The Prompt

Hello reader,

Last week, Nvidia teamed up with BlackRock, Apollo, and other Wall Street firms to mobilize US$500 billion for an AI infrastructure initiative that aims to make it easier for companies to buy Nvidia GPUs and build AI data centers.

The process is similar to how someone would take out a loan to purchase a car. They would pay it back over time, with the car itself being part of the security.

In this case, a cloud company looking to buy thousands of GPUs may not need to raise all the money upfront. It could finance the hardware and repay the loan with the revenue it earns from renting out compute.

Rather than deep-pocketed hyperscalers, the biggest beneficiaries could be neoclouds and co-location companies, especially in India and Southeast Asia, where AI infrastructure fundraising is picking up. India's Neysa became a unicorn in February after raising US$600 million, while Singapore-based DayOne raised more than US$2 billion in a series C round in January.

Still, lenders will want to see proof that the GPUs will keep generating revenue. The question may become less about balance-sheet strength and more about who can promise demand for their compute.

Nvidia's shares fell after the announcement, and it's easy to see the market reading the deal as yet another example of circular financing in AI.

It also comes as competition grows. I wrote last week about startups across Asia trying to build alternatives to Nvidia's GPUs.

By making its own chips easier to finance, Nvidia is giving the next generation of AI clouds one more reason to stay in its ecosystem.

Later in this edition, we speak to Ashish Dsa, co-founder and CTO of New York-based Arbor. Most voice AI is built to speak to customers, but this firm is using it to get employees to share more useful feedback with top management.

Pranav Balakrishnan, AI reporter

TOP AI READS FROM OUR DESK

1️⃣ What's stopping Asia's chip makers from taking on Nvidia?

There is a gold rush underway in AI chips. Across Asia, conglomerates and scrappy startups alike want in. But getting a slot at Taiwan-based TSMC, a monopoly in the production of advanced chips, is close to impossible.

Some South Korean startups are showing a way through. Their advantage? Having powerful friends. Will Indian and Southeast Asian chip startups follow South Korea's playbook or find a path of their own?

2️⃣ The AI infrastructure boom is built on data nobody actually has

Guest writer Matt Walker, a telecom industry analyst, argues that the AI boom is largely proceeding based on emotion, not logic. Without hard, vetted industry data on what AI traffic costs, how fast it's growing, and who measures it, the industry is betting hundreds of billions on numbers no one can verify.

3️⃣ Graas buys Temasek-backed Trustana in agentic commerce push

Graas' platform lets businesses build their own AI agents for retail and commerce use cases. It will use the funds to create a database where agents can access accurate customer and product information, with the customer data coming from Graas and the product data coming from the newly acquired Trustana.

WHAT WE THINK

1️⃣ Meta's open-weight push is a business opportunity for Asia

Last week, Meta released Muse Glimmer, an open-weight model designed to run AI agents locally on a Mac or PC with a consumer GPU. It can handle coding, tool use, file management, and document analysis.

More importantly, Meta is returning to open-weight models after keeping Muse Spark closed. It also said that it will release the weights of a stronger Muse Spark model later.

An Indian edtech founder I spoke with last week was excited by the prospect. If more AI tasks can run on a device instead of the cloud, companies may not have to rely as much on expensive cloud inference. That could lower compute costs and improve performance.

It could also be good news for chip startups in India and Southeast Asia. As we reported last week, few regional companies are trying to build alternatives to Nvidia's general-purpose GPUs. Many are instead building chips for edge use cases: cameras, robots, industrial systems, and other devices that need to run AI locally.

Meta's move will not create that market on its own. But more capable models that can run on-device could push more inference to the edge, creating a bigger opportunity for the chip startups trying to serve it.

2️⃣ American AI clouds are increasingly drifting to Asia

Together AI has given Indian infrastructure company L&T an order worth up to US$1.6 billion to build an AI data center in Chennai with 10,000 Nvidia B300 chips.

It is a significant deal on its own, but it makes more sense alongside a development from last month. Several Nvidia-backed neoclouds - including CoreWeave, Together AI, Nebius, Lambda, and Crusoe - were reported to be in talks to lease capacity from Indian data center operators.

The next generation of American AI clouds may not build all their capacity in the US. They can lease or commission it in India and other markets while continuing to sell compute to customers globally.

The timing is interesting, as building data centers in the US is becoming harder. Communities are pushing back over power consumption, water use, and noise. Lenders are beginning to treat local opposition and permits as a real risk.

India will have to deal with these issues too, but it has not yet seen anything close to the same level of resistance.

FOUNDER FOCUS

In this edition, Tech in Asia speaks with Ashish Dsa, co-founder and CTO of Arbor, a New York-based startup that runs anonymous AI-enabled voice interviews with frontline workers. Arbor's bet is that people will say more to a voice bot than to a survey or a superior.

The interview has been edited for brevity and clarity.

What does Arbor actually do?

Companies already run employee surveys, but they are usually one-to-five ratings, and people do not reveal much. There is fear around saying something negative. There is also hesitation around saying something positive, because people do not want to come across as condescending.

We run anonymous AI voice interviews with employees, customers, and contractors. We begin with a consent screen, make clear that their response cannot be traced back to them, and then let them have a conversation.

The AI can also sense when someone is uncomfortable. It knows when to move on instead of pushing them. That is not something a traditional survey can do.

Voice AI is mostly associated with customer service and sales calls. What has changed to make it useful for employee research?

Voice AI has only recently become good enough for this. Earlier systems worked like an assembly line: speech was converted to text, passed to a language model, and then turned back into speech. Each step added a pause. It did not feel like a conversation.

Real-time models now combine those steps, so the AI can respond quickly enough to hold a natural exchange. That matters when you are asking someone about their workday. If the system pauses after every answer, people stop opening up.

But the bigger shift is behavioral. People are getting more comfortable speaking to AI. They may still use it to book an appointment or solve a customer-service issue, but there is now an opportunity to use voice AI for something more open-ended: finding out what people really think.

Why would people speak more openly to an AI than to a manager?

People are more vulnerable with AI companions than with friends, sometimes even therapists. A human can judge you. With an AI, you feel you have a free pass.

The setting matters too. In an unstructured conversation, people speak more. Someone who may have given a two-minute answer in a survey can speak for 20 or 30 minutes if they are guided politely.

Trust also builds over time. In the first interview, someone may say a little. By the third or fourth, they are more likely to tell you what is actually going on.

Can you give me a real-world example?

We worked with a large freight operator in Los Angeles that had warehouses and terminals operating in isolation. It was hard for management to know what was going wrong across all of them.

We interviewed workers across a few hundred terminals and found that the fastest teams had something very simple: free cans of WD-40, a lubricant oil. Their loading equipment would jam, and instead of waiting for a mechanic, workers would spray the hinges and keep going.

That was one small detail from one conversation. But once the company made WD-40 available across its terminals, loading and unloading times came down.

A consulting firm would have sent people terminal by terminal, and workers called into a room may not have said the same thing anyway.

Does this risk turning employee feedback into workplace surveillance?

That is everything for us. If people think we can trace a response back to them, they will stop talking.

We do not build a system for continuous listening. That would be an anonymity nightmare. We use selective listening: a worker is invited to take an interview and chooses whether to participate.

There are exceptions for serious safety issues. If someone indicates that they may harm another person, the company gets an alert. But even then, it does not get a name.

What does the technology look like behind the scenes?

For these conversations to feel natural, latency matters a lot. If the AI pauses too long after every response, it stops feeling like a conversation and becomes a monologue.

We often use real-time models, where speech comes in and a voice response comes out through one system. That gives us less flexibility than stitching together different speech and AI models, but it is much faster.

Where does Arbor go from here?

Most AI is still one-shot. You ask a question, it answers, and that is it. We want to build something that gets better with every interview, correction, and piece of feedback.

The people at the top of a company often do not know what is happening on the ground. The people on the ground usually do. The question is whether companies can finally hear them.

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.