BofA Raises Capex Forecasts for Alphabet, Amazon, and Meta
Miles Bennett
BofA Securities raised capital-expenditure forecasts for Alphabet, Amazon, and Meta, with 2027 estimates up by a combined $195 billion-plus; memory-price inflation and next-generation AI model compute demands are the twin drivers as all three turn to debt markets to fund expansion.
How much did each forecast rise?
Alphabet's 2026 capex estimate rose from $187 bn to $195 bn; its 2027 estimate jumped from $257 bn to $290 bn — the largest two-year increase of the three.
Meta's 2026 estimate rose from $130 bn to $145 bn; 2027 from $157 bn to $185 bn.
Amazon's 2026 figure held at $159 bn, but 2027 climbed from $196 bn to $230 bn. This means → Amazon's acceleration is back-loaded into 2027, with near-term spending relatively restrained.
Why is Amazon issuing $25 billion in bonds?
On the same day as the forecast upgrade, Amazon disclosed plans to sell at least $25 billion in bonds across eight tranches for general corporate purposes.
Amazon told underwriters this would be its last U.S. public debt deal of 2026; it has already tapped Swiss-franc, Canadian-dollar, and other currency markets this year.
In plain terms = AI data-center spending now outpaces what operating cash flow alone can cover, so all three hyperscalers have turned to borrowing to fund expansion.
Where is the cost pressure coming from?
Memory: DRAM spot prices have risen roughly 40% from the previous quarter, pushing up server procurement costs directly.
Compute: Meta revealed that its next-generation AI model, "Watermelon," requires 10× the compute of Muse Spark, the frontier model released in April.
This means → cost pressure is not single-sourced — hardware prices are climbing while models are growing larger, squeezing budgets from both sides.
What does it cost to build an AI data center?
BofA estimates that 1 GW of AI data-center capacity costs between $25 bn and $45 bn to build.
AI servers and GPUs account for $14 bn to $28 bn; the remainder covers power infrastructure and networking.
This reflects the fact that GPUs and servers are the single largest line item, but power and networking combined are equally hard to ignore — they make up over 40% of total cost.
What do the 2030 capacity plans look like?
BofA projects that by 2030, installed capacity will reach 58.1 GW for Amazon, 32.4 GW for Alphabet, and 22.8 GW for Meta.
Alphabet last month announced plans to sell $80 bn in stock to fund AI infrastructure, including a $10 bn commitment from Berkshire Hathaway.
Put simply = whether memory prices keep rising and whether next-gen model compute demands materialize as expected are the two variables that will determine if this round of capex forecasts gets revised upward again.
Content is for reference only, not financial advice.