Bottleneck Tracker

Which AI bottleneck is binding right now

Six layers. One framework. Free for the surface, paid for the signal archive.

3 critical
3 constrained
last updated
next update

Power (generation)

Critical

Existing dispatchable thermal cannot be replicated on a 12-month timeline. The bet is on a regime change in how power demand is priced through the late 2020s.

What changed: Behind-the-meter co-location accelerating: MSFT-Three Mile Island restart, GOOGL $4.75B Intersect Power, MSFT $10B/10GW Brookfield. 46 BTM facilities now in flight, 56 GW capacity, 90% announced in last 12 months.

CEGVSTNRGBWXTPOWL

Power thesis remains strongest in the vault by falsification audit. No falsifiers firing. POWL margins expanded 19% → 30%. Siemens Energy sold out through 2030.

Full signal history & alerts for this layer → paid tier (launching)

Grid (transformer + interconnect)

Critical

High-capacity transformer lead times at 200+ weeks. FERC interconnect queue ~2 TW (5-8 year wait). NERC Level 3 "Essential Actions" alert May 4, 2026 — first ever for computational loads.

What changed: 71% Gallup opposition to new data centres in their area (May 13, 2026). $64B in DC projects blocked or delayed across 24+ states. The DC moratorium narrative was corrected May 16: no state has enacted a broad ban into law — the real constraint is transformer/grid, not politics.

POWLETNGWHEOSEFLNC

High-capacity transformer lead times now 200+ weeks (4+ years). Materially exceeds prior thesis estimate of 80-128 weeks. New US manufacturing capacity not expected until 2027 at earliest.

Full signal history & alerts for this layer → paid tier (launching)

Memory (HBM + DRAM + NAND)

Critical

Memory super-cycle absorbing fab capacity at advanced nodes. AI demand creating contagion into non-AI sectors (automotive ECUs, PLCs, switches, medical devices).

What changed: Kioxia entire 2026 NAND output sold out. DRAM prices forecast +58-75% Q2 (TrendForce). DDR5 relief not until 2028-2030. 2D NAND planar flash prices spiked 2-3x.

MUNVDATSM

Memory pricing crisis confirmed accelerating. The Samsung strike risk added a compound supply shock; partially reduced May 18 by court injunction. Helium shortage from Qatar Ras Laffan strikes (March 2026) destroyed ~33% global helium supply — affects EUV cooling.

Full signal history & alerts for this layer → paid tier (launching)

Compute (GPU)

Constrained

NVDA still captures 79% of AI profits after two years of explosive growth. The "compute commoditises soon" thesis is two years past its sell-by.

What changed: Q1 FY2027 consensus revised to $78.5-78.8B (Citi at $80B). Beat-is-priced-in dynamics elevated. Hyperscaler capex run rate now ~$805B, up from $765B.

NVDAAMDAVGO

Altimeter Capital: AI ecosystem grew 5x in 2 years, semis still claim 79% of profits (down from 87%). At this pace, apps need 15+ years to reach cloud-level economics.

Full signal history & alerts for this layer → paid tier (launching)

Optical interconnect

Constrained

InP epiwafer foundry capacity is the gating constraint. 12-24 month qualification cycles. NVDA $2B equity into COHR + multi-year CPO deal through 2030.

What changed: LITE Nasdaq-100 inclusion executed May 18; sell-the-news pattern in effect. LightCounting May 2026: high-speed transceivers >60M units shipped 2025, approaching 100M. LPO/CPO market doubling $5B → $10B+ in 2 years.

LITECOHRAVGOIQEALMU

IQE -26.5% intraday on UK Huawei ban — selloff mispriced the Wireless segment risk and ignored the Photonics segment entirely (42% of revenue, growing, AI-driven).

Full signal history & alerts for this layer → paid tier (launching)

Cooling

Constrained

GB200 NVL72 at 120-140 kW/rack creates thermal density that air cooling cannot serve. Liquid cooling capture accelerating across hyperscale.

What changed: $14.25B in liquid-cooling M&A in 14 days in May 2026. Vertiv (VRT) the publicly-traded primary expression. Thermal throttling at 95-100°C cited in GB200 production briefs.

VRTENS

GB200 thermal/assembly difficulties are a Law II signal — the difficulty is what creates the moat for cooling specialists. 5,000+ NVLink cables per rack causing assembly errors. KeyBanc maintains cautious Q1 <1K, Q2 ~3K rack ramp.

Full signal history & alerts for this layer → paid tier (launching)

Migration signals — last 7 days

  1. ⇄ cross-layer Compute

    NVDA Q1 FY2027 earnings tonight (after market close). Six watch items: Q2 guide vs $84B threshold, DC revenue QoQ trajectory, China revenue disclosure, GB200 ramp, optical supply chain language, enterprise GPU utilisation commentary.

  2. ⇄ cross-layer Optical

    DDOG backs Dust $40M Series B (multiplayer AI agent platform). DDOG extending observability into AI-agent workflow layer. Verification-thesis-confirming; cross-cuts to optical-demand narrative through AI infrastructure breadth.

  3. ▼ easing Memory

    🔒 Samsung court decision partially accepted injunction against striking unions. Strike risk …

  4. ▲ tightening Compute

    🔒 Enterprise GPU utilisation reported at 5% across enterprise K8s clusters (Cast AI). If gen…

  5. ▲ tightening Grid

    🔒 DC moratorium narrative corrected: no state has enacted broad DC ban into law (Maine vetoe…

  6. ▲ tightening Memory

    🔒 Kioxia entire 2026 NAND output sold out. DRAM +58-75% Q2 forecast. DDR5 relief not until 2…

The full signal archive — 60+ dated, sourced, layer-tagged signals — is the paid tier. Get notified when it ships

Tiers (paid launching soon)

Bloomberg Terminal is £24,000/year. Tegus institutional is £50,000+. Stockopedia is £400. The Tracker is £490/year for Operator — same structural framework, different access tier. Built by one person, priced for operators, not enterprises.

Free

Browse the framework

£0

Forever. No card.

View free tracker
  • Current state across 6 layers
  • Last 7 days of signals
  • Ticker exposure tags
  • Weekly Tuesday updates

Team

For investment teams & funds

£165/mo

£1,990/yr — save £398

Contact for Team
  • Everything in Operator
  • Unlimited seats per firm
  • Shared watchlists
  • Priority signals (24h early)
  • Weekly office hours
  • Quarterly thesis review

Pricing in GBP. Stripe handles FX automatically. 30-day full refund, no questions. After 30 days, cancel any time — access stays until end of billing cycle.

From the analyst

I built this because I couldn't find what I needed. AI infrastructure analysis is either institutional (Bloomberg, Tegus, £20k+) or opinion (FinTwit, Substack, mostly noise). The Tracker is the middle: which layer of the AI stack is binding right now, dated and sourced, with the falsifiers that would prove me wrong.

Built on the same five-law framework I write about in The Durability Curve. Maintained by one person — me. Which is why founding-member pricing is locked for the first 100 subscribers, and why the analyst-hours tier on Team isn't a marketing line — there's literally one analyst.

— Harry Floyd

Common questions

When does the paid tier actually launch?

When 50+ subscribers join the launch list AND the signal-history + alerts features ship. Realistic window: 2-6 weeks from now. Founding-member pricing is locked for everyone on the list at launch — that's the promise.

What happens when I sign up to the launch list right now?

You join the free Substack. When paid launches, you get an email with a private link and founding-member pricing locked in. No surprise charges, no automatic enrolment — you choose to upgrade.

Is this investment advice?

No. This is research framing. Every signal is dated, sourced, and falsifiable. Each layer carries explicit kill criteria. Your positions and decisions are yours. The Tracker tells you which layer is binding, not what to buy.

How is this different from the free Substack?

The Substack is the long-form argument — essays explaining why a thing is true. The Tracker is the live dashboard — current state, dated signals, alerts when things move. Different shapes, same framework. Most subscribers will want both.

How accurate is the underlying data?

Signals are sourced from primary disclosures (earnings, SEC filings, regulator publications), cross-confirmed where possible. Ghost-agent-sourced signals carry a documented 77% strict / 91% directional accuracy baseline — every secondary-source figure gets flagged for verification before high-stakes use.

Why GBP not USD?

I'm UK-based. Stripe handles FX automatically on launch. £49/mo is roughly $62 at current rates. Annual: £490 ≈ $620.

What if I cancel?

Cancel any time. First 30 days you get a full refund, no questions asked. After 30 days, cancellation keeps your access until the end of the billing cycle. No auto-renew traps.

The framework

Value migrates when lower layers commoditise. The bottleneck never disappears — it moves. Right now the binding constraints are stacking: grid and memory are critical, power and optics are tight, cooling is tightening, compute is still rent-extracting longer than anyone predicted.

This tracker is the live state of that migration. It's the same vault-laws framework used in The Durability Curve, applied to one question: which layer pays the rent this week?

This is research, not investment advice. Every signal is dated, sourced, and falsifiable. Treat verified Ghost-sourced data with 77% strict / 91% directional accuracy until independently confirmed.