Revolut has reportedly reduced its artificial intelligence inference costs by as much as 90% after moving away from expensive commercial AI services and adopting self-hosted, open-weight models developed by China’s Z.ai. The move highlights a growing shift among major businesses looking for more affordable ways to integrate artificial intelligence into their daily operations without sacrificing productivity.
The London-based digital banking giant has not abandoned AI. Instead, it has changed how it accesses and runs the technology, replacing some of its reliance on Anthropic’s Claude with models hosted on its own infrastructure.
The results appear significant. Revolut says the transition has allowed developers to use AI more freely, without constantly worrying about token limits or unexpected bills. The company has also reported improvements in development productivity, suggesting that controlling AI infrastructure can offer benefits beyond simply reducing expenses.
The decision reflects a broader question facing businesses worldwide: Is it really worth paying premium prices for commercial AI models when open-source alternatives can deliver much of the functionality at a fraction of the cost?
You might want to read this: Revolut Hits a Stunning $115 Billion Valuation as U.S. Banking Expansion and a Potential $200 Billion IPO Loom
Why Revolut Is Moving Away From Anthropic’s Claude
Revolut has been experimenting with artificial intelligence for some time, testing different models and providers as it looks for the most effective tools for its engineering teams.
The company previously explored Google’s Gemini before adopting Anthropic’s Claude for certain development workloads.
However, the financial implications of widespread AI adoption soon became difficult to ignore.
According to David Abitbol, Revolut’s infrastructure lead, some employees were spending thousands of dollars every month on AI tokens without generating a comparable return on investment.
Speaking at a conference last month, Abitbol explained that the company was initially cautious about the financial impact of introducing AI tools.
“We were highly skeptical about the cost of AI,” he said, according to The Register. He added that Revolut introduced the technology with strict spending limits.
These restrictions were intended to prevent excessive usage and keep AI-related expenses under control.
However, limiting access also created a potential productivity problem. Developers who rely heavily on AI coding assistants can quickly consume their allocated token budgets, particularly when working on complicated software projects that require repeated interactions with AI models.
Rather than continuing to impose strict limits or accepting increasingly expensive commercial subscriptions, Revolut decided to explore another approach.
Its solution was to bring more AI workloads in-house.
Revolut Turns to China’s Z.ai and Its Open-Weight GLM Models
Instead of relying primarily on externally hosted AI services, Revolut has adopted open-weight models developed by Chinese AI company Z.ai.
The company has been using its GLM model family, including GLM-5.1, to support workloads previously handled through commercial AI providers.
Open-weight models offer an important advantage over traditional subscription-based AI services: organizations can obtain the model weights and deploy the technology within their own computing environments.
This gives businesses greater control over how their AI systems operate, where data is processed, and how much computing capacity is allocated.
For Revolut, the change means it can operate its own inference infrastructure rather than paying an external provider every time employees generate AI outputs.
The approach also provides greater flexibility when managing workloads.
Instead of worrying about individual usage limits or unpredictable token consumption, developers can access the company’s internal AI resources according to available infrastructure capacity.
Abitbol has also highlighted the privacy and auditability benefits of Z.ai’s GLM models.
For a financial services company handling sensitive customer and business information, having greater control over AI infrastructure and data processing can be particularly valuable.
This does not automatically eliminate every security or privacy risk, but it gives Revolut more direct oversight of its deployment environment.
AI Inference Costs Have Reportedly Fallen by 90%
The most striking result of Revolut’s transition is the reported reduction in inference expenses.
According to the figures discussed in reports about the company’s AI infrastructure strategy, moving to self-hosted models has reduced inference costs by as much as 90% in less than a year.
The economics become clearer when looking at the reported pricing of GLM-5.1.
Inference costs have reportedly fallen to approximately $0.30 per million input and output tokens.
That is around one-tenth of the cost of comparable commercial AI services.
For companies running large numbers of AI requests every day, the difference can become substantial.
A developer occasionally asking an AI assistant to explain a piece of code may not generate significant expenses. However, when hundreds or thousands of employees use AI tools continuously, token consumption can grow rapidly.
Complex coding tasks can involve extensive conversations, repeated requests, large codebases, and substantial context windows.
Under a conventional commercial pricing model, every additional interaction can contribute to the monthly bill.
Self-hosting changes this financial equation by allowing companies to manage their own computing resources and distribute the costs across their internal workloads.
Revolut can now determine how much infrastructure it needs and allocate that capacity according to demand.
The company also benefits from more predictable expenditure, avoiding some of the billing uncertainty associated with pay-per-token services.
Developers Are Seeing Major Productivity Improvements
Lower costs are only part of the story.
Revolut has also reported substantial improvements in developer productivity following the introduction of its self-hosted AI infrastructure.
Speaking at GTC Berlin, Abitbol said the transition delivered a 70% improvement in lead-time productivity and a 60% improvement in coding-time productivity.
These figures suggest that removing restrictions on AI usage can have a meaningful effect on how engineering teams work.
When developers are constantly concerned about token budgets, they may avoid using AI for certain tasks or limit the number of interactions they have with coding assistants.
With internal infrastructure, those restrictions can become less significant.
Developers can experiment more freely, request additional code reviews, explore alternative solutions, and use AI assistance throughout more stages of the development process.
This can make AI a more integrated part of everyday engineering rather than an expensive tool that employees must use selectively.
It also illustrates an important point about enterprise AI adoption: the cheapest model is not necessarily the best option if usage restrictions prevent employees from getting meaningful value from it.
By reducing the cost of each interaction, Revolut appears to have made broader AI adoption economically more attractive.
The Hidden Costs of Running AI Models In-House
Despite the impressive reported savings, self-hosting AI models is not a completely inexpensive solution.
Moving away from commercial AI services eliminates certain usage fees, but it also transfers responsibility for infrastructure, hardware, and maintenance directly to the company.
Running sophisticated AI models requires substantial computing resources, particularly high-performance GPUs with large amounts of memory.
Revolut has reportedly assembled an infrastructure setup containing 56 Nvidia B200 GPUs to support its internal AI workloads.
That represents a significant hardware investment.
The GLM-5.1 model itself reportedly requires approximately 380GB of GPU memory, while additional memory is needed to maintain user context during interactions.
Estimates suggest that context retention can require another 6GB to 26GB of memory per user, depending on usage.
These requirements demonstrate why self-hosting is not necessarily a practical option for every organization.
Companies must consider several additional expenses, including:
- GPU hardware: Purchasing or leasing powerful processors can require substantial upfront investment.
- Electricity: Running large GPU clusters continuously can generate significant energy bills.
- Maintenance: Infrastructure needs monitoring, repairs, upgrades, and ongoing technical support.
- Cooling and facilities: High-performance computing equipment generates considerable heat and requires suitable data center environments.
- Engineering expertise: Businesses need specialists capable of deploying, maintaining, securing, and optimizing AI models.
Consequently, a 90% reduction in direct inference costs does not necessarily translate into a 90% reduction in total AI expenditure.
The actual financial benefit depends on hardware utilization, electricity prices, infrastructure efficiency, maintenance requirements, and the number of users supported.
For a company operating AI at Revolut’s scale, however, investing in dedicated infrastructure may make financial sense.
Why More Companies Are Considering Open-Source AI
Revolut’s decision reflects a wider shift in the artificial intelligence industry.
Businesses are increasingly evaluating whether they should continue paying premium prices for commercial AI models or move certain workloads to open-weight alternatives.
The appeal is straightforward: not every business task requires the most advanced AI model available.
A company may need a frontier model for complex reasoning, sophisticated research, or particularly demanding programming assignments.
But many routine activities can be handled by smaller or less expensive models.
These might include summarizing documents, generating basic code, classifying information, answering internal questions, or assisting with repetitive administrative tasks.
Using a premium commercial model for every request can therefore create unnecessary costs.
A more efficient strategy is to match the model to the complexity of the task.
Organizations can reserve expensive frontier models for situations where their advanced capabilities provide a meaningful advantage, while using open-weight models for more routine workloads.
Self-hosting can make this strategy even more attractive by giving companies direct control over computing resources and usage patterns.
For Revolut, open-source and open-weight AI have reportedly become a priority because the financial economics are more favorable.
As Abitbol explained: “The billing is flat, we don’t have any bill shock at the end of the month, and the frontier is the exception.”
The Bigger Lesson: Businesses Need a Smarter AI Spending Strategy
Revolut’s experience highlights an increasingly important challenge for companies adopting artificial intelligence.
The technology can deliver impressive productivity gains, but those benefits do not automatically justify unlimited spending.
Businesses need to understand which AI workloads generate genuine value, how frequently employees use them, and whether expensive commercial models are necessary for every task.
Simply giving employees access to powerful AI tools without considering the financial consequences can quickly become an expensive experiment.
Revolut’s approach offers a different strategy: use open-weight models where they make financial sense, operate infrastructure internally when usage justifies the investment, and reserve premium commercial models for tasks that genuinely require them.
However, the decision to self-host should not be based on token prices alone. Companies must compare the full cost of infrastructure ownership against commercial alternatives.
For organizations with substantial AI workloads, dedicated computing infrastructure may deliver considerable long-term savings. Smaller businesses, meanwhile, may find that renting computing capacity remains more practical.
Is Self-Hosted AI the Future of Enterprise AI?
Revolut’s reported 90% reduction in inference costs could encourage other large organizations to reconsider how they deploy artificial intelligence.
As open-weight models continue to improve, businesses may find fewer reasons to depend exclusively on expensive commercial AI platforms.
The shift could also create a more competitive market, putting pressure on established AI providers to offer better pricing, greater flexibility, and more efficient models.
Still, commercial frontier models are unlikely to disappear. Their advanced capabilities remain valuable for complex tasks where performance matters more than the cost of individual requests.
The more likely outcome is a hybrid AI strategy in which companies combine commercial models, open-weight alternatives, and self-hosted infrastructure.
For Revolut, that transition has already produced encouraging results, with significantly lower reported inference costs and measurable improvements in developer productivity.
The company’s experience sends a clear message to the wider technology industry: successful AI adoption is not simply about using the most powerful models available. It is about choosing the right models, controlling infrastructure costs, and making sure every AI investment delivers measurable business value.
