The marginal dollar at every layer earns less. Except the network.

Key takeaways

• The marginal dollar at every layer of the AI stack earns less than the prior dollar — except the network and interconnect layer, the only place where pricing power compounds rather than compresses.
• Open weights crossed the closed frontier on coding: DeepSeek V4 (MIT-licensed, frontier-class) plus Kimi K2.6 and GLM-5.1 leading SWE-Bench Pro move the on-prem question from 'can we?' to 'which workload first?'
• Anthropic committed ~8.5+ GW of forward compute across NVIDIA, Trainium, and TPU — the largest single-counterparty compute commitment ever recorded, deliberately spread across three platforms.
• Hyperscalers are deploying ~$700B of 2026 capex against ~$120B of AI-attributable revenue — a ratio that gets tested when Microsoft, Google, Meta, and Amazon all print Q1 between April 29 and May 6.
• The binding constraint is power, packaging, and HBM4, not GPUs: PJM capacity cleared at 11x the prior auction price and time-to-power in the busiest US markets stretched to 36-48 months.
• Watch the Q1 hyperscaler earnings window (Apr 29 - May 6): higher capex pushes the capex / AI-revenue ratio past the 6.0 threshold; lower is the first sign of ROI re-rating risk.

By the numbers

Aggregate 2026 hyperscaler capex against ~$120B of AI-attributable revenue

~$700B — Tested when all four hyperscalers print Q1 between April 29 and May 6

Anthropic's forward compute commitment across NVIDIA, Trainium, and TPU

8.5+ GW — The largest single-counterparty compute commitment ever recorded

CoreWeave revenue backlog, up 4x YoY

$66.8B — Includes the new Meta $21B six-year commitment through 2032

PJM 2026/27 capacity auction $/MW-day — 11x the prior auction's $29

$329 — Power is the new binding constraint, not chips, capital, or land

Custom silicon share of incremental AI compute, up from ~22% in three months

~30% — The threshold where merchant-GPU pricing power starts to compress

Kimi K2.6 on SWE-Bench Pro vs GPT-5.4 at 57.7 — first open lead over closed

58.6 — The on-prem demand catalyst for the open-weights hypothesis

Flywheel arc · all-three

The durable winner of the buildout is the network and interconnect layer — the only layer where pricing power compounds rather than compresses.

• Capital is still flooding the AI stack, but the marginal dollar earns less than the prior dollar at every layer except one: the network and interconnect layer connecting cloud, neocloud, and on-prem.
• Hyperscalers are deploying ~$700B of 2026 capex against ~$120B of AI-attributable revenue — a ratio that gets tested when Microsoft, Google, Meta, and Amazon all print Q1 between April 29 and May 6.
• Frontier labs are committing forward compute at multiples of their disclosed revenue.
• The on-prem floor moved up sharply with DeepSeek V4 (open weights, frontier-class on coding): the question is no longer whether to self-host frontier capability, but which workload to move and on what fabric.

What this means

Open weights crossing the closed frontier on coding moves the on-prem question from 'can we?' to 'which workload first?' Procurement should pull DeepSeek V4 and the leading open coding models into pilot this quarter; merchant-vs-self-host is now a real fork, not a hypothetical. Anthropic withholding Mythos on cyber-risk is the new diligence signal — vendor-risk frameworks need a capability-gate criterion, not just an availability SLA.

• Open weights crossing the closed frontier on coding moves the on-prem question from 'can we?' to 'which workload first?' — merchant-vs-self-host is now a real fork, not a hypothetical.
• Pull DeepSeek V4 and the leading open coding models into pilot this quarter.
• Anthropic withholding Mythos on cyber-risk is the new diligence signal: vendor-risk frameworks need a capability-gate criterion, not just an availability SLA.

Apr 22-24

DeepSeek V4 released — MIT-licensed, 1.6T MoE, 1M context, frontier-class on coding

Sources HuggingFace, deepseek.com

Apr 23

Open-weight models lead closed on SWE-Bench Pro — Kimi K2.6 58.6 and GLM-5.1 58.4 vs GPT-5.4 57.7 and Opus 4.6 57.3

Sources Artificial Analysis, HuggingFace model cards

Apr 21

Anthropic withholds 'Mythos' flagship on cyber-capability grounds; UK AISI confirms autonomous offensive capability

Sources anthropic.com, red.anthropic.com, aisi.gov.uk

What this means

Vera Rubin samples landing 12 months after Blackwell Ultra means refresh planning needs to shift from a 24-month to a 14-16-month cycle; standard depreciation schedules are now wrong for AI silicon. The binding constraint is power, packaging, and HBM4, not GPUs themselves — buyers should secure HBM allocation and grid interconnect alongside chips, not after. Custom silicon at ~30% of incremental compute is the threshold where merchant-GPU pricing power starts to compress; expect single-vendor leverage to fade through 2026.

• Vera Rubin samples landing 12 months after Blackwell Ultra shift refresh planning from a 24-month to a 14-16-month cycle — standard depreciation schedules are now wrong for AI silicon.
• The binding constraint is power, packaging, and HBM4, not GPUs — secure HBM allocation and grid interconnect alongside chips, not after.
• Custom silicon at ~30% of incremental compute is the threshold where merchant-GPU pricing power starts to compress; expect single-vendor leverage to fade through 2026.

Apr 24

NVIDIA Vera Rubin engineering samples shipping to customers within 12 months of Blackwell Ultra GA — half the prior generation gap

Sources NVIDIA developer channels, multiple OEM disclosures

Apr 22

Custom silicon share of incremental AI compute reaches ~30%, up from ~22% three months ago

Sources TrendForce, SemiAnalysis

Apr 18

HBM4 validation milestones at SK Hynix, Samsung, and Micron — primary risk for H2 2026 hyperscaler tail

Sources TrendForce, vendor advisories

What this means

Equinix Fabric Intelligence formalizes AI-native interconnect as its own product category — investors should track cross-connect and fabric revenue as a leading indicator that decouples from raw colo capacity. NVLink Fusion plus UALink means architects can stop betting on a single vendor's scale-up fabric; multi-vendor, mixed-silicon clusters are now a real design target. With 1.6 Tbps optics shipping ahead of the compute they serve, the network is no longer the bottleneck for hybrid AI — designs targeting multi-DC training in 2027 should plan around fabric headroom, not fabric scarcity.

• AI-native interconnect is now its own product category — track cross-connect and fabric revenue as a leading indicator that decouples from raw colo capacity.
• NVLink Fusion plus UALink makes multi-vendor, mixed-silicon clusters a real design target — architects can stop betting on a single vendor's scale-up fabric.
• With 1.6 Tbps optics shipping ahead of the compute they serve, designs targeting multi-DC training in 2027 should plan around fabric headroom, not fabric scarcity.

Apr 15

Equinix Fabric Intelligence launch creates AI-native networking as a distinct product category

Sources Equinix press, Light Reading

Apr 17

NVLink Fusion + UALink standardization moves cross-vendor scale-up fabric toward an interoperable baseline

Sources NVIDIA, AMD, Intel joint statements; OCP working group

Apr 19

1.6 Tbps and co-packaged optics shipping in volume ahead of the compute they serve — Gilder's Law in motion

Sources Marvell, Broadcom press; OIF working notes

Frontier Labs detail

Frontier labs are the largest single source of forward compute commitments this week. Anthropic alone disclosed 8.5+ GW across NVIDIA, Trainium, and TPU — distributed across three platforms specifically to reduce single-vendor dependency. The other labs are following: OpenAI's Stargate restructure, Google DeepMind's TPU absorption, and xAI's Memphis expansion. The capital is committed faster than disclosed revenue can absorb it; that gap is the entire reckoning.

Capital in value

$50B

Revenue out value

$30B

• Apr 7 · Anthropic-Google / Broadcom 3.5 GW TPU deal · $100B / 10 yr
• Apr 20 · Anthropic-AWS expansion to 5 GW · undisclosed
• Apr 22 · Stargate Abilene capacity restructure passed back to Microsoft · $21B / 1.6 GW

Sources TechCrunch, Broadcom 8-K · anthropic.com, AWS news · DCD, GlobeNewswire

Hyperscaler-Hosted detail

Hyperscaler capex deployment this week is dominated by Microsoft picking up Stargate's orphaned 1.6 GW capacity. Azure's deceleration on guidance is a wedge against the broader thesis but doesn't yet contradict the aggregate ~$700B 2026 capex trajectory. Q1 prints (Apr 29 - May 6) test whether the platform-pull-through narrative holds under disclosed numbers.

Capital in value

$48B

Revenue out value

$10B

• Mar 6 / Mar 27 · Microsoft absorbs cancelled Stargate Abilene capacity · $21B / 700 MW + 900 MW
• Apr 25 · Azure guidance commentary heading into Q1 print · n/a

Sources DCD, GlobeNewswire · Microsoft IR commentary

Neoclouds detail

Neocloud backlog is up 4x year-over-year, but conversion velocity is the unresolved question. CoreWeave's $66.8B total backlog now includes the new Meta $21B six-year commitment through 2032; Crusoe-MS expansion adds another lever; Nebius and Applied Digital are tracking similar scaling. The capital-in / revenue-out spread is the widest in the stack and depends entirely on customers ramping into committed capacity.

Capital in value

$15B

Revenue out value

$1.6B

• Q4 2025 · CoreWeave revenue backlog (cumulative) · $66.8B
• Apr 20 · Meta-CoreWeave six-year through 2032 · $21B
• Apr 22 · Crusoe-MS expansion announcement · undisclosed

Sources CoreWeave Q4 25 press · CoreWeave press · Crusoe press

On-Prem / Hybrid detail

Indirect spend through GPU shipments, sovereign programs, and enterprise on-prem deployments. DeepSeek V4 crossing the closed frontier on coding is the demand catalyst this week — sovereign programs aggregate to $80B+ of forward commitments across UAE, Germany, France, Mistral-Sweden, GMI Japan, IndiaAI, and the UK AI Growth Zones. The category has no consolidated revenue line; spend shows up in NVIDIA's data-center revenue, OEM hardware shipments, and government program disbursements.

Capital in value

$42B

Revenue out value

Indirect

• Apr 22-24 · DeepSeek V4 release (open weights, frontier-class on coding) · n/a (catalyst)
• Q1 2026 · Sovereign AI commitments aggregate across 8 countries · $80B+
• Apr 21 · UK AI Growth Zones aggregate · $38.5B

Sources HuggingFace, deepseek.com · various government announcements · UK government

Signal score 5/5

Anthropic has committed to ~8.5+ GW of forward compute capacity across NVIDIA, AWS Trainium, and Google TPU.

Real. The largest single-counterparty compute commitment ever recorded. Distributed across three platforms reduces single-vendor dependency and is critical evidence for the all-three-lenses thesis.

Sources

anthropic.com (Apr 20 + Apr 7), AWS news, Verge, Broadcom 8-K, multiple SEC filings

Signal score 4/5

Aggregate 2026 hyperscaler capex tracking ~$700B (Microsoft + Google + Meta + Amazon + Oracle).

Directionally correct. Will likely revise upward on Q1 2026 prints (Apr 29 - May 6). Caution: the aggregate hides per-company variance — Microsoft is decelerating Azure guidance while Google capex still accelerates.

Sources

Composite of Q4 2025 earnings guides; not a single sourced number

Signal score 2/5

DeepSeek V4 was trained end-to-end on Huawei Ascend silicon, eliminating NVIDIA dependency.

Noise. Significant if true (sovereign-AI hardware decoupling milestone) but unverified. Watch for any DeepSeek primary source. If confirmed, this advances the on-prem-via-non-Western-stack thesis materially.

Sources

Chinese-language tech blogs only. No primary DeepSeek confirmation.

Signal score 4/5

Anthropic withheld 'Mythos' on cyber-risk; UK AISI confirmed autonomous offensive capability.

Real. First major lab to deliberately withhold a flagship on capability-risk grounds. Procurement diligence implications: cyber-capability gating becomes a vendor-risk criterion.

Sources

anthropic.com, red.anthropic.com, aisi.gov.uk, Cloud Security Alliance

Signal score 5/5

Open-weight models now lead closed models on credible coding benchmarks.

Real. SWE-Bench Pro: Kimi K2.6 58.6, GLM-5.1 58.4 vs GPT-5.4 57.7, Opus 4.6 57.3. First time open leads closed on a credible code benchmark. The on-prem demand catalyst for the open-weights hypothesis.

Sources

HuggingFace model cards (Kimi K2.6, GLM-5.1), Artificial Analysis, multiple independent evals

Frontier lab cash position (avg months runway, top 3)

Top 3 frontier labs (OpenAI, Anthropic, Google DeepMind) by disclosed runway, computed from cash on hand divided by trailing-12-month operating burn. The drop from ~26 to ~22 months reflects accelerating compute commitments outpacing revenue growth. <18 months triggers re-rating risk — investor models start pricing dilution or strategic exit; <12 months forces consolidation, acquisition, or revenue reset.

Hyperscaler capex / AI revenue ratio (top 4 weighted)

Top 4 hyperscalers (MSFT, GOOG, META, AMZN) weighted aggregate of total capex divided by AI-attributable revenue. ~5.5 means $5.50 of capex deployed per $1 of AI-attributable revenue. >6.0 is the threshold where sell-side analysts stop accepting the platform-pull-through narrative; sustained 5.5x+ is the wedge against capital-efficiency claims.

CoreWeave revenue backlog

Booked but unrecognized revenue. The 4x YoY jump reflects the new Meta $21B/6yr commitment plus Microsoft and OpenAI extensions. The metric to actually watch is conversion velocity: how quickly capacity comes online and customers ramp. If Meta recapitalizes its in-house compute and pulls back, this number compresses materially.

NVIDIA Q-over-Q data center revenue

Data center segment revenue, sequential Q-over-Q. Captures Blackwell Ultra ramp and the start of Vera Rubin sample shipments. Beat or miss against the $62B run-rate sets the slope for H2 hyperscaler capex; a miss flattens Huang's-Law trajectory and re-rates the whole AI hardware narrative.

Open vs closed gap on SWE-Bench Pro (coding)

Coding benchmark differential between top open-weight model (Kimi K2.6 at 58.6) and top closed model (GPT-5.4 at 57.7). Sustained open lead reshapes enterprise procurement because open weights run on-prem with predictable cost-per-token and no rate limits. The flip from closed +6 to open +0.9 is the inflection that makes self-hosting frontier-class AI a real procurement option this quarter.

Sovereign AI commitments (count / aggregate $)

Count of named national AI infrastructure programs and aggregate forward commitments. Current set: UAE Stargate (1 GW), Germany National DC Strategy, France IA program, Mistral-Sweden, GMI Japan, IndiaAI, UK AI Growth Zones aggregating $38.5B, plus the Saudi-PIF AI commitment. Reflects de-Westernization of compute provisioning and creates new merchant-silicon demand outside the hyperscaler axis.

PJM 2026/27 capacity auction price ($/MW-day)

Most recent PJM auction cleared at 11x the prior auction price, reflecting tightness in the capacity market driven by data-center load growth outpacing new generation build-out. Behind-the-meter generation is now mainstream; FERC compliance milestones in May-June will determine whether 2027 clearing prices repeat or normalize.

Time-to-power, busiest US markets (months)

Months from new-load interconnection request to energization in the most constrained US power markets (Northern Virginia, Phoenix, Dallas-Fort Worth, Columbus). Lengthening despite grid investment because demand is outpacing transmission build-out by 2-3x. The constraint is no longer chips, capital, or land — it is electricity delivery.

Cost-per-task, frontier reasoning model

Falling cost expands workloads (Jevons), not contracts demand

Median cost across the frontier-tier reasoning models for a benchmark complex task. The 38% drop reflects continued pricing pressure from open-weight competition and provider economies of scale. Falling cost expands the addressable workload set rather than contracting demand — this is the Jevons signature in AI inference.

Custom silicon share of incremental AI compute

Approximate share of newly deployed AI compute capacity using custom silicon (TPU, Trainium, MTIA, Maia, Granite Rapids AI) versus merchant silicon (NVIDIA, AMD). The 30% threshold is structurally important: at this share, merchant GPU pricing power begins to compress as customers gain credible alternatives. Past 35% the compression is material and re-rates merchant gross margins.

Capital · 70% confidence

At least one frontier lab announces a customer-funded compute commitment greater than 2 GW.

ID

p1-2gw-customer-funded

Deadline

By June 30, 2026

Trigger

Earnings cycle commentary on RPO from Oracle and AWS; announcements at Google I/O, Microsoft Build, AWS Summit.

Software · 60% confidence

At least one Fortune 500 enterprise discloses an on-prem AI workload greater than $100M annual using open-weight models.

ID

p2-f500-on-prem

Deadline

By September 30, 2026

Trigger

Enterprise architecture announcements; bank, insurer, or pharma F500 first-mover; DeepSeek V4 reference deployment.

Capital · 80% confidence

Aggregate 2026 hyperscaler capex revises upward by 10% or more from the $700B baseline.

ID

p3-capex-revise

Deadline

By October 31, 2026

Trigger

Q1 2026 earnings (Apr 29 - May 6) and Q2 2026 earnings cycle.

Networking · 70% confidence

At least one major colocation or interconnect operator reports cross-connect or interconnect revenue growth outpacing compute capacity revenue growth for two consecutive quarters.

ID

p4-interconnect-outpaces

Deadline

By July 31, 2026

Trigger

Equinix Q1 (May 7) and Q2 (early August) earnings; broader colo and IX operator reporting.

Capital · 35% confidence

At least one neocloud loses an anchor tenant or sees backlog growth turn negative quarter-over-quarter.

ID

p5-neocloud-anchor-loss

Deadline

By September 30, 2026

Trigger

Quarterly disclosures from CoreWeave, Nebius, Applied Digital.

Hardware · 55% confidence

Custom silicon (TPU + Trainium + Maia + MTIA + Granite Rapids AI) reaches 35% of incremental AI compute share, up from ~30% today.

ID

p6-custom-silicon-35

Deadline

By July 31, 2026

Trigger

TrendForce / SemiAnalysis quarterly mix breakdown; Q2 2026 hyperscaler earnings color on internal silicon vs merchant GPU mix.

Apr 29 - May 6

Q1 2026 hyperscaler earnings

Capex direction signal that hits all four hyperscalers in eight days. Watch: aggregate revising higher, AI-segment margin disclosure, RPO commentary. Higher pushes the capex / revenue ratio past 6.0 (a lever threshold); lower is the first sign of ROI re-rating risk and tests prediction p3.

Late May 2026

NVIDIA Q1 FY27 earnings

First disclosure on Vera Rubin ramp velocity, HBM4 supply, and custom-silicon competitive pressure. Watch: data-center revenue vs the $62B run-rate, gross margin trend, customer-concentration commentary. A miss flattens the doubling slope; a beat resets the H2 hyperscaler capex tail.

May 2026

Google I/O + Microsoft Build

Frontier model and platform announcements that directly test prediction p1-2gw-customer-funded. Watch: any disclosed customer-funded compute commitment greater than 2 GW, RPO disclosures, agent-platform pricing. A 2 GW+ commitment scores p1 a hit before its June 30 deadline.

Through Q2 2026

HBM4 memory supply ramp

Validation status at SK Hynix, Samsung, and Micron is the binding constraint on Vera Rubin shipments. Watch: yield or qualification slips at any of the three. A slip compresses the H2 2026 hyperscaler capex tail materially; a clean ramp protects it and keeps Huang's slope intact.

April - June 2026

FERC PJM compliance milestones

Behind-the-meter generation materiality threshold takes effect, reshaping unit economics of large loads in the busiest US power market. Watch: FERC commentary on capacity attribution, ERCOT and MISO reactions. The outcome determines whether 2027 PJM clearing repeats the $329/MW-day or normalizes.

• Inaugural issue. No revisions to the thesis. Future issues append entries here when evidence shifts a hypothesis.