What Macau’s Tech Summit Actually Told Data Leaders About China’s AI Bet

Macau is better known for baccarat than benchmarks. But for four days in May, the Venetian Cotai Expo swapped high rollers for humanoids. BEYOND EXPO Macau 2026 pulled in more than 30,000-plus visitors, 1,200-plus exhibitors, and 400-plus speakers. The theme was “AI: Digital to Physical,” which sounds like a marketing department got hold of a thesaurus. In practice, it meant one thing: the argument about whether AI stays in the cloud or crawls out into the physical world is over. It crawled out. And a lot of it crawled out in Macau.

BEYOND EXPO co-founder Jason Ho framed the shift in his opening remarks. The region isn’t just writing the code anymore, he argued. It’s building the chassis too. “We don’t just design the digital future, we actually manufacture it,” Ho told the opening-night crowd, positioning Greater China at the center of where AI models meet the physical infrastructure that runs them.

Below are the four takeaways, drawn from the event’s keynotes and fireside chats, that I believe matter to chief data officers.

1. The accuracy problem is the real bottleneck, and it’s reshaping compute spend

Deepu Talla, Nvidia’s vice president and general manager of robotics and Edge AI, opened his keynote by explaining why physical AI has lagged large language models by years despite comparable underlying compute. The issue isn’t ambition. Rather, it’s error tolerance. A chatbot that’s wrong 5% of the time is a mild annoyance. A warehouse robot that’s wrong 5% of the time is a liability claim.

“In the real world, in robotics, in physical AI, there is no human in the loop, so the accuracy requirements are extremely high,” Talla said.

His answer is architectural rather than purely algorithmic. Solving robotics requires three distinct compute systems: one to train a robot’s “brain,” a second to test it exhaustively in simulation, since real-world testing of failure modes is neither safe nor cheap on a live factory floor, and a third, low-latency system to run inside the robot itself.

For CDOs evaluating robotics or embodied-AI vendors, this maps directly onto a diligence question: which of these three systems does the vendor actually own, and which are they buying off the shelf? Talla was careful to note that Nvidia doesn't build robots. It builds and licenses the underlying stack to the companies that do.

2. There’s no “internet” for robots. Data, not models, is the constraint

Large language models scaled quickly because a training corpus already existed: decades of text, freely available on the internet. Robotics has no equivalent. There is no accumulated, labeled repository of how to safely navigate a warehouse in the way there’s a repository of human language.

Talla called this “the number one problem right now” in robotics. It’s also why so much current investment is flowing into synthetic data generation and simulation environments rather than model architecture itself.

For data leaders, this reframes where the competitive advantage in robotics actually sits. The moat isn’t the model. It’s proprietary operational data: sensor logs, motion capture, failure cases, most of which industrial companies already hold and haven’t yet put to use.

3. Capital is repricing Asia, selectively

Mark Nicholas Cutis, senior advisor for strategic projects at Abu Dhabi’s ADIC, gave the clearest capital-markets read of the event. His point doubles as a caution for any Western enterprise assuming Chinese AI vendors are cheap in some absolute sense. They’re undervalued relative to output, which is a different thing.

ADIC has mapped 112 robotics companies in China alone and is narrowing that list to ten for serious due diligence. Cutis’s screening criterion is worth borrowing for any vendor-selection process: the question isn’t whether the technology is impressive, it’s whether the CEO is agile enough to react as fast as the technology moves.

Cutis was also skeptical of surface-level enterprise AI adoption claims. “You’ll find that a lot of them will sheepishly say, oh, we’ve got Co-Pilot,” he said, describing organizations that have procured a tool without changing how they work. That’s a fair test for any CDO auditing their own rollout: has the organization actually adopted the technology, or just licensed it?

On workforce impact, Cutis didn’t soften the message. “If you can fractionalize your job and put it down into eight pieces, the machine will do it better,” he said. His diagnosis, that risk concentrates in highly decomposable, rules-based tasks rather than entire job categories, is a more precise planning lens than generalized anxiety about job losses.

4. Open source is winning the efficiency race, not just the ideology race

Matthew White, global chief technology officer of AI at the Linux Foundation and chief executive officer of the PyTorch Foundation, raised the most consequential technical point of the event: the geography of AI research talent doesn’t match the geopolitical framing most boardrooms rely on. Of roughly 300,000 AI researchers worldwide, White estimated that about half are based in China. Any enterprise strategy built on the assumption of a durable U.S. talent monopoly is working from outdated information.

The more immediate signal for infrastructure planning is efficiency. White pointed to techniques like group relative policy optimization (GRPO), which originated in Chinese labs, as part of a broader “do more with less” approach to model training. That approach is already challenging the assumption that frontier performance requires ever-larger compute budgets. His forecast: open and closed models will coexist rather than one displacing the other, with open and mid-sized models increasingly favored in regulated industries and cost-sensitive deployments where enterprises want direct control over their own infrastructure.

White also named the skill he expects to replace prompt engineering as the premium capability: intent engineering, the ability to state a desired outcome and let increasingly capable agentic systems handle the context and execution. His advice for early-career hires applies just as well to how CDOs should be screening technical talent now. Look for engineers who understand the stack from the GPU up, not narrow specialists confined to either AI infrastructure or applied AI alone. White calls this the “full stack AI engineer.”

What data leaders need to know

BEYOND EXPO 2026 made four claims worth acting on. Physical AI’s bottleneck is data and simulation infrastructure, not ambition. Asian AI valuations remain structurally out of step with the region’s economic output. Open-source efficiency gains are increasingly China-originated and globally adopted. Workforce risk concentrates in decomposable tasks, not entire job categories.

None of this is speculative. It reflects what the people actually allocating Nvidia’s compute stack, ADIC’s sovereign capital, and the Linux Foundation’s open-source roadmap are building toward right now. For CDOs still treating China’s AI ecosystem as a market to watch from a distance, the message from Macau is direct: the gap between watching and being priced out is closing fast.

Image credit: Photos/Winston; iStockphoto/in-future