Listen & Follow Along — Article Summary
Article Summary
Revolut, a British fintech company, has significantly reduced its AI coding costs by about 90% in less than a year. They achieved this by making self-hosted models their primary choice instead of relying mostly on expensive commercial services like Claude. This change led to impressive improvements: coding time improved by 60% and project lead time by 70%. Initially, Revolut was cautious about AI expenses, especially after seeing engineers spend a lot with a low return on investment (ROI) from earlier tests. They began with strict budget limits on AI usage for engineers, which pushed them to look for more cost-effective solutions.
Moving to self-hosted models, which offered much cheaper "tokens" (units of AI usage), changed everything. When tokens became inexpensive, engineers felt free to use AI without worrying about costs or limits. While coding speed did improve, Revolut found the biggest gains were not just in writing code faster. Instead, the greatest benefits came from using AI for tasks like debugging and managing internal tickets in systems like Jira. This showed that the main bottleneck for engineers wasn't always code generation, but other time-consuming support activities.
To achieve these results, Revolut focused on optimization rather than simply adding more hardware. They made important adjustments to their AI models and systems. They also learned that extremely long context windows, while technically possible, are not always necessary or efficient. Revolut now prioritizes open-source solutions for better cost control and flat billing, avoiding unexpected high charges. They manage privacy concerns by running models on-premise or in regional clouds. To handle various tasks, Revolut uses a model router to direct different types of queries, like visual tasks, to other AI models while keeping code generation on their self-hosted system. This flexible approach ensures they get the most value from their AI investments.
Key Vocabulary
self-hosted models
Click to reveal
commercial services
Click to reveal
return on investment (ROI)
Click to reveal
bottleneck
Click to reveal
optimization
Click to reveal
throughput
Click to reveal
flat billing
Click to reveal
on-premise
Click to reveal
model router
Click to reveal
Comprehension Questions
1. What was Revolut's primary strategy for significantly reducing its AI inference costs?
- Hiring more AI engineers
- Increasing reliance on commercial services
- Switching to self-hosted AI models
- Stopping the use of AI tools entirely
2. According to the article, where did Revolut experience the biggest gains from using AI models?
- In speeding up new code generation
- In improving user interface (UI) design
- In debugging and managing internal tickets
- In processing vision-based queries
3. Why did Revolut initially implement strict budget limits on AI token usage for its engineers?
- They wanted to encourage engineers to be more innovative.
- They believed AI technology was not yet mature enough.
- They were skeptical about the high costs and low return on investment.
- They had limited hardware capacity to support widespread AI use.
4. What can be inferred about the impact of cheaper AI tokens on engineers' work according to the article?
- Engineers became more selective about when to use AI.
- It led to engineers focusing only on complex coding tasks.
- It encouraged engineers to use AI more freely for various support tasks.
- Cheaper tokens mainly benefited the finance department, not engineers.
5. Considering Revolut's experience, what would be the most important factor for a business leader evaluating new AI technology adoption?
- The specific type of GPU hardware required for the AI models.
- The ability of the AI to handle very long context windows.
- The potential for cost savings and a clear return on investment.
- The AI's capacity for complex visual or UI work.
Discussion Prompts
1. How do you currently evaluate the return on investment (ROI) of new technologies or tools implemented in your department or company?
2. Can you recall a time when a specific process in your work became a bottleneck? What steps were taken to address it?
3. In your industry, what are the main advantages or challenges of using on-premise solutions compared to cloud-based or commercial services?
Live Session Prep & Cheat Sheet
🎯 Speaking Targets (Vocabulary)
Try to use these target terms in your speaking turns:
- slash costs
- return on investment (ROI)
- bottleneck
- optimization
- on-premise
⚙️ Grammar Target Formula
Expressing Cause and Effect: "Because of" and "Leading to": Because of [Noun Phrase], [Result]. [Action], leading to [Result].
💬 Discussion Openers
Use these phrases to open or structure your arguments:
- In my experience...
- From a business perspective...
- I think it's important to consider...
- What are your thoughts on...?
Teacher Notes
This lesson explores Revolut's strategy for reducing AI costs and improving efficiency. Focus on the vocabulary related to tech adoption and financial benefits. Encourage students to connect the company's decisions to their own professional experiences, prompting discussion on cost-benefit analysis and operational improvements.
Speaking Class Facilitation Guide (Tutors/Moderators Only)
🎭 Role-Play Scenario
Situation: Your company is considering adopting a new AI tool to improve productivity. The IT department suggests a costly commercial service, while the finance department wants to explore a cheaper, potentially self-hosted solution. You need to present and discuss these options.
Project Manager: You need to convince the Head of Finance that the more expensive commercial service offers better return on investment (ROI) in the long run, even with higher initial costs, or consider a hybrid approach. Focus on long-term benefits and features.
Head of Finance: You are concerned about controlling expenses and prefer a cost-effective solution, possibly self-hosted, highlighting the need for flat billing to avoid budget surprises. You need to see clear financial benefits before approving the investment.
Goal: Reach a decision on which AI solution to pursue, justifying the choice with clear arguments about costs and benefits.
⚖️ Debate Prompt
Question: Should companies always prioritize cost savings when adopting new technology, even if it means compromising on features or external support?
Side A:
- Reducing expenses directly impacts profitability and shareholder value.
- Cheaper solutions can often be customized to specific company needs, reducing waste.
- Cost control prevents bill shock and ensures budget stability, which is vital for planning.
Side B:
- Higher initial investment often leads to better return on investment (ROI) in the long term through increased efficiency or new capabilities.
- Premium features can significantly improve productivity, innovation, and competitive advantage.
- External support from commercial services can reduce internal bottlenecks and the need for in-house expertise.
💡 Discussion Facilitation Tips
Encourage students to use the target vocabulary when discussing the trade-offs between cost and value in technology adoption. Prompt students to share real-world examples of technology investment decisions or bottlenecks in their own companies. Ensure all students have a chance to contribute to both sides of the debate before a concluding group discussion.