AI Capex Cycle — Infrastructure Spending, Energy Constraints & Beneficiaries

contested
Horizon: n/a Evidence: 291 Contributors: 61 Updated: 2026-04-10

Verdict

The AI capex cycle remains broadly intact but faces growing investor skepticism about overinvestment. BofA's Fund Manager Survey shows a record net 23% of investors now believe corporations are 'overinvesting,' with CIOs pivoting toward balance sheet improvement over capex increases (35% vs. 20%) as of April 2026 [E1641]. Yet bullish participants argue the 1990s fiber overbuild analogy is misplaced because GPU-powered inference generates immediate revenue [E1643], and the trade is evolving into second-order beneficiaries—optical connectivity, memory testers, hybrid bonding, and silicon photonics supply chains—where Citrini's optics basket has tripled with +35% YTD returns [E2035]. The tension between Jevons paradox reasoning (efficiency gains driving more usage and more capex [E2051]) and the risk that concentrated dependencies like Microsoft's 45% OpenAI-tied backlog [E1838] prove uneconomic keeps this theme actively contested rather than cleanly directional, while Goldman flags that tech ROE has nearly doubled since the GFC even as capex intensity rises [E4069].
What would falsify this thesis:
Evidence Balance
0.75
Velocity
accelerating
Consensus
61 contributors
Contestation
5%
Confidence
62%
Market

Quantitative Context

Tech Concentration: QQQ vs SPY (30d)
0.9%
neutral

🟢 Supporting (212)

[E2050] Short Micron thesis based on efficiency gains potentially reducing semi/power consumption and capex projections. Not extrapolating past gains without accounting for efficiency changes. 270-230 looks interesting level, looks toppy and susceptible to narrative-driven move.
@Stuart Hardy · 2026-04-07 · slack
[E2035] Citrini update: Optics thesis validated - connectivity basket tripled (+35% YTD). Now crowded in pluggable leaders, shifting to deeper CPO/SiPh supply chain. Photons solve electron heat/dissipation/IO limits. Pluggable demand >$10B 2026. Flagship idea: Himax (HIMX) for wafer-level micro-lens arrays/prisms for TSMC COUPE, 2026 first mass production >$100M annualized. Supporting basket: AIXTRON, Soitec, SUSS MicroTec, Nokia, Furukawa.
@thibault · 2026-04-07 · slack
[E2009] Data centers need land, natural gas and water. TPL at $523/share.
@Mark Tetreault · 2026-04-07 · slack
[E1833] Taking a put spread on AAPL (thesis they're running out of steam) paired with long Micron position. MU targets: above $500 take profit sell half, above $550 close rest, below $320 stop loss.
@Will B · 2026-04-07 · slack
[E1831] Citrini memo covers semis/memory trade evolution from HBM manufacturers to HBM4 and hybrid bonding. Second-order winners: memory testers, subsystem suppliers, and silicon photonics. Also covers cybersecurity thrown out with SaaS meltdown and SpaceX supply beneficiaries. Dispersion remains key at macro (IWM>QQQ) and vertical (memory) level.
@thibault · 2026-04-07 · slack
[E1639] Question broad brush commentary on Tech stocks. Not with capex projections and physical AI just getting started.
@Scott Leavitt · 2026-04-07 · slack
[E1442] Using AI extensively for research and trade idea generation - allows covering much more ground. Found incredible winners from AI research. Has properly engineered a prompt pipeline with comprehensive specific prompts.
@Antonio Furtado · 2026-04-07 · slack
[E1441] Claude for Excel built complete valuation model in 20 minutes - read earnings reports including recent press releases, presented assumptions, made adjustments. Mind blown by capability.
@James S · 2026-04-07 · slack
[E1433] Expects massive overbuild in utilities and when all power comes online, energy prices will get crushed. Agrees with running local infrastructure given this trajectory.
@Antonio Furtado · 2026-04-07 · slack
[E1429] CRWV relies on leases and companies they lease from showing stress. One data center already delayed and CRWV cut guidance - equipment sitting idle depreciating.
@Antonio Furtado · 2026-04-07 · slack
[E1420] AI trade moving from chips to: what has a moat that will never be disrupted. Will Walmart still exist in 30Y? Yes. Investors shift to hard-to-replicate physical infrastructure that AI will supercharge.
@thibault · 2026-04-07 · slack
[E1316] Per Gavekal: coal and hydroelectric power sources will exert a larger 'gravitational pull' on data center capex expenditures if radical Islam is successful in disrupting long term hydrocarbon energy supply expectations.
@Mark Tetreault · 2026-04-07 · slack
[E1124] We should follow Alex Wang if we don't already. It's becoming really disruptive to my investment decision making process. Could not sleep last night thinking about the impact AI can have on private credit markets.
@Nicky Adam · 2026-04-07 · slack
[E1120] Alex Wang is the reason Inner Loop is called Inner Loop - the most powerful force in AI is recursive self-improvement, the innermost loop. His daily substack is astonishing. What's happening in AI is massively underappreciated in terms of impact and velocity. His posts are like peeking into the future.
@Will B · 2026-04-07 · slack
[E1114] Paired AAPL put spread with long Micron position. Set alerts: MU above $500 take profit (sell half), above $550 close rest, below $320 stop loss.
@Will B · 2026-04-07 · slack
[E1113] Off the back of Citrini analysis, have taken small positions in GFS, UCTT and ICHR. Would have done more in last two but they ran away. Been in and out of MU over Jan.
@thibault · 2026-04-07 · slack
[E1080] Stock price punishment only affects shareholders, not companies. Real corporate punishment plays out in the bond market with a time delay dictated by bond ladder maturities.
@Mark Tetreault · 2026-04-07 · slack
[E1079] AI capex is like fracking NG in the Permian. If you don't recover your sunk costs in 18 months your business model fails. AI pricing power erodes very quickly.
@Mark Tetreault · 2026-04-07 · slack
[E1078] Kevin's podcast argues MAG7 FCF forecasts have collapsed even as NTM EPS keeps rising, driven by insane AI datacenter capex boom now funded by unprecedented debt. Big Tech plans $660bn on AI this year, 60% rise from $410bn in 2025. Amazon capex $200bn this year, $50bn above expectations. Microsoft down 18% since results with 45% of $625bn cloud backlog tied to OpenAI.
@Stuart Hardy · 2026-04-07 · slack
[E747] DB call comment: Anthropic Enterprise Agents briefing suggests model providers unlikely to displace software incumbents - positioning as orchestration layer on top of existing systems instead.
@Stuart Hardy · 2026-04-07 · slack
Show 192 more

🔴 Challenging (20)

[E2052] Andreas loves MU. Never managed to get in as overbought for long time. Dangerous short.
@Gaetan Warzee · 2026-04-07 · slack
[E2051] Increasing efficiency will reduce price which will incentivize exponentially more usage and more capex than currently projected (Jevons paradox).
@Jesse · 2026-04-07 · slack
[E1642] BofA: Most obvious catalyst to reverse the AI-awe to AI-poor rotation is an AI hyperscaler announcing a capex cut.
@Stuart Hardy · 2026-04-07 · slack
[E1641] BofA FMS: Net 23% of investors now say corporations are 'overinvesting'—a new record. CIOs telling CEOs to improve balance sheets (35%) rather than increase capex (20%, down from 34%).
@Stuart Hardy · 2026-04-07 · slack
[E1640] Capex and earnings projections are lagging indicators and just as reflexive as price. When times are good they make deals and project huge future capex, when times are bad they cancel deals.
@Jesse · 2026-04-07 · slack
[E1108] Space datacenter economic viability is pure speculation. If we improve energy production on earth faster than we decrease launch costs, we'll never get there. The big step change requires Rocket Lab, Blue Origin, and SpaceX all achieving full rapid reusability and competing for launch costs around $1M per trip.
@Jesse · 2026-04-07 · slack
[E1081] If it was 1 company spending on datacenter capex, maybe they're delusional. But all top tech companies in the world saying they could monetize more capacity immediately? Believe in Bezos's early comments about investing heavily for the future - if you don't like it, don't invest. Zero issue with the build outs.
@Scott Leavitt · 2026-04-07 · slack
[E770] Disagrees with Jordi Visser that large enterprises are structurally limited in AI adoption. Enterprise adoption will happen fairly quickly. RIF drives AI adoption, not the other way around — Jack Dorsey's 40% RIF demonstrates this.
@Mark Tetreault · 2026-04-07 · slack
[E594] NBIS had great earnings call Thursday night — management still delivering everything from 3 months ago. Happy holding NBIS despite data center stress narrative.
@James S · 2026-04-07 · slack
[E326] Massive overbuild in utilities coming. When all this power comes online, energy prices will get crushed. Short-term yes higher prices, long-term probably energy price collapse.
@Antonio Furtado · 2026-04-07 · slack
[E311] NBIS had strong Q4 earnings - nothing slipped, everything delivered inline with guidance. Happy holding NBIS. Also holds Terawulf (WULF) which builds datacenter shells for 10-20yr tenants including Google.
@James S · 2026-04-07 · slack
[E4275] Wood argues there is growing risk that 2026 marks peak US AI capex this cycle — a view not discounted in semiconductor stocks. US hyperscaler capex is projected to grow 65% to $620bn in 2026, but supply constraints (energy) and demand risks (questioning ROI, Meta's Avocado model delay, OpenAI losing market share to Gemini) create downside catalysts. AI capex is increasingly debt-funded with IIF estimating $450bn annualized bond issuance.
@Christopher Wood (Jefferies) · 2026-03-20 · r2
[E4280] Wood questions the sustainability of US AI capex at current levels. The Financial Times highlighted the energy bottleneck as one reason supply constraints may cause AI capex not to grow as fast as anticipated. Meanwhile, Gillian Tett's article warns 'Power failure could undermine America's AI ambitions.' Wood has much greater confidence that China AI capex will continue increasing versus US potentially peaking.
@Christopher Wood (Jefferies) · 2026-03-20 · r2
[E4007] Pal/Mattin explicitly reject the bubble narrative. CapEx is backed by $1T+ backlog orders growing faster than spending. This is demand-led investment. The comparison to dotcom is flawed because in 1999-2000 the Fed hiked 175bps; today policy is easing not tightening. Better analogues are 2021 or early 1990s — consolidation within broader uptrend.
@Raoul Pal (Global Macro Investor) · 2026-03-03 · r2
[E3553] Hanmi Semiconductor (042700 KS) currently holds ~90% share in HBM4 TCB equipment but faces significant risk from large-scale hybrid bonding adoption in HBM4E through HBM5. SK Hynix is jointly developing second-generation hybrid bonding equipment with Hanwha Semitek, threatening Hanmi's dominant position.
@Citrini Research · 2026-02-09 · r2
[E3175] Nicoletos challenges the AI capex overinvestment concern. The current spending is justified as Schumpeter's 'credit inflation' preceding productivity gains. Warsh explicitly believes genuine prosperity comes from innovation and capital investment, not central bank printing. If AI adoption is faster than markets expect, deflationary pressures will be larger and arrive sooner than modeled. The gap between spending and productivity pricing is the opportunity.
@Michael Nicoletos · 2026-02-03 · r2
[E3067] Hartnett recommends shorting IG tech bonds as 'bonds new regulators of AI capex…saying slow it down.' Big Tech's flip from asset-light to asset-heavy business model creates credit stress. US IG tech credit spreads and Oracle 5-year CDS shown widening. This positions bonds as the constraint mechanism on hyperscaler spending.
@Michael Hartnett (Bank of America Global Investment Strategy) · 2026-01-31 · r2
[E2287] Wood argues the key risk for US equities is whether markets can look through an AI trade unwind. His base case for 2026 is that investors will start to question the lack of returns on AI capex. However, he acknowledges this concern is being ignored while capex projections continue rising and the cycle can persist longer.
@Christopher Wood (Jefferies) · 2026-01-26 · r2
[E4795] Hyperscalers cutting capex expectations due to ROI uncertainty and bottleneck constraints. Meta hiring freeze after investor pressure on AI compensation costs. Free cash flow at risk if capex-to-revenue doesn't improve Q2-Q3 2025.
@Jordi Visser · 2025-05-04 · transcript
[E4878] Nanell reveals Microsoft cutting data center projects despite capital commitments. Chinese chips making competitive progress against Nvidia. CoreWeave forced to cut IPO valuation sharply. Market recognizing that unlimited AI spending may not be sustainable without ROI visibility.
@Jordi Visser · 2025-03-30 · transcript

🟡 Contested (14)

[E4069] Christian Mueller-Glissmann notes ongoing anxieties around hyperscalers' massive AI capex spend. The big spenders are becoming more capital-heavy, which might weigh on ROE in the medium term. However, tech sector's ROE has nearly doubled since the GFC, with TMT commanding largest weight within US equities comparable to Tech Bubble levels but with substantially greater earnings contribution than then.
@Goldman Sachs Global Investment Research (Allison Nathan, Jenny Grimberg, Ashley Rhodes et al.) · 2026-03-10 · r2
[E3811] Kevin Warsh (Trump's Fed nominee) argues AI will be 'a significant disinflationary force' like the internet in the 1990s, justifying lower rates. However, an FT analysis counters that current productivity growth at 1.9% is nowhere near the 3.9% seen before Greenspan's 1999 speech. High AI expectations are already reflected in equity prices, boosting demand ahead of supply gains — risking inflation, not containing it.
@Grant Williams · 2026-02-16 · r2
[E3733] Goldman acknowledges several caveats to their wholesale price projections: short-term supply curves may prove more flexible over a full year, policy responses could limit demand growth, and the relationship between peak/non-peak capacity and prices could change. These factors could moderate the price impact.
@Goldman Sachs (Manuel Abecasis, Hongcen Wei) · 2026-02-13 · r2
[E3725] Goldman acknowledges significant uncertainty in their wholesale price estimates. Supply may prove more flexible over a full year than daily markets suggest, policy responses could limit demand growth, and capacity measures may shift between peak and non-peak periods. More energy-efficient data centers, faster supply growth, or 'island' self-generation could reduce the drag; accelerated AI adoption or power outages would increase it.
@Goldman Sachs (Manuel Abecasis, Hongcen Wei) · 2026-02-13 · r2
[E3358] The author positions defensively against AI narrative, preferring sectors that 'can stand on fundamentals and balance sheets vs. narrative' until sustained broad productivity gains translate into earnings outside hyperscalers and semiconductor stack. This implies skepticism about near-term AI capex returns.
@J. King (ChartWizardsNFT / Peter L Brandt) · 2026-02-09 · r2
[E3362] The authors acknowledge the AI capex cycle is concentrated in US and China, but argue price action has been driven more by infrastructure build-out than realized productivity gains. They explicitly prefer sectors that 'can stand on fundamentals and balance sheets vs. narrative' until broad productivity gains translate into earnings outside hyperscalers and semiconductor stack.
@J. King / Peter L. Brandt (ChartWizardsNFT) · 2026-02-09 · r2
[E3512] Memory bears have salient points — supply gluts are likely to happen at some point (author guesses 2027) as Samsung and SK Hynix historically lean in, increase capacity and inevitably glut the market. Human behavior is predictable and greed wins out even with cartel-like dynamics. However, the authors argue this misses how new capacity comes online — new capacity requires new capex spending, which is another player's revenue.
@Citrini Research · 2026-02-09 · r2
[E3523] Memory bears argue 'this time isn't different' — Samsung and SK Hynix historically lean in, increase capacity, and glut the market. Citrini acknowledges they 'probably aren't wrong' and expects oversupply 'sometime in 2027.' However, new capacity requires capex spending which creates fab equipment revenue opportunities before oversupply materializes.
@Citrini Research · 2026-02-09 · r2
[E2526] MS acknowledges downside scenario risk: American LLMs could hit a 'scaling wall' where capabilities plateau despite massive compute increases. Slowness in AI adoption could cause heightened concerns with AI infrastructure capex ROIC, pressuring LLM developers to reduce capex — in such outcome, AI Infrastructure and AI Enabler stocks would underperform.
@Morgan Stanley Research (Stephen Byrd, Michelle Weaver, Daniel Blake, et al.) · 2026-01-26 · r2
[E2286] Wood's base case remains that AI will turn out like the capex-intensive airline industry rather than a winner-takes-all network effect. The key question is when investors will start questioning returns on AI capex, with the cycle already three years old. Despite evidence of market broadening, Nvidia and four hyperscalers still account for 44% of S&P500 gains since 2023 (50% including Broadcom, Palantir, Oracle).
@Christopher Wood (Jefferies) · 2026-01-26 · r2
[E2342] Bubble-like characteristics exist in AI private markets — Thinking Machines Lab and Safe Superintelligence each raised $2B without released products. Circular financing rising among AI companies, with Nvidia $100B investment in OpenAI for 10 gigawatts of systems. OpenAI CFO discussed need for 'backstop' and 'guarantee' financing, prompting White House AI czar Sacks to say 'no federal bailout for AI'.
@Goldman Sachs Investment Strategy Group · 2026-01-26 · r2
[E2490] MS acknowledges market will oscillate between AI bullishness and skepticism on adoption pace in 2026. Key risk: American LLMs hit 'scaling wall' where capabilities plateau despite massive compute increases. Also risk that Chinese LLMs keep pace via innovative approaches. Slowness in adoption could cause AI Infrastructure and Enabler underperformance.
@Morgan Stanley Research (Stephen Byrd, Michelle Weaver, et al.) · 2026-01-26 · r2
[E6359] The AI/datacenter boom is shifting from equity to debt financing, which increases bust risk similar to the telecom bubble of the 2000s and shale oil sector in the 2010s. While AI investment shows strong returns (JPM 100% ROI), the transition to leverage-funded buildout represents a critical risk factor for the sustainability of the AI capex cycle.
@Luke Gromen · 2025-12-06 · ka
[E8081] Munger's characterization of AI as having 'a lot of crazy hype' while being genuinely important suggests the massive AI capex cycle may be partially justified but likely includes significant overinvestment. He favors narrow, proven AI applications like insurance underwriting over broad transformative claims.
@Charlie Munger · 2025-12-06 · ka
💬 Commentary (45)
[E2054] Wife got into MU and SNDK very early - source of much daily gloating.
@Scott Leavitt · 2026-04-07 · slack
[E2053] Dale Pinkert said MU looks like it's got some downside maybe to half its runup.
@Jesse · 2026-04-07 · slack
[E1918] If this is correct, maybe we should think about effects on K-shaped economy and whether this prevents it becoming E-shaped. Similar to UK where fuel inflation post-GFC hit lowest earners hardest.
@Stuart Hardy · 2026-04-07 · slack
[E1854] Then you go talk to 99.999% of everyone else and they're like 'Yeah, tell me again how the chatbots are going to take over the world?!'
@Will B · 2026-04-07 · slack
[E1839] Or they treat AI capex as infrastructure for other internal brands/activities. Their cash flow supports current borrowing. Once overextended, won't be market that punishes them, it will be the banking system. The 'punishment' of the company plays out in the bond market with time delay dictated by bond ladder maturities.
@Mark Tetreault · 2026-04-07 · slack
[E1838] The cashflow these guys have maybe gives them a bit longer runway (certainly for Mag7, less so CRWV), but the market will punish them if they don't see a decent ROI. 45% of Microsoft's backlog tied to OpenAI is shocking.
@Stuart Hardy · 2026-04-07 · slack
[E1644] Journeyman view: We're in capex overspend fear right now but long-term it's overblown. A barbell of AI (bits) and capex receivers (atoms) will do well. We're very early in the AI adoption cycle.
@thibault · 2026-04-07 · slack
[E1643] The 1990s fibre over-build analogy is nonsense: if there's power the chips can be monetised immediately, even if it's just for inference. The revenue attributable to GPUs is already enormous.
@Will B · 2026-04-07 · slack
[E1421] Keep in mind most hyperscalers are also early adopters. Suggests JPM and LLY as potential picks.
@Scott Leavitt · 2026-04-07 · slack
[E1317] Shared analysis of hydroelectric power as % of GDP. Norway (~1.9%), Iceland (~1.5%), Paraguay (~1.2%), Ethiopia (~0.9%), Costa Rica (~1-1.5%) rank highest. Small, water-rich economies dominate.
@Mark Tetreault · 2026-04-07 · slack
[E1123] Alex Wang is a regular on the Moonshots podcast with Peter Diamandis. Very good.
@James S · 2026-04-07 · slack
[E1109] First 2 satellites for space-based power going up this year. Obviously no mention of cloud cover issues. Not exactly efficient due to atmospheric loss compared to LEO datacenters, still not convinced of viability for reasons previously discussed.
@Stuart Hardy · 2026-04-07 · slack
[E1083] Grok analysis: Recent MAG7 selloff created cheaper entry points with forward P/Es in mid-20s. If monetization ramps up quickly post-build-out, this fear-driven dip might look like a bargain. But path to ROI isn't guaranteed to be smooth or swift.
@Scott Leavitt · 2026-04-07 · slack
[E1082] C-suite bonuses move behavior but heavy insider ownership is the only thing that truly aligns management with shareholders.
@Stuart Hardy · 2026-04-07 · slack
[E771] Jordi finds AI stocks using AI to analyze earnings calls for management commentary and revisions, but doesn't do further due diligence. EOS Energy was on his list but just had a -39% day on 4th consecutive earnings miss.
@James S · 2026-04-07 · slack
Show 30 more

Events Reckoned With (10)

Material events in this theme's relevance window. A theme page is only as fresh as the events it has reckoned with — unreckoned events signal the analysis may be stale.

First 2 space-based power satellites scheduled for launch in 2026 reckoned
2026-03-27
Bill Gates' TerraPower receives approval to build new nuclear reactor reckoned
2026-03-06
TerraPower receives NRC approval for new nuclear reactor reckoned
2026-03-06
EOS Energy down 39% on 4th consecutive earnings miss despite high AI theme score reckoned
2026-02-26
Amazon shares fell 11% after hours on $200B capex guidance, $50B above expectations reckoned
2026-02-20
Blue Owl fails to secure $4B financing for CoreWeave datacenter citing overcapacity concerns reckoned
2026-02-20
Reports of Big Tech planning $660bn AI capex for 2026, triggering $900bn combined market cap loss reckoned
2026-02-20
Microsoft down 18% since results with 66% surge in quarterly datacenter spending reckoned
2026-02-20
NBIS strong Q4 earnings - management delivering on guidance reckoned
2026-02-13
NBIS Q4 earnings: management delivering on guidance reckoned
2026-02-13