Erliao Fei Si | From Ao Fei Temple
QuantumBit | Public Account QbitAI
No way! Silicon Valley is completely crazy! in its race for AI computing power.
Even A Company CEO Dario Amodei pleaded humbly in front of Meta’s gate:
Bro, give me some power...
What do you think happened? Xiaozha refused, and Musk took advantage of it:
SpaceX acts as a second-hand dealer of NVIDIA GPUs, reselling computing power for Anthropic, making huge profits—
The first Q3 financial report in August shows promising results. The AI business revenue alone reached 2.561 billion US dollars, a 247% increase year-on-year and twice as much compared to the previous quarter.
Wow! For computing power, Silicon Valley is going all out. Dalian Hengshidai is coming at full force.
No matter how fierce the criticism was online before, in reality, everyone has to bend down for GW.
It won’t work if it doesn’t bend, look at Google next door that has fallen behind.
The top scientist Noam Shazeer, who was brought back with a huge $2.7 billion payout, becauseDifference in computing powerIn less than two years, Thomas moved to OpenAI for his spin.
Other people spend money just to hear a response, while you do Google purely for charity. (doge)
Anyway, Silicon Valley is now getting into its stride:
Whether talent can be retained, whether the model can remain competitive, and whether AI can generate real profits... These issues depend increasingly on how much computing power is left.
GPU is a hard currency in Silicon Valley
Previously, there was a shortage of both money and people; now what is taking the lead is computing power economy.
Even tech giants like Google cannot guarantee that star researchers will get enough GPU resources.
When thinking of Noam Shazeer, the first thing that comes to mind is that he is one of the Transformer eight. He left Google in 2021 to start Character.AI, focusing on developing generative AI chatbots.
In 2024, Google spent 2.7 billion US dollars to facilitate Shazeer’s return through a transaction, and he serves as the co-head of Gemini.
Such researchers should be those that giants compete to secure resources for.
But according to latest reports from foreign media, his team also faced difficulties in having their computing power reassigned. Shazeer was forced to switch to the opposing side, OpenAI.
By coincidence, Hasabis also expressed dissatisfaction with the insufficient computing power of the project for a long time before, which also influenced his decision to step down as DeepMind CEO.
Currently, there are significant disagreements within Google regarding resource allocation. Sergey Brin sometimes directly intervenes to reallocate resources to projects he believes are important.
But ultimately, it’s because of insufficient computing power resources; we can only save costs and reduce spending.
Competition outside the company makes industry relationships more complex.
Top startups like OpenAI and Anthropic not only compete in model capabilities, but also for data centers, servers, and power supply capabilities.
Take Anthropic as an example; the computational power requirements of its models are so high that they need to discuss supply with another model company.
The cooperation agreement with SpaceX in May this year stipulates that a new capacity of over 300 terawatts will be achieved that month, involving 220,000 NVIDIA GPUs, and the usage limits for Claude Code and API will also be increased.
In other words, the purpose of doing this is to allow the backend to receive more computing power for the frontend. With more computing power and a stronger model, users will stay.
Thus, the competition among model companies began to feature two sets of logic operating simultaneously:
At the product level, people still compete for users and development talent; at the infrastructure level, whoever can provide resources promptly may become a potential partner.
So what about the small AI startups that survive in the gaps?
Their situation is more dramatic.
These companies have difficulty building their own data centers, but in the past, they could rely on flexible rental of GPUs from cloud platforms such as AWS.Pay-as-you-go。
However, as many leading AI labs lock down computing resources, some cloud service providers start requiring multi-year contracts, with advance payments amounting to 30% of the contract value, and even ask startups to find a payment guarantor.
All of these factors have put start-ups in a difficult situation.
Buying too much computing power means that if future user growth falls short of expectations, the continuous cash expenditure will be wasted; buying too little may result in missing growth opportunities due to the lack of computing resources to meet demands.
This situation has instead given rise to a new business: Global computing power mining.
That is, finding underutilized AI server resources across the world, and renting out this computing power to enterprises in need.
For example, the representative San Francisco Compute company once went to old chicken coops in rural America to collect servers.Musk also and **TSMC **discussed further computing power cooperation.
Top VC firms in Silicon Valley are also starting to purchase GPUs directly. YC launched a dedicated GPU cluster for its startups in July, and Radical Ventures will help the invested companies coordinate supplier contracts.
This slightly solves the immediate problem for small teams that are training or deploying models.
AI calculation competition becomes out of control
But draining is better than blocking. The shortage of computing power occurs, and the reasons behind it are quite complex.
OpenAI, Meta, and Google are all actively building new data centers, promising to provide massive GPU resources in the future.
But the demand for rapid model growth is too urgent; this time gap, or the huge promises made, is one of the reasons why the company still has to continue looking for computing power elsewhere.
The way users use AI also adds to the supply pressure in an invisible manner.
Starting this year, Agent begins to replace the question-and-answer approach as the core paradigm used by users. Continuous long-term work tasks force servers to handle more parallel computing.
After the demand increased, the rental market for chips also underwent drastic changes.
Data released by SemiAnalysis in April this year show that the H100 one-year lease contract price rose from $1.70 per GPU per hour at the low point in October 2025 to $2.35 in March 2026, a increase of nearly 40%.
The claim that GPUs are comparable to gold isn't baseless.
There is a lack of real, immediate, and stable computing power, and new computing power is still far in the future. All these factors together create the difficulties in the computing power market.
Thus, computing power hegemony becomes concrete.
The narrative of start-ups challenging industry giants will gradually wither away; even if someone has a path, it will fail if there is not enough computing resources.
It is visible to the naked eye that computing power is becoming a barrier to entry in the AI era, and it reinforces the advantages of large companies with their own infrastructure.
I can say that I have computing power, and strength is in my hands.
For those working on AI products, there are more things to consider. One has to worry about the product not being used by anyone, and another about the user growth rate exceeding the company’s computing power expansion.
But difficulties are also opportunities; The intermediate layer selling shovels is thriving.
Is that so, Old Huang?
Reference links:
[1]
https://www.wsj.com/tech/ai/ai-computing-power-demand-a63da9b9?st=tzDj41
[2]
https://www.anthropic.com/news/higher-limits-spacex
[3]
https://newsletter.semianalysis.com/p/the-great-gpu-shortage-rental-capacity
— End —