AI capex is not over; the math argues boom, not bust. Extreme fear, weak tape in MU and peers, and record-wide AI credit spreads are set against $835B of 2026 capex, a $2.4T backlog, and a claim that the spend is already sold out into 2027-2028. The negative case leans on capital-cycle risk, data-center bottlenecks, higher rates, and delayed IPOs at Anthropic; the positive case leans on exploding inference demand, hyperscaler revenue growth, 5-6 year GPU lives, and enough free cash flow at the big buyers to keep funding rolling. Final verdict: the AI trade is still alive, with casualties, but the big four can keep paying if growth and depreciation stay favorable.
Fear and greed is at 19, extreme fear, while MU is down 14% on the week; MARV is down nearly 17%, ALAB nearly 16%, ASML 7%, TSLA 7%, SpaceX 11.6%, PLTR 6%, Cerebras 22%, Eshinx nearly 19%, SALANA 7%, and MICRO 12%.
Rumors say MU is being booted from the MSCI World index fund; yet MU is still up more than 200% YTD, AMD 150%, MARV 170%, ALAB 72%, and ASML 54%.
BofA’s bear case is a historic boom-bust pattern like the dot-com bust, with cheerleading instead of investment creating capital-cycle risk.
Nearly 70% of the U.S. population is said to oppose AI data centers, and the buildout is described as physically constrained by trades, connectivity, substrates, land, power, hookups, and multi-year permits.
OpenAI and Anthropic are said to want 5 GW each, while another 43 GW of new AI compute is targeted next year despite regulatory and physical bottlenecks.
AI credit is said to have been smashed this week: neoclouds brutalized, the credit market flinching, the Fed up 25 bp, the 10-year Treasury above 100 bp higher to near 6%, and CDS on CORE and ORCL at record-wide spreads.
Anthropic is rumored to be pushing its IPO out past November after first targeting October, then November, with valuation chatter cut to $1.5T from $2T.
The big four capex totals are said to be AMZN $220B, GOOGL $205B, MSFT $175B, and META $145B, for $735B-$750B; adding ORCL at another $100B lifts 2026 capex to $835B.
$835B of 2026 capex is set against a $2.4T backlog, with AI infrastructure spend projected at $3T-$4T per year by 2030 and backlog covering 2026 capex by nearly 3x.
The bullish case starts with a Jevons-style paradox: as intelligence gets cheaper, demand for intelligence rises; GPT-1 training is framed as falling from $20M to $100, a 1,000,000x compression.
A 10,000-agent, 4M-compute math problem is cited as proof that small compute can still solve hard problems, but scale then explodes with agents and a claimed billion-x rise in demand.
Jensen’s Nvidia split is cited as one-third classic computing, one-third across five to six frontier labs, and one-third split across startups and neoclouds on open-source models.
Inference is called where the money is: SemiAnalysis is cited for improving GPU efficiency and falling prices, while frontier-model inference margins are said to have risen from 40% to over 70% in a few months and to ultimately consume 99% of compute.
Hyperscaler revenue growth is used as proof the spend is paying off: Azure up 43%, GOOGL cloud up 82% YTD, and AMZN up 37%.
NVDA is said to have given notice that it expects to double growth next year, while AI hardware demand for Google TPUs is described as off the hook, with TPU sales forecast at $2B in 2026, $18B in 2027, $43B in 2028, $52B in 2029, and $60B in 2030.
Memory stocks are said to have been clobbered this week, with Samsung, SK Hynix, and MU hit; Michael Dell is cited saying the memory crunch gets worse, there is no memory, and supply is sold out for next year.
The supply-side bull case says there is no compute available for 2027 or 2028 because it is already sold out, and SpaceX is said to want 2.7 GW of compute by year-end, equal to about $100B in GPUs and roughly $100B ARR.
SpaceX is said to plan to use 60% of its compute for itself and lease 40% out, with a CFO claim that each gigawatt turns on with ROI in under a year; Elon is then expected to exceed 10 GW by end-2027, implying nearly 2 GW per quarter.
OpenAI and Anthropic are said to plan to add 10-12 GW combined next year, roughly 5-6 GW each, while OpenAI and Anthropic combined revenue is projected to rise from about $100B this year to over $500B in the next 12 months.
The cash-flow verdict says over half of the players still generate enough free cash flow to fund the AI capex; AMZN is slightly negative for one quarter, while MSFT, GOOGL, and META stay free-cash-flow positive, and big players should keep access to credit.
The final call is that the AI trade is not dead: 'Are we going to go bust? No' and 'Are we going to go boom? Yes,' with casualties expected but the big four still able to pay if revenue, GPU depreciation, and funding hold.