Zha Nan Concedes: Major AI Firms Begin to "Divert Talent"

Mark Zuckerberg has just admitted that the progress of Meta's AI agents is slower than expected.

According to Reuters, he said at an all - hands meeting within Meta that in the past four months, the development of AI agent technology "has not accelerated as we expected," and the organizational restructuring around AI at the beginning of the year by the company "has not delivered" the expected results. On the other hand, Meta's capital expenditure this year could reach up to $145 billion, with a large part directed towards AI infrastructure and data center expansion.

The money has been spent, but the product returns are still on the way.

Almost at the same time, Bloomberg reported that Meta is preparing a cloud business and plans to rent out its excess AI computing power. This puts Meta in a delicate position: it was once one of the most aggressive hoarders of computing power in Silicon Valley, and this hoarding was for its own radical AI R & D plans. Now, it is considering monetizing its own AI infrastructure.

Elon Musk has taken the lead on this path. SpaceX rented the computing power of Colossus to Anthropic and also signed a multi - year cloud service agreement with Google. Now, Zuckerberg seems to be following in Musk's footsteps.

A new turn is emerging in the AI computing power competition. In the past, the competition was about who could grab more GPUs. Now, the question is who can fully utilize, rent out, and turn these GPUs into revenue after getting them.

Is Zuckerberg following in Musk's footsteps?

The news that Meta is going to sell computing power came out on July 1st local time.

Bloomberg reported that Meta is preparing a cloud business and plans to sell its excess AI computing power.

According to the report, this plan is still in development, and the strategy may still change. Meta also declined to comment. But the general direction is clear: Meta may allow developers to access AI models hosted on its infrastructure, such as Muse Spark, and charge based on the computing power required to run these models. At the same time, it is also considering directly renting out raw AI computing power, similar to neoclouds like CoreWeave and Nebius.

This is not a small move. For a long time in the past, Meta was not a cloud service provider. Its main business was social platforms and advertising, and its AI infrastructure was mainly for internal models, recommendation systems, advertising systems, and product features.

The outside world immediately began to speculate - does this mean that Meta is ready to shift its focus away from self - developed/full - stack AI? Is Meta looking for a more realistic way to recoup its huge AI investment?

This speculation is not unfounded. Meta has invested heavily in AI in recent years, but the narrative of its cutting - edge models has not been very smooth. It released Muse Spark in April, which was the first model launched by Meta's newly formed high - cost AI team. However, its main feature is lightness, and the more core "flagship large - model" has not appeared yet.

The problem is that the money has already been spent. In its first - quarter earnings report, Meta raised its capital expenditure forecast for 2026 to between $125 billion and $145 billion, higher than the previous range of $115 billion to $135 billion. The reasons include rising component prices and increased data center costs for future computing power capacity.

When AI infrastructure spending swells to the level of hundreds of billions of dollars, investors will naturally be concerned about when these GPUs, data centers, and electricity can turn into revenue.

So, the market's reaction to the news that "Meta is going to rent out computing power" was straightforward. After the report came out, Meta's stock price rose by more than 10%, which relieved the pressure of its previous underperformance against the S&P 500 index. At the same time, CoreWeave and Nebius fell by 10.8% and 12.4% respectively.

The most direct benefit of Meta selling computing power is to "recover cash."

Elon Musk has taken the lead on this path.

In May this year, SpaceX rented the computing power of its Memphis data center, Colossus, to Anthropic. Anthropic said at the time that it would use the full computing power of the SpaceX Colossus 1 facility. This facility has more than 220,000 NVIDIA processors and will bring an additional 300 megawatts of capacity to Anthropic within a month.

Subsequently, SpaceX also signed a multi - year cloud service agreement with Google. According to the document, Google will pay SpaceX $920 million per month from October 2026 to June 2029 in exchange for the computing power provided by about 110,000 NVIDIA GPUs and related components such as CPUs and memory.

Reuters calculated that the two computing power access agreements between SpaceX and Anthropic and Google amount to about $26 billion per year; if neither contract is terminated early, the total scale will exceed $70 billion.

Elon Musk was also one of the most aggressive players in the AI computing power arms race. Colossus was originally for the narrative of xAI and Grok and was an important infrastructure for Musk to catch up with OpenAI and Anthropic. But before his own AI business could fully digest this computing power, SpaceX rented the computing power to Anthropic and Google, which were more in need of computing power and could immediately digest it.

More importantly, Musk emphasized the short - term nature and flexibility of this cooperation.

Regarding the arrangement of Anthropic renting Colossus, he later clarified that SpaceX first signed a 180 - day lease, and then both parties could cancel with 90 - day notice. He also said that the short - term arrangement was proposed by SpaceX. If the computing power becomes very tight in the future, SpaceX may need to take back the computing power.

This statement shows the real intention of renting out computing power: it does not mean giving up self - developed AI, nor does it mean admitting that one's own AI business has no future. The company chooses to maintain flexibility between internal demand and external monetization. The computing power can be rented to more eager customers first to get cash flow; when its own models, products, and user needs catch up, the computing power can be taken back.

The same principle applies to Meta. It doesn't need to announce its transformation into a cloud provider, nor does it need to admit that its self - developed AI route has been frustrated. By taking out a part of the excess or temporarily undigested computing power, it can tell two stories at the same time: internally, AI infrastructure remains the future strategy; externally, these infrastructures can start generating revenue.

Thus, a subtle turn has emerged in the AI computing power competition.

In the past few years, Silicon Valley giants have been competing to see who can grab GPUs faster and build larger data centers. Now, the question has become who can fully utilize these GPUs after getting them. And this "full utilization" can be for one's own use or for others' use.

The Two Extremes of Computing Power Redundancy and Shortage

However, the fact that Meta and SpaceX are starting to rent out computing power does not mean that the entire AI industry has entered a state of computing power surplus.

On the contrary, almost at the same time, computing power shortage is still the most frequently mentioned keyword in AI news.

When Anthropic announced the completion of a $65 billion Series H financing in May this year, it said that the company's post - investment valuation reached $965 billion, and its annualized revenue had exceeded $47 billion at the beginning of May. Anthropic said that this financing would be used to expand computing power to meet the growing demand for Claude.

Reuters also mentioned in the report that Anthropic had difficulty meeting the demand in the past few months and had to set usage limits during peak hours and regulate the computing power pressure by encouraging users to use during off - peak hours.

The subsequent computing power cooperation announced by Anthropic further illustrates this problem. In addition to SpaceX, Anthropic also listed several other computing power arrangements: an agreement with Amazon for up to 5 gigawatts, including nearly 1 gigawatt of new capacity by the end of 2026; a 5 - gigawatt agreement with Google and Broadcom; a $30 billion Azure capacity cooperation with Microsoft and NVIDIA; and a $50 billion investment in AI infrastructure in the United States with Fluidstack.

For Anthropic, more computing power means a higher usage limit for Claude Code, a larger API call capacity, and more enterprise customers can be onboarded.

The situation is similar for OpenAI. On April 29th, when introducing the progress of Stargate, OpenAI said that to meet the growing demand for AI from consumers, enterprises, developers, and the government, the company is continuing to expand its computing power map and put new capacity into use faster. OpenAI also mentioned that when it announced "Stargate" in January 2025, it promised to secure 10 gigawatts of AI infrastructure in the United States by 2029. More than a year later, this goal has been exceeded ahead of schedule, with more than 3 gigawatts added in the past 90 days.

So, it cannot be simply said that the AI computing power bubble has burst. A more accurate statement is that computing power is starting to be misaligned.

On one side are companies like Anthropic and OpenAI. They already have clear product entrances, developer ecosystems, and enterprise customers. User demand is directly translated into token calls, subscription revenues, and API bills.

For them, the more computing power they have, the more demand they can handle, and the higher their revenue ceiling will be.

On the other side are players like Meta and xAI/SpaceX. They are also the most aggressive hoarders in the AI computing power arms race, but their AI products and business closed - loops have not been verified to the same extent.

Meta has Meta AI, AI glasses, AI features in Instagram and WhatsApp, and new models like Muse Spark. xAI has Grok and the Musk ecosystem. However, compared with the growth achieved by Anthropic through Claude Code, APIs, and enterprise customers, the AI investments of Meta and xAI have not formed a similarly clear revenue closed - loop.

Currently, the people who need computing power the most are not necessarily the ones who hoarded computing power the earliest and most aggressively.

Computing power shortage and computing power redundancy are happening simultaneously. Companies that can truly turn computing power into revenue still lack computing power, while those that built up computing power first but whose internal product demand has not been fully released are starting to treat computing power as an asset that can be rented, traded, and used to recover cash flow.

The AI computing power competition is entering the "monetization stage" from the "hoarding stage."

The New Pattern of Computing Power Circulation

After the computing power starts to circulate, the positions of players in the AI infrastructure market are also being rearranged.

It is still the traditional cloud providers that were the first and most natural to benefit from this wave of demand. Microsoft, Google, and Amazon AWS are originally cloud service providers, and renting out computing resources is their main business. In the AI era, the core resources of cloud services are shifting from CPUs, storage, and databases to GPUs, TPUs, model calls, and AI development platforms.

Google is a typical example. At the first - quarter earnings conference call of Alphabet, Sundar Pichai said that Google Cloud's revenue in the first quarter increased by 63% year - on - year, exceeding $20 billion for the first time; the unfulfilled orders, that is, the contract amount of signed but unrecognized revenues, almost doubled quarter - on - quarter, exceeding $460 billion. The growth of Google Cloud comes from the demand for enterprise AI products and AI infrastructure.

Google also said that the number of tokens processed by its first - party models through direct customer API calls increased from 10 billion per minute in the previous quarter to more than 16 billion per minute. At the same time, Google continued to package its self - developed TPUs, Axion CPUs, and NVIDIA GPUs into cloud services and launched the eighth - generation TPU at Cloud Next.

Microsoft didn't suddenly start selling AI computing power. Azure is already one of the largest cloud platforms in the world, but AI has changed its growth and cost structure.

The third - quarter earnings conference call of Microsoft's 2026 fiscal year showed that Microsoft Cloud's revenue reached $54.5 billion, a 29% year - on - year increase; the revenue of Azure and other cloud services increased by 40%. But Microsoft also emphasized that customer demand still exceeded the available capacity. That is to say, for Microsoft, the computing power is still not enough.

However, AI computing power is also changing Microsoft's business model. Microsoft said that the company's gross profit margin decreased year - on - year, including reasons such as continuous investment in AI infrastructure and the growth of AI product usage. GitHub Copilot is an example. Microsoft said that nearly 140,000 organizations are using GitHub Copilot, and the number of enterprise subscription users has almost tripled year - on - year; but it also announced that starting from June 1st, it will shift GitHub Copilot to a pricing model that is more in line with actual usage and costs.

This shows that the problems faced by traditional cloud providers are different from those of Meta and SpaceX. Microsoft and Google already have cloud platforms, enterprise customers, sales systems, and billing systems, and AI computing power can be naturally integrated into their existing businesses. What they need to solve is how to continue to turn the expensive investments in GPUs, TPUs, and data centers into high - quality revenues.

When the computing power of big players like Meta and SpaceX changes from an internal strategic resource to a market supply, the business of the original cloud providers - especially small and medium - sized cloud providers - will inevitably be affected.

Directly affected are neoclouds like CoreWeave and Nebius. Neoclouds are essentially a group of new - type cloud service providers that emerged around the AI computing power gap. They obtain GPUs from NVIDIA, build data centers specifically for AI training and inference, and then rent the computing power to customers like OpenAI, Microsoft, Meta, xAI, and Anthropic.

In the past two years, AI companies were short of computing power, and traditional cloud providers were also queuing up for expansion, so neoclouds had a large living space.

But as soon as the news that Meta is considering renting out computing power came out, the market immediately realized that this space might be squeezed again. MarketWatch reported that Meta currently has a $35.2 billion infrastructure contract with CoreWeave; if Meta builds its own "Meta Compute" department, which provides hosted models to developers and directly rents out raw computing power, it will directly compete with existing neoclouds.

This is the reason for the decline in the stock prices of CoreWeave and Nebius. For neoclouds, they originally played the role of middlemen: obtaining resources from chip companies and the capital market and then selling computing power to AI companies. Now they may be squeezed from both sides, on one side by established cloud providers like Google, Microsoft, and AWS, and on the other side by non - cloud giants that originally only used computing power for themselves but now also start to rent it out.

NVIDIA stands at a more upstream position. It not only sells chips but also tries to make these chips enter a smoother circulation system.

In May last year, NVIDIA launched the Lepton platform, trying to establish an AI computing power market. Reuters reported at the time that Lepton allows cloud computing companies to sell GPU capacity on a platform, and companies such as CoreWeave, Nebius, Crusoe, Firmus, Foxconn, GMI Cloud, Lambda, Nscale, SoftBank, and Yotta have all joined this platform. Alexis