Another hard-core evidence in the industrial chain counteracts the “AI excess computing power theory”! New cloud leader Nebius (NBIS.US) is about to lock in $1 billion until 2029

The Zhitong Finance App learned that in recent years, European technology leader Nebius Group NV (NBIS.US), which has switched to providing cloud-based AI computing power resources, has agreed to sell AI computing power infrastructure resources worth at least 1 billion US dollars to Reflection AI, a world-renowned AI model developer. Reflection AI was founded by two former senior Google DeepMind researchers. The two companies (Nebius and Reflection AI) said in a statement on Tuesday that the agreement will last until 2029 and will enable Reflection to use a high-performance AI server computing power cluster built exclusively by Nvidia around the GB300 AI chip.

Before the opening of the New York market, Nebius stock showed strong gains. At one point, it rose sharply by more than 5% before the market. Nebius will sell $1 billion worth of AI computing power resources to Reflection. In addition, there was news on Monday that Meta's investment and construction scale for its Hyperion data center park in Louisiana was expanded from the initial announcement of $10 billion to more than 50 billion US dollars, jointly countering the pessimism that some AI pessimists had previously interpreted Facebook's parent company Meta as simply interpreting the sale of idle AI computing power resources as “oversupply of AI computing power infrastructure.”

This deal is an important part of the “Neverending AI Computing Power Demand Narrative” where AI startups and large mature tech companies compete to obtain AI computing power infrastructure. Last month, Reflection and SpaceX signed a multi-billion dollar AI computing power supply agreement to obtain the right to use a computing power cluster around the same type of AI chip.

Reflection has received financial support from Nvidia and other large institutional investors in recent years. Recently, there were media reports that the company had discussed financing 2.5 billion US dollars at a valuation of 25 billion US dollars.

Nebius was spun off from Russian internet service giant Yandex in 2024 and is one of a group of cloud computing technology companies known as “new cloud computing service providers” (Neocloud). These companies rent out artificial intelligence computing capabilities. The Amsterdam-based company has reached a long-term computing power infrastructure cooperation agreement with Microsoft and Meta Platforms.

Nebius and CoreWeave, an AI cloud computing power leader headquartered in the US, are both part of the “AI-specific cloud computing vendor” Neocloud circuit, focusing on faster delivery, optimized training/inference stacks and flexible contracts to serve the huge AI workloads of large-scale model developers and enterprises. Microsoft cooperated with both companies at the same time to help ease supply bottlenecks.

Nebius is not just a “hosting room,” but an integrated AI cloud leader from hardware, network to scheduling stack. It emphasizes flexibility from a single card to a “pre-optimized cluster of thousands of AI GPUs” to reduce integration costs and delivery uncertainty on the Microsoft side. Regional production capacity that can be delivered quickly is also an important logic of cooperation between Microsoft and Nebius. In particular, Nebius has huge Neocloud infrastructure sites in both the EU and the US.

“Selling AI computing power” is not a sign of excess, but rather turns an idle window into strong cash flow

Nebius's billion dollar order went to Meta's five-gigawatt superpark. These latest developments together countered the argument that Meta's sales computing power is simply interpreted as “there is already structural excess AI infrastructure.”

Meta plans to provide external model interfaces and raw computing power, which is closer to turning phased idle, intermittent training, or advanced deployment capacity into revenue, increasing the utilization rate and return on capital of expensive assets; at the same time, its Hyperion campus in Louisiana has also expanded from the initial announcement of $10 billion to more than 50 billion US dollars. The planned computing power will increase from more than 2 gigawatts to 5 gigawatts, and plans to expand the company's overall computing capacity to 14 gigawatts in 2027. A company that actually judges that future demand is insufficient will not launch cloud services while drastically expanding long-term capacity; the more reasonable explanation is that Meta is shifting from a simple computing power user to an AI computing power infrastructure operator that combines “personal use, scheduling, and commercial output.”

Reflection signed a GB300 computing power agreement with Nebius worth more than 1 billion US dollars, and previously reached another multi-billion dollar computing power arrangement with SpaceX, indicating that cutting-edge model developers are still competing for next-generation GPU clusters through multi-year contracts, rather than waiting for the spot price to collapse due to excess. Similar signs that demand for AI computing power continues to expand include a five-year agreement worth 17.4 billion US dollars between Microsoft and Nebius, a long-term contract between Meta and CoreWeave of about $14.2 billion, and Nebius's earlier limited contract size with Meta due to its limited usable capacity; these facts suggest that there may be generational changes in AI computing power hardware, mismatches in location, or when it was put into operation, but advanced computing power that can be delivered on schedule and is still scarce.

Nomura, a well-known investment institution on Wall Street, released a research report to refute the “semiconductor peak theory,” and the latest research report released by Bank of America (BofA) this week shows that by 2027, against the backdrop of a strong trend where AI inference computing power continues to surge under the big wave of AI agents, global capital expenditure on cloud computing and artificial intelligence related infrastructure will reach 1.5 trillion US dollars, and points out that the current summer correction of AI semiconductors, including memory chip stocks, is a healthy reset trajectory, rather than any structural changes at the level of AI computing power requirements.

Brian Nowak, a senior analyst from Wall Street financial giant Morgan Stanley, led the analysis team and released the latest research report on July 12, which once again significantly raised the 2027/2028 capital expenditure forecasts for the five largest hyperscale cloud computing and vendors (Meta, Amazon, Microsoft, Google, SpaceX) in the global market, to about $1.2 trillion and $1.4 trillion, respectively. The agency's capital expenditure forecast for major US tech giants in 2026 was drastically raised from 433 billion US dollars a year ago to 805 billion US dollars.

Morgan Stanley's latest study raised Meta's 2027 and 2028 capital expenditure forecasts by 29% and 22%, respectively, to US$225 billion and US$250 billion; Amazon's corresponding forecasts were raised 15% and 29% to US$308 billion and US$318 billion. Morgan Stanley said that the capital expenditure supercycle is not over yet, but 2026 and 2027 are probably the steepest years of growth. After 2028, what determines the stock price will no longer be just “who spends the most money,” but “who can quickly turn AI computing power resources into revenue, profit, and free cash flow.”

According to a recent research report released by SemiAnalysis, a well-known semiconductor research institute, the agency's rebuttal to “Meta selling computing power = excess computing power = downturn in capital expenditure” is to pull the market narrative back from “idle capacity” to “ability to monetize multiple exports.”

SemiAnalysis said that Meta's data center and computing power procurement will “accelerate rather than slow down”, and capital expenditure will be “surprisingly high” in 2027; in the first half of this year alone, Meta has signed up for more than 5 GW of cloud leasing and hosting capacity, and the two largest parks under construction have a total capacity of 2.5 GW, which directly weakens the pessimistic narrative of “large-scale delays in US data center construction, and only 5GW under construction.” More importantly, SemiAnalysis emphasizes that Meta's potential outbound computing power is not equivalent to low-margin bare metal leasing, but can also flow to high-value scenarios such as super intelligent laboratories (MSL), advertising recommendation systems (RecSys), Bedrock-like model services, and SpaceX-style short-term rental large customer contracts; this means that computing power has become a strategic asset that can be dispatched, resold, and arbitraged internally and externally, rather than a sunk cost for a single use.

According to SemiAnalysis, Meta's initiative to sell AI computing power resources and Nvidia's new “AI computing power for revenue share” model form a highly consistent positive signal in the AI computing power industry chain: demand for AI computing power is still strong. What is really scarce is GPU/ASIC/TPU, DRAM/NAND/HBM storage components, power resources, liquid cooling equipment, network infrastructure, and even a complete set of low-cost data center park delivery capabilities and financing channels including data center CPUs, high-speed optical interconnection systems, and data center transformer systems Combinations.