Microsoft, Meta, Amazon, and Alphabet guided to roughly $732.5 billion in combined 2026 capital expenditure, 79% growth year over year and 158% above the forecasts issued in September 2024 [S2].
That figure embeds about $432 billion in second-half 2026 outlays alone, with consensus for the same four names reaching $934.5 billion in 2027, Google $284.8B, Amazon $256.5B, Microsoft $207.6B, Meta $185.6B [S3]. Even a flat 2027 versus the back half of 2026 keeps the run rate above $860 billion annualized.
Where 2026 capex actually stands by hyperscaler
First-half 2026 capex across the four hyperscalers hit $301 billion, with updated full-year guidance implying $432 billion of spending in the remaining two quarters [S2]. For comparison, the 2025 outturn of $410 billion was itself 62% above the $253 billion forecast that Goldman Sachs published in September 2024, and 64% growth year over year [S2]. Each upward revision has come from the operators, not the sell side: guidance now exceeds the November 2025 consensus by 39% [S2].
JP Morgan and Goldman Sachs both model 2027 AI capex at $1 trillion or more, and the semiconductor revenue base is keeping pace: Gartner's August 2026 forecast puts the chip market at $1.56 trillion in 2026, up 92% year over year, with memory tripling to $837 billion [S3]. IDC's base case is $1.29 trillion for 2026, climbing to $1.75 trillion by 2030 [S3].
What a slowdown would actually hit first
Capex is not a single line; the order book is sequenced. Long-lead items, silicon wafers, advanced packaging, HBM stacks, are placed 9 to 18 months ahead of deployment, and 2027 GPU shipments are already largely committed against the $934.5 billion consensus. A pullback would land first on the shorter-cycle spend: 800G/1.6T optical transceivers, switch ASIC slots, custom NIC volumes, and the rack-scale power and cooling balance-of-plant. [S3]
The second-order hit is memory and high-bandwidth packaging. Gartner sees memory revenue tripling to $837 billion in 2026 and AI data centers rising from 36.5% of industry revenue in 2026 to over 53% by 2030 [S3]. HBM3E and HBM4 capacity at the leading IDMs is sold out well into 2027, so a moderation would show up first in HBM bit-allocation cuts and in CoWoS / SoIC wafer starts, not in the GPUs themselves. Networking gear is the most elastic line: 51.2T switch ASICs, 1.6T DR8 optics, and 400G/800G NICs are the order books hyperscalers trim first when they want to protect GPU and memory allocations.
The training-to-inference pivot and what it changes
Sandra Rivera, chair of VSORA and former head of Intel's Data Center and AI Group, frames the next leg of the cycle as an inference-efficiency problem: "Power efficiency, cost per token, and certainly much lower latency and more determinism than you're able to get from more general purpose computing architectures" [S3]. Training dominated the 2023 to 2025 build-out because it is throughput-driven and latency-tolerant; inference is perpetual, every token carries a power and cost line, and latency binds to user experience [S3].
This matters for capex composition. A 10% cut that hits the front of the build (new sites, new campuses) hits construction cranes, MV/LV switchgear, transformers, and UPS systems before it hits any silicon. The composition of the cut matters more than the headline number.
Signals worth tracking for an actual inflection
Three indicators would show up before a capex reset hits earnings. First, the ratio of H2 to H1 spend: 2026 needs $432 billion in H2 versus $301 billion in H1, a 43% sequential step-up, and any miss in the Q3 capex line is the first warning [S2]. Second, foundry utilization at the 3 nm and 5 nm AI nodes, where hyperscaler pull dominates the load. Third, lead times on 1.6T optical modules and 800G/1.6T switch ports, which have been the tightest constraint in 2026 alongside CoWoS [S3].
The Reuters report on September 15, 2026 captured the market framing: investors are watching for any sign that spending near $800 billion in 2026 could lose steam [S1]. Through the second half of 2026, the data still points the other way: guidance has risen, not fallen, and the four hyperscalers collectively need only 36.5% growth in 2027 to clear $1 trillion [S2]. The most useful early indicator is not a capex cut but a composition shift away from training cluster expansion and toward inference-optimized rack density, which changes which industrial suppliers feel it first.
Related reading on the build side: Ethernet-APL vs 4-20 mA HART for temperature transmitters: a 2026 spec-level comparison and Dual Power Input Redundancy in Industrial Ethernet Switches: Design, Sourcing, and cover the cabling and switch redundancy that hyperscale buildouts actually pull through.
Spec-level background on the components involved: first aid kit, and construction machinery and equipment.