Open-Source AI: Why Corporate America Is Switching

Corporate America is increasingly turning to open AI models as the cost of running advanced Artificial Intelligence systems rises and companies look for more flexibility, lower prices and greater control over their technology.

AT&T is one of the clearest examples of that shift. The telecom giant previously relied heavily on closed models from companies such as OpenAI and Anthropic for tasks including customer service, call transcription and coding. Those systems offered powerful capabilities but required companies to pay for access through subscriptions, API charges or other usage-based fees.

AT&T has now moved significantly toward open models. According to Andy Markus, the company’s chief data and AI officer, open models accounted for about 40% of AT&T’s AI usage after representing roughly 20% earlier in the year. The company expects the share could rise further, potentially reaching 60% or more.

The reason is largely economic. Markus said AT&T was saving as much as 80% on AI costs compared with earlier in the year by moving more workloads to open models and using smarter model-routing strategies.

The shift reflects a larger change in the AI industry. Companies once competed primarily to gain access to the most powerful closed model. Increasingly, businesses are asking a different question: does every AI task really require the most expensive model available?

For many routine and specialised workloads, the answer appears to be no.

Why companies are moving toward open AI models

The corporate appeal of open AI is built around three advantages: cost, customisation and control.

Closed AI models are usually accessed through a company’s API or a hosted application. Businesses do not control the underlying model and may have limited ability to modify its behaviour beyond the tools and settings provided by the vendor.

Open models work differently. Depending on the licence and the specific model, organisations can download model files, run them on their own Infrastructure, fine-tune them and integrate them into internal applications.

That can substantially change the economics of deploying AI.

Instead of paying a model provider each time an application generates a response, a company can run certain models on its own servers or through lower-cost infrastructure and pay primarily for computing resources.

The economics can become especially attractive when the same AI system is used millions of times for repetitive tasks.

AT&T’s AI strategy shows the economics changing

AT&T’s experience offers a useful example of how large enterprises are approaching the transition.

The company is not abandoning advanced closed models altogether. Instead, it is using different models for different jobs.

AT&T has been developing a model-routing system that selects an appropriate model based on the task. A computationally expensive frontier model does not need to handle every simple request when a smaller or more specialised model can produce a sufficiently good result.

The company is also training open models for telecom-specific use cases.

That approach can make AI substantially cheaper because the company is no longer treating one premium model as the default engine for every workload.

AT&T has publicly discussed its use of open-source technology and efforts to create models tailored specifically to telecommunications operations, showing that open AI is becoming part of its production strategy rather than merely an experiment.

Open models are gaining ground across corporate America

AT&T is not alone.

Companies including Airbnb and Deloitte are also exploring open models as businesses look for ways to deploy AI at greater scale without allowing model costs to rise faster than the value generated by the technology.

Data from OpenRouter, a platform that provides access to multiple AI models, illustrates the broader change. US usage data showed that open models accounted for about 58% of AI model usage in August 2026, compared with roughly 10% a year earlier.

The precise mix differs from company to company, but the direction is clear. Open models are no longer confined to hobbyists and machine-learning researchers. They are increasingly becoming part of enterprise AI strategies.

That represents an important shift from the early ChatGPT era, when businesses typically began with a small number of closed services and built applications around them.

Nvidia’s $12.93 billion Hugging Face deal sends a major signal

The rise of open AI has also attracted one of the most important companies in the technology industry.

Nvidia agreed in September 2026 to acquire Hugging Face for approximately $12.93 billion, one of the chipmaker’s largest deals. Hugging Face is a major platform for sharing, discovering, evaluating and deploying AI models, datasets and applications.

The significance of the transaction extends far beyond the purchase price.

Hugging Face has become one of the central meeting points for developers working with open models. Nvidia’s acquisition therefore gives the world’s most influential AI-chip company a much deeper position in the software and developer ecosystem surrounding open AI.

Nvidia has said Hugging Face will remain an open platform and will continue to support different models, frameworks, cloud providers and computing platforms.

The message from Nvidia CEO Jensen Huang is clear: open AI is no longer a side market. It is becoming a strategic layer of the entire AI Economy.

What is the difference between open-source AI and open-weights AI?

The phrase “open AI” can be misleading because not every model described as open is open in exactly the same way.

There are two broad categories.

An open-source AI model generally makes important parts of its software and code available for inspection, modification and redistribution according to its licence.

An open-weights model makes the model’s trained parameters, or weights, available so that users can download and run the model themselves. The source code used to create or train the system may not necessarily be fully open.

That distinction matters because access to model weights gives developers significant control, but it does not automatically provide full transparency into the entire training process, data or development pipeline.

For businesses, however, open weights can still be extremely valuable because they allow companies to deploy and customise models without relying entirely on a hosted API.

Why the distinction matters to businesses

Enterprises care less about ideological definitions of open source than about what they can actually do with a model.

Can the model be downloaded? Can it run inside a private Environment? Can it be fine-tuned? Can the company inspect its behaviour? Can it be deployed on different hardware?

Those questions determine how much control a business has over its AI infrastructure.

Open-weights models can therefore offer many of the commercial benefits companies associate with open source even when the entire software stack is not publicly available.

This flexibility is especially important for companies handling sensitive information. Running an AI model within a controlled environment can reduce some concerns around sending proprietary information to an external AI provider, although security and privacy risks do not disappear simply because a model is self-hosted.

Chinese AI companies have become major players in open models

One of the most important features of the current open-model boom is the strong contribution from Chinese AI companies.

Models from companies and laboratories such as Moonshot AI, DeepSeek and Alibaba have become widely used by developers around the world.

This creates an unusual competitive dynamic. American companies remain dominant in many of the most powerful closed AI systems, while Chinese developers have become increasingly influential in open and open-weights models.

Developers can download these models, run them on their own infrastructure and build products around them without paying the same access fees associated with leading proprietary systems.

That gives Chinese AI developers a route to influence the global AI stack even when American companies retain an advantage in some frontier systems.

DeepSeek changed the economics of open AI

The rapid growth of open models accelerated after Chinese startup DeepSeek released DeepSeek-V3.

The model attracted international attention because it demonstrated competitive performance while operating with significantly lower computing requirements than many leading frontier systems.

The significance was not simply that DeepSeek had built a strong model. It was that developers increasingly saw evidence that useful AI did not always require the enormous cost structure associated with the most advanced closed models.

That changed the industry’s discussion from “which company has the smartest model?” to another question: how much intelligence do businesses actually need for each task?

That distinction is now influencing enterprise purchasing decisions.

Open models can be good enough for many tasks

For a large corporation, not every AI application needs to write sophisticated computer code, generate high-quality video or solve complex scientific problems.

Many enterprise workloads are comparatively narrow.

A model may need to classify customer messages, summarise calls, extract information from documents, route support tickets or answer questions about internal procedures.

Those tasks can often be handled effectively by smaller or specialised models.

That is where open AI becomes economically attractive.

If a company can get most of the required performance at a fraction of the cost, paying for the most powerful model on every interaction makes little commercial sense.

The performance gap with closed models is narrowing

Open models do not necessarily match the very best closed models across every benchmark or workload, but the gap has narrowed significantly.

Jerry Tang, CEO of Atlas Cloud, has argued that some Chinese open models can reach roughly 80% to 90% of the performance of leading systems from companies such as OpenAI and Anthropic while costing substantially less to operate.

His comparison is essentially that businesses do not always need a luxury vehicle for every trip.

That logic is increasingly resonating with corporate technology buyers.

In an enterprise environment, the best model is not necessarily the one with the highest benchmark score. It may be the one that delivers sufficient accuracy at a manageable cost and can be deployed at the scale the business requires.

Closed AI still has important advantages

The rise of open AI does not mean that closed models are becoming obsolete.

Most US companies currently use a hybrid strategy, combining open and closed models according to the task.

Premium closed models can remain valuable for difficult coding, sophisticated reasoning, image generation, video creation and other workloads where quality matters more than cost.

Closed providers also take responsibility for infrastructure, updates, security controls, model scaling and other operational requirements.

For a business that does not want to maintain AI infrastructure internally, paying an external provider may still be more convenient than managing an open model.

The likely future is therefore not a total replacement of closed AI with open AI. It is a much more fragmented model ecosystem in which businesses choose the appropriate system for each workload.

Open models can be customised for specific industries

One of the biggest advantages of open models is the ability to adapt them to specialised environments.

A general-purpose AI model may know a great deal about the world but lack detailed knowledge of a company’s internal terminology, processes and systems.

An enterprise can take an open model and customise it for a particular industry.

AT&T, for example, has been working on AI models designed specifically for telecommunications use cases. The company’s Open Telco AI initiative is focused on models and resources tailored to telecom operations rather than general-purpose consumer applications.

This approach allows companies to turn AI from a generic service into a more specialised internal tool.

Why customisation matters more than raw intelligence

For many businesses, a specialised model that understands the company’s own terminology can be more useful than a much larger general-purpose model.

A telecommunications company may care about network diagnostics, customer plans, service outages and engineering documentation. A legal firm may care about contracts and case files. A manufacturing company may need models trained around machinery, maintenance and production workflows.

The ability to adapt a model to those specific environments can create value that is difficult to measure through general AI benchmarks.

This is one reason open models are attractive to corporations: they allow companies to shape the AI around the business rather than forcing the business to adapt to a fixed AI service.

AI costs are becoming a strategic issue

The economics of AI have changed rapidly since generative AI became mainstream.

Early experimentation often involved relatively small numbers of users asking models occasional questions. Enterprise deployment is different.

Once AI is built into customer service, software development, employee workflows and internal systems, usage can explode.

Every additional interaction can create computing costs. A company processing billions of tokens cannot approach AI spending in the same way as an individual consumer using a chatbot a few times a day.

That is why model efficiency has become a boardroom issue.

Companies are now looking for intelligent routing, smaller models, caching, fine-tuning and open alternatives as ways of controlling the cost of AI at scale.

AT&T is using model routing to cut expenses

AT&T has described a strategy in which a routing layer determines which AI model should handle a particular task.

A simple task can be sent to an inexpensive model, while more complex work is routed to a premium system.

This resembles a broader trend in computing: businesses do not send every problem to the most powerful available Processor if a cheaper resource can handle it.

The same logic is now being applied to AI inference.

Combined with open models, intelligent routing can substantially reduce the amount of work performed by expensive frontier systems.

Open AI also reduces vendor dependence

Cost is not the only consideration.

Companies also care about vendor lock-in.

When an application is built around one proprietary model provider, changing providers may require significant engineering work. APIs, model behaviour, pricing structures and technical requirements can change.

Open models can provide an alternative because the model itself can be downloaded and moved between infrastructure environments.

That does not eliminate dependencies altogether. Companies still need hardware, cloud capacity, engineering talent and model updates.

But it can reduce reliance on one AI company controlling the entire software layer.

Why some US companies are cautious about Chinese models

The success of Chinese open models presents a strategic dilemma for American businesses.

On one hand, Chinese models can offer strong capabilities at relatively low operating costs. On the other, US companies may have concerns about data privacy, regulation, geopolitical risk and supply-chain dependence.

AT&T, for example, has researched Chinese AI models but said it is not using them in its production systems.

Instead, the company is working with American alternatives, including Google’s Gemma family and Meta‘s Llama models.

This creates an important distinction between technological openness and geopolitical trust. A model can be technically available to everyone while still raising concerns for a company operating under strict data and regulatory requirements.

America is responding with its own open models

Chinese leadership in several open-model categories has encouraged American companies to accelerate their own efforts.

Meta has continued to promote open-weight AI, while Google has released open model families such as Gemma. Nvidia has also become an increasingly active participant in the open-model ecosystem.

Meta’s strategy is particularly significant because it places the company in opposition to the closed-model strategy pursued by some of its biggest technology rivals.

Meta CEO Mark Zuckerberg has argued that restricting access to foreign open models would be less effective than ensuring American open models become the strongest globally.

That argument turns open AI into a matter of technology competition and national strategy rather than simply a question of developer preference.

Open AI has become part of the geopolitical competition

The debate over open versus closed AI now extends beyond software architecture.

Washington and Silicon Valley are increasingly discussing AI through the lens of National Security, technological leadership and global influence.

One side argues that unrestricted access to powerful models can create risks, including misuse and national-security concerns. Others argue that open systems are necessary to prevent AI development from becoming concentrated in a small number of companies.

Open models can also help countries and organisations build their own AI capabilities without relying entirely on foreign companies.

That makes the international competition over open AI strategically important.

Anthropic and OpenAI face a different economic model

The rise of open AI could create a challenge for companies whose business models depend heavily on selling access to proprietary systems.

OpenAI and Anthropic spend enormous amounts on model development, computing infrastructure and research. Their ability to sustain those investments depends partly on monetising access to increasingly sophisticated systems.

If businesses discover that an open model can handle a large share of their workloads at a fraction of the price, demand for premium proprietary models could become more selective.

That does not necessarily mean closed AI companies will lose their market. They may continue to dominate the highest-performance segment while open models handle simpler and cheaper tasks.

But the pricing power of closed providers could face pressure if open alternatives continue improving.

The AI market could split into premium and efficient layers

One possible outcome is a two-tier AI market.

At the top would be expensive frontier models used for difficult reasoning, complex coding, advanced media generation and specialised research.

Below them would be a much larger ecosystem of smaller and open models handling everyday enterprise workloads.

That would resemble the broader computing market, where high-end systems coexist with enormous numbers of lower-cost processors designed for specific applications.

Businesses would then build AI stacks rather than selecting one universal model.

AT&T’s routing strategy provides an early glimpse of how that future could work.

Hardware is becoming part of the open-model race

The growth of open AI also benefits companies that sell computing infrastructure.

Open models often need to be run somewhere, even when the model itself is available at no charge.

That means developers still require GPUs, servers, cloud infrastructure or other accelerators for training and inference.

This is one reason Nvidia has such a strong strategic interest in open models. A freely available model can create more demand for computing hardware because developers are able to deploy it in a wider range of environments.

Hugging Face is particularly valuable in this ecosystem because it connects models, developers and infrastructure.

Nvidia’s acquisition therefore gives the chipmaker a much deeper relationship with the software layer surrounding AI deployment.

Hugging Face’s role goes beyond hosting models

Hugging Face is often described as a library of AI models, but its role is broader.

The platform supports models, datasets, applications and tools that developers use to train, evaluate and deploy AI systems.

That makes it an important distribution and collaboration layer for open AI.

By owning Hugging Face, Nvidia gains visibility into what developers are building and which models are becoming popular.

The strategic significance of the deal therefore extends beyond simply acquiring a repository of open models. It gives Nvidia a stronger position in the workflow through which AI developers discover and deploy technology.

What could go wrong with open AI?

The open-model revolution comes with trade-offs.

Making models easier to download and modify can also make them easier to misuse. Companies deploying open models are responsible for more of the security, monitoring and infrastructure around the system.

Open models may also vary in licensing terms, documentation quality and support. “Open” does not automatically mean easy to deploy safely at enterprise scale.

Businesses have to evaluate model provenance, security vulnerabilities, privacy implications and legal restrictions before putting an external model into production.

These challenges are particularly important for companies operating in regulated industries.

Data privacy can be an advantage of self-hosted AI

For companies handling sensitive information, self-hosted AI can offer an important benefit: greater control over where data is processed.

A business can potentially keep confidential documents or customer information inside its own infrastructure rather than sending every request to a third-party provider.

That does not automatically make a self-hosted model secure. Poorly configured infrastructure can create serious vulnerabilities.

But companies gain more control over data flows, access permissions and internal security policies.

For industries such as telecommunications, healthcare, finance and law, that control can be a major reason to consider open models.

AI can now run on smaller devices

Another major development is the increasing efficiency of smaller AI models.

Researchers and companies have developed techniques that allow sophisticated models to operate with less memory and computing power than earlier generations required.

Quantisation, model compression and other optimisation approaches can make some models practical on laptops, workstations and even mobile devices.

That has significant implications.

If useful AI can run locally, users and businesses do not necessarily need a constant connection to expensive cloud infrastructure.

That can reduce latency, operating costs and, in some cases, exposure of sensitive data.

The cloud is not going away

Despite the growth of local AI, cloud computing will remain critical for many workloads.

Large models require substantial memory and processing capacity, especially when serving many users simultaneously.

Businesses also benefit from the scalability and managed infrastructure provided by cloud platforms.

The likely result is another hybrid arrangement: some AI workloads will run locally or on private infrastructure, while heavier tasks will remain in public clouds.

Open models fit neatly into that hybrid world because they can be deployed across different environments.

Why “open” could become the default for ordinary AI tasks

The biggest long-term shift may be that businesses stop thinking of AI models as a single product category.

Instead, AI could become an infrastructure layer with a wide range of interchangeable models.

A customer-service application might use one open model. A coding assistant could use another. A complex research tool might call a premium proprietary model only when necessary.

The software application would decide which model to use based on cost, speed, accuracy and privacy.

This would transform model providers from monopolistic platforms into components that compete for particular workloads.

Why corporate AI adoption is different from consumer AI

The consumer AI market tends to reward convenience. A person is often willing to pay for access to the best chatbot if the product is easy to use.

Corporate buyers operate under very different incentives.

A large enterprise can spend millions or even billions of dollars on AI infrastructure. At that scale, a modest reduction in inference costs can translate into enormous savings.

That is why enterprises are more willing to deploy open models, build internal tooling and maintain AI infrastructure when doing so lowers long-term costs.

Corporate AI is therefore likely to become much more cost-sensitive than consumer AI.

What this means for OpenAI and Anthropic

The rise of open models does not immediately threaten OpenAI or Anthropic’s entire businesses, but it does challenge the assumption that premium AI will capture every enterprise workload.

The most expensive models will have to justify their prices through significantly better performance or through additional services surrounding the model.

At the same time, closed-model providers can respond by making their systems more efficient, reducing prices, introducing smaller models or offering better tools for enterprise deployment.

That competition could ultimately benefit customers.

As open and closed models compete, companies may gain access to more capable systems at lower costs.

What happens to AI pricing next?

Open models could accelerate a broader decline in the cost of AI inference.

When a company can run a model internally rather than paying per request, the marginal cost of generating an additional response can become primarily a function of computing efficiency.

That encourages developers to optimise models aggressively.

Competition can then shift from simply building larger models to building models that deliver the best performance for the least amount of computing.

This is a significant change because AI’s future may depend as much on efficiency as on raw intelligence.

Corporate America is building a hybrid AI future

The growing popularity of open AI does not mean the closed-model era is ending overnight.

Instead, businesses are building hybrid AI systems in which different models perform different jobs.

Premium proprietary models remain useful for complex, high-value tasks. Open models increasingly handle repetitive, specialised or cost-sensitive workloads.

AT&T’s move toward open models provides a clear illustration of the strategy: use a mixture of systems, route each task intelligently and reserve expensive models for workloads that actually need them.

That could become the standard architecture for large corporate AI deployments.

The China factor makes the shift more important

The popularity of Chinese open models adds another dimension to the transformation.

DeepSeek, Moonshot AI, Alibaba and other Chinese developers have demonstrated that open AI can be a powerful vehicle for global technological influence.

American companies now face a strategic choice. They can try to restrict access to foreign open models, or they can develop domestic alternatives that outperform them.

Meta’s leadership has explicitly argued for the second approach.

That suggests the open-model debate will increasingly be treated as a competition over technological standards and developer ecosystems rather than simply a question of licensing.

Why the Hugging Face acquisition matters so much

Nvidia’s decision to spend nearly $13 billion on Hugging Face is one of the clearest signals that open AI has become strategically valuable.

A chip company would not need to make such a large investment in an AI developer platform if open models were merely a niche interest.

Nvidia clearly sees an opportunity to position itself across more layers of the AI stack: hardware, software, developer tools and model distribution.

The acquisition also demonstrates how the AI industry is evolving beyond a simple contest between chatbot companies.

The next phase could involve competition to control the entire infrastructure through which models are created, deployed and improved.

The future of AI may be more open and more fragmented

Corporate adoption of open models is changing the AI market from a small group of closed systems into a much broader ecosystem.

Developers can now choose among models created by US technology companies, Chinese laboratories, independent researchers and specialised startups.

They can run those systems in public clouds, on private servers or increasingly on local devices.

That diversity creates competition but also complexity.

Companies will need new skills to evaluate models, compare costs, manage deployments and maintain security across multiple AI systems.

But the payoff can be substantial: lower costs, reduced vendor dependence and models better suited to specialised business needs.

Open-source AI is becoming a corporate strategy, not a developer hobby

The most important change in the AI market may be happening quietly inside corporate technology departments.

Businesses that initially treated AI as an expensive service from a few frontier labs are discovering that many workloads can be handled by cheaper, adaptable models.

AT&T’s rapid increase in open-model usage is an especially strong example of that transition. The company’s strategy shows that the question is no longer whether enterprises will use open AI, but how much of their AI workload should be open.

Other companies are reaching similar conclusions, while Nvidia’s acquisition of Hugging Face indicates that the infrastructure surrounding open models has become strategically important enough to justify one of the industry’s biggest acquisitions.

Chinese companies have played a major role in pushing the technology forward, forcing US firms to strengthen their own open-model offerings.

At the same time, premium closed models will likely continue to dominate the hardest AI tasks.

The result could be an AI economy that looks less like a race toward one universally dominant model and more like a vast computing market with different tools optimised for different jobs.

For corporate America, that may be the most important lesson: the smartest AI is not always the most valuable AI. When a cheaper model can do the job well enough, businesses have a powerful reason to choose it.

That simple economic calculation is turning open AI from a technology popular with developers into a serious corporate strategy and it could reshape who controls the next phase of the artificial intelligence industry.