Trends›AI demand outruns expectations

AI demand outruns expectations

AI compute and inference demand keeps outrunning consensus expectations through 2030, so the build-out (compute, memory, power, thermal, photonics, packaging) never plateaus and the spend-driven book stays long.

78 ↑3 confidence, computed 23 Oct 2026

Economic
High weight

Fig. 1The evidence leans for, about 3 to one by weight, and the newest evidence points the same way.

Rests onfacts underneath

Three firms hold leading-edge DRAM

What rests on thismarkets it underwrites or threatens

Evidencethe schedule the figure draws

The evidence, newest first

Oct 2026
for · claim Record revenue $8.22B (+22% q/q), non-GAAP GM 52.4%, operating margin 40.2% (first time above 40%); December-quarter guide $8.6B +/- $0.4B at 40.5% operating margin. Lam Research results for the quarter ended 27 Sep 2026 (newsroom release, tier 1) + Q1 FY27 call coverage (Motley Fool transcript / BigGo, tier 5)
Oct 2026
for · observable Q3 revenue NT$1,494.24B / US$46.72B (+50.9% y/y, +17.6% q/q), gross margin 66.4%, operating margin 58.7%, net income NT$763.06B (+68.7% y/y). TSMC Q3 2026 results press release (pr.tsmc.com 3346, 15 Oct 2026) + Q3 call coverage (CRN Asia, xenospectrum, 15-16 Oct)
Oct 2026
against · claim The same event, read the other way and logged separately. SpaceX $40B Nvidia-chip financing — debt-funded, vendor-linked demand (Reuters via BNN Bloomberg, 7 Oct 2026; Nvidia's ~122.8M-share SpaceX stake per Aug 2026 filings coverage)
Oct 2026
for · claim A single non-hyperscaler buyer is arranging about $40B of debt to buy Nvidia hardware exclusively for its datacentres. SpaceX seeks ~US$40B financing (≈$10B bank loans + ≈$30B investment-grade debt; Apollo expected to lead, PIMCO in talks) to buy Nvidia chips, sources say (Reuters via BNN Bloomberg, 7 Oct 2026)
Jul 2026
against · observable Private source
Jul 2026
for · primary Datacentres now take ~70% of all memory chips produced worldwide; DRAM rose ~98% in Q1 2026 and 80-90% in Q2 (Counterpoint); hyperscalers locking supply via long-term agreements; shortage forecast into 2027+. Datacentres consume ~70% of global memory output; DRAM +80-90% in Q2 2026 (IDC / Counterpoint, Thursday Four cost-curve issue verification)
Jun 2026
against · observable Gartner forecasts that by 2030, running inference on a one-trillion-parameter model will cost providers more than 90 percent less than in 2025, indicating major efficiency step-changes that could reduce future capex demand intensity. Gartner forecast on inference cost reductions by 2030
May 2026
for · primary Q1 FY27 (ended 26 Apr 2026) record revenue $81.6B (+85% YoY), record Data Center revenue $75.2B (+92% YoY), next-quarter guide $91.0B. NVIDIA Q1 FY2027 results (8-K)
May 2026
against · observable Bear case: US AI capex ~$500B/yr 2026-27 vs single-digit-billions consumer + ~$25B (OpenAI)/~$19B (Anthropic) enterprise revenue implies ~$8-10 capex per $1 AI revenue; Burry argues ~$176B understated depreciation 2026-28. Hidden financial bubble in AI infrastructure / Burry depreciation (coverage)
May 2026
for · primary Google CEO Sundar Pichai stated on Q1 2026 earnings call: 'We are compute-constrained in the near term. Q1 2026 AI Earnings: 6 Numbers That Confirm the Compute Shortage Is Now the Only Constraint That Matters
Apr 2026
for · observable Microsoft says Azure remains supply-constrained and won't ease until after June 2026; Azure grew ~40% YoY. Microsoft Q1 FY26: Azure supply-constrained through June 2026 (Directions on Microsoft)
Apr 2026
for · primary Google Cloud Q1 2026 revenue >$20B (+63% YoY), backlog doubled to $462B; Pichai: 'we are compute constrained... Google Cloud >$20B, capacity-constrained (TechCrunch)
Mar 2026
for · claim At GTC (16 Mar 2026) Huang put cumulative Blackwell+Vera Rubin visibility/orders at ~$1 trillion through 2027. Nvidia GTC 2026: ~$1T Blackwell+Rubin orders through 2027 (CNBC)
Mar 2026
for · observable Oracle RPO (booked but undelivered backlog) reached $553B at end of Q3 FY26, up 325% YoY, driven by large-scale AI contracts — a forward promise, not delivered demand. Oracle Q3 FY2026 results
Feb 2026
against · claim TD Cowen flagged Microsoft walking from ~2GW of combined developments and leases (revised up from ~200MW in Feb); AWS paused a portion of its leasing discussions — early digestion signals, though AWS called it routine capacity management. Microsoft datacenter freeze + AWS leasing pause (SemiAnalysis / DCD via Wells Fargo, TD Cowen)
Feb 2026
for · primary Aggregate annual AI infrastructure commitment from five largest US cloud companies increased from approximately $380 billion in 2025 to projected $660–690 billion in 2026, with all hyperscalers reporting markets as supply-constrained rather than demand-constrained.
Feb 2026
for · primary Amazon, Alphabet, Microsoft and Meta collectively plan ~$725B of 2026 capex, up ~77% on 2025's ~$410B; guidance raised across the board. Tech AI spending approaches $700B in 2026 (CNBC)

What would move itthe measures watched

watchAggregate hyperscaler capex guidance (MSFT+GOOGL+AMZN+META+ORCL), direction of revisions. Sum the big spenders' guided capex; sustained RAISES = demand outrunning expectations, a CUT or 'optimizing/digesting' = plateau signal
watchNVIDIA Data Center revenue (QoQ + YoY) and forward guidance. Decelerating YoY with a flat guide = early plateau; sequential step-ups and a raised guide = demand still ahead of supply
watchBacklog / RPO / purchase commitments (Google Cloud backlog, Oracle RPO, CoreWeave backlog, NVIDIA supply commitments). Rising multi-year backlog = locked future demand; watch for the first quarter backlog growth stalls
watch'Capacity-constrained' vs 'digestion/optimization/pause' language. Constraint language across all clouds = supply is the binding limit (bullish); 'digesting'/lease cancellations = the bear inflection
watchCapex-to-AI-revenue gap and depreciation/ROI scrutiny. A widening gap with no monetization inflection is the structural bear; rising AI revenue run-rates support the thesis