It's not a surplus of computing power! Zuckerberg selling computing power reveals the biggest truth about AI

On July 1, Bloomberg broke the major news that under Mark Zuckerberg’s leadership, Meta has launched a new internal project codenamed Meta Compute, planning to sell its idle AI computing power and Llama series large model API access to the public, officially entering the cloud computing sector.

As early as this May during Meta’s annual shareholders’ meeting, Zuckerberg publicly stated: once surplus data center computing power appears, selling it externally is a backup commercialization option. Moreover, almost every week external companies are willing to pay a premium to procure Meta’s computing resources.

According to Meta’s official financial statements, the company will increase its AI infrastructure capital expenditure for 2026 to $125–$145 billion, nearly doubling last year’s amount, making heavy investments to stockpile GPUs and build massive AI compute clusters. However, industry public data shows that Silicon Valley tech giants’ average GPU utilization rate is only 10%–30%. After large model training is over, daily inference doesn’t even use up the computing power, causing substantial high-end hardware to remain idle and depreciate at a loss over time.

This time Meta has clarified two monetization paths: first, leasing idle computing power by time, benchmarking against leading AI cloud providers; second, opening up commercial Llama large model APIs, charging based on usage flow. On the day the news landed, Meta’s market cap skyrocketed by $100 billion in a single day, while small and medium computing power rental companies collectively plummeted, instantly reshaping the market’s competitive landscape.

Many people mistakenly think this is an AI compute surplus, but in fact, the core logic is the complete opposite—here are three official and crucial truths you must understand:

First: Not selling core cutting-edge computing power, only revitalizing existing assets

The computing power Meta is renting out is all older, surplus machines from previous generations. The latest high-end compute power used for next-generation top-level model training remains for internal use only, and AI R&D progress has not slowed in the slightest. What Zuckerberg is really doing is transforming money-burning fixed assets into cash-flow-generating assets, optimizing financial reports and easing the burden of capital expenditures.

Second: This has become a standard practice in Silicon Valley, as AI enters a commercialization and payback cycle

Not only Meta, but Musk’s xAI had already opened its computing power to external rentals and secured large long-term orders. All tech giants now face the same challenge: huge AI investments with long payback cycles, and limited in-house business consumption. Revitalizing idle compute to boost profits has become the standard operation for Silicon Valley giants, signaling that AI is moving from pure expansion into a steady monetization phase.

Third: Track competition is upgrading, and AI is entering an integrated era of compute + models + cloud

In the past, Meta only did models and not cloud services, but now by adding computing power rental, it is directly benchmarking against Amazon, Microsoft, and Google Cloud. In the future, the AI industry won’t just compete on model parameters, but on who owns self-built compute infrastructure, self-developed model systems, and commercial delivery capability. Full-chain players will hold absolute dominance.

In conclusion: Zuckerberg’s move to commercialize computing power is not a sign of AI cooling down, but a landmark signal that the AI industry is maturing. There is no compute power surplus in the market—only a complete shift in the money-making logic. Ordinary investors must distinguish between revitalizing old computing power and saturation of high-end computing power; the high-growth AI cycle is still ongoing.