Moonshot Has a 20,000 Nvidia Chip Cluster

ATP is presented by:

Welcome back to The Asia Tech Podcast. Today we discussed Moonshot AI’s massive Nvidia chip cluster, why tokenized stocks are booming on crypto exchanges, how robots learn social manners, AI avatars that ditch realism on purpose, agents that can finally learn from their mistakes, and the hidden money rails behind Australian retirement savings.

• 31:14 Alex Svanevik - CEO and Co-Founder, Nansen

• 42:53 Alisher Abdulkhaev - Co-founder & CTO, Kanaria Tech

• 1:02:52 Kurt Liu - CEO, XRSPACE

• 1:23:57 Sonam Pankaj - CEO, StarlightSearch

• 1:42:29 Ethan Lee - Founder, Smalt AI

• 2:02:49 Ben Styles - Chief Commercial Officer, SuperAPI

• 2:23:01 Bhagwat Pant - Consultant, CXGLOBE

Here is an overview of the topics we discussed today:

Michael dug into reports that Moonshot AI is running roughly 20,000 Nvidia Hopper-class chips through a compute deal with Alibaba, even though US export rules are supposed to block this. The math is wild: that cluster could deliver up to 80 exaflops of AI compute, enough to train Moonshot’s 2.8-trillion-parameter Kimi K3 model. The real story isn’t whether the deal breaks the rules. It’s that companies can now rent compute instead of owning chips, which makes export bans nearly impossible to police.

Nansen built its name by unmasking anonymous crypto wallets and tracking what “smart money” actually does, not just what it holds. Alex shared something surprising: on Hyperliquid, the top decentralized trading platform, 10 of the top 15 products traded aren’t crypto at all. They’re tokenized versions of the S&P 500, gold, and even SpaceX stock. He compares it to how stablecoins gave people around the world access to US dollars, except now it’s brokerage accounts.

Alisher’s company builds the “social brain” that helps robots navigate around people, not just avoid bumping into them. That’s trickier than it sounds. A hand wave that means “come here” in Japan looks like “hello” in the US, so the robot has to know which country it’s in. He shared a lovely example: a guide robot named Wakko-chan at a Japanese city hall that plays quiz games with kids and became a local mascot.

Kurt’s team makes AI avatars for customer service, and they learned an unexpected lesson: people don’t actually want photorealistic avatars. Even Foxconn’s chairman chose a cartoon version of himself for an internal town hall, because employees felt too intimidated to question a “serious” avatar. Kurt also described a bad case where an avatar kept smiling while an angry customer yelled at it, so now the system reads emotional tone and responds with more empathy.

Sonam argues today’s AI agents have a blind spot: they take actions without really understanding the consequences, which is why they hit a performance ceiling around 85%. She compares it to a kid touching a hot cup once and never doing it again, something most agents still can’t do. Her fix is “outcome-informed” agents that actually learn from what happens after they act, not just from more data.

Ethan’s platform turns messy AI drafts into work that’s ready for internal stakeholders, not just a chatbot answer. One striking example: an investment firm’s two-month due diligence project got a comparable analysis from Smalt in five minutes. He’s careful to note humans still have to judge whether the output is trustworthy, especially since studies show AI can make people twice as confident while making more mistakes.

Ben’s company moves the mandatory retirement contributions for 16 million Australian workers, which adds up to roughly $3.8 billion every week. SuperAPI now onboards 34,000 new employees a month and works with 7 of the top 10 Australian super funds. Notably, they keep AI out of the actual money-moving decisions and use it instead to make their engineering team faster, not smaller.

Bhagwat helps Western companies expand into Asia, and his biggest warning is simple: winning in Singapore tells you almost nothing about winning in Indonesia or the Philippines. He vets companies by asking if they have real customers and a genuine reason to be in the region, not just a “dot on the map” for investors. He also pushed back on a stat claiming Gen Z workers sabotage AI at work, saying older employees are usually more resistant to change.

That’s just a taste of everything we covered today. Watch the full conversation for all the details:

Watch the episode

Twilio is the secure and scalable platform to build amazing customer engagement without the complexity, with the AI-driven workflows you need to connect.

Visit Twilio.com to start building today

Boomi, the data activation company for AI, powers the agentic enterprise by bringing data to life across the business. Trusted by over 30,000 customers and supported by a network of 800+ partners, Boomi helps organizations of all sizes achieve agility, efficiency, and innovation at scale.

Visit www.boomi.com to learn more

Download Boomi’s latest market research: AI Ambition Meets Data Reality: APAC Technology Priorities and Challenges 2026

HCLTech is supercharging progress with AI that moves beyond ambition and delivers real business impact. A recognized AI leader, trusted by global enterprises to turn technology into transformation. HCLTech: Supercharging Progress

Visit https://www.hcltech.com/ai to learn more