Microsoft cuts orders by 1.98GW for a four-year period, equivalent to an actual annual cut of 500MW. The distributed data centers include US Atlanta, US Wisconsin, US San Antonio 2 facility, US Kansas City, US Cedar, US Illinois, and EU 390MW.
At that time, the expert mentioned that Deepseek is rapidly increasing GPU utilization, mainly because customers are experiencing Double Purchase situations, meaning they retain the use of GPT/Llama while trying Deepseek, requiring two months to observe the final impact.
<ol>
<li>MSFT has abandoned a total of 1.98GW leasing demand:</li>
</ol>
· Among them, 390MW is in Europe, and the rest is in the United States.
· This demand cut is related to OpenAI, but also to the weakening of the Jevons Paradox of Deepseek.
· After the launch of Deepseek, customers using both Deepseek and GPT/Llama have begun to transition from simultaneous trial/testing to the final model.
· Some downstream customers shifting from GPT to Deepseek have led to a slowdown in the originally expected demand growth, compounded by the fact that some customers using mid-to-low-end Llama models do not have such high demand for high-end computing GPUs.
<ol>
<li>MSFT may therefore reduce CAPEX:</li>
</ol>
· Experts estimate that MSFT FY26 CAPEX may decline by USD 500 million to 1 billion.
· Experts estimate that MSFT FY27 CAPEX may remain flat or slightly increase to USD 82-84 billion, mainly depending on Rubin's demand situation.
· Experts estimate that MSFT FY28 CAPEX may see a year-on-year decline, as MSFT hopes to adopt more self-developed ASICs by then.
· Experts believe that the slowdown in MSFT's CAPEX is mainly due to the lack of need for self-developed model investments and changes in the relationship with OAI.
3. META and GOOG may not experience similar situations:
· META and GOOG currently still have cutting-edge model investments, and the slowdown in MSFT's CAPEX cannot be used as a basis for the subsequent CAPEX trends of META and GOOG.
<ol>
<li>Considering MSFT's demand cuts and the impact of Deepseek on OpenAI, MSFT's leasing revenue for CY25/26 may decline:</li>
</ol>
· CY25 is expected to decline from USD 4.4 billion to USD 4.1 billion.
· CY26 is expected to decline from USD 9 billion to USD 8 billion.
· The demand for OpenAI API Chatbot types is slowing down due to the emergence of Deepseek, but financial/government/healthcare enterprises are still accelerating their investments
·The impact of GB200's launch on the rental prices of H cards that have already been leased is very limited, mainly because most of the current prices are long-term contracts; however, the rental prices for newly signed H cards will continue to decrease, leading to a marginal decline in profit margins, and utilization may also decline by about 15%.
·Currently, Azure has launched 20,000 GB200 cards, and 60,000 to 70,000 cards will be gradually launched in May, consistent with the previous progress of the Phoenix City IDC.
<ol>
<li>The reduction in Coreweave's CAPEX may also be related to Deepseek:</li>
</ol>
·Coreweave's CAPEX is also affected by its IPO financing capability, and it has been adjusted downwards for cash flow considerations.
·The situation with downstream customers is similar to that of MSFT, as demand for GB200 has marginally decreased due to the impact of Deepseek and supply chain issues with GB200 itself. Therefore, Coreweave plans to reduce demand for GB200 and shift to GB300 in the future, temporarily increasing some H200 cards.
·It is uncertain whether CAPEX will be adjusted upwards or downwards later, pending the situation with GB300.
<ol>
<li>OpenAI's CAPEX remains aggressive:</li>
</ol>
·Oracle's FY26 CAPEX is expected to be 25 billion, of which 16-17 billion comes from OpenAI's demand.
·OpenAI will subsequently build 1-2 large IDCs for its own ASIC research and development needs.
The market has risks, and investments should be made cautiously. This article does not constitute personal investment advice and does not take into account the specific investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their specific circumstances. Investment based on this is at one's own risk

MSFT

Microsoft announced a reduction of 1.98GW in leasing demand, phased over 4 years, approximately 500MW per year, involving multiple data centers in the United States and Europe. This move is related to the usage of OpenAI and Deepseek, as customers shift from GPT to Deepseek, leading to a slowdown in demand expectations. Experts predict that Microsoft's FY26 CAPEX will decline by USD 500 million to 1 billion, FY27 may remain flat or see a slight increase, and FY28 may decline year-on-year. Unlike META and GOOG, which still have cutting-edge models in deployment

- Microsoft has reduced its leasing demand by 1.98 GW over four years, translating to an annual cut of 500 MW, primarily affecting data centers in the US and Europe.  
- This decision is linked to the impact of Deepseek on GPU utilization and a shift in customer preferences from GPT/Llama to Deepseek, leading to a slowdown in expected demand growth.  
- Consequently, experts predict a potential decrease in Microsoft's CAPEX for FY26 by $500 million to $1 billion, with further implications for rental income in CY25 and CY26.

Microsoft cuts orders and the effect of the Jevons Paradox seems to weaken