From Google Chips to Humanoid Robots: The AI Race Is Now Entering the Era of “Cheaper and More Efficient Brains”

Two major developments in the technology industry currently appear to come from different worlds.
In Silicon Valley, Google and Marvell are expanding their partnership to develop custom AI chips. Meanwhile, in China, Unitree founder Wang Xingxing warned that humanoid robots still have relatively low efficiency, although their use could expand widely once the technology reaches a higher level of maturity.
In fact, the two stories share one common thread:
The future of AI will not be determined by intelligence alone. It will be determined by efficiency.
Google No Longer Wants to Depend on a Single Path
The partnership between Google and Marvell demonstrates the growing importance of custom silicon strategies in the AI race.
Marvell granted Google warrants to purchase up to approximately 58.97 million shares with a potential value of around US$12.2 billion. The agreement is related to technological cooperation on custom AI chips and, if certain procurement targets are met, could potentially generate cumulative revenue of up to US$120 billion for Marvell through fiscal year 2033.
However, the core of this agreement is not merely the US$12.18 billion investment.
Google is seeking to expand its control over the entire AI computing chain, from accelerators and networking to memory and storage technologies.
As long as AI requires more computing power, technology companies face a simple problem: GPUs and AI infrastructure are extremely expensive.
That is why custom chips are becoming increasingly strategic.
The goal is not merely to make AI faster.
The goal is to make each unit of intelligence cheaper to run.
The agreement is also viewed as supplier diversification, rather than a total replacement for partners such as Broadcom. The market reacted quickly: Marvell shares surged after the announcement, while Broadcom came under pressure temporarily as investors assessed that competition for the AI chip supply chain was becoming more open.
Wang Xingxing Reveals a More Difficult Problem: AI Must Learn to Live in the Real World
If data centers face the problem of computing costs, humanoid robots face a far more complex problem.
The real world is not like a data center.
Robots must understand changing environments, maintain balance, recognize objects, respond to commands, and make decisions when faced with situations they have never encountered before.
Wang Xingxing said the robotics industry is still waiting for a kind of “ChatGPT moment” for physical intelligence, or embodied AI. He estimated that this major breakthrough could still take around two to ten years, even as robot technology continues to advance.
This is why humanoid robots have not yet been adopted at scale like smartphones.
Hardware can be built.
Robots can walk.
Robots can even perform impressive demonstrations.
But performing tasks consistently, efficiently, and autonomously in new environments remains a major challenge.
AI Is Facing the Same “Cost Problem” in Two Worlds
Google and Marvell are trying to make digital intelligence more efficient.
Unitree and the robotics industry are trying to make physical intelligence more efficient.
Both face the same economic question:
Can the cost of running intelligence decline faster than its usage grows?
If the answer is yes, major changes could occur.
The history of technology shows that mass adoption often does not begin when a technology is first successfully created.
Mass adoption begins when the technology becomes good enough and cheap enough.
Computers did not immediately enter every home.
The internet was not immediately used by billions of people.
Smartphones only took off after devices, connectivity, and applications reached the right combination.
Humanoid robots will likely follow a similar pattern.
Wang Xingxing believes robot use could expand widely as technological capabilities become more mature. However, limitations in efficiency and the ability to work generally in real-world environments remain the main obstacles at present.
Conclusion: The AI Winner May Not Be the Smartest
So far, the AI race has often been portrayed as a competition to create the most intelligent model.
However, the next phase may be more interesting.
Who can make AI the cheapest?Who can run it with the greatest energy efficiency?Who can take that intelligence out of data centers and into the physical world?
Google and Marvell are betting on the first side.
Unitree and humanoid developers are pursuing the second.
Custom chips may make AI increasingly cheaper to think. Humanoid robots may finally enable AI to work.
When the two revolutions meet, AI will no longer be merely a technology that answers questions on a screen.
It could become a digital and physical workforce.
And perhaps, that is the greatest opportunity the market is building today.
#SummerCreationCamp