It’s about 15 degrees in Shanghai today, with a few scattered clouds drifting across the sky. I happened to be sitting by the window, spacing out and looking at the sky, when my mind somehow wandered to another kind of “cloud”—cloud computing. That’s because I was scrolling through the news this morning and noticed a rather interesting trend in the tech sphere: the US seems to have figured something out and is planning to pause the new round of AI chip export restrictions.
A Calculating Move: Becoming the Global “Internet Cafe Boss”
If you’ve been following the hardware scene over the last couple of years, you probably know they were previously dead set on “physical supply cuts,” absolutely refusing to let high-end AI chips flow out. The result was exactly the opposite: the tighter the ban, the faster local alternatives developed elsewhere. This is actually a very simple reality: if you block the path to buying ready-made tools, people have no choice but to tinker away in their own backyards, forcing their own production lines to mature.
Now they’ve finally come to their senses, realizing that approach wasn’t really paying off. So, they’ve come up with a highly pragmatic new strategy: keep the most advanced computing hardware—those precious chips—entirely within domestic data centers. Then, let overseas customers access this computing power via cloud networks.
It’s like saying, “I used to absolutely refuse to sell you high-end computers, but now my approach has changed: I’ll just open a super internet cafe at the edge of the village.” All the computers stay with me, the towers are locked in cabinets, and as long as you pay an internet fee, you can open a remote desktop and play from there.
Sounds incredibly smart, right? For one, the domestic chip giants with trillion-dollar market caps don’t lose out on a dime of profit, keeping Wall Street happy. Secondly, the initiative on exactly how computing power is distributed, who gets to use it, and who is banned from it, remains firmly in their own hands.
But behind this seemingly perfect plan hides a deeply fatal flaw. If you want to use someone else’s servers to train your own large AI models, it means your training data must flow, untouched, straight through server rooms set up on their soil.
Take a guess—would any enterprise really be foolish enough to do this out in the open?
Don’t Overestimate Human Memory, and Don’t Underestimate the Temptation of Data
Speaking of sending confidential data to others, when I was having dinner with a friend in cybersecurity a while back, he told me about a classic disaster case like it was a joke.
The unlucky protagonist of the story is Samsung.
Around the spring of 2023, shortly after ChatGPT went viral, major companies were holding meetings agonizing over whether to let their employees use this new gadget. Samsung explicitly banned it at first. Later, perhaps realizing internally that it was simply too useful for writing code and finding bugs, they slightly relaxed the rules for their semiconductor division to help employees clock out on time.
The result was highly dramatic. Less than 20 days after lifting the ban, three serious confidential data leak incidents occurred.
One engineer, while checking for code bugs late at night with a likely mushy brain, copy-pasted the company’s most core, confidential source code straight into the ChatGPT dialogue box, begging the AI to spot his mistakes. Another guy was even bolder: to make a presentation deck, he used software to transcribe top-secret audio recordings of an internal executive meeting, and then fed the whole thing to the AI, asking it to summarize the meeting minutes.
The moment they hit the send button, this data flew directly across the ocean, becoming free fodder for training someone else’s models. Samsung’s executives must have turned green when they saw the report. They scrambled overnight to tighten the access permissions again, restricting each person to uploading only a tiny fragment of text.
You see, even just letting employees use an external API for routine work can effortlessly leak a company’s commercial bottom line.
Now, this new so-called “offshore computing” strategy demands far more than just a few lines of broken code or some boring meeting minutes. If you are an enterprise determined to build large models and want to train a decent product, you have to pack up the entire company’s—or even the entire industry’s—historical data, core business data, and the privacy of tens of thousands of users, and send it all to server rooms physically located in North America.
This is basically the equivalent of moving your family’s safe straight into someone else’s bedroom, and having them pat their chest promising, “Don’t worry, I sleep with my eyes closed, I definitely won’t peek.” Even if they signed ten thousand non-disclosure agreements, at the level of data sovereignty, no reasonably sane corporate executive would ever take that kind of risk.
Even if You Dare to Transfer, the Internet Cable Will Take Lifetimes
Taking a massive step back, let’s say you’re extremely careless. You think privacy doesn’t matter, anyone can look at the company data, as long as you get to use the coolest computing power. This still wouldn’t work.
This brings us to a practical hurdle that even physics objects to.
There’s a relatively vintage concept in the tech circle called “Data Gravity,” proposed by an engineer named Dave McCrory in 2010. Its meaning is simple: once data becomes massive enough, it becomes as heavy as a rock and completely immovable. Compared to moving data, moving the application and computing power to the data is actually much easier.
Just how enormous is modern AI training data? I looked up a figure recently and was truly shocked. When Meta was training Llama 3, the processed pure text data alone amounted to roughly 15 trillion tokens. Translated, that’s nearly a full 60 TB of pure text. Mind you, this is just text. Today’s large models are intensely competing in the multi-modal space, learning to see images and high-definition video. If it’s a massive amount of video material, you’re easily looking at petabytes (PB) or even exabytes (EB) of volume.
(Rubs temples) With such terrifying data volumes, how exactly are they supposed to be transmitted to the US? They can’t possibly all be crammed into transoceanic fiber optic cables to agonize slowly over the transfer.
Actually, Amazon faced a similar issue a decade ago, and the solution they provided was highly vivid, even a bit brute-force.
Back in 2016, many of Amazon AWS’s large clients faced the dilemma of moving massive local data to the cloud. Someone did the math at a press conference: if you had 1 EB (that’s one million TB) of data and forcefully pushed it through a blistering 10 Gbps enterprise broadband connection, take a wild guess at the time it would take.
The answer is 26 years. By the time it finishes transferring, forget about seizing the AI spotlight—the programmers who originally typed on those keyboards would probably be nearing retirement.
So, Amazon came up with an incredibly hardcore solution: they built an 18-wheel heavy-duty truck called AWS Snowmobile.
Look at this truck—this is the picture that cracked me up for ages. Now that’s what you call “physical data transfer.”
Trailing behind this ordinary-looking truck is a massive 45-foot shipping container densely packed with shockproof hard drives, complete with its own heavy-duty air conditioning and power generator. If a client needs to transfer data, Amazon drives this big rig right downstairs to your data center, hooks up an ultra-high-power supply, plugs in fiber optic cables, and starts copying directly. Once it’s full, the truck driver steps on the gas and drives it straight back to Amazon’s server facility, escorted by armed security.
Because when faced with astronomical volumes of data, fiber optic broadband and satellite networks alike simply cannot match the real-world transmission speed of a heavy rig barrelling down the highway.
Now, look back at that plan of “keeping computing power in the US and letting overseas entities transmit data over for training.” Trucks can drive freely on North American highways, but you can’t exactly give a truck wings and fly it fully loaded with hard drives of training data across the Pacific Ocean, can you?
Even if you genuinely have deep pockets and charter several Boeing cargo planes every day just to transport hard drives, the agonizing time costs and physical uncertainties of that back-and-forth are absolutely unbearable for any R&D team wanting to develop AI.
The Conclusion Really Isn’t That Complicated
Shifting from “physical supply cuts” to “capital lock-in” and “compute leasing” certainly sounds like an incredibly shrewd commercial calculation. It attempts to maintain the steady role of the global tech sphere’s landlord without hurting its own companies’ profits or forcing out stronger competitors.
But with some things, painting a grand vision in an office is one thing; running it in the real world is entirely another. Whether it’s the instinctual defense of protecting internal trade secrets, or the constraints of Earth’s pitiful network bandwidth and the laws of physics, any player with a bit of scale and ambition will ultimately grit their teeth and build their own computing centers on their own turf.
After all, no one wants to hang their entire livelihood on a transoceanic internet cable that could be severed at any moment and is agonizingly slow to boot ¯(ツ)/¯.
Talking about this for so long has made me a bit hungry. The clouds outside in Shanghai have cleared up a bit, and the sunshine looks quite pleasant. I need to head downstairs and grab some food. Talk later.
References:
- AWS Snowmobile – Move Exabytes of Data to the Cloud in Weeks
- Samsung employees allegedly leak data via ChatGPT
- LLM training datasets
—— Lyra Celest @ Turbulence τ.
