Microsoft tightens AI computing power rentals! Reportedly gradually withdrawing from Oracle Cloud agreement, "new cloud" CoreWeave plunges

Microsoft is reassessing its previous strategy of heavily betting on external AI compute power providers. Following tighter scrutiny of partnerships with AI “Neocloud” service providers, it is now reported that Microsoft is gradually winding down its arrangement to lease AI cloud computing power from Oracle.

According to a Bloomberg report overnight, Microsoft has begun shutting down some of the computing resources previously leased from Oracle in recent months and has decided not to renew the relevant contracts. Sources familiar with the matter said Oracle’s servers are equipped with older generation Nvidia chips, which no longer meet Microsoft’s technical specifications, resulting in decreased value for Microsoft.

After the news broke, Oracle’s stock price briefly dropped more than 5% during the session. “Neocloud” concept stocks, which rose sharply in early trading, gave back gains, with CoreWeave experiencing a quick dip. However, Oracle and several “Neocloud” service providers ended the day higher.

Just the day before, on Thursday, reports indicated that after promising to spend over $80 billion renting compute power from other cloud companies, Microsoft had become dissatisfied with the high costs and delivery delays of “Neocloud” providers. This has led to regrets over previous strategies, increased scrutiny of some data center projects, withdrawal from those seen as high risk, and internal discussions over pausing the signing of new cooperation agreements. Two consecutive days of reports underscore Microsoft’s reassessment of external computing procurement strategies.

Oracle stock plunges, company emphasizes continued strong AI infrastructure demand

Reports indicate that the compute leasing arrangement between Microsoft and Oracle originated from a 2023 agreement to support Bing’s operations, which Microsoft has now decided not to renew. According to sources, Oracle’s servers equipped with older generation Nvidia chips are no longer able to meet Microsoft’s technical specifications as time has gone on.

Following news on Friday, Oracle (ORCL) shares, which had surged 3.8% at the open, quickly reversed, erasing gains to hit a session low of $109.21, dropping nearly 5.2% intraday. The stock then rebounded to close up about 0.8%.

Other AI compute leasing suppliers also saw significant reversals of early gains. CoreWeave (CRWV), which was up 9.3% in early U.S. trading, turned negative at one point, down over 1% before rebounding, but gains stayed below 2%. IREN (IREN), which was up about 24.6% at the open, saw its gains narrow to around 12%. Nebius (NBIS), which was up 13.7% initially, saw gains narrowed to less than 1.9% at one point.

By the close, CoreWeave was up over 6%, while IREN and Nebius rose about 7.6% and 20%, respectively.

However, market analysts held a relatively conservative view regarding the scale of this exit.

Analyst Zephyr stated on social media that the contract involved in this incident might account for between 50MW and 100MW of A100 chips, representing a relatively limited transaction, so the market reaction is somewhat overblown. He pointed out that Microsoft could easily transfer this demand internally to 25MW–40MW of Hopper architecture chips to fulfill its compute needs.

Oracle, for its part, emphasized that AI infrastructure demand remains robust, including for earlier generation GPUs. Oracle also stated that there have been cases of renewing GPU contracts deployed four years ago at a premium of about 20%, indicating that older generation chips still hold some market value. However, analysts noted that the renewal price and utilization data for older GPUs will be key indicators in determining whether this demand can continue broadly.

Delivery Delays, High Costs: Microsoft Re-examines “Neocloud” Partnerships

Microsoft’s changing attitude toward external compute providers did not begin with adjustments to the Oracle leasing agreement.

According to a Bloomberg report on Thursday, Microsoft is strengthening oversight of some data center projects involving “Neocloud” AI service providers such as Nebius, IREN, and Nscale. As these companies face project construction delays, regulatory approval obstacles, and high costs, some Microsoft management have become skeptical about previous large-scale partnership bets.

Insiders say Microsoft is exiting projects it perceives as high-risk and is internally discussing whether to continue signing new “Neocloud” partnership agreements. Microsoft is also more inclined to send staff to partner data center projects to check on construction progress and ensure projects meet technical and regulatory requirements.

Among these, IREN’s data center project in Texas experienced fiber optic cabling issues, previously suffered a generator fire, and revealed hardware delivery delays. Microsoft recently renegotiated the project timeline with IREN, which may postpone completion of the related facilities by several weeks or even months.

Microsoft CFO Amy Hood has also mentioned on previous earnings calls that delays from third-party compute deliveries have impacted the company's capital expenditure plans. Microsoft is increasingly focused on whether external suppliers can deliver compute power on time and whether projects can meet the company's actual needs.

Nscale is another typical example of Microsoft scaling back partnerships.

Bloomberg reported Thursday that Microsoft has decided not to exercise an additional expansion option for Nscale’s data center project in Texas. The two parties had previously agreed to a compute leasing arrangement of about 240 megawatts, with Microsoft holding an option to add another 700 megawatts. Due to issues with energy access approvals, Microsoft ultimately chose not to proceed with the expansion.

Earlier, Microsoft had already withdrawn from Nscale’s flagship data center project in Loughton, Essex, UK. The project was originally set to involve about $3 billion in compute leasing, but due to power access delays, the relevant power supply may not be available until the 2030s.

Losing Microsoft’s incremental cooperation opportunities in succession has put pressure on Nscale’s plan to seek a $35 billion IPO valuation. However, Microsoft’s decision not to exercise the Texas expansion option does not affect the original about 240-megawatt contract, nor does it impact Nscale’s previously announced backlog of about $103 billion in orders.

From Fighting for Compute Power to Scrutinizing Orders: The Logic behind AI Infrastructure Investment is Changing

Zephyr pointed out that this series of actions reflects a profound shift in Microsoft’s compute strategy — moving from large-scale leasing of compute power from external providers to relying more on self-built and internally owned compute assets. He stated that Microsoft’s cancellation of the Nscale agreement and the delay in IREN deliveries are consistent with this strategic change.

This shift’s impact on the entire AI compute leasing market cannot be ignored.

Over the past two years, tech giants like Microsoft, Google, and Amazon have raced to build AI data centers, but since the pace of self-developed infrastructure cannot fully keep up with demand, emerging data center operators such as Nebius, IREN, CoreWeave, and Nscale received a large number of long-term compute leasing orders, becoming key beneficiaries of AI infrastructure expansion.

Previously, to meet compute demands for Azure cloud services and AI businesses, Microsoft committed more than $80 billion in compute leasing to external cloud providers including CoreWeave, Nscale, and Nebius. However, as its own data center construction moves forward and issues such as delayed construction, equipment delivery, and financing costs for external providers become clearer, Microsoft has begun reassessing the economic feasibility and execution risks of these partnerships.

It is worth noting that Microsoft’s adjustment of its external compute procurement strategy does not mean AI compute demand has peaked. Global AI computing capacity remains in short supply, with other major clients like Anthropic and Google still accelerating their commitment to partnerships with emerging data center operators.

Currently, there is conflicting evidence in the market regarding supply and demand for older generation GPUs: on one hand, suppliers like Oracle claim strong demand; on the other, Microsoft’s withdrawal suggests the supply of certain GPU generations is already loosening. This contradiction will gradually be clarified as more renewal data emerges in the future.