Morgan Stanley: Hyperscale Cloud Vendors' Capital Expenditure to Exceed 2 Trillion in Two Years, Accelerating the Arms Race in Computing Power

Morgan Stanley's U.S. Internet Analyst Brian Nowak recently presented a comprehensive cross-team deep dive in the latest edition of "What We Learned, What Matters Next," covering the costs of computing power, types of chips, and the capacity expansion of hyperscale cloud providers.

"Hyperscale companies' total capital expenditure will exceed $1 trillion by 2027, with accumulated expenditures over the past years amounting to $2 trillion." Brian Nowak opened the show with this figure, setting the tone for the core topic of the report: The race for computing power is far from over.

How Much Will Computing Power Cost?

Morgan Stanley collaborated with semiconductor, networking, and power teams, employing a bottom-up estimation logic based on server and rack delivery data from hyperscale manufacturers, including products from NVIDIA and custom ASICs, to systematically map out the relationship between supercomputing capital expenditure and the scale of computing power for the first time.

The report breaks down the cost of computing power per GW into two major modules: Rack Costs (including pricing for GPU/TPU chips, HBM, DRAM, and CPUs) and Off-Rack Costs (including backup power, cooling systems, power enclosures, network transceivers, and other supporting facilities). Even if hyperscale providers purchase entire racks of products from NVIDIA in bulk, Morgan Stanley has independently disaggregated and tallied these detailed costs.

The calculation results indicate significant differences in the cost per GW for three mainstream chip systems: GB300 systems are about $33 billion, TPU systems are about $25 billion, and other chips are about $15 billion. This framework can be adapted to dynamically adjust to component inflation and chip price fluctuations.

NVIDIA's High PriceTag is Justified: Performance per Watt is 2 to 8 Times Higher than Custom Chips

In the face of the market narrative that "cloud providers will massively shift to custom ASICs," the report provides a counterargument.

Morgan Stanley cites semiconductor analyst Joe Moore's research indicating that NVIDIA's chip performance per watt is 2 to 8 times higher than that of custom chips. This is the core reason why hyperscale providers are willing to pay a higher per-GW cost for NVIDIA chips—the higher energy efficiency means lower per-token inference costs. As the workload of large models shifts from training to inference scenarios, the cost value of this energy efficiency advantage will become even more prominent.

Custom chips are not without a path to catch up, but the key is not a simple stack of native silicon performance. The report points out that the key direction to narrow the gap in performance-to-power ratio with NVIDIA includes: network architecture optimization, memory bandwidth improvement, software stack optimization (such as AWS's internal optimization for EFS), and access to NVIDIA's NVLink technology. Cloud providers actually output entire systems to customers, not individual chip products.

Regarding critical junctures, Trainium 3 will be launched by the end of this year, and Trainium 4 will be launched in 2027, which will be an important window to observe whether the gap between custom chips and NVIDIA has substantially narrowed.

34GW Added in Two Years, the "Second Derivative" of Computing Power Expansion is Still Accelerating

The report provides the most specific quantitative forecast to date for the scale of computing power expansion of the four major cloud providers (Google, Amazon, Microsoft, Meta):

• 2026 total additional computing power 14GW

• 2027 total additional computing power 20GW

• Total additional 34 to 35GW over the next two years

Looking at the providers: Google has the largest increment, with a total of about 13GW over two years, covering both NVIDIA chips and its own TPUs; AWS and Azure each add 8 to 10GW, with both adding about 5GW in 2026 and about 4 to 5GW in 2027; Meta adds several GW in 2026 and about 3 to 4GW in 2027.