Thoughts on the Market
2026-05-08 · Hosted by Mike Wilson · Morgan Stanley
Executive Summary
Morgan Stanley’s Tom Wigg (Head of Specialty Sales, Americas, TMT) interviewed Stephen Byrd (Global Head of Thematic and Sustainability Research) on what Byrd calls a non-linear improvement in AI capabilities. Byrd argued the market is mispricing AI as linear when frontier-model capabilities follow an exponential scaling law (10x training compute = 2x capability). His core thesis: token economics for hyperscalers are excellent, compute is severely scarce, and pricing power is widespread. Byrd identified 2026 as the revenue-inflection year for hyperscalers. He flagged that token usage has exploded (a couple hundred percent in two months per OpenRadar), driven by agentic workflows that consume ~10x more tokens than query-based use.
Key Stories & Changes
1. Non-Linear AI Capability Improvement
Scaling law: 10x training compute → 2x capability
Frontier models in spring/summer expected to perform “much greater percentage of the economy at better levels of accuracy at incredibly low cost”
“Things are about to get weird” — Byrd
2. Token Economics Are Excellent
Morgan Stanley’s token economics model views from hyperscaler/LLM perspective
Returns on fully-loaded data center costs (including financing) are “excellent”
Pricing power favoring those with compute and power
3. Adopter ROI Mathematics
Average enterprise LLM use: replaces ~1.5 hours of human work = ~$55 saved
Cost: $5 per million tokens (input/output)
Typical agentic enterprise task: tens of thousands of tokens today (well below 1M)
Net result: “economics are a home run for adopters”
4. Token Usage Explosion
OpenRadar weekly token usage up a couple hundred percent in 2 months
Power industry contacts saw “sense of urgency” jump in AI infrastructure demand
Agentic AI workflow uses ~10x more tokens than query-based usage per occupation
Knowledge-based jobs moving to agentic = compute demand explosion
5. Bullishness Trajectory
Byrd: “The bullishness is going to get more bullish over the coming months”
2026 = revenue inflection year for hyperscalers
Fast adopters showing massive economic benefit (anecdote: friend’s software company stopped writing code after PhDs finally adopted AI)
Average enterprise lagging fast adopters
Trends Identified
1. Compute as Strategic Bottleneck
The market understands compute is scarce, but Byrd argues it isn’t pricing in just how scarce. Hyperscalers control pricing because demand-side adopter economics work at any reasonable token cost — meaning hyperscalers can extract value without elastic adopter pushback. The investment implication: every node in the AI value chain (memory, optical, semi-cap) sees volume increases.
2. Two-Speed Adoption Curve
The “fast adopters” and “average enterprise” gap creates a wave structure. Fast adopters demonstrate economics; average enterprise lags but eventually follows. Tom Wigg’s question (when do hyperscalers see revenue inflection?) Byrd answered as 2026, suggesting the second wave hits this year. This frames the hyperscaler stock weakness YTD as misaligned with imminent revenue catalysts.
3. Agentic Workflow as Token-Demand Multiplier
The agentic shift is the most underappreciated multiplier in the AI build-out. Coding workflows are token-intensive, yes — but Byrd argues every knowledge-work occupation transitioning to agentic will see ~10x usage growth. This supports unlimited compute demand across non-coding sectors over time. —-
Sentiment Analysis
Overall Market Sentiment: Very Bullish on AI Infrastructure
Byrd’s framing is unambiguously positive on AI infrastructure plays; hyperscaler underperformance YTD viewed as misaligned with fundamentals.
Risk Factors Highlighted
Investor disruption misread: Selling everything is wrong; some businesses are immune, some enabled, some disrupted
Coding-as-token-intensive bias: Bear case that agentic non-coding work won’t drive same token demand
Hyperscaler free cash flow pressure: Implicit in Tom Wigg’s question; stocks pressured by investment burden
Power, memory, compute constraints: Triple bottleneck cited multiple times by hyperscaler/lab CEOs
Adopter lag: Average corporate user not keeping up with capability improvements
Agentic workflow cost surprise: Speaker’s friends “set agents loose overnight” with $5K credit card bills illustrates parameter-setting risk for adopters
Capability uncertainty: Whether new frontier models meet expectations through 2026 spring/summer
This episode was covered in today’s The Market Signal — 2026-05-08, a cross-source synthesis of multiple podcast reports.