Negotiations between Oracle and OpenAI have collapsed, primarily due to financing issues and OpenAI's frequently changing requirements, leading to the cancellation of plans to expand their flagship artificial intelligence data center campus in Texas. The breakdown in talks has created an opportunity for Meta to step in. Meta is now considering leasing the planned expansion site in Abilene, Texas, from developer Crusoe, with NVIDIA assisting in advancing Meta's negotiations. NVIDIA's role is particularly noteworthy; to secure Meta as an expansion tenant, NVIDIA has paid a $150 million deposit to Crusoe. This prepayment is intended to ensure that NVIDIA's products, rather than those of competitor AMD, will populate the expanded data center.

The partnership between Oracle and Crusoe has faced challenges, including data center outages lasting several days due to energy and reliability issues, prompting broader industry reflection on future gigawatt-scale data center construction. The globally watched Stargate project has encountered another setback. According to foreign media reports, OpenAI and Oracle have canceled plans to expand their Abilene flagship data center campus. Developed by Crusoe at the Lancium Clean Campus in Texas, this project was the first Stargate initiative to go live following last year's announcement of a $500 billion joint venture. The infrastructure is operated by Oracle for OpenAI's use. Two data center buildings commenced operations last September, with six new buildings scheduled for this year, bringing total capacity to approximately 1.2GW. Subsequently, the companies planned to expand the data center to 2GW. However, with the collapse of expansion talks, these plans have now halted abruptly. This marks a new phase of uncertainty and strategic reassessment in hyperscale computing deployment, suggesting a potential paradigm shift in AI infrastructure evolution over the next five years—from a singular focus on "scale centralization" towards "geographic diversification" and "energy sovereignty."

The expansion project has been fraught with complications. As a grand AI infrastructure blueprint initiated by OpenAI with participation from SoftBank and Oracle, boasting a total investment of $500 billion, Stargate has attracted significant attention since its announcement. Oracle agreed in July of last year to develop 4.5 gigawatts of data center capacity for OpenAI. While that agreement remains on track, challenges have persisted as individual projects advance. The Crusoe-developed Abilene data center, a core node of Stargate, has been one of the most closely watched facilities. The 1,000-acre site is still under construction, with portions already operational. Discussions among Oracle, Crusoe, and OpenAI about expanding the facility from 1.2 GW to approximately 2.0 GW began in mid-2025. However, as negotiations progressed and various factors converged, the decision was ultimately made not to proceed with the large-scale expansion plan.

The曲折of the Abilene data center highlights the complexity of building gigawatt-scale AI data center projects. Firstly, the immense computational power required for training and deploying AI models has driven an unprecedented surge in project scale, but also necessitates capital investments amounting to tens of billions of dollars. To support OpenAI's computing needs, Oracle's capital expenditure is projected to reach $35 billion in fiscal year 2026. Capital markets have expressed concern over this aggressive expansion premised on extreme financial cost. Moody's and S&P have subsequently revised Oracle's credit rating outlook to negative, further complicating financing for the Abilene expansion.

Secondly, power supply for data centers is becoming a critical issue. The total volume of large power load interconnection requests in Texas, where Abilene is located, surged to 230 GW in 2025, nearly four times the 2024 figure. Despite Texas' abundant energy resources, upgrades to grid infrastructure have not kept pace with the trend of AI data center construction. For instance, following the delivery of the 1.2GW capacity at the Abilene data center, power suppliers faced significant challenges in providing subsequent electricity. A planned natural gas peaker plant, intended to meet demand, encountered equipment shortages and environmental permit delays, preventing it from delivering sufficient power within the required timeframe.

Thirdly, frequent changes in OpenAI's requirements have also stalled the expansion project. OpenAI continuously challenges the design of AI compute clusters. According to OpenAI infrastructure executive Sachin Katti, a single-site data center scale of 1-2 GW is currently the "sweet spot" for network interconnection and training stability. Exceeding 2 GW for a single cluster introduces non-linear complexity in heat management, power outage recovery, and cross-rack synchronization. Furthermore, the rapid iteration speed of NVIDIA's AI chips is a major consideration in OpenAI's cluster design. While the originally planned Blackwell clusters for the Abilene project were not yet fully operational, NVIDIA had already announced and begun promoting its next-generation Rubin architecture chips. Consequently, OpenAI showed a preference for deploying Rubin architecture chips at entirely new sites rather than within the Abilene expansion project. OpenAI infrastructure executive Sachin Katti stated on social media, "The Stargate flagship project is one of the largest AI data center campuses in the US. We considered further expansion but ultimately chose to allocate additional capacity to other locations."

NVIDIA's Role as a "Mediator" Following the change in plans, NVIDIA moved swiftly into action. According to informed sources, NVIDIA paid a $150 million deposit to Crusoe and began efforts to secure Meta as an expansion tenant. This substantial prepayment was explicitly designated as a fee to lock in future capacity priority. A core stipulation requires Crusoe to prioritize the deployment of NVIDIA's next-generation GPUs (such as the Blackwell series and its successors) in the subsequent 600-800MW expansion, ensuring these products populate the expanded data center instead of competitor AMD's products. Sources indicate that negotiations between Meta and Crusoe regarding the Abilene site expansion are ongoing and subject to change. Meta is currently building several large data centers in Louisiana and Indiana. Last month, Meta reached an agreement with AMD to deploy AMD equipment in its data centers.

NVIDIA's evolving role in the Abilene project is indeed milestone-setting. When gigawatt-scale data center projects encounter challenges like chip supply, financing, and power infrastructure, associated capital market valuations and project expenditures fluctuate. Relying solely on cloud service providers like Oracle or AWS to complete projects of this scale becomes difficult. NVIDIA is no longer content with merely selling chips but is acting as a "mediator," actively facilitating connections between Meta and Crusoe. This proactive referral not only mitigates the developer's tenant vacancy risk but also ensures that this expensive GPU capacity will be quickly absorbed by a financially capable buyer. This signifies that as the Stargate project advances, NVIDIA is deepening its involvement in the global spatial allocation and lease locking of AI computing power, effectively becoming a "compute real estate broker."

Future Trends in AI Infrastructure Evolution The scaling back of the Abilene project does not signal the end of the AI boom but rather marks a transition from the "great leap forward" of GW-scale data centers to a more robust and diversified "maturity phase." Based on this event, core trends in global AI infrastructure construction over the next three years can be anticipated.

Trend 1: Geographic Shift from "Super Hubs" to "Distributed Hedging" The industry-wide recognition of risks associated with single-site complexities exceeding 1 GW is growing. Future compute deployment will increasingly adopt a "Hub-and-Spoke" model, deploying medium-scale modules of 300 MW to 600 MW across multiple nodes rich in power and friendly regulations, interconnected by high-bandwidth, low-latency fiber backbones to form virtual "large clusters." This distributed approach not only effectively spreads grid risk but also leverages regional peak/off-peak electricity price differences to reduce long-term operational costs.

Trend 2: Data Center Energy Competition Igniting a "Nuclear Power Boom" Oracle's reliance on natural gas peaker plants in Texas, which encountered setbacks in Abilene, is conversely boosting market demand for direct "nuclear power + data center" models. AWS's $20 billion agreement with Talen Energy for 1.9 GW of nuclear power, and Microsoft's plan to restart the Three Mile Island nuclear plant, indicate that top AI firms are seeking to bypass public grids by securing exclusive access to baseload power sources like nuclear energy to ensure computational supply continuity.

Trend 3: Financial Restructuring and the "Asset-Light" Trend As individual project costs approach tens or even hundreds of billions of dollars, traditional cloud providers' balance sheets cannot bear the burden alone. Future projects will see increased use of Special Purpose Vehicle (SPV) models involving sovereign wealth funds (like Abu Dhabi's MGX), infrastructure funds (like Blackstone), and chip suppliers (like NVIDIA). OpenAI's proposed "asset-light" model—where OpenAI focuses on models and operations while partners handle the heavy asset construction—is set to become a standard paradigm for collaboration between AI startups and infrastructure giants.

Trend 4: Shift from "Pre-training Priority" to "Globalized Inference Nodes" The Abilene site was originally designed for large-scale pre-training. However, with the proliferation of AI applications, inference traffic is becoming the primary data center load. Inference workloads do not require extremely high inter-cluster interconnect bandwidth but demand proximity to end-users. Consequently, the next two years will witness a significant flow of compute power from remote large centers like Abilene, Texas, towards medium and small-scale edge data centers located near urban areas with good network connectivity.

Trend 5: Tightening Regulatory Environment for Large-Scale Loads The introduction of Texas's SB6 bill is a precursor to this trend. As data center electricity consumption occupies an increasing share of the grid (projected to reach 12% of US electricity by 2030), governments will mandate that large loads possess flexible dispatch capabilities—specifically, the ability to rapidly reduce load during peak demand periods to safeguard residential power supply. This "load-shedding requirement" will compel AI companies to develop more sophisticated computational orchestration systems to dynamically pause or migrate non-critical training tasks during grid stress.

In summary, the halt of the Oracle-OpenAI Abilene expansion plan represents a significant course correction in AI infrastructure investment and construction, warranting continued industry attention. It reveals three fundamental realities: First, physical world constraints (like supporting power infrastructure, geography, and climate) evolve much slower than AI algorithm iteration. Second, financial markets' willingness to fund the AI narrative has shifted from "blind optimism" to "cash flow verification." Finally, computational sovereignty no longer depends solely on who possesses the most GPUs, but increasingly on who controls the most stable power resources and the most flexible geographic distribution.