Deutsche Bank: Meta computing power expansion plan may not necessarily push up capital expenditures; cloud business may unlock AI investment returns

Deutsche Bank: Meta computing power expansion plan may not necessarily push up capital expenditures; cloud business may unlock AI investment returns.

Meta’s new round of computing power expansion plan has once again touched the market’s sensitive nerves about the return on AI investment of large technology companies.

According to media reports, Meta plans to increase its computing infrastructure capacity from 7GW in 2026 to 14GW in 2027, and mass-produce its self-developed AI chip Iris as soon as September 2026.

If the construction cost of a traditional GPU data center is estimated, the new 7GW capacity may correspond to an investment of approximately US$245 billion, which is significantly higher than market expectations. Wall Street's previous consensus estimate for Meta's capital expenditures in 2027 was approximately US$160 billion to US$165 billion.

However, Deutsche Bank said in its latest research report that the market may have overestimated the pressure this round of expansion will have on Meta’s balance sheet and profits, while underestimating the company’s ability to convert idle computing power into third-party cloud revenue.

Deutsche Bank maintains a Meta "buy" rating and sets a target price of $810. Why does the 14GW capacity plan make the market uneasy?

The market's first reaction came from the scale of Meta's capital expenditures, and investors' concerns about excessive capital expenditures have been suppressing the stock prices of hyperscale cloud vendors such as Meta and Microsoft since last year.

Deutsche Bank used the cost of "Neocloud" company CoreWeave as a reference and estimated that each additional 1GW of computing capacity would cost about US$35 billion. If Meta adds 7GW in 2027, the related AI infrastructure investment could theoretically reach US$245 billion. CoreWeave mainly uses NVIDIA GPUs to build data centers.

Although some spending may occur earlier in 2026, this plan still means that Meta's 2027 capital expenditures may exceed $200 billion. That compares with Wall Street's consensus forecast of just $160 billion to $165 billion, and Deutsche Bank's previous forecast of about $185 billion.

In the past two years, large technology companies have continued to increase investment in AI servers, data centers and power infrastructure.

Investors initially focused more on computing power shortages and AI growth potential, but as capital expenditures continue to rise, market pricing focus is shifting to depreciation pressure, free cash flow and investment return cycles.

For Meta, the greater the scale of infrastructure investment, the more obvious the impact of future server and chip depreciation on profit margins.

However, Deutsche Bank estimates that buyer institutions’ previous actual expectations for Meta’s capital expenditures in 2027 have reached approximately US$220 billion, which is much higher than the public consensus expectations.

This means that some professional investors may have already factored in large-scale expansion in advance, and the incremental impact brought by the 14GW target may not be as large as the superficial number.

Off-balance sheet financing could reduce Meta's direct construction capacity to 5.5GW Deutsche Bank believes that all the new 7GW cannot be simply regarded as being independently funded by Meta. Meta has brought in partners and off-balance sheet financing for its Louisiana-based data center project, Hyperion.

The project plans to build about 5GW of capacity in the long term, of which about 1GW to 1.5GW may be put into operation in 2027. Since the relevant construction adopts a joint venture or partnership investment structure, Meta's direct capital investment is relatively limited, and part of the funds it should bear may have already been invested.

If the approximately 1.5GW occupied by the Hyperion project is deducted from the 7GW added, the capacity Meta will need to directly finance and build in 2027 may only be 5.5GW. Calculated at US$35 billion per GW, 5.5GW corresponds to an investment of approximately US$190 billion to US$195 billion.

While the size was still substantial, it did not significantly exceed Deutsche Bank's estimated buy-side expectations. This arrangement reflects a new trend of large technology companies building AI infrastructure.

As investments in individual data center projects reach tens of billions of dollars, technology companies are increasingly using joint ventures, infrastructure funds, asset managers and long-term lease arrangements to spread construction costs and reduce balance sheet pressure.

Iris self-developed chips are expected to reduce unit computing power costs The second important variable comes from Meta’s self-developed chips. According to reports, Iris took about six weeks to complete the test and no major problems occurred. Meta collaborated with Broadcom on the design and TSMC is responsible for manufacturing.

The company plans to start mass production in September and accelerate the chip iteration rhythm to 2027. Iris is part of Meta's training and inference accelerator project MTIA. Its goal is not to immediately replace Nvidia or AMD's GPUs, but to shift some internal workloads to chips customized for Meta's business.

Deutsche Bank estimates that servers and chips account for about 60% to 65% of total data center costs. Nvidia's gross profit margin is about 75%. If the procurement or manufacturing costs of Meta's self-developed chips are about 50% lower than Nvidia's GPUs, the overall data center construction cost may drop by 30% to 35%.

Assuming that the 5.5GW capacity directly built by Meta uses 50% Iris chips and 50% external GPUs, the average cost may drop from about US$35 billion per GW to about US$30 billion, corresponding to a total investment of about US$165 billion.

This means that even if Meta achieves its 14GW capacity target, actual capex will likely still be within the range that investors have already accepted. However, it is difficult for self-developed chips to completely replace GPUs in the short term.

Large model training still requires general-purpose accelerators with stronger performance and a more mature software ecosystem. Meta will continue to be an important customer of Nvidia and AMD in the future. The main significance of MTIA is to optimize the chip portfolio and reduce marginal costs, rather than completely getting rid of external suppliers.

Excess computing power can be converted into cloud revenue However, in an environment where the supply of computing power is still tight, Meta can also lease older generation, temporarily idle or non-core computing capacity to third parties.

Deutsche Bank predicts that Meta will have approximately 2GW of capacity at the end of 2025, add 5GW in 2026, and add another 7GW in 2027, reaching 14GW by the end of 2027. Under the pessimistic, baseline and optimistic scenarios, Deutsche Bank estimates Meta's total capacity available for sale to be 1.95GW, 2.575GW and 3.2GW respectively.

If 75% of them can be successfully leased, the actual sold capacity will be approximately 1.463GW, 1.931GW and 2.4GW.

Based on an estimate of annualized revenue of US$10 billion to US$15 billion per GW, Meta's third-party cloud business may contribute approximately US$14.6 billion to US$36 billion in incremental revenue in 2027, which the report summarizes as US$15 billion to US$36 billion.

Among them, the incremental revenue of the base scenario is approximately US$24.1 billion, which is equivalent to 8% higher than Wall Street’s consensus expectation of Meta’s revenue of US$301.8 billion in 2027. The revenue upside potential corresponding to the pessimistic and optimistic scenarios is 4.8% and 11.9% respectively.

Deutsche Bank's previous baseline scenario only expected about US$17 billion in third-party cloud revenue. As Meta's capacity target rises to 14GW, the baseline revenue forecast increases to approximately $24 billion, indicating that greater capital expenditures lead to greater potential external realization space.

Deutsche Bank assumes that the incremental profit margin of the computing power rental business is 50% to 75%, taking into account related costs such as sales and customer support. According to this assumption, the incremental operating profits corresponding to the three scenarios are US$7.3 billion, US$15.1 billion, and US$27 billion respectively.

Wall Street currently expects Meta's 2027 GAAP operating profit to be approximately $105 billion, with an operating margin of 34.8%.

After including third-party cloud revenue, operating profit may increase to $112.4 billion to $132 billion, and operating profit margin may rise to 35.5% to 39.1%, an increase of 70 to 428 basis points from current expectations. In terms of earnings per share, the market currently expects Meta's 2027 GAAP diluted earnings per share to be $35.12.

Deutsche Bank estimates that after adding third-party computing power business, EPS may reach 37.47 to 44.17 US dollars, corresponding to an upside of 6.7% to 25.8%. Under the base case scenario, Meta's 2027 earnings per share are approximately $40.11, 14.2% higher than consensus estimates, equivalent to an increase of approximately $5.

The pace of AI model release is accelerating, and infrastructure construction is beginning to be transformed into products Meta recently released Muse Spark 1.1, further expanding its product scope to agent programming.

Deutsche Bank believes that Meta’s recent accelerated pace of model releases indicates that its AI team may have been mainly building training, inference, model and software infrastructure from scratch in the past 9 to 10 months.

As the underlying technology stack gradually improves, the company is expected to increase the frequency and predictability of new model releases in the future. (

#Stocks #Nvidia #Microsoft #Meta #AMD