AI isn't a bubble. It might be hitting an air pocket.
Two of Wall Street's biggest houses are aligned: this AI cycle is driven by real spend and real earnings, not mania. The debate isn't "bubble or not." It's how to price the time gap between massive investment and monetization.
That reset matters. For investors, returns may depend less on narratives and more on who converts capex into cash flow fastest.
Micro is macro: capex at a historic scale
BlackRock sees AI spending plans between $5 trillion and $8 trillion globally through 2030, largely in the US. The firm argues the build-out is big enough to push US growth above the long-run ~2% trend.
The catch: can revenues scale to match that ambition? The firm's read is clear-front-loaded spend is required, but revenues will lag deployments. AI remains the core driver of US equities, yet the payoff cadence will be uneven.
Constraints are physical, not just financial. Compute availability and the grid are the bottlenecks. By decade's end, AI data centers could consume 15%-20% of US electricity, making siting, interconnects, and equipment lead times a central risk factor. For background on load growth and data center demand, see the International Energy Agency's brief on data centres and networks here.
From bubble talk to an "air pocket" risk
Bank of America agrees this isn't 2000. But they expect leadership to wobble as capital outlays outpace near-term revenue-an "air pocket," not a collapse.
Hyperscaler capital expenditures have climbed to 60% of operating cash flow (from ~30% a decade ago), still well below the ~140% peak during the dot-com era. BofA expects hyperscaler spend to reach $400B in 2025 and $510B in 2026, as compute, networking, and data center build-outs continue.
Why this cycle is different from 2000
- Valuations are supported by earnings growth, not just multiple expansion.
- Public equity allocations sit well below late-90s extremes.
- IPO sizes are smaller; speculation in unprofitable names is less severe.
- Balance sheets are stronger; capex is high but internally financed at healthier levels.
Your dashboard: what to watch
- Capex-to-sales and OCF coverage at hyperscalers and AI-first software vendors.
- AI service pricing (tokens, seats, inference rates) vs. unit costs (GPUs, networking, energy, cooling).
- Grid and infrastructure: interconnection queues, transformer and switchgear lead times, substation upgrades. Track US interconnection data via LBNL's queue tracker here.
- Supply signals: GPU availability, memory pricing, and delivery backlogs.
- Utilization: data center occupancy, throughput, and model usage KPIs that map to revenue.
Positioning: pro-risk, but selective
- Own the infrastructure chain: semiconductors (compute, memory, networking), AI-focused equipment vendors, and grid-enabling manufacturers.
- Utilities with load growth and credible capacity plans, plus equipment suppliers tied to transmission and distribution upgrades.
- Quality growth with clear AI monetization-contracts, pricing power, and expanding margins-not just "AI exposure."
- Trim names with stretched multiples and unclear revenue timing; the air pocket can compress valuations even if the thesis is intact.
- Mind rates and duration: long-duration growth is sensitive; barbell with cash-generative cyclicals linked to the build-out.
Timeline and expectations
Base case: a soft patch as spend runs ahead of revenue in 2025, then a stronger revenue ramp in 2026 as deployments come online and pricing stabilizes. Market leadership likely broadens even if the AI theme stays central.
On the index: one major house pegs the S&P 500 at 7,100 by end-2026, while peers sit higher (e.g., 7,750 and 8,000). The spread reflects one question: how fast does investment convert to earnings?
Practical moves for CFOs, PMs, and corp dev
- Stage investments: gate AI infra and model spend by ROI proofs (utilization, payback, incremental gross margin) each quarter.
- Procurement: secure long-lead equipment and electricity early; location strategy should prioritize interconnect timelines and cooling efficiency.
- Disclosure: report clean AI KPIs-revenue per compute unit, per enterprise seat, or per workflow-to reduce the valuation gap during the air pocket.
- Treasury: align duration with capex windows; consider hedges tied to electricity costs where feasible.
- M&A/partnerships: co-develop with utilities, chip suppliers, and data center operators to derisk capacity and delivery dates.
Bottom line
This isn't a speculative frenzy. It's a capital cycle with real constraints and a messy middle. The edge goes to investors who track capacity, pricing, and conversion to cash-then size positions to survive the air pocket and participate in the ramp.
If you're building an AI roadmap on the finance side, here's a curated index of tools and courses that can speed up evaluation and implementation: AI tools for finance.
Your membership also unlocks:
