What did the report that smashed AI computing stocks actually say?

Overnight, a report revealing significant changes in the computing power market caused a collapse in U.S. AI concept stocks.

On March 26, TD Securities' TD Cowen released a report stating that Microsoft has abandoned new data center projects in the U.S. and Europe that collectively consume 2GW of electricity, attributing this to an oversupply of computer clusters.

As a result, the decline in U.S. chip stocks widened, with the Philadelphia Semiconductor Index closing down about 3.3%, NVIDIA falling over 5.7%, leading the decline among the seven major tech companies, and AI concept stocks broadly declining, with AMD dropping nearly 8.9%.

TD Cowen's channel survey found that the cancellation and postponement of Microsoft’s data center leases is more widespread than expected, not only in the U.S. but also extending to the European market. However, Google and Meta quickly filled the gap, with total demand for data centers still growing year-on-year. Meanwhile, supercomputing architecture is transitioning to higher density, leading to a short-term slowdown in equipment orders, which poses challenges for equipment suppliers.

Microsoft Cancels Leases, Is There an Oversupply of Data Centers?

TD Cowen's research shows that Microsoft has recently canceled and postponed data center leases on a large scale in the U.S. and Europe. In the past six months, Microsoft has abandoned over 2GW of data center capacity while canceling and postponing several existing leases in the U.S. and Europe.

Analysts believe that Microsoft's contraction in new capacity leasing is primarily due to the decision to no longer support incremental OpenAI training workloads. Additionally, the cancellation and postponement of leases indicate that Microsoft's existing data center supply is relatively oversupplied compared to current demand forecasts. This move aims to reserve core market capacity for its own cloud and inference workloads while canceling leases that exceed its updated mid-term capacity needs.

However, this adjustment creates opportunities for other hyperscale data center operators. The report found that Google is filling the capacity abandoned by Microsoft in international markets, while Meta is taking over related capacity in the U.S. Both companies are currently experiencing significant year-on-year growth in data center demand.

OpenAI Has the Potential to Start Building Data Centers Independently in the Medium to Long Term

The report points out that OpenAI is increasingly seeking to obtain data center capacity directly from third parties (including GPU-as-a-service providers and third-party data center operators), with a recent deal announced with CoreWeave as an example.

Furthermore, OpenAI has demonstrated significant long-term ambitions for data center capacity. Surveys indicate that OpenAI plans to create multiple "Stargate" projects, each representing 800MW to 1.5GW of capacity, with a potential total cumulative long-term capacity demand exceeding 6GW. To meet its capacity needs, OpenAI is recruiting personnel with design, construction, and capacity planning experience from other hyperscale operators, indicating that OpenAI may begin to build its own data centers in the medium to long term. **

Orders for Hyperscale Data Center Equipment Slow Down

The report emphasizes that orders for data center equipment are slowing down, a trend that began in January of this year, stemming from hyperscale operators redesigning their data centers to support higher rack densities. For example, Microsoft has finalized liquid-to-liquid cooling solutions (as opposed to air-assisted liquid cooling solutions used for data center retrofitting), and other hyperscale operators are undergoing similar redesigns (Google seems to be making the fastest progress in this regard).

The derivative impact of these redesigns is that hyperscale operators and the third-party data center operators supporting them are essentially unable to place orders for equipment until the new designs are finalized, as the specific types of equipment required will vary based on the final design. Supply chain surveys indicate that procurement decisions for data center equipment have been delayed, with leased hyperscale capacity orders postponed by one to two quarters.

This short-term slowdown in equipment orders, combined with significant equipment warehousing by third-party data center operators to accelerate future market deployment, may negatively impact Vertiv's equipment order volume in the first half of 2025.

The Path to Inference Begins with Data Center Retrofits

The report reviews that Microsoft notified selected data center operators in October 2024 of its intention to retrofit existing data centers to support liquid cooling. Recent surveys have found that Microsoft has an internal plan to deploy air-assisted liquid cooling through auxiliary equipment to support higher rack densities in traditional cloud data centers.

Interestingly, the Azure Infrastructure blog highlighted in October (around the same time the retrofit notification was provided to selected operators) the need for "independent liquid-air heat exchangers to support traditional data centers, which typically lack the infrastructure to support direct-to-chip (DTC) liquid cooling." Surveys indicate that this form of air-assisted liquid cooling is a more cost-effective way to retrofit traditional data centers, as it allows for "localized cooling" of high rack density loads in certain areas of the data center without retrofitting the entire facility to support liquid cooling.

Importantly, a key motivation for these retrofits is the rapid entry into the inference market. By collecting power through liquid cooling in major markets, hyperscale operators have begun to release power capacity available for inference, which is much faster than acquiring capacity from third-party data center operators. At least two hyperscale operators are pursuing this recent approach to support inference.

The following is the full report, translated by AI:

Our channel checks at NVIDIA GTC and DCD Connect indicate that while Microsoft's lease cancellations/postponements are more widespread than initially thought, overall demand for data centers still shows year-on-year growth. The redesign of hyperscale data centers for higher density has led to a slowdown in data center equipment procurement, and we hold a negative outlook on Vertiv's orders for the first half of 2025.

Year-on-year growth in hyperscale data center demand; Microsoft's lease cancellations create opportunities for Google and the Metaverse

A clear conclusion from last week's NVIDIA GTC held in San Jose and this week's DCD Connect in New York City is that the overall demand for hyperscale data centers has increased year-on-year, which aligns with the factors we emphasized in late February. Specifically regarding Microsoft, since we released our initial report on the cancellation of Microsoft leases, further channel checks indicate that the list of third-party data center operators affected by Microsoft's lease cancellations has expanded, with leases being terminated in both the United States and Europe. In addition to lease cancellations, our channel checks also point to lease extensions by Microsoft. Combining this with our findings at PTC, Microsoft has (1) abandoned over 2 GW of capacity in the ongoing leasing process in the United States and Europe, and (2) delayed and canceled existing data center leases in the United States and Europe over the past month. We believe that Microsoft's withdrawal from new capacity leases is primarily due to the decision not to support additional Open AI training workloads. However, we still believe that the lease cancellations and delays indicate an oversupply of data centers relative to its current demand forecast. Therefore, we think the lease extensions are intended to provide Microsoft with a mid-term capacity runway in key markets to support cloud computing/inference workloads, as Microsoft has canceled capacity leases beyond its updated mid-term capacity needs.

Favorable for third-party data center operators, our checks show that Google is filling the capacity abandoned by Microsoft in international markets, while our checks indicate that the metaverse is filling capacity in the U.S. market, as both hyperscale data center operators are experiencing significant year-on-year growth in data center demand. We increasingly believe that Google's demand growth is due to a global capacity shortage, as its internal demand increased after withdrawing from the market by the end of August 2024 (which we emphasized in September 2024), stemming from an internal plan aimed at improving the utilization of its existing data center fleet. As for the metaverse, the demand growth is due to its significant increase in data center capacity to support Llama. Additionally, our checks show that Open AI is increasingly seeking to procure data center capacity directly from third parties, including GPU-as-a-service providers and third-party data center operators, with its recent deal announced with CoreWeave being an example. Furthermore, we observe that Open AI has significant long-term ambitions for data center capacity, as our checks indicate its intention to create multiple Stargate projects, each representing 800 MW to 1.5 GW of capacity, with potential long-term capacity demand totaling over 6 GW. As it seeks to meet its capacity needs, our checks show that Open AI is recruiting personnel with design and construction experience as well as capacity planning experience from other hyperscale data center operators, which we believe indicates that Open AI has the potential to begin building data centers independently in the medium to long term Our checks indicate that data center equipment orders have slowed due to hyperscale data center operators redesigning their future data centers to support higher rack densities, which began in January, as NVIDIA continues to iterate its roadmap, pushing for ever-increasing rack densities. For example, our checks show that Microsoft has finalized a liquid-to-liquid cooling solution (separate from the air-assisted liquid cooling solution used in the data center renovations highlighted below), and other hyperscale data center operators are undergoing similar redesigns (Google appears to be the most advanced in this regard). However, the derivative impact of these redesigns is that hyperscale data center operators supporting hyperscale data center operators and third-party data center operators are largely unable to place equipment orders until the new designs are finalized, as the specific types of equipment required will vary based on the final design. In this regard, our supply chain checks indicate that data center equipment procurement decisions have been delayed, and orders for hyperscale data center leasing capacity have been postponed by one to two quarters. Additionally, when considering the considerable degree of equipment warehousing undertaken by third-party data center operators to expedite future go-to-market speed, we believe the result is an NT air pocket in equipment orders, which we believe may impact Vertiv's order volume in the first half of 2025.

The path to inference has begun; we believe this path starts with hyperscale data center renovations.

Recall that we emphasized in January that Microsoft notified selected data center operators in October 2024 of its intention to retrofit existing data centers to support liquid cooling. In our subsequent checks connected to the subway, we noted that Amazon also has similar renovation plans, which we pointed out relate to harvesting power in major markets and the desire for workload interchangeability within its fleet. Through our recent checks, we learned that Microsoft has an internal plan to deploy air-assisted liquid cooling via sidecars to support higher rack density in legacy cloud data centers. Interestingly, a blog post on Azure infrastructure released around October 2024 (approximately the same time as the notification to selected operators regarding renovations) emphasized the need to use "independent liquid-to-air heat exchangers to support legacy data centers that typically lack infrastructure support for direct chip (DTC) liquid cooling." Our checks indicate that this form of air-assisted liquid cooling is a more cost-effective way to retrofit legacy data centers, as it allows for "point cooling" of high rack density loads in certain parts of the data center without the need to retrofit the entire data center to support liquid cooling. Importantly, our checks show that a key motivation for these renovations is the speed of inference go-to-market. By harvesting power through liquid cooling in major markets, hyperscale data center operators have begun to unlock power capacity for inference faster than obtaining additional capacity from third-party data center operators, with at least two hyperscale data center operators supporting inference in NT. Therefore, we are increasingly convinced that the initial indicators of inference demand will be the speed of hyperscale renovations rather than the speed of leasing third-party incremental sub-deployments within existing available zones in major markets (related to parent-child architecture)