Alvin Wang Graylin

Author “Our Next Reality”| Digital Fellow, Stanford HAI | Senior Fellow, ASPI-Center for China Analysis | Prof, UW | Chairman, VWS | 4X Founder | Speaker🎙️

[Alvin Wang Graylin @AGraylin]

@AshleyDudarenok Thanks for sharing. We are headed into rocky times.

[Alvin Wang Graylin @AGraylin]

Full episode here: https://t.co/Fsf7aLAU8J

[Alvin Wang Graylin @AGraylin]

🚨The #SpaceX #IPO today is likely the beginning of the process for deflating the massive AI #bubble that’s been built over the last few years. I’m not saying #AI isn’t a valuable technology, but I am saying that the expectation that AI labs will reap all its value and the overbuilding of data centers are misguided ideas that will be proven false in the near future.

More details in the full episode. #BigBangTechReport (Spotify)

#OpenAI #Anthropic @superai_conf

[Alvin Wang Graylin @AGraylin]

So much to see and so many to meet at @superai_conf in Singapore!🇸🇬 10k #AI industry experts in one place! 👍 Congrats @peternoszek on another successful show.

[ATP @ATPinsights]

Within two to three years, most AI will run locally on phones and laptops rather than through the cloud. That is the argument @AGraylin, Digital Fellow | Sr. Fellow at Stanford HAI | Asia Society Policy Institute, makes based on the current trajectory of model compression.

Models are shrinking and becoming more capable each year. A model that runs on a phone today without an internet connection matches the performance of the best online models from a year ago. Alvin expects that to continue, with frontier-level models running on laptops within another year or two.

Apple is the only company among the Magnificent Seven to have spent zero dollars building cloud AI infrastructure. While other major players are each spending $100 to $200 billion annually on cloud compute, Apple has focused entirely on on-device delivery, announcing at WWDC that it intends to make all its AI available on device.

Chinese open source models, including MiniMax and Kimi, are small enough to download and run locally and are free. That is accelerating the shift. Open Router data shows that one year ago, 1% of AI traffic ran on open source models. Today that number is 60%.

This category is called edge AI. Alvin sees it as the dominant delivery mechanism for AI within two to three years, and the speed of the transition is outpacing most people's expectations.

[Alvin Wang Graylin @AGraylin]

Full episode here https://t.co/qfqmnrreAS

[Alvin Wang Graylin @AGraylin]

The #SpaceX IPO is coming tomorrow. What does this mean for the #AI industry and the world? Have a listen to the latest #BigBangTechReport 🎙️🚀

@JeBuhr @DigEconLab

[Alvin Wang Graylin @AGraylin]

Read this to better understand what I spoke about.
https://t.co/6TSTlSpjQw

[Alvin Wang Graylin @AGraylin]

What a unique experience to speak in such a spectacular venue. Thanks @BalzerSB and #TEDxBerlin for the opportunity. The AGI windfall for the AI labs is a mirage, but the oasis of a better world can be real…if we take the right actions now. ✊🌍

@VRWorldSociety @TEDxBerlin1

[Alvin Wang Graylin @AGraylin]

@bookwormengr Thanks for sharing my friend! Action is needed!

[Alvin Wang Graylin @AGraylin]

@RajaKorman Thx for sharing

[Alvin Wang Graylin @AGraylin]

Link to full essay. Please share it with others. https://t.co/6TSTlSpRG4

[Alvin Wang Graylin @AGraylin]

Are we pouring trillions into an optical illusion?
In my latest piece for #Abundanist, I take a hard look at the core assumption driving the current frontier AI race: the false belief my many lab leaders (and their VCs) that the first lab to achieve #AGI will capture a multi-trillion-dollar monopoly on global white-collar labor value.

Look at the recent SpaceX/xAI S1 filings, where "Enterprise Applications" (wage replacement) represents ~80% of its massive $28.5T targeted market. The logic seems simple: spend a few hundred billion now to capture tens of trillions later. 🚀

But drawing from my 35 years across the AI, semiconductor, investing and tech policy space, I argue that this "AGI Windfall" is a mirage.

The massive technology shifts are real, but value capture is notoriously difficult. Consider these counterintuitive realities:

1 The Cisco Lesson: Cisco was 100% right about the internet revolution, yet lost 86% of its market cap when the infrastructure capex cycle decoupled from immediate ROI.📉

2 The Supercar vs. Moped Dilemma: Frontier AGI models are supercars, but the global economy runs on buses and mopeds. Cheap, "good enough" open-source and diffused intelligence will commoditize the very market frontier labs expect to monopolize. 🏎️

3 The Nominal GDP Paradox: Massive AI productivity gains will drive down costs dramatically, benefiting consumers and human quality of life, but potentially causing nominal GDP—and direct vendor revenue—to stagnate or fall. ⁉️

4 There’s clear research showing that when just 3.5% of the population of a country protests, regimes fall. The current path we are on will disenfranchise an order of magnitude more. If the regulators don’t step in soon, the stability of the country is at stake. 🚨

The short-term race to monopolistic AGI is self-defeating, capital-destructive, and structurally destabilizing for our economy and country.

The true, sustainable path forward isn't winning this week's benchmark—it's driving the broad-based diffusion that unlocks true global abundance.

This essay is a bit long, but the topic warrants it. I promise you’ll learn something new. 💡 Please share it with others that can benefit from a deeper understanding of the topic.

See link below. 👇

@DigEconLab @VRWorldSociety @AsiaSociety @AsiaPolicy @citrini

[Alvin Wang Graylin @AGraylin]

@Prakharmis39951 @superai_conf Read this if you have time. https://t.co/LR6Gce12Rl

[Alvin Wang Graylin @AGraylin]

Look forward to insightful discussions about the future of #AI at the coming @superai_conf in #Singapore. Please join if you’re in town June 10-11th.

[peter @peternoszek]

SuperAI speaker list.

Spot who’s flying in from SF/NY/Boston vs Shanghai/Beijing/SZ.

[Alvin Wang Graylin @AGraylin]

Details here. https://t.co/AIdI7pYuFU

[Alvin Wang Graylin @AGraylin]

Ironic that @AnthropicAI is now distilling from @Alibaba_Qwen and @deepseek_ai for its Opus 4.8 #AI model. 😅 Two months ago, Anthropic was complaining about Chinese labs distilling from them… 🤷‍♂️

[Max For AI @MaxForAI]

笑死了，Claude Opus4.8蒸馏了阿里巴巴Qwen啊🤣

通过API用中文问你是谁，会很大概率回答

我是通义千问（Qwen），是阿里巴巴集团旗下的统义实验室自主研发的超大规模语言模型。

[Alvin Wang Graylin @AGraylin]

@quxiaoyin Yes your results are consistent with the findings from this #Stanford paper on Enterprise AI. https://t.co/0wWZVz47ij

[Jukan @jukan05]

Huawei's latest announcement carries real significance, because China has, in effect, shown the direction in which advanced technology needs to move. And it has done so in cutting-edge semiconductors, no less.

China has long been a follower. In semiconductors, Western technology played the role of the pioneer, while China was preoccupied with simply keeping pace.

But by banning EUV exports to China, the U.S. manufactured a bottleneck at the lithography tool — and in doing so, it effectively forced creativity onto China.

To circumvent the sanctions, China was pushed toward approaches the West had never needed to take.

That is exactly what today's announcement represents.

Where Nvidia co-designs memory, packaging, and logic to optimize TCO at the system level — doing it rack by rack — Huawei is doing the same thing at the chip level.

I'll say it again: this is a genuinely striking approach. Memory makers are already struggling with cost scaling. As linewidths shrink, the resources required to keep shrinking them — capital, manpower, time — are climbing exponentially.

So the day will come when the West, too, must make packaging, logic, and memory collaborate from the node-design stage. And it won't be far off.

China, through the paradox of sanctions, has been driven to do this ahead of the West — unintentionally.

This is what genuinely frightens me. As YMTC has already demonstrated, U.S. sanctions pushed China to skip the incumbent standard and jump straight to the next-generation one.

The result? YMTC carved out a meaningful presence in hybrid bonding — and even Samsung, the king of NAND, ended up licensing YMTC's patents.

I believe the West may well find itself licensing this Huawei technology a few years down the road. And I believe cases like these will multiply, spreading China-style standards in their wake.

[Alvin Wang Graylin @AGraylin]

Interesting read on how @deepseek_ai innovations in #AI models were driven by China’s HW constraints and why its solutions will ultimately negate those constraints. 💡🐳

[GDP @bookwormengr]

https://t.co/KHnPhxJBiz