The race to reduce the cost of enterprise AI usage is starting to influence actual buying decisions. DeepSeek, the Chinese company that entered strongly into the open models and low-cost API market, led Ramp’s trending software vendors list in June, a U.S.-based platform for managing corporate expenses. This doesn’t mean it has surpassed OpenAI or Anthropic in total adoption, but it does confirm that some companies are seeking more affordable alternatives to the major U.S. models.
This movement is accompanied by a second signal, this time from China. Authorities in the country have issued warnings about security risks related to intermediary services that enable access to foreign models through unofficial platforms. Both news items point to the same core issue from different angles: in the new AI economy, price matters, but the journey of data matters even more.
Cost pressures are beginning to change AI procurement
Ramp places DeepSeek at the top of its “trending software vendors” ranking for June. This list doesn’t measure total market share but rather identifies providers gaining new corporate payments—that is, when a company starts paying for a specific software for the first time.
This insight is significant because it’s not about technical curiosity or GitHub downloads, but about corporate spending. According to analysis by Ramp Economics Lab, some U.S. companies are paying directly to DeepSeek. This suggests that not all are running open models on their own servers but are instead using services hosted by the Chinese provider.
That distinction greatly influences the analysis. Self-hosting a model offers better control over where workloads run and what data leaves the organization. Using a direct API, however, can be faster and cheaper but involves sending information to external infrastructure, raising questions about privacy, compliance, jurisdiction, and security.
| Usage Option | Main Advantage | Main Risk |
|---|---|---|
| Direct API of DeepSeek | Lower cost and quick deployment | Data sent to a third-party servers |
| Self-hosted open model | More control over data and infrastructure | Requires technical capacity and hardware |
| OpenAI or Anthropic | Established ecosystem with broad enterprise adoption | Higher cost in certain use cases |
| Inference platforms | Flexibility to use multiple models | Additional layers of intermediaries |
| Unofficial relays | Cheap or alternative access | High security and compliance risks |
DeepSeek had a prior cycle of attention in January 2025, when its corporate adoption in Ramp’s index rose to 0.3%. It then receded to 0.1% and remained at that level through April 2026. To put this into perspective, at that time, Anthropic and OpenAI dominated Ramp’s AI adoption index with 34.4% and 32.3%, respectively.
The difference between a one-time growth and market dominance is important. DeepSeek may be growing from a small base, while OpenAI and Anthropic continue to be the most present providers across many companies. But the signal is clear: AI spending is becoming a key decision factor.
Price is no longer just a technical detail
During the early adoption phase, many companies tested AI models without focusing heavily on unit costs. The priority was understanding use cases: text generation, customer support, internal copilots, document analysis, coding, semantic search, and automation.
When those pilots move into production, the conversation shifts. Calls to models multiply, agents perform more steps, development teams integrate AI into daily workflows, and departments start receiving recurring bills. At that point, the cost per token ceases to be a technical detail and becomes a financial variable.
DeepSeek is entering precisely at this stage. Its offering is perceived as more cost-effective for certain workloads, especially compared to reference commercial models. For tasks where the most powerful model isn’t always necessary, many companies are considering more selective architectures: premium models for complex work, economical ones for repetitive tasks, and self-hosted solutions for sensitive data.
| Workload Type | Most Likely Model |
| Sensitive or regulated data | Internal or audited provider model |
| Basic writing or classification | Economical model |
| Complex reasoning | Premium model |
| Mass automation | Cost-optimized model |
| Critical software development | Traceability and continuous evaluation |
| Rapid prototyping | External API or multi-model platform |
This trend isn’t necessarily favoring a single winner. Instead, it may fragment the market. Companies are reluctant to always pay the highest price if a cheaper option can solve 80% of the work. However, they also want to avoid unnecessary risks with critical data. That’s why AI governance is increasingly resembling cloud policies: workload classification, provider rules, data limits, and continuous auditing.
China warns about foreign model intermediaries
While some U.S. companies are testing DeepSeek for cost reasons, China has focused on a different concern affecting its own developers: AI relay or intermediary services. These platforms act as gateways that aggregate APIs from various models, both domestic and foreign, to provide unified access through a single service.
The appeal is clear. Developers can access multiple models without contracting directly with each provider, avoiding technical restrictions, simplifying integrations, or reducing costs. But the problem is that many of these services add an opaque layer between the user and the actual model.
Chinese authorities have warned about privacy risks, data leaks, resale of information, the use of lower-quality models sold as advanced, potential backdoors, and unauthorized cross-border data transfers. In other words, while relays promise convenient, cheap access, they can turn into security weak points.
This warning from China isn’t just a local issue. The global AI market is full of intermediaries: gateways, aggregators, proxies, model routers, inference platforms, and services promising cost, latency, or availability optimizations. Many are legitimate and necessary; others may operate without sufficient transparency.
The emerging risk: not knowing which model is responding
One of the most serious technical and compliance challenges is traceability. When a company directly uses a model from OpenAI, Anthropic, Google, Mistral, or DeepSeek, it can review the contract, data policies, location, and retention rules. But when it goes through an opaque intermediary, the question becomes more complex: which model actually responded, where was it executed, who saw the prompt, and what data was stored?
This has practical implications. A team might believe they are using an advanced model and receive responses from a cheaper one. They could send internal data to a platform that stores it without authorization. They might include API keys, credentials, or code snippets into a relay that offers no guarantees. And they could unintentionally violate internal policies.
| Risks of AI relays | Potential Consequences |
| Lack of transparency about the model used | Less reliable or misleading results |
| Prompt storage | Internal data exposure |
| Resale of information | Unauthorized use for training or analysis |
| Uncontrolled international transfer | Regulatory issues |
| Backdoors or malicious code | Theft of credentials or system access |
| Lack of auditing | Difficulty investigating incidents |
For regulated companies, these risks are even greater. Banking, healthcare, industry, public administration, insurance, and telecoms cannot treat all AI models as consumer tools. They need to understand where data travels, who processes it, what controls are in place, and how to demonstrate compliance during audits.
Geopolitics enters the inference layer
The DeepSeek case also highlights that competition among models is no longer purely technological. It’s economic and geopolitical. That U.S. companies pay Chinese providers directly for AI services may raise questions internally within legal, security, and compliance departments. Likewise, Chinese developers seeking indirect access to foreign models concern Beijing’s authorities.
The inference layer is becoming a strategic infrastructure. It’s no longer just about where data is stored but where questions, documents, conversations, and intelligent agent actions are processed. AI doesn’t just move text; it moves intent, business context, and operational knowledge.
Therefore, AI sovereignty isn’t only about having a proprietary model. It also involves controlling the entire flow: interface, API, model, inference provider, cloud, logs, encryption, retention, evaluation, and auditing. An inexpensive model can be useful, but if the data flow isn’t governed, cost savings can turn into risk.
What companies should do
The answer isn’t banning all foreign models nor always opting for the most expensive provider. The market is moving toward a mix of models. The key is to set clear rules before each team starts contracting independently.
Organizations should classify their data, define which models can be used for each type of information, prohibit unauthorized relays, require clear retention and training contracts, record API usage, and periodically assess quality, cost, and risks. When justified by use volume or data sensitivity, it’s also wise to explore self-hosted alternatives.
DeepSeek can be a compelling option to lower costs in certain scenarios. OpenAI and Anthropic will remain references for many enterprise workloads. Open models will gain traction in private environments. intermediaries are useful when they offer transparency and control. The problem isn’t the diversity of providers but using it without proper governance.
This news delivers a clear conclusion for the tech sector: affordable AI has arrived in the enterprise but not alone. It raises questions about data, jurisdiction, security, traceability, and trust. The next competitive advantage will not just be selecting the most powerful or cheapest model but knowing when and how to use each without losing control.
Frequently asked questions
Why is DeepSeek now on the U.S. business radar?
Because it led Ramp’s trending software vendors list in June based on new corporate payments. This suggests some companies are testing cheaper alternatives to OpenAI and Anthropic.
Does DeepSeek already surpass OpenAI or Anthropic?
Not according to available data. In April 2026, OpenAI and Anthropic still dominated Ramp’s AI adoption index. DeepSeek is trending but not leading in market share.
What are AI relays?
Services that aggregate access to various AI models via a single platform. They can simplify integration but also introduce risks if lacking transparency or security.
What should companies watch out for when using cheaper models?
Check where data is processed, whether prompts are stored, if data is used for training, which jurisdiction applies, which provider responds, and if proper auditing is in place.