The Prompt

Hello reader,

For much of the AI boom, the story has been about compute. Nvidia became the world's most valuable company by supplying the GPUs that power today's AI models.

But GPUs do not work alone. They rely on memory chips to move massive volumes of data fast enough to keep AI systems running at full speed.

That made what happened in Asia this past week particularly revealing.

CXMT, China's leading memory-chip maker, surged 466% in its Shanghai debut on July 27, becoming the country's most valuable listed company. Just two days later, South Korea's stock market suffered a sharp chip-led selloff, with Samsung Electronics and SK Hynix falling despite the large profits AI memory has brought them. It has since rebounded, climbing 17% on Friday.

It's not hard to see the connection. Samsung and SK Hynix dominate the market for the advanced memory chips that have become essential to modern AI. CXMT is not yet on par with the South Korean giants, but its blockbuster listing can give it a much larger pool of capital to expand its memory-chip ambitions.

The timing is also worth noting. Reports emerged that a Chinese state-backed company had begun producing key lithography equipment used to make these chips. What rattled investors was the possibility that China could close its gap with the US faster than expected.

For the past year, China's AI progress has largely been judged by its models. From DeepSeek to Moonshot AI, Chinese companies have emerged as serious challengers, forcing a fresh debate over open vs. closed AI in the US.

But its advances in hardware may prove to be the bigger story in the months ahead.

That shift is also playing out in Southeast Asia. My colleague Glenn's latest map of the region's physical AI ecosystem shows that local startups are focusing less on frontier AI models and more on putting AI to work in the physical world.

Later in this edition, we speak with Zuzanna Stamirowska, co-founder and CEO of Palo Alto-based AI lab Pathway, about an alternative to today's compute-heavy model architecture - one that also puts memory at its center.

Pranav Balakrishnan, AI reporter

TOP AI READS FROM OUR DESK

1️⃣ The startups behind Southeast Asia's physical AI push

Investors are looking for physical AI's ChatGPT moment. Robotics and physical-AI startups raised US$16.3 billion across 492 deals globally in the first quarter of 2026, according to PitchBook. Asia accounted for 54% of these deals, though only about a quarter of their value.

Southeast Asia is joining in. Our tracker identifies 35 physical-AI startups across six categories, which have collectively raised US$496.5 million. Rather than trying to replicate the scale of the US and China, most Southeast Asian players are tackling narrower commercial jobs or building the infrastructure needed to train and deploy machines.

2️⃣ Firms are spinning off AI tools, but are clients interested?

Asian consumer tech companies and marketplaces have collected data over the years, plenty enough to build their own AI tools. Now, they are trying to commercialize these tools. Can an AI side-hustle turn them into serious players?

3️⃣ AI fever leaves SEA's consumer startups fighting for attention

Southeast Asian VCs have long backed high-growth consumer businesses, from ecommerce and fintech to digital platforms. But these bets, which depend on startups winning significant market share, are getting harder to underwrite as customer acquisition costs rise and AI firms take the spotlight.

Slower economic growth, softer consumer spending, and higher oil prices stemming from the US-Middle East conflict add even more hurdles to getting these investments.

WHAT WE THINK

1️⃣ Moonshot reveals how AI is taking over more of the work behind AI

Moonshot AI and US-based inference provider Together AI held a public discussion recently, and one detail stood out.

Together AI research engineer Austin Silveria described an internal agent system that prepares new models for deployment. It runs tests on GPU clusters, finds performance bottlenecks, and tries fixes in parallel or overnight. If it needs a tool it doesn't have, it builds one and reuses it later.

Moonshot did something similar. Its technical report says an early version of Kimi K3 handled most of the company's late-stage work on optimizing how the model uses AI chips.

To be clear, humans still set the direction in these two scenarios. Engineers determine the goals, build the safeguards, design the tests, and decide which changes to keep. What's different is their level of involvement: less running experiments by hand, more supervising a system that runs many of them at once.

Open weights don't make a model easy to deploy on their own. Silveria said Kimi K3 needs at least eight of Nvidia's newest data-center GPUs to run, and older hardware needs multiple servers wired together. For most companies, renting Kimi K3 through a specialist provider remains a more practical option than running it themselves.

Minako Kojima, Moonshot's head of developer relations, also floated the idea that cheaper models with K3-level capability could show up soon, though she gave no timeline.

Either way, this development makes it clear that the AI race now includes automating the engineering work needed to keep these models running.

2️⃣ Sarvam is trying to own every layer of India's AI stack. But its flagship model is still missing

At Epoch, Sarvam's inaugural developer conference, the company presented a version of sovereign AI that reaches far beyond a sovereign LLM. Its biggest announcement was still a promise: a trillion-parameter model expected within six months.

Everything else focused on expanding beyond the model. Sarvam launched an India-hosted inference platform, introduced AI agents for research and coding, upgraded its speech and vision models, and unveiled new products ranging from a dictation app to services for voice agents and smart glasses.

The unifying idea is "token sovereignty": capture more of India's AI usage, whether through their own models, some other model, applications, or the infrastructure underneath them. This puts Sarvam up against frontier labs, neoclouds, Cursor, Wispr Flow, and Meta, alongside the many other voice AI upstarts.

The question is whether that is a coherent sovereign-AI strategy or a broad hedge while Sarvam waits for a flagship model.

FOUNDER FOCUS

We speak with AI founders and executives about AI, startups, and whatever else is on their minds.

Nearly every leading chatbot you use today runs on the same piece of math. The transformer, published by Google researchers in 2017, does one thing: it predicts the next word. Do that well enough, billions of times over, and you get ChatGPT.

Now, a Palo Alto AI lab is trying to build an alternative. Pathway calls it the post-transformer world. We spoke with co-founder and CEO Zuzanna Stamirowska to know more.

The interview has been edited for brevity and clarity.

Why do you think AI needs to move beyond transformers?

This is math made to predict sequences. In language, you're trying to guess the next word. And this math is limited.

Right now, transformers have no memory after they are trained. You train a model, the weights are set. As you use it, you deal with context, but that doesn't impact the weights anymore. So what transformer models do is relive their Groundhog Day, or the Memento movie. They wake up every day with a fresh mind, without having internalized anything that happened to them.

Acquiring new skills, acquiring judgment, is something you do through experience. Transformers fundamentally don't do that.

If transformers are limited, what does a better approach to reasoning look like?

Reasoning was bolted on to the transformer with a chain of thought, a process in which every step of thinking has to be verbalized in language. That constrains thinking to language.

What we discovered at Pathway is that thinking doesn't happen in language. It happens in some sort of abstract space. Once you release reasoning from the constraints of language, you can do way more powerful things way faster.

Sudoku is a good example. It's difficult to express in language sequentially, but it's obvious as you see it. Your grandmother can solve it over coffee, and yet it's deeply difficult for LLMs to do natively. They do it by finding tricks.

This matters because compute is exploding. Anthropic would have had even larger growth if it had more compute. A lot of that compute is going toward reasoning, because the chain of thought requires so much of it.

If you move to latent reasoning, you cut these costs dramatically. For us, solving Sudoku - both training the model and running it - costs so little it was unworthy of reporting.

Can you explain the alternative architecture Pathway is building?

We published a paper in October titled "The dragon hatchling: the missing link between transformer and the brain." It was the second most popular paper on Hugging Face in 2025.

Everything is organized as neurons, small computational units connected through synapses, forming a network like a brain. Not everybody is connected to everybody. You have your neighbors, a bit like in a social network. When a signal comes in, the neuron that gets activated pushes it only to its neighbors. If they find it relevant, they fire up too. And if both neurons at the ends fire up, that connection becomes stronger. This forms a memory.

So memory is stored directly on the synapses within the model. In transformers, you have fixed context windows. Here, you have access to the entire context all the time, but you only trigger the synapses relevant to your query. That gives us incredible efficiency.

Does moving beyond transformers mean changing the hardware stack?

That was the key part. Transformers sort of won the hardware lottery, and that's great, because it allowed scale. But we made this architecture work on any GPU. We mostly work on H200s.

Will it work better one day on specialized chips? Yes, but you don't need any special chip. If we had to change the chip and wait for everything to align, that would be a difficult path.

Where has the most interesting post-transformer research been coming from?

Quite the opposite of what you'd expect. The US is the most transformer-religious place. There was a time when most of the papers I was reading were from China, because there was no religion about doing just the transformer, and it was far more innovative.

The first interesting work on sparsity that we saw came out of China. The US was very focused on brute-force scaling for a long time.

How close are you to bringing this to market?

Right now, we are a lab. We invest in our moat. The more time we spend being dramatically differentiated, the bigger the impact.

We have a partnership with AWS and Nvidia, and we're working with design partners on personalized AI and long-horizon reasoning. On cost, it's at least 10x less than open models for advanced reasoning. We should have public numbers soon.

Strategically, it makes little sense to build yet another OpenAI or yet another Anthropic. A breakthrough is inevitable - the markets just don't see it yet.

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.