Where AI Actually Goes Next: What It CAN Do ... vs What People Are ALLOWING
AI's software is still getting cheaper, faster, smarter. AI's physical infrastructure is hitting walls.
$156B
in US datacenter projects blocked or delayed in 2025 (Data Center Watch)
Stargate Abilene
600 MW expansion cancelled — first major hyperscaler AI project killed
Apr 27, 2026
OpenAI now multi-cloud, not entirely Microsoft-dependent
>1M chips
Anthropic training frontier on Trainium + Ironwood — CUDA lock-in broke
The Two Curves Are Diverging
Read across 28 industries at once, the 2026 AI economy splits cleanly in two. Software capability accelerates. Physical capacity bottlenecks. Both are true. Both compound. The resolution determines the next investment cycle.
Capability: Accelerating
Software wins keep arriving, faster than markets know how to price
DeepSeek visual processing just shipped with 90% fewer tokens per task (May 2026)
Model costs dropped ~99% in one year (ARK Big Ideas 2026; inference costs)
Anthropic frontier training running on >1M Ironwood TPUs + >1M Trainium2 at GigaWatt scale
Walmart Sparky grows shopper baskets +35% among users (~50% of app users have tried it)
AIG Lexington underwriting: 55% faster, 40% higher bind rate — first incumbent disclosing AI productivity at line-of-business granularity
Hugo Boss generative design: 85% design-time reduction, 30%+ sample reduction at scale
AI traffic conversion flipped 80 points in 12 months: from 38% worse than human shoppers (March 2025) to 42% better (March 2026; Adobe)
Google Cloud margin doubled 17.8% to 32.9% in one year — owning silicon collapses margin compression
Capacity: Bottlenecking
Physical and political walls now bind harder than chips or models
Stargate Abilene 600 MW expansion cancelled — the first major hyperscaler AI project killed
Prince William Digital Gateway abandoned April 2026 — would have been the largest datacenter corridor in the world ($24.7B)
$156B in US datacenter projects blocked or delayed in 2025; 188+ active opposition groups; 300+ state bills in 30+ states in the first six weeks of 2026 alone
Ohio data center requests collapsed 30 GW to 13 GW after PUCO 85% minimum-demand rule — minus 57% in one regulated market
Copper deficit math: ICSG flipped 2026 from +200K tonnes surplus to -150K tonnes deficit — a 350K-tonne swing in one year
TSMC refused High-NA EUV before 2029 ("too expensive!") — first time TSMC isn't first to a major lithography step
CoWoS advanced packaging tighter than wafer starts — single global AI supply-chain bottleneck
Large power transformers: 2-4 year lead times from only 3 US manufacturers; PJM interconnection queue: 7+ years
The Five Bottleneck Families
Every industry's AI ceiling traces back to one of five constraint families. The names of the constraints are physical, even when the consequences are financial.
Power
Transformers 2–4 year lead time
PJM interconnect 7+ years
Ohio −57% DC demand
Hyperscaler off-grid stuck on siting
Gas turbines backlogged through 2028
Materials
Copper +350K tonne deficit swing
SK Hynix HBM sold out through 2026
HBM4 forward-contracted exclusive
DC needs up to 50K tonnes Cu each
Industrial-gas chip-fab tightness
Packaging / Silicon
CoWoS tighter than wafers
TSMC refused High-NA EUV
Custom silicon cracking CUDA
Anthropic on Trainium plus Ironwood
AWS won largest training contract ever
Regulatory
Maryland HB 895 grocery pricing ban
WA bans AI insurance denials
VA take-or-pay datacenter rate class
CMS WISeR prior-auth pilot
NAIC AI tool live in 12 states
Mobley v. Workday class certified
300+ state bills filed in 6 weeks
$156B datacenter projects blocked
Capital
Private-credit redemption crisis
$265B PE-firm market cap erased
SaaSpocalypse: $285B in 48 hours
BCRED redemptions hit 7.9%
Stargate Abilene cancelled
First AI-driven financial-stress event
Five binding constraint families. Each one re-prices a specific group of tickers — see the cascade map below.
Q1-Q2 2026: Inflection Events
Six months that flipped the AI narrative. Every event below is sourced; together they make "capability fast, capacity slow" a measurable claim rather than a thesis.
Q1 2026 — January through March
Jan 1CMS Cat I CPT 75577 live — first explicit AI-software CMS code
Feb 3SaaSpocalypse — $285B wiped in 48 hours (~$2T Jan–Feb)
Feb 4Epic AI Charting launches — threat to Abridge / Ambience / Suki
Mar 2NAIC AI tool live in 12 states — from guidance to live exam
Mar 14Private-credit redemption crisis named — Fortune flags AI fear
Mar 18CA Kaiser nurses strike — explicitly an AI-deployment grievance
Mar 29Lilly + Insilico $2.75B deal — soft acquisition of an AI platform
Q2 2026 — April through May
Apr 20Anthropic $100B / 5GW Trainium — largest AI training contract ever
Apr 26Sora shuttered — Disney walked from $1B deal
Apr 27OpenAI–Microsoft exclusivity ENDED — OpenAI now multi-cloud
Apr 30Bot Auto first no-observer commercial freight — Houston to Dallas
May 1Crown Castle fiber sold $8.4B — CCI out of the AI-RE narrative
May 7Roche acquires PathAI $1.05B — diagnostic-AI M&A wave starts
May 12CME launches compute futures — compute commodified on a major exchange
May 13Gallup: 71% oppose local AI DC — exceeds nuclear opposition
May 19Google Genie + Street View — Maps Imagery Grounding launches
May 21Retatrutide 30.3% weight loss Phase 3 — bariatric-equivalent pill
May 22This map published — capability-vs-capacity bifurcation public
Six months of inflections. Stargate Abilene cancellation, dates by spring 2026, sat against equal-sized capability wins on the software side.
Bottleneck to Industries to Tickers
The cascade is mechanical. A power constraint translates to specific utility ticker repricings. A copper constraint translates to specific materials names. The map below traces the dominant chain from each bottleneck family to its first-order beneficiaries and victims.
Power
transformers / grid / gas
Utilities — CEG, NEE, VST, DUK Beneficiary
Capital Goods — ETN, GEV, HUBB, VRT Beneficiary
Real Estate / Data Centers — EQIX, DLR Beneficiary
Beneficiary tags benefit from the bottleneck. Headwind tags pay for it. Ambiguous = tail risk if a circular-financing knot tightens.
The Bifurcation Map
The engine's 28-industry calibration sorts into clear winners and losers when you read it through the capability-vs-capacity lens. Multipliers below are mid-term engine projections (12-36 months) reflecting both AI tailwinds and the constraint layer above.
Structural Winners
Mid-term multipliers 1.5x to 5x
Cloud Platforms2.5-4.0x
Vertically integrated AI stacks; Google Cloud margin doubled. MSFT / GOOGL / AMZN.
Chip Design & Fabless3.0-5.0x
Demand inelastic for two more cycles even as CUDA moat cracks. NVDA + custom silicon (AVGO, MRVL).
Foundry & Equipment3.5-5.5x
CoWoS is the binding global constraint. TSM, ASML, AMAT, KLAC, LRCX.
Capital Goods (DC electrical)1.5-2.0x
Eaton's 228 GW pipeline = 12 years of forward demand. ETN, GEV, HUBB, VRT.
Amazon walled-garden play vs UCP coalition. Mid-tier department store decline accelerating.
What This Means For Your Portfolio
The clean line of "AI inevitable triumph" was a low-resolution map. The high-resolution version is that capability is running faster than capacity, and the gap is widening for at least the next 18-36 months. Both curves are real. Neither curve cleanly invalidates the other.
Three things follow from holding both curves in your head at once:
Picks-and-shovels beats software incumbents through 2027. The capacity layer (foundry, advanced packaging, datacenter electrical, copper, transformers, nuclear, gas peaker, datacenter REITs with sited capacity) re-rates first because it has measurable backlog. Horizontal enterprise SaaS without a data moat compresses first because its pricing model is failing.
The middle of every consumer category dies. Retail, apparel, CPG, restaurants, media — in every consumer vertical, AI personalization plus agentic commerce splits outcomes into a barbell. The premium end with data moats wins; the low-cost end with AI-native cost structures wins; the middle gets cannibalized from both sides.
Watch the regulatory wave by state count, not federal headlines. Maryland HB 895 is the template. NAIC's 12-state evaluation tool is live. Ohio's PUCO 85% minimum-demand rule produced a -57% datacenter demand collapse in one regulated market. If 3-5 more states replicate any of these patterns, every load-growth and AI-deployment forecast needs a 30-50% downward revision — with very specific ticker consequences.
See How The Capability-vs-Deployment Gap Affects Your Holdings
I'm Scott Covert — an independently curious person and the person who built everything here, including the 28-industry cross-effect engine — the “AI Revolution Cascade Matrix”. I'm not a fund, a broker, or a newsletter reselling someone else's research. I built the systems that take my ideas and sources and turn them into opinion pieces with machine-verified reasoning and sources, all shown so you can argue with me (I am, after all, trying to predict the future of the stock market, through a series of continual deep research loops into everything affecting stocks).
My edge is pattern recognition across fields (an involuntary feature of ADHD), not a Wall Street pedigree. Everything here is directional synthesis meant to help you think, not financial advice. (If you're a publication or fund and want to license or collaborate, that lives over here.)