Interpreting Meta's Sale of Computing Power: Does this Indicate an Excess of Computing Power?

TL;DR

· Bloomberg reports that Meta is planning its cloud business, looking to sell excess AI compute power and model access.

· The market is trading on the path to externalize AI capital expenditure, rather than realized cloud revenue.

· Related Tickers: META, NVDA, CoreWeave, Nebius, GOOGL, AMZN, MSFT.

According to Bloomberg's July 1 report, Meta is planning its cloud infrastructure business, aiming to sell excess AI compute power to external customers and considering offering access to models hosted on Meta's infrastructure. Reuters stated it could not independently verify the news, and Meta did not immediately comment.

The market reacted swiftly. In pre-market trading, Meta briefly surged, while shares of emerging GPU cloud companies like CoreWeave and Nebius came under pressure. Investors are not suddenly believing that Meta will soon become the next AWS but are instead reevaluating an old question: whether the money invested in AI data centers and servers over the past few years, apart from internal use, can become rentable and recyclable assets.

The core concept is CapEx (Capital Expenditure). Companies first spend a lot of money to buy servers, chips, and build data centers, which are then depreciated over many years. Meta's cloud plan offers a new interpretation: AI compute power is not just an internal production cost but could also become an ability to sell externally.

Stock Prices Are Trading the ROI Narrative

This news initially impacted stock prices because it struck at the core valuation pressure Meta has faced over the past year: too much AI investment with insufficient quantifiable returns.

Following Meta's Q1 2026 earnings report, the full-year capital expenditure guidance was raised to $125-145 billion. Publicly disclosed expenditures mainly involve AI infrastructure, servers, data centers, and networking, with rising costs also from component prices and data center construction.

This magnitude is no longer a regular tech upgrade; it's a balance sheet-level wager. The market is willing to value the AI story but is unwilling to accept a "spend now, talk about it later" approach in the long term. As capital expenditures continue to rise, investors worry about compressed profit margins and question whether these servers and data centers can be fully utilized.

In a May shareholder meeting, Zuckerberg mentioned that the cloud computing business is "definitely within the consideration scope," and external companies often inquire about API or compute power purchases. However, he also noted that Meta has not done so yet as the company believes the compute power still has internal purposes.

Bloomberg's report has moved this matter from a management contingency plan to a more specific business expectation. It did not prove that revenue has arrived, but it allowed the market to see that Meta's massive AI investment may have a second recycling path.

Costs Reinterpreted as Assets

The business significance of selling AI computing power is not that Meta has an additional ordinary cloud service line, but that it has changed the market's interpretation of capital expenditure.

If a company buys a large number of GPUs that can only be used for internal model training, recommendation systems, and AI assistants, then the return on investment must be indirectly demonstrated through advertising efficiency, user engagement, or new product revenue. This chain is long, making it difficult for investors to confirm in short-term financial reports.

If some computing power can be rented out externally, the logic becomes more straightforward. Data centers that were originally seen as sunk costs begin to have external cash flow attributes. Even if the revenue scale is initially small, it can alleviate concerns about "this money is just being burned."

This is also the market language behind the divergence in stock prices of META and new GPU cloud providers. For Meta, the cloud plan has added a second purpose to AI infrastructure. For companies like CoreWeave and Nebius, the potential supply may increase, and the additional supply comes from giants with lower funding costs and stronger chip procurement capabilities.

This change will also impact Nvidia's narrative. In the short term, Meta is still heavily investing in AI infrastructure, and the demand for high-end GPUs has not disappeared. However, if more giants commercialize surplus computing power in phases, the computing power market price will become more important.

Selling Computing Power Does Not Directly Indicate Surplus

This clue needs to be given boundaries. Meta selling computing power does not automatically mean that it has a severe surplus of computing power, nor does it mean that Llama's self-developed model route has failed.

But when looking at this news in the broader AI industry chain, it does point to another change: computing power is concentrating on top models. Projects with insufficient performance and unclear commercialization in models are increasingly finding it difficult to absorb all their computing power alone, while top model companies are still seeking more training and inference resources.

One example is xAI. Grok, as a Musk-owned model product, has not received market feedback establishing a clear advantage comparable to top models. Subsequently, Musk has leased some computing power to Anthropic. This action itself illustrates the problem better than a slogan: when proprietary models cannot fully absorb the investment, computing power can flow to top model companies in higher demand.

Meta has now exhibited a similar signal. Llama is still being updated, and Meta officially launched Llama 4 Scout and Maverick in April 2025, adopting a native multimodal and MoE (Module-on-Demand) architecture. At least from the roadmap at the time, Meta is still investing in self-developed models rather than exiting the model race. However, from the results perspective, the market still questions the return on investment of Meta's large-scale model roadmap, and the cloud business plan implies that some of its AI infrastructure may be repackaged as a service.

Another clue comes from Google's relationship with Meta. In June of this year, there were reports that Google had capped Meta's use of its Gemini AI model as Meta's computational demands exceeded what Google could provide. Reports indicated that Google informed Meta around March that it could not meet all of its requested Gemini computational power, stating that this gap disrupted and delayed some of Meta's internal AI projects.

The implications of this case are more complex.

As Meta builds its own large-scale AI infrastructure, it still requires external top-tier models and computational support. However, when Google's own computational supply is also limited, Meta's internal projects are affected. It is not simply "Meta's excess computational power" but illustrates that the AI computational supply-demand is becoming more hierarchical: the strongest models and most defined use cases are prioritized to receive computational power, while other projects may be reallocated between self-use, outsourcing, and renting.

Meta has also been reported in the past to have signed a multi-billion-dollar multi-year agreement with Google Cloud. This narrative resembles a mismatch between construction cycles and demand growth: building long-term capacity on one hand and meeting immediate needs or peak demands through external clouds on the other.

A more prudent understanding is that Meta is designing its AI infrastructure as a set of schedulable assets. When internal demand is high, it is primarily used internally, and during surplus phases, it is sold externally. If model access and hosted inference services mature, the Llama ecosystem may also be packaged into cloud products.

Big Tech Squeeze Shrinks GPU Cloud Premium

For emerging GPU cloud companies, the pressure comes from the sensitivity of their business model to pricing.

These companies typically rely on financing to purchase GPUs, then recover costs through long-term contracts or spot rentals. As long as AI companies are willing to pay a premium for scarce computational power, these companies can expand. Once the tech giants start competing by releasing marginal computational power, customer choices, rental fees, contract terms, and gross margins may all come under pressure.

More importantly, the additional supply may not necessarily come from traditional cloud providers, but also from companies with in-house super-scale computing power that may not perform as expected or be absorbed internally at the desired rate. Instances such as the external rental of xAI computing power after Grok and Meta planning to sell excess AI computing power indicate a shift in the supply side of the GPU cloud market: computing power is no longer just a commodity of professional cloud providers but may also become an adjustment item on the balance sheets of large model companies.

This will weaken the scarcity premium of some GPU cloud companies. In the past, the market was willing to assign a high valuation to GPU clouds because of the scarcity of computing power, customer queues, and price support. However, if large tech companies and large model companies start releasing temporary excess computing power, the market will reevaluate rental levels, contract terms, and long-term utilization rates.

Meta's advantage lies not only in scale but also in cost of capital, chip procurement capabilities, data center operational experience, and the Llama model ecosystem. If its cloud service includes not only card rental but also model access, fine-tuning, and inference hosting, then its competition with pure GPU clouds is not on the same level.

However, this does not mean that Meta will soon challenge AWS, Azure, or Google Cloud. Cloud business requires a sales system, enterprise customer support, security compliance, service stability, and a global network. Meta was not a traditional enterprise cloud player in the past and is more likely to enter through AI computing power and model services.

The most direct short-term impact of this news is a change in relative valuation. Meta gains a narrative bonus for "asset monetization," while new GPU cloud companies are forced to face a discount for "giant supply entering the market." Furthermore, it also reflects the redistribution of resources within the AI industry: computing power is still scarce, but scarcity is increasingly concentrated on a small number of leading models and high certainty applications.

Pricing Power Determines the Bottom Line

Whether Meta's cloud plan can truly change valuation ultimately depends on two variables: how much computing power can be stably sold externally and at what price.

If the so-called excess is merely a temporary window in the construction pace, with internal products quickly consuming the additional capacity, then the cloud business is more of an investor communication tool than a substantive revenue curve. It can enhance the narrative, but it is difficult to redefine the profit model.

If Meta can disclose clearer external computing power revenue, customer types, utilization rates, and gross margins in its financial reports, the market will begin to consider some AI CapEx as a priceable asset rather than just a cost prepayment for future advertising efficiency.

The key question is not whether Meta wants to do cloud, but whether this business can prove that: AI investment at the hundred billion dollar level is not just a ticket to an internal competition, but can also find a valuation in the external market. Until the answer is revealed, Meta's rise appears more like a reassessment of option value, while the decline of the new GPU cloud seems to be anticipating that the pricing power may be diluted.

Welcome to join the official BlockBeats community:

Telegram Subscription Group: https://t.me/theblockbeats

Telegram Discussion Group: https://t.me/BlockBeats_App

Official Twitter Account: https://twitter.com/BlockBeatsAsia