Recent reports that Meta is exploring leasing or selling excess AI compute capacity have sparked concerns about a potential oversupply, rattling the AI infrastructure investment theme and contributing to a global tech stock selloff. Against this backdrop, Softbank Group Corp, led by legendary investor Masayoshi Son, and its telecom unit plan to begin leasing their extensively deployed AI compute infrastructure to U.S. companies starting next fiscal year. This move aims to monetize the company's expanding portfolio of AI data center projects for stronger profit growth.
However, this strategy has intensified market fears that an "AI compute glut" may be becoming a grim reality. The dual developments of Meta founder Mark Zuckerberg's exploration of selling excess capacity and Softbank's planned entry into the U.S. AI cloud infrastructure services market are, in the short term, severely impacting AI semiconductor stocks, compute supply chain leaders, and new cloud (or "Neocloud") providers closely tied to AI infrastructure build-out. However, Wall Street analysts suggest a more accurate interpretation is not that "massive compute buyers like Meta and Softbank are conceding defeat" or that "global compute is in oversupply," but rather a transformation of massive AI capital expenditures from a pure cost center into a monetizable asset on the balance sheet.
Softbank's establishment of SB Neo, its plan to lease AI chips and cloud services in the U.S., and its long-term target to expand AI data center capacity to approximately 10 gigawatts (GW) largely indicate that tech giants are still betting on a long-term explosion in demand for training and inference compute. The U.S. Department of Energy has also disclosed that Softbank's SB Energy will build 10GW of new power generation in Ohio to support a 10GW data center project. This essentially represents an integrated expansion of "compute infrastructure + power resources," not a collapse in AI-related training/inference compute demand.
Regarding Meta's latest move, Wall Street analysts widely expect that its plan to sell excess AI compute is an effort to convert substantial AI capital expenditures into external revenue elasticity. This reflects both tech giants' desire to improve AI investment returns and will intensify competition for new cloud providers like CoreWeave and Nebius. As for the sharp single-day declines in U.S. and Korean stock markets, analysts view this more as a selloff driven by concerns over AI infrastructure investment returns, overbuilding, heightened competition, and the unwinding of record leverage and crowded positions—with the latter factor being particularly dominant in the Korean market—rather than proof that global compute supply has already entered an oversupply state.
From Telecom Operator to AI Compute Supplier: Softbank's U.S. Neocloud Plan Launches
The two companies (Softbank Group Corp and its telecom subsidiary) stated in a declaration on Thursday that they will establish SB Neo Inc. this month to prepare for plans to offer comprehensive AI cloud computing services—including access to AI chips and high-performance network infrastructure—to large companies, including hyperscale cloud service providers and compute leaders like OpenAI. Junichi Miyakawa, head of Softbank's telecom division, stated that this major joint venture focused on the AI "new cloud" space plans to gradually expand its cloud computing resources, ultimately targeting a massive AI data center capacity of around 10GW by approximately 2030 to meet the enormous demand for large-scale AI model training and the immense cloud inference compute needs driven by AI agents.
According to informed sources, offering new cloud operational services in the U.S. could easily drive the annual operating profit of mobile operator Softbank Corp., part of the Softbank group, to increase two to threefold, reaching a scale of 3 to 4 trillion yen ($18.5 to $25 billion). Miyakawa noted in an interview that this substantial new cloud business, with mobile operator Softbank Corp. holding a 51% stake and its parent Softbank Group Corp holding 49%, has the potential to generate profits of a "different order of magnitude." He stated in the interview, "We view the full-scale launch of this cloud computing business in the U.S. market as a second startup for our company." This Japanese mobile operator, the third largest in the country, once played a crucial foundational role for its parent company, helping Masayoshi Son secure massive funding for his early venture capital career.
In recent years, the billionaire's operational focus has fully centered on AI compute infrastructure, as part of his AI ambition to capitalize on the seemingly endless global demand for AI compute resources and pave the way for the widespread adoption of cutting-edge AI technology. Softbank Group Corp is already the majority shareholder (holding nearly 90%) of Arm Holdings Plc, whose architecture is widely used across the consumer electronics industry and increasingly serves as the foundation for server CPU product lines in hyperscale data centers. Softbank's substantial investments in recent years in Arm, Graphcore, and Ampere Computing, combined with its latest massive investments in OpenAI and the U.S. "Stargate" AI infrastructure project, demonstrate its full-stack layout from the most fundamental AI hardware architecture to AI compute infrastructure clusters and up to the AI application layer.
The AI platform Softbank is building is no longer a single-point bet but an increasingly complete "AI stack": at the bottom layer are the Arm architecture and Arm's self-developed data center CPUs; in the middle layer are UK AI chip design company Graphcore Ltd. (fully acquired by Softbank) and recently acquired Ampere; at the upper layer are OpenAI, Stargate, and enterprise-grade cloud AI compute platform collaborations with Oracle. Son himself has publicly stated that Softbank's goal is to become the largest provider of AI compute and application-layer foundational platforms in the era of "super artificial intelligence" (ASI) over the next decade, and has adopted an almost "all-in" aggressive investment stance towards ChatGPT developer and global AI model leader OpenAI.
The wave of "Neoclouds"—a new category of cloud infrastructure providers renting out massive compute capacity tailored for AI training/inference—has fully emerged globally to meet the持续飙升 demand for computing capacity. New cloud companies like CoreWeave Inc. and Nebius Group NV are rapidly leasing complete AI compute infrastructure resources, including specialized AI chips needed by clients to train and run AI models. Softbank's new cloud operations may have a ready-made customer in OpenAI; its parent Softbank Group Corp has pledged total investments in OpenAI that could reach approximately $65 billion by October. However, this field is becoming increasingly crowded. Beyond specialized Neocloud service providers, traditional cloud computing leaders like Amazon Web Services (AWS), Microsoft Azure, and Google Cloud are also actively selling access to AI compute resources. More significantly, according to media reports, Facebook parent Meta Platforms Inc. is also formulating similar plans to enter this arena. Miyakawa stated that a major advantage for Softbank's upcoming U.S. cloud computing business in the American market is its unique ability to secure robust power supply sources, primarily from the extensive natural gas power plant system within the Softbank investment group ecosystem.
Masayoshi Son's Softbank Group Corp envisions a massive, data-center-focused AI infrastructure construction project in Ohio with a scale reaching $500 billion and a capacity of 10GW, which would rank among the world's largest AI data center projects. Domestically in Japan, Softbank's telecom unit is building data center campuses in Hokkaido in northern Japan and Sakai City, Osaka.
Transforming Compute Resource Supply: From GPU Scarcity to Strategic Monetization
Although "compute glut" concerns have triggered a collective selloff in U.S. and global tech stocks, Wall Street analysts believe that Meta selling compute and Softbank establishing SB Neo to enter the U.S. AI cloud market are not essentially bets on "compute oversupply," but rather bets on the long-term expansion of AI training and inference demand into the gigawatt-scale infrastructure era. Softbank plans to launch its U.S. Neocloud services in the new fiscal year ending March 2028, offering compute for large model training and inference to U.S. large enterprises and hyperscale cloud providers, with capacity gradually expanding based on the 10GW-scale energy and AI infrastructure Softbank is developing. Coupled with the U.S. Department of Energy's earlier disclosure that Softbank and SB Energy plan to build 10GW of new power generation capacity—with at least 9.2GW from natural gas—to support 10GW of data center development, this indicates that the real competitive focus has shifted from "who has GPU resources" to "who can package power, land, data centers, AI chips, network infrastructure, and customer contracts into a financeable and actively monetizable super compute platform."
Meta exploring the sale of excess AI compute impacts chip stocks and new cloud providers in the short term, but a more accurate understanding is not that "Meta is conceding" or that "global compute is in oversupply," but rather the transformation of massive AI capital expenditures from a pure cost center into a monetizable asset on the balance sheet. Meta is studying selling excess AI compute and may offer model access services similar to AWS Bedrock; the news boosted Meta's stock price but put pressure on new cloud companies like CoreWeave and Nebius, as Meta transitions from a potential customer to a potential competitor. The core of the debate among institutions like Bank of America, UBS, and Morgan Stanley lies here: for Meta shareholders, selling compute is an "EPS bridge" and cash flow buffer that clarifies AI return on invested capital (ROIC); for new cloud companies, this is a stress test on their business model where pricing power and renewal capabilities face pressure from超大科技 clients.
In the view of Wall Street analysts, the sharp single-day declines in U.S. and Korean stock markets resemble a deleveraging event triggered by crowded positions, leveraged ETFs, momentum trading, and the "compute glut" narrative, rather than a real collapse in AI demand. South Korea's KOSPI Composite benchmark previously plunged nearly 10% after regulators warned of risks in chip stock leveraged ETFs, with Samsung Electronics and SK Hynix both falling over 12%, triggering temporary trading halts. On the same day, SK Hynix still announced plans to invest 100 trillion won (approximately $64.38 billion) by 2029 to build new large-scale DRAM/NAND/HBM memory chip manufacturing plants. Therefore, in analysts' view, the price action suggests the market is trading on concerns about whether "the AI hardware chain has risen too much in this phase, leverage is too high, and whether capital expenditures can deliver returns," not proving that HBM, AI server clusters, and massive cloud AI inference demand have suddenly vanished causing a compute glut.
For investors focused on the AI super-bull market, the true investment revelation from Meta and Softbank establishing channels to sell/lease AI compute infrastructure is this: the AI compute infrastructure investment theme is transitioning from a phase of "absolute shortage, indiscriminate buying of hardware" into a second phase where "shortages still exist, but utilization rates, rental prices, customer contracts, power supply costs, and free cash flow must be validated." CoreWeave recently raised its 2026 capital expenditure lower bound to $31 billion due to strong AI cloud demand and disclosed a revenue backlog as high as $99.4 billion, indicating that high-quality AI compute infrastructure is not unwanted. However, when Meta, Softbank, SpaceX, traditional cloud providers, and new cloud providers all rush into compute leasing simultaneously, future winners will shift from those who "simply possess AI compute resources like GPUs" to those "super AI compute platforms" with comprehensive advantages in low-cost power, stable customers, advanced scheduling software, model/API access points, and cost of capital. In other words, this is not the end of the AI super-bull market, but rather the AI compute infrastructure investment market entering a new phase of price discovery, ROIC scrutiny, and the clearing of financial market leverage and crowded positions.
Comments