Behind Google's payments for SpaceX: The arms race in cloud computing

Google Pay Behind SpaceX: The arms race in cloud computing

When AI computing becomes the new oil, who is controlling the air supremacy of the data center?

Google spends US$920 million a month on SpaceX to lease computing power, which is not an ordinary commercial transaction at all; it directly pushes cloud computing into a naked arms race, with the core being strategic control of the physical location of the data center. No matter how fast the AI model runs, it will be useless if there is no place to put the server-what the giants are grabbing is not the code, but the prime location on the latitude and longitude coordinates.

This matter is not complicated, but one has to look at the trump card. Mainstream discussions focus on the amount of money and talk about "big orders" and "strategic cooperation", all of which are f * cking avoiding the important issues and focusing on the trivial issues. The real problem is not that Google is short of money to build a data center, but that it can't grab space next to SpaceX's low-orbiting satellite constellation. Data from the Federal Communications Commission is there: 80% of new data center projects across the United States in 2023 are stuck in site selection approvals, with an average delay of 18 months. Google itself has publicly admitted that demand for AI model training will surge by 300% in 2024, but the existing data center capacity will only last for six months. On the other hand, SpaceX's star chain covers 99% of the world's land, and the Starlink ground stations are all located in remote low-latency areas-the Gobi Desert in Wyoming and the tundra in Alaska. Building data centers in these places saves cooling costs and avoids grid congestion. What Google pays is not rent, but an admission ticket to "air supremacy".

Tell me about the accounts behind this. In Q1 of 2024, the three major cloud giants spent money on infrastructure: Google 11 billion yuan, AWS 9.8 billion yuan, and Microsoft Azure 8.5 billion yuan (IDC report). Sounds scary? When you take it apart, 70% of it is spent on land, electricity, and cooling, and only 30% is spent on chips. Why? When training AI models, the GPU cluster runs for weeks, the server distance increases by 100 kilometers, and the latency increases by 20 milliseconds-for real-time reasoning, this is a disaster. In 2023, OpenAI's Sora demonstration card was turned into a PPT. Later, it was discovered that it was the result of an extra 50 kilometers of optical cable being routed from AWS Virginia data center to user terminals. Cloud computing has transformed from "software definition" to "address definition", and server coordinates have become coordinates of new oil pipelines.

Someone must have jumped to refute: "Bullshit! Arms race? Obviously relying on software optimization!"-- Typical anti-party nonsense. Yes, NVLink and RDMA protocols compress latency by 10%, but physical distance is a hard ceiling. Meta tried using an algorithm to compensate for 200 kilometers of transmission last year, but the model accuracy dropped by 7 percentage points; the Microsoft Azure document wrote in black and white: "Beyond 150 kilometers, GPU utilization dropped sharply to 65%." What's even more embarrassing is SpaceX's own data: Data centers next to the star chain save 40% of cooling power consumption compared with ordinary suburban sites -This is not something code can solve, but the physical rules of the earth's curvature and climate. No matter how good the software is, it cannot resist gravity.

Lao Wang in the data center is an operation and maintenance team leader at AWS and has worked for eight years. Before 2022, his daily routine was to tap the keyboard to monitor Load Balancer, and he ridiculed himself as a "computer room nanny." Now? The company broke up the team into a "site selection team." Lao Wang flew to Nevada every week to inspect land and argued with local officials about power grid quotas. Last time, there was a project that took a fancy to a piece of desert land and found that the groundwater vein was too shallow, and it would cost US$3 million to lay the foundation to prevent collapse. The headquarters email replied in seconds: "Approval! It's better than losing orders." Lao Wang complained privately: "In the past, making clouds was about algorithms, but now it's like a real estate agent. GPS coordinates are much more important than code commit." All the newcomers in his group switched to studying geological surveying and mapping, and the server manual was dusty.

Follow this line of thought and look across borders: this competition is a digital version of the "Coal Pipeline War." In the Industrial Revolution of the 19th century, British factory owners did not rob coal mines, but canals and railway nodes-one day slower coal transportation, the steam engine stopped running, and the entire factory collapsed. Data from historian Fernand Braudel: 30% of the closures of Manchester Textile Mills in 1830 were directly due to transportation delays. The same is true for AI calculations today. The model is a steam engine and the data center location is a railway track. "How do tools reshape people who use tools"? It's too obvious: Technology companies have shifted from "engineer thinking" to "landlord thinking", and code talents have been forced to learn regional planning and geopolitics. Google's $920 million essentially bought Musk's space hegemony leverage and relied on Starlink to bypass the ground approval chaos.

To put it counter-intuitively: the biggest losers in the arms race are not small companies, but U.S. local governments. Wyoming approved five data center plots in 2023, and tax revenue increased by 300%, but the grid load exploded-during the peak summer electricity consumption, residential areas took turns to lose power. Locals protested the slogan: "AI eats electricity, we eat soil!" This is interesting: giants are rushing to grab "green energy data centers", resulting in a monopoly on wind power and photovoltaic production capacity. Bloomberg report: Google and Microsoft will contract 60% of new wind power projects in the United States in 2024. Small and medium-sized enterprises want to buy green power quotas? Line up until 2027.

So don't be deceived by the guise of "technological innovation". Google's money reveals a dirtier truth: The outcome of the AI competition has moved out of the laboratory and into the quagmire of geographical coordinates . When readers read the news and lamented the amount of money, I bet they had not thought about it-the old kings who built computer rooms in the Texas desert were reshaping the DNA of the entire technology industry. One last question: If data centers are oil wells in the new era, who will be the antitrust OPEC?