Big-5 hyperscaler 2026 capex landed at roughly $775–800B after Q1 2026 earnings, about 64% above 2025, with Goldman Sachs projecting US data-centre power demand to climb from 41 GW in 2026 to 66 GW in 2027 [S2]. PwC's central case, built with Oxford Economics, puts global data-centre capex at $31.6T through 2050 in a range from $22T to nearly $50T, driven by 4–6 year refresh cycles on GPUs, networking, and storage rather than by buildings [S1].
Allianz frames the same flows as roughly $575B in 2026 hyperscaler capex (about +50% year-on-year) and models a doubling of the data-centre pipeline to around 200 GW by 2030, while warning that energy volatility, not aggregate demand, is the variable that can reshape the spend mix [S3]. The market signal in March 2026: cash-flow visibility and revenue realisation matter more to investors than headline capex, even with the supercycle intact [S3].
Where the 2026 hyperscaler dollars are being booked
Allocating the $775–800B 2026 envelope by announced guidance, Amazon runs near $200B, Microsoft around $190B, Alphabet about $185B, Meta roughly $135B, and Oracle near $50B, with downstream captures led by NVDA GPU silicon at approximately $250B, Vertiv power and cooling at $85B, Equinix colocation at $55B, Constellation nuclear at $28B, and HBM memory at $25B [S2]. Moody's March 2026 update projects Big-5 capex reaching about $820B in 2027, a 21% year-on-year step on top of 2026, while Morgan Stanley floats Alphabet alone at up to $250B in 2027 [S2].
The 2025 to 2028 trajectory runs $429B actual (2025), $789B guided (2026), $935B estimated (2027), and $1,080B projected (2028), a roughly 20% compound that Dell'Oro pegs as a 21% CAGR through 2028 [S2]. Goldman Sachs expects Meta capex near $144B in 2027, and CreditSights lifted the 2026 aggregate estimate to about $750B post-Q1 earnings, above company midpoints [S2].
Overbuild risk: the fibre and canal parallels
The historical anchor is the early-2000s telecom buildout, where roughly $250B in global capex fed fibre and 3G wireless networks and ended in a capacity glut once the dot-com bubble burst [S5]. Pre-railway Britain drew a smaller but instructive parallel: more than 60 canal companies raised over £12 million in new capital (about $12B in today's money) before demand caught up, and the consolidation was brutal [S5]. For AI, the demand side has more near-term visibility (cloud migration, hyperscaler expansion, AI-native workloads), but the spread between optimistic and pessimistic AI adoption paths is roughly the size of US GDP in PwC's cone [S1].
Mark Zuckerberg's framing, "meaningful chance that a lot of the companies are over-building now", was paired on Google and Meta calls with the opposite view: "the risk of underinvesting is dramatically greater than the risk of overinvesting" [S5]. Allianz's AI Bubble Risk Monitor in March 2026 reads moderate pressures, with exuberant positioning cooled but credit spreads widening, and a 30% negative correlation between US tech capex and severe energy-price moves in the historical record [S3]. The base-rate read: the supercycle survives a 2026–2027 stress, but a 50% jump in chip costs, as seen in Q1 2026, can shift dollars from build to lease, with leasing models offering 20–30% capital-intensity savings [S3].
Underbuild risk: power, packaging, and substrates
JLL's year-end 2025 figure of about 1.4% North American colocation vacancy is a historic low, while Morgan Stanley projects a 49 GW US capacity shortfall by 2028 against a 207 GW requirement by 2030 (up from 82 GW in 2025) [S2]. Germany's plan to double data-centre capacity by 2030 shows public-sector capital substituting where grid headroom is missing, with the European 200 GW pipeline x2 anchored by both private and sovereign spend [S3].
On the silicon side, TSMC's CoWoS throughput was set to scale from about 40,000 wafers per month in 2024 to 140,000–150,000 by 2026, almost a 4× expansion in two years, with ASE's advanced-packaging revenue up over 60% in 2024 and a new Malaysia plant running, while Amkor built a $2B Arizona packaging facility backed by about $400M in federal grants and an Apple anchor order [S4]. ABF substrate makers Ibiden, Shinko, Unimicron, Nan Ya PCB, and Kinsus are adding capacity with government support, and Ajinomoto, the sole source of ABF build-up film, plans roughly 50% more output by 2030 [S4]. Photomask capacity from Toppan (rebranded Tekscend), Dai Nippon Printing, and Photronics is being lifted to relieve EUV reticle shortages at the 5nm and 3nm nodes [S4].
Physical plant side: what a control engineer actually specs
On the brownfield side of this build, the same supply tightness that hits CoWoS and ABF substrates is hitting process instrumentation. Cooling-plant flow metering on hyperscaler campuses is moving toward Coriolis and ultrasonic flow meters for chilled-water and dielectric fluids, with custody-transfer uncertainty under 0.5% and no moving parts to service in the warm-aisle return. Chilled-water isolation now routes through metal-seated industrial valves rated for sustained below-ambient duty cycles, because the same plant that absorbs $789B in compute capex still loses days of revenue to a single failed two-inch butterfly seat. [S1]
The same asset-replacement cadence that drives GPU refresh every 4–6 years applies to the controls layer, so PLC racks in new data-centre buildouts are being replaced on a 7–10 year cycle rather than the historical 15–20. Where cooling pumps and CRAH fans are driven by variable-speed servo motors for soft-start and harmonic mitigation, the spec is typically 1.5–3.0 kW with IP65 enclosures and integrated drives rather than separate VFD panels. Field data from industrial monitoring rollouts, including transport-fleet pickups detailed in accelerometer vs 6-axis IMU for transport monitoring, show the same six-month refresh window showing up on the mechanical side that PwC's 4–6 year model applies to compute silicon [S1].
Comparing the two camps on hard criteria
Overbuild thesis and underbuild thesis on five criteria: demand visibility (overbuild: low for 2028+, underbuild: high through 2027 given 1.4% colocation vacancy and 64% YoY 2026 step), supply elasticity (overbuild: high in fibre-style scenarios, underbuild: constrained in CoWoS, ABF, and grid interconnect), leading indicator (overbuild: credit-spread widening, underbuild: forward power commitments and 200 GW European pipeline), historical base rate (overbuild: 2000s telecom, underbuild: none in last 30 years for an ICT cycle of this scale), and capex-to-cash conversion (overbuild: declines with leasing shift, underbuild: pressured by Q1 2026 50% chip-cost step but partly hedged by hyperscaler cash reserves) [S2][S3][S5]. The PwC central case effectively assumes neither pole dominates, with 2026 at $800B, 2030 at $1.1T, and 2050 at $1.8T in annual data-centre capex, and a $22T–$50T range that lets the overbuild and underbuild scenarios coexist inside one cone [S1].
SSGA's November 2025 framing, that "the AI capex cycle may have more staying power than you think", leans on the same hyperscaler cash buffers Allianz cites, plus 4–6 year refresh cadences that mechanically re-anchor spend every half-decade, and revenue backlogs like Microsoft commercial RPO of $678B (+84% YoY) that pre-subscribe a slice of future compute [S2][S6]. The unhedged tail is power: a 49 GW US shortfall by 2028 is not a capex line item, it is a permitting and grid-interconnect queue, and that queue, more than any bubble call, decides whether 2026 lands closer to overbuild or underbuild [S2][S3].
Limits of the overbuild vs underbuild frame
Both camps tend to treat "AI capex" as a single line item, but Allianz's March 2026 note flags that value creation across the stack is uneven, with semiconductors and telecom equipment as relative winners and software and consumer electronics as more challenged, so aggregate $775–800B headlines can mask divergent sub-cycle outcomes [S3]. The same note flags a likely tilt in capex mix from software and computing services toward critical equipment and raw-material production outside Asia, driven by helium and LNG supply risks in South Korea and Taiwan [S3].
The 30% negative correlation between US tech capex and severe energy-price moves means the underbuild case is most exposed not to a demand shock but to an oil or LNG shock that lifts discount rates and inflates chip-fab power costs [S3]. On the silicon bottleneck, ABF substrate, photomask, and CoWoS expansions all sit on multi-year ramps that do not flex with quarterly demand, so a 2027 demand air-pocket would still leave substrate and mask capacity on the build path defined in 2024–2026, exactly the kind of mismatch that produced the 2000s telecom glut [S4][S5].
Trackable signals into Q4 2026: Q3 2026 hyperscaler earnings for any 2027 guide revisions, North American colocation vacancy (the 1.4% YE 2025 base is the line to watch), Morgan Stanley and Goldman updates on the 49 GW / 66 GW power-demand paths, and any second wave of Q1-2026-style 50% chip-cost moves that would tip capex toward 20–30%-savings lease models [S2][S3].