Understanding the AI Selloff: The Infrastructure vs. Software Divide
Software Faces Disruption. Infrastructure Faces a $500 Billion Boom.
The AI stock market just lost nearly $1 trillion in value over the past week. Software stocks have been hammered, with the S&P 500 software index down over $800 billion since late January. Even chip giants like Nvidia (-9%), Broadcom (-7%), and AMD (-15%) got caught in the downdraft.
But here’s the thing: the fundamentals driving AI infrastructure spending haven’t changed at all.
The Fear: AI Will Replace Software Companies
The selloff was triggered by Anthropic’s release of new legal automation tool, a four-paragraph announcement that somehow sparked what traders are calling the “SaaSpocalypse.” Investors panicked that AI agents will automate away entire software categories, from legal tech to CRM to data analytics.
One analyst noted that while legal tools were disrupted today, “tomorrow it might be sales or marketing or finance.”
That fear is real for software companies. But it’s completely irrelevant, or even bullish, for the infrastructure providers that make AI possible.
The Reality: Hyperscalers Are Doubling Down
While software stocks cratered, something remarkable happened this week: Alphabet just announced it will spend $175-$185 billion on AI infrastructure in 2026, more than double its 2025 spending of $91.4 billion.
This is a company going all-in.
The broader numbers are staggering:
Microsoft: $140 billion in capex projected for 2025
Amazon/AWS: ~$125 billion for 2025, increasing in 2026
Meta: $115-$135 billion for 2026 (nearly double prior year)
Oracle: Up 249% year-over-year in AI infrastructure spending
Add it up and hyperscalers are on track to spend over $500 billion on AI infrastructure in 2026, up 36% from 2025. UBS forecasts global AI capex will hit $423 billion in 2025, $571 billion in 2026, and $1.3 trillion by 2030.
The Disconnect: Software Gets Disrupted, Infrastructure Gets Built
Here’s what the market is missing in its panic: AI can’t disrupt software without massive physical infrastructure.
Every dollar spent on AI agents requires:
Physical GPUs that can’t be automated away (Nvidia, AMD, Broadcom)
Chip fabrication at advanced nodes (TSMC, ASML)
Power distribution systems for energy-hungry data centers (Eaton, Vertiv)
Cooling infrastructure for heat-intensive GPU clusters (Vertiv, nVent)
High-speed networking to connect it all (Arista, Marvell)
Unlike software seats that can be compressed or eliminated by AI, you cannot train Claude or GPT-5 without servers, power, and cooling. The infrastructure layer has zero substitutes.
The Evidence: Demand Signals Remain Explosive
Despite the selloff, the actual business fundamentals for infrastructure plays are accelerating:
Alphabet’s cloud backlog: Hit $240 billion in Q4, up 55% sequentially and more than double year-over-year. That’s committed future revenue.
Vertiv’s backlog: Exited Q3 with $9.5 billion in orders and a 1.4x book-to-bill ratio, providing revenue visibility well into 2026. The stock is up 43% in 2025 after gaining 190% in 2024.
TSMC’s earnings: Crushed expectations with 68% market share in chip foundry, benefiting from AI chip demand described as “insatiable.”
H100 rental pricing: At 8-month highs, signaling continued supply constraints and pricing power despite concerns about GPU obsolescence.
The Market Discrepancy
JPMorgan’s enterprise software analyst called fears that AI plugins will “replace every layer of mission-critical enterprise software” an “illogical leap.” Even Nvidia CEO Jensen Huang said predictions that AI would replace software tools are “illogical.”
However, even if software does face disruption, that dynamic would likely accelerate infrastructure spending. Every new AI capability requires more compute, more memory, more power, more cooling.
The selloff has affected companies with different fundamental profiles:
Strong revenue visibility from hyperscaler commitments
Pricing power from supply constraints
Post-selloff valuations that may not reflect projected growth rates
Companies in the Infrastructure Layer
The companies directly supplying the physical infrastructure for AI buildout fall into several categories:
Core Compute:
Nvidia (NVDA): Trading at 24x forward P/E with 45% projected EPS growth
Taiwan Semiconductor (TSM): 22x P/E with 28% growth; supplies all major chip designers
Broadcom (AVGO): Custom silicon provider for hyperscalers at 25x P/E, 22% growth
Micron (MU): HBM memory supplier at 12x P/E with 297% growth forecast
Infrastructure Support:
Vertiv (VRT): Power and cooling systems with direct NVIDIA partnership
Arista Networks (ANET): Networking infrastructure for AI clusters
Applied Materials (AMAT): Semiconductor manufacturing equipment
Key Observations
Software companies face potential disruption from AI agents. Infrastructure companies face a fundamentally different situation driven by unprecedented capital deployment.
The selloff has treated these two categories similarly despite divergent underlying fundamentals. Software companies face questions about AI-resistant competitive advantages. Infrastructure suppliers have committed demand from hyperscalers planning to spend half a trillion dollars annually.
The market is pricing in fear across both categories. The actual business trends suggest different trajectories for each.