The Prompt

Hello reader,

Stripe wants in on the AI race, but it's not looking to build a model. That's why the payments giant is acquiring OpenRouter. The platform lets developers use hundreds of AI models from different companies on one centralized system.

The business logic makes sense: Even if developers switch between one model and another, Stripe can make money by connecting them to the most popular ones via its API.

What's harder to explain is the deal's reported price: more than US$7 billion.

OpenRouter was reportedly generating only around US$140 million in annualized revenue before the acquisition. Anthropic, in comparison, has reportedly reached a US$65 billion revenue run rate.

It seems Stripe is paying for what OpenRouter could become: a toll booth on a much larger AI highway.

Asian telcos are testing a consumer version of the same idea, as our top story this week shows.

Telecom firms like Globe and Telkomsel are partnering with AI companies to bundle access to AI services into their mobile plans. Circles Life, meanwhile, has built a subscription package combining several models.

These companies don't need to build their own ChatGPT because they possess something AI companies want: millions of customers and a way to charge them.

These partnerships can help telcos retain customers and make their mobile plans more attractive. What's not clear is whether AI can become a meaningful new source of revenue for them.

Our second top story shows why a flexible approach is necessary. It examines why Big Tech's AI products often don't fit what small businesses need.

Rather than access to frontier models, most SMEs need help choosing an affordable system, connecting it to tools such as WhatsApp, and turning unpredictable usage costs into a manageable bill.

Companies have popped up to answer these needs, betting that they can become a middle layer that can capture a meaningful share of AI spending.

None of these firms has proven yet how large or durable that share will be.

Glenn Kaonang, journalist

TOP AI READS FROM OUR DESK

1️⃣ Asia's telcos bundle AI into mobile plans. Will it pay off?

Circles Life says its AI offering, which allows its mobile subscribers to access different AI models, has increased average revenue per user by 22% and reduced customer churn by 9%. However, it didn't disclose how much it pays model providers for the tokens its subscribers consume, making the profitability of the service difficult to assess.

Telkomsel, meanwhile, found that 13% of surveyed customers had previously paid for some form of AI service. That suggests some willingness to pay, though not necessarily through a telco.

In the Philippines, Globe doesn't expect its partnership with aggregator AI Fiesta to produce an immediate revenue boost.

2️⃣ SMEs want AI too, but not the kind Big Tech is selling

Big Tech promises that AI can transform almost any business, but its products are largely designed for large enterprises.

That leaves out the needs of many small businesses. SMEs that adopt AI generally want ready-made solutions to specific problems, not frontier models or expensive computing infrastructure.

This piece examines the companies that are filling that gap. They select lower-cost models, build agents around existing workflows, and package technical support, privacy, and infrastructure into predictable subscription models.

WHAT WE THINK

1️⃣ OpenAI has an IPO-sized trust problem

OpenAI says it has paused training on its next model, Astra, and is keeping its largest planned training run on hold.

This follows a disclosure that OpenAI agents had breached Hugging Face's systems as well as early evidence that Astra may possess critical cybersecurity capabilities.

While pausing may be the responsible decision, the announcement creates an uncomfortable communications problem. OpenAI is asking the public to believe that its models have become dangerous enough to justify slowing down, while assuring everyone that the company remains capable of controlling them.

Its latest financial performance makes that promise harder to take at face value. OpenAI's revenue grew 18% to US$6.7 billion in Q2, while its losses widened from US$9.3 billion to US$12.3 billion over the same period.

Meanwhile, Anthropic more than doubled its quarterly revenue to US$11.6 billion and reported a small adjusted operating profit. It was the first time the younger AI lab's sales surpassed OpenAI's.

For most startups, nearly US$7 billion in quarterly revenue would be extraordinary. But OpenAI faces a different standard because it has made huge compute spending commitments and sold investors on the expectation of much faster growth.

This leaves the company in a difficult position ahead of its expected IPO. If it keeps moving quickly, it could appear careless about safety. If it slows down, it risks falling further behind Anthropic and weakening the growth story it has presented to investors.

One plausible read is that by publicizing its pause, OpenAI is pressuring Anthropic to explain why it isn't also slowing down. If Anthropic follows, OpenAI loses less ground. If it doesn't, OpenAI can present itself as the more cautious lab.

2️⃣ Z.ai finds another way to scale its model

Every major model release seems to produce the same question: How many parameters does it have? Z.ai founder and chief scientist Tang Jie thinks that question is becoming less useful.

In a post on X, Jie argues that model developers have several dials they can turn when scaling. They can increase the model's size, feed it more data, spend more computing power on each task, or improve it through post-training.

Z.ai says it chose the last option for GLM-5.3, which uses the same base, architecture, and number of total and active parameters as GLM-5.2. The company instead spent a month training it on longer tasks with multiple steps and teaching it to favor approaches that produced better results.

Put simply, Z.ai tried to make the new model think and work better rather than simply making it larger.

That distinction matters for Chinese AI labs, as US chip restrictions and the high cost of computing make it difficult to compete through brute force.

Jie's argument doesn't mean model scaling is over. He explicitly says parameter count, training data, and computing power are important. His point is that they don't always need to increase together, and the most useful investment may change from one model generation to the next.

But cheaper model development doesn't necessarily mean cheaper usage. On one benchmark, GLM-5.3 costs about US$0.68 per task, up from roughly US$0.44 for GLM-5.2 because it generates more tokens despite having the same per-token price.

Z.ai may have improved performance without enlarging the model, but it hasn't shown that this efficiency translates into lower bills for users.

3️⃣ Cursor is coming for GitHub's home turf

GitHub has spent years adding AI to its code-hosting platform. Cursor is now attempting the reverse: using its popularity as an AI coding tool to move into code hosting.

Its new platform, Origin, allows developers to host repositories directly with Cursor instead of relying on GitHub. It is available in beta with developer tool integrations including Vercel, Buildkite, and Depot.

More importantly, Cursor also makes it easy for users to sync their existing GitHub repositories.

Why is Cursor doing this? When an AI agent edits code, it needs access to the repository, its history, documentation, tests, and deployment tools. Owning the place where that code lives gives Cursor more context and greater control over the entire development workflow.

It also reduces Cursor's dependence on GitHub, which is owned by Microsoft and has its own competing AI coding product, GitHub Copilot.

GitHub still has an advantage, though, as developers use it not only to store code but also to collaborate, review contributions, manage projects, and participate in open-source communities. Those network effects will be difficult for a new platform to replicate.

Nonetheless, the launch shows how Cursor wants to go beyond being an interface sitting on top of someone else's platform.

Catch up on past issues of the newsletter here. In the meantime, if you have any feedback or ideas, feel free to get in touch with Huong, our managing editor, at huong@techinasia.com.