From Meta's cloudification to NVIDIA's "computational power sharing", optimists see the cycle of "AI computational power as capital".

Just as news of Meta's parent company Facebook preparing to sell idle AI computing power has triggered concerns of AI computing power oversupply, causing a massive drop in global tech stocks associated with AI computing power infrastructure, on the same day, the "AI Chip Superpower" - NVIDIA Corporation (NVDA.US), the world's highest market value company, announced on Thursday that the company is entering revenue-sharing agreements with AI startups in the rapidly expanding market. This move will allow these AI startup clients to exchange part of their future profits for access to the vast NVIDIA Corporation AI GPU computing power cluster. In the eyes of Wall Street analysts, NVIDIA Corporation's move is the strongest refutation at the industry's highest level of the pessimistic argument of "AI computing power infrastructure demand surplus," highlighting that as AI applications and AI intelligent agents become popular worldwide, the demand for AI computing power is virtually endless. Wall Street analysts unanimously believe that Meta's sale of computing power, along with SoftBank's establishment of SB Neo to enter the US AI cloud, and NVIDIA Corporation's offer to AI startups to exchange computing power for revenue-sharing opportunities are essentially not betting on "excess computing power," but rather on the long-term expansion of AI training and inference demand to the gigawatt-level infrastructure era. A recent research report by well-known semiconductor research institute SemiAnalysis shows that the institute's rebuttal of the "Meta selling computing power = computing power surplus = capital expenditure reduction" argument focuses on shifting the market narrative from "idle capacity" back to "excess monetization capabilities." In addition, Goldman Sachs Group, Inc.'s latest investment outlook for the second half of the year shows that the institute predicts that "tech giants will continue to fall out of favor, while the semiconductor sector will continue to reign," emphasizing that the global bull market surrounding the AI computing power supply chain is far from over. Computing power is capital, GPU is oil: NVIDIA Corporation initiates a revenue-sharing model, leading the narrative of AI financing in the age of AI The artificial intelligence chip leader NVIDIA Corporation stated in a press release that its new partner program announced on Thursday would provide Token credit limits to fast-growing AI startups to support their accelerated development of cutting-edge AI applications. AI companies based on cloud computing resources, builders focused on AI large models, and other enterprises will share product and cloud business revenue with NVIDIA Corporation, which will position itself not only as the AI chip giant but also as a new AI era intermediary agency providing full-stack artificial intelligence computing power driven by NVIDIA Corporation's cutting-edge AI chips to help startups directly access it. In this press release, NVIDIA Corporation mentioned two Australian AI computing power infrastructure supply companies that are early participants in this grand plan, both of which will provide massive AI computing power infrastructure resources to the program. Sharon AI plans to deploy as many as 40,000 NVIDIA Corporation's exclusive AI GPUs, while Firmus Technologies, a leader in AI computing power infrastructure companies, stated that it is building a large AI data center in Batam, Indonesia, expected to expand to 360 megawatts and accommodate as many as 170,000 NVIDIA Corporation AI GPUs. NVIDIA Corporation's move highlights the importance of obtaining scarce computing power for AI research and development-oriented startups. Some analysts have repeatedly compared AI chip components such as AI GPUs/AI ASICs/TPUs to oil, even suggesting that they may be linked to futures contracts, as large tech giants as well as AI application leaders like Anthropic and OpenAI are facing cost fluctuations and issues in accessing AI computing resources. Meanwhile, more and more AI-related tech companies are entering revenue-sharing and equity agreements with chip companies like NVIDIA Corporation, AMD, and Broadcom Inc., to bypass the liquidity issues plaguing the industry. Media reports in January previously indicated that OpenAI had signed several significant financing deals, some of which include considering investments or stock purchases from partners such as Amazon.com, Inc. and AMD in exchange for access to larger-scale AI computing resources in the future. Earlier this month, NVIDIA Corporation announced plans to issue bonds, with sources indicating that the amount could reach at least $20 billion. The company intends to use the proceeds from this issuance for general corporate purposes, including repaying and refinancing existing debt. Pessimists see signs of oversupply, while bulls see opportunities to monetize AI resources: AI computing power infrastructure entering a new growth cycle dominated by "computing power as capital" NVIDIA Corporation's arrangements with Sharon AI and Firmus are not traditional chip sales, but rather tying GPU, data center capacity, and future cloud business revenue bases: Sharon AI plans to add 72 megawatts of data center capacity in Australia, deploy up to 40,000 Grace Blackwell GB300 GPUs, and sell cloud services driven by NVIDIA Corporation's cutting-edge AI chips through revenue-sharing and credit support models; Firmus has announced the construction of a large AI factory campus in Batam, Indonesia, with a scale of 360 megawatts and capacity for up to 170,000 GPUs. The agreement extends to 2034, with NVIDIA Corporation receiving standard product revenue and cloud revenue sharing. When AI chip and data center capacity providers are willing to support downstream customers with a "computing power quota + revenue sharing" model, it indicates that the bottleneck in AI computing power infrastructure has shifted from "do we have actual demand" to "who can access large-scale AI GPUs in a timely manner, who can accelerate financing, who can efficiently convert computing power into chargeable tokens." The increasing leverage and crowding of positions in the AI semiconductor trading theme, along with the pressure for price increases from consumer electronics leaders like Apple Inc., have led to the Philadelphia Semiconductor Index dropping by as much as 7.9% in a single day, with multiple instances of sharp fluctuations exceeding 5% in the past month. This highlights that the AI computing power industry chain associated with semiconductors has entered a high-volatility, leveraged, overcrowded position, and high expectations of cashing in, especially with Meta shifting towards selling computing power resources. That's why there have been increasing voices from institutional investors in recent days to emphasize that the "AI semiconductor trading frenzy has bottomed out" and that the "AI bubble is gradually bursting," indicating overly pessimistic bearish narratives. However, well-known Wall Street investment firm Nomura released a new research report on Wednesday refuting the "semiconductor peak theory." Nomura's key rebuttal to the theory that the semiconductor market has peaked is not simply that AI chips will rise, but rather that the demand for AI cloud infrastructure is shifting from a shortage of single-point GPUs to systemic component mismatches. According to Nomura's research framework, AI server revenue is expected to grow by 78% and 76% in 2026 and 2027, with global data center projects increasing from 240 to 280, including around 50 gigawatt-level projects and visibility of 32GW in new computing power deployments in 2027, rising to 23GW in 2028; the real bottleneck is shifting from GPU capacity and Taiwan Semiconductor Manufacturing Co., Ltd. Sponsored ADRCoWoS advanced packaging to wafer-level substrates, AI PCBs, copper-clad laminate (CCL), electronic cloths, MLCCs, glass substrates/ABF substrates, IC carriers, high-end capacitors, power management chips, and data center optical high-speed optoelectronic components. Additionally, renowned market research firm ABI Research predicts that the active data center capacity worldwide for AI workloads will expand from 11.5GW in 2026 to 43.6GW in 2031, with the AI data center capacity in the US expected to increase from 8.2GW in 2026 to 21.4GW in 2031. This structural expansion trend in power terms is completely at odds with the pessimistic argument of "systematic AI computing power oversupply." SemiAnalysis notes that Meta's data center and computing power procurement will "accelerate rather than slow down," with capital expenditures expected to be "astonishingly high" by 2027; in just the first half of this year, Meta has signed leases for over 5GW of cloud capacity, and two of the largest under construction campus projects have a combined capacity of 2.5GW, directly undermining the pessimistic narrative of "large-scale delays in US data center construction, with only 5GW under construction." More importantly, SemiAnalysis emphasizes that Meta's potential external sale of computing power is not equivalent to low-margin bare-metal leasing, but can also flow to high-value scenarios such as Super Intelligence Labs (MSL), ad recommendation systems (RecSys), Bedrock model services, SpaceX-style short-term large customer contracts, etc.; this means that computing power becomes a strategic asset that can be scheduled, resold, and internally and externally arbitrage, rather than a sunk cost for a single purpose. In SemiAnalysis's view, Meta's move to sell AI computing power resources, along with NVIDIA Corporation's new revenue-sharing model of "computing power in exchange for revenue-sharing," signals a highly consistent positive signal in the AI computing power industry chain: the demand for AI computing power remains strong, with true scarcity lying in GPUs/ASICs/TPUs, DRAMs/NANDs/HBMs storage components, power resources, liquid cooling equipment, network infrastructure, and even a complete set of data center delivery capabilities including data center CPUs, high-speed optical interconnect systems, and data center transformer systems and a low-cost financing channel. The latest research report from the globally renowned research firm IDC shows that the world's highest market value company - the AI chip superpower NVIDIA Corporation (NVDA.US) has become the world's largest supplier of Ethernet switches in data centers on a revenue basis for the first time. IDC's latest report is consistent with the views of Wall Street giants such as Morgan Stanley, Goldman Sachs Group, Inc., and Bank of America Corp., indicating that leaders in the AI computing power industry chain like NVIDIA Corporation are expanding the coverage of the computing power chain from "single-point control of GPUs/AI chips" to "GPU cabinet clusters + networks + DPUs + optical interconnection systems + software ecosystem + data center power chain," forming a closed-loop AI factory system-level. Morgan Stanley notes that the AI computing power arms race has entered a stage of systemic expansion, with the institute revising upward its capital expenditure expectations for the large US tech giants in 2026 from $433 billion a year earlier to $805 billion, and expecting capital expenditures to reach $1.1 trillion in 2027, up from the previously predicted $950 billion. Goldman Sachs Group, Inc.'s derivatives expert Brian Garrett's leading team of investment strategists pointed out in their latest research report that investors are reducing their risk exposure to the seven giant tech companies and the market is still rewarding tech companies that are "profiting from AI capital expenditure," such as beneficiaries in the semiconductor and other AI computing power infrastructure chains. Goldman Sachs Group, Inc.'s investment outlook for the second half of the year shows that the institute predicts "tech giants will continue to fall out of favor, while the semiconductor sector will continue to reign." The latest expectations from Morgan Stanley and Goldman Sachs Group, Inc., as well as other major Wall Street institutions, clearly highlight that the supply chain bottleneck in the AI computing power infrastructure has shifted from "large GPU/ASIC purchases" to "simultaneously addressing data center power equipment, liquid cooling, data center CPUs, DRAMs/NANDs/HBMs, data center optical communication/optical interconnection, high-performance Ethernet network infrastructure, transformers, gas turbines, and the complete chain of AI data center delivery processes." In addition, Morgan Stanley predicts that by 2028, nearly $3 trillion in AI-related infrastructure investments will flow through the global economy, with over 80% of expenditures still ahead.