organisation can also switch between different models more easily than if the entire application were tightly coupled to one proprietary provider.
For many enterprises, this may be the most practical way to incorporate open-weight AI.
Self-Hosting Makes More Sense At Scale
There are still situations where running the models yourself makes sense.
If AI is central to the company's product, the organisation may eventually reach a scale where owning and optimising the infrastructure becomes economically worthwhile.
Large request volumes can make external inference fees expensive enough that investing in GPUs and specialist teams starts to pay for itself.
The company may also need extremely strict control over data or model behaviour.
At that point, self-hosting becomes a strategic infrastructure decision rather than simply an attempt to avoid API costs.
Most Companies Probably Should Not Start There
For smaller organisations, building an internal AI infrastructure team immediately may be unnecessary.
The company first needs to know whether the application actually provides enough value to justify that investment.
Using hosted inference allows teams to experiment quickly without committing millions to hardware and operations.
If the workload eventually becomes large enough, the organisation can reconsider whether bringing infrastructure in-house makes sense.
This is similar to how many companies begin in the cloud before selectively moving certain workloads onto dedicated infrastructure later.
Frontier Models Still Have Important Advantages
Open-weight models are improving quickly, but proprietary frontier models still hold meaningful advantages.
When an application requires the strongest reasoning, multimodal understanding or general-purpose capability, models from the leading AI providers may still perform better.
Businesses also gain the benefit of infrastructure that has already been heavily optimised.
Instead of spending months tuning inference systems, developers can call an API and immediately access highly capable models.
You Are Paying For Reliability As Well As Intelligence
The cost of a frontier API does not only pay for the model.
It also covers the enormous infrastructure required to serve that model reliably.
The provider handles GPU clusters, failover, performance optimisation and software updates.
For applications where downtime or poor performance would directly affect customers, that managed reliability can be extremely valuable.
The API may appear more expensive per token while still being cheaper once internal operational costs are included.
Open Models Can Move The Cost Somewhere Else
This is why comparing inference pricing alone can be misleading.
An open-weight model may have a lower theoretical compute cost.
But the organisation may then need additional engineers, infrastructure management and optimisation work.
Latency may also be worse if the serving stack has not been tuned carefully.
The cost does not disappear. It simply moves to another part of the system.
Customer-Facing Applications Have Less Room For Latency
Latency becomes particularly important when users are waiting directly for the AI response.
A customer support chatbot that takes several additional seconds to answer may create a noticeably worse experience.
In those situations, paying more for a highly optimised hosted frontier model may make sense.
Performance consistency can be more important than minimising the cost of every individual request.
Internal Workloads Can Tolerate More Trade-Offs
Internal applications have very different requirements.
If an AI system is summarising customer support tickets overnight, an extra second or two of latency may not matter at all.
That makes cheaper open-weight models particularly attractive.
The company can prioritise cost over response time because no customer is sitting in front of the screen waiting.
Different workloads therefore justify different model strategies.
Open Models Have Not Completely Caught Up
Claims that open-weight AI has effectively reached complete parity with proprietary frontier models should also be treated cautiously.
Stanford's AI Index reportedly found that the performance gap between the best open and closed models actually widened slightly during the year.
That does not mean open models are becoming worse.
Both sides are improving, and frontier providers continue investing enormous resources into pushing capability forward.
The important question is whether that remaining gap matters for the workload you are actually running.
You Do Not Always Need The Best Model
A business application does not necessarily benefit from using the most capable model available.
If a cheaper model achieves 98% of the required result while costing a fraction as much, it may be the better engineering choice.
The final few percentage points of capability can become extremely expensive.
Frontier models make sense when those additional capabilities matter.
When they do not, smaller models can provide better economics.
Model Routing Could Become Increasingly Important
One increasingly practical architecture is routing different requests to different models.
Simple queries could go to inexpensive open-weight models.
Complex reasoning tasks could be sent to a frontier model.
Highly specialised tasks could use dedicated classifiers or embedding models.
This allows the application to optimise cost without sacrificing quality where advanced capabilities are genuinely needed.
A Diversified Model Strategy Reduces Lock-In
Using several models also makes it easier to change providers later.
If the entire application has been designed around one proprietary API and its unique behaviour, switching can become expensive.
A model abstraction layer makes the application less dependent on one vendor.
Prompts, evaluation systems and data pipelines can be designed so different models can be substituted more easily.
That architectural flexibility may ultimately be more valuable than saving a few cents on inference today.
Vendor Lock-In Is The Bigger Strategic Risk
Model providers will continue changing prices, capabilities and terms.
A company deeply dependent on one provider may eventually discover that switching requires rebuilding large parts of its AI stack.
That creates negotiating weakness as well.
If the provider raises prices substantially, the business may have little practical alternative.
Supporting multiple models from the beginning gives organisations more leverage and resilience.
Evaluation Becomes Essential In A Multi-Model World
The downside is that organisations need reliable ways to determine which model performs best for each workload.
Marketing claims and benchmark scores are not enough.
Companies need evaluations built around their actual data and tasks.
A model that performs brilliantly on general benchmarks may struggle with a company's specialised terminology.
Another supposedly weaker model may perform extremely well after targeted tuning.
Build Your Own Tests
A good evaluation set should include real examples from the intended application.
For a customer support system, that means genuine support questions.
For document extraction, use representative documents.
The models can then be compared for accuracy, latency and cost using the same test cases.
This produces information far more useful than relying entirely on public leaderboards.
Open Models Make Experimentation Easier
Because many open-weight models can be deployed through several providers, businesses gain more freedom to experiment.
The same model can potentially be tested through managed inference first and self-hosted later.
Companies can also evaluate several models without negotiating separate enterprise contracts with every creator.
That flexibility encourages competition and makes the model layer increasingly interchangeable.
Models Could Eventually Become Commoditised
The broader trend resembles what happened with operating systems and infrastructure.
At one point, the operating system itself was one of the most commercially valuable layers of computing.
Over time, that layer became increasingly commoditised.
Open-source software helped make cloud infrastructure possible, shifting enormous economic value towards services and applications built above the operating system.
AI models could follow a similar path.
The Model May Eventually Matter Less Than The Application
As open-weight models become more capable, choosing between models could gradually become less important.
Businesses may increasingly differentiate through their own data, workflows and user experience rather than through exclusive access to one particular model.
A customer does not necessarily care which language model generated an answer.
They care whether the application solved their problem accurately and quickly.
The competitive advantage therefore moves above the model layer.
Proprietary Data Becomes More Important
If models become easier to substitute, unique data becomes more valuable.
A company may use the same base model as several competitors but achieve better results because it has better internal information and retrieval systems.
The model provides reasoning capability.
The company's proprietary context makes that reasoning relevant.
That combination can become much harder for competitors to replicate than simply selecting the same model API.
Context Engineering Could Become A Core Skill
This is why context engineering is becoming increasingly important.
Applications need to provide models with the right information at the right time rather than expecting the model to know everything inherently.
Retrieval systems, permissions, user history and structured company data all become part of the AI product.
A cheaper open model with excellent context can sometimes outperform a more capable frontier model operating without the relevant information.
The quality of the surrounding system therefore matters enormously.
Applications Become The Real Competitive Layer
If model capabilities continue converging, users may eventually care much less about whether the application uses OpenAI, Anthropic, Llama or another model underneath.
They will choose the software that integrates best with their workflows.
The companies building strong applications, proprietary datasets and useful context systems could therefore capture more value than the companies simply offering another general-purpose model.
This is similar to how most users today care more about cloud applications than the operating system running on the server underneath them.
Open And Closed Models Can Coexist
The future therefore does not need to involve one side defeating the other.
Frontier proprietary models can remain valuable for the hardest and most latency-sensitive tasks.
Open-weight models can dominate specialised workloads where cost and control matter more.
Businesses can use both simultaneously.
That mixed environment may ultimately become the normal enterprise architecture.
The Right Choice Depends On The Workload
There is no universal answer to whether a company should use an open-weight or proprietary model.
The decision depends on data sensitivity, latency requirements, model capability, operational expertise and cost.
A healthcare organisation processing confidential information may prioritise local deployment.
A consumer chatbot may prioritise the strongest possible reasoning and lowest latency.
A high-volume classification system may care almost entirely about inference cost.
Each workload creates a different answer.
Do Not Self-Host Just Because You Can
Modern tooling has made running local models remarkably easy.
That accessibility is useful, but it can also create the impression that production self-hosting is equally simple.
It is not.
Running a demonstration on one GPU is very different from serving thousands of reliable requests every minute.
Companies should therefore self-host because there is a clear economic, privacy or strategic reason, not simply because downloading the model is possible.
Managed Inference May Be The Sensible Default
For many organisations, using an inference provider for open-weight models provides the most balanced starting point.
It avoids direct dependence on one frontier model vendor while also avoiding the operational complexity of maintaining GPU infrastructure internally.
The company can evaluate open models, compare costs and gradually build expertise.
If volume eventually justifies self-hosting, the transition can happen later.
That creates flexibility without requiring an enormous infrastructure commitment on day one.
Final Thoughts
Open-weight models are getting cheaper and increasingly capable, but that does not mean every company should immediately start buying GPUs and operating its own AI infrastructure.
Their biggest advantage is not simply lower inference cost. It is control: control over data, customisation, deployment and dependence on external vendors.
That control comes with responsibility. Running an open-weight model means taking greater ownership of safety, infrastructure, scaling and reliability. The model weights may be freely available, but production operations certainly are not.
For narrow workloads such as embeddings, classification and PII detection, specialised open models can provide excellent performance at significantly lower compute cost. For demanding customer-facing applications or complex reasoning, proprietary frontier models may still justify their higher prices through stronger capability and highly optimised infrastructure.
For most enterprises, the practical answer is probably neither complete cloud dependence nor immediate self-hosting. Managed inference providers offer a middle ground where businesses can use open-weight models without taking on the entire GPU operations burden themselves.
More importantly, companies should avoid building their entire AI strategy around one provider. A diversified model architecture makes it easier to choose the right model for each workload and reduces the risk of painful vendor lock-in later.
As models continue improving, the question of which model "wins" may eventually become much less interesting. The real competitive advantage will increasingly sit above the model itself — in the applications, proprietary data, workflows and context that turn generic intelligence into something genuinely useful.
Comments 0