Reasoning Engine

Conducting research…

Step 1 / 5
  1. Discovering sources
    Identified 6 candidate sources across 3 publication types.
  2. Analyzing sources
    Extracted atomic claims; scored credibility, recency, and bias on each.
  3. Cross-referencing
    Detected contradictions and reconciled overlapping claims.
  4. Synthesizing findings
    Compressed claim graph into structural themes and a working thesis.
  5. Generating intelligence
    Drafted executive brief, evidence map, risks, and recommendations.

Research Telemetry

live · demo
Reasoning
86/100
Confidence
87/100
Evidence
89/100
Depth
82/100
Diversity
84/100

Synthesized Answer

GPU and power constraints

The AI infrastructure market is the largest capex cycle in tech history — $400B+ hyperscaler spend in 2026, dominated by Nvidia GPUs, HBM memory, advanced packaging, and the increasingly binding constraint of grid power.

Key Points

  • Hyperscaler capex ~$400B in 2026, +60% YoY
  • Power, not chips, is the binding constraint
  • Nvidia retains ~85% accelerator share; AMD and custom rising
  • Nuclear PPAs are the new strategic asset class
  • Cumulative AI infra spend $1.5–2.0T through 2028

Knowledge Graph

11 nodes · 13 edges
topicconceptcompanyentity
AI infrastructureGrid powerNvidiaAMDTSMC / CoWoSHBM memoryUltra EthernetNuclear PPAsMicrosoftGoogleAmazon

Auto-generated Insights

Trend

Power has overtaken chips as the binding constraint on data-center buildouts.

Finding

Custom silicon hits 15% of inference share; first credible challenge to Nvidia dominance.

Signal

Nuclear PPAs are emerging as the new strategic asset class for hyperscalers.

Contradiction

Capex disclosures imply 60% YoY growth; grid data caps physical buildout at 30–40%.

Structured Data

Extracted from sources

Hyperscaler 2026 capex

$400B+

60% YoY

Microsoft+Google+Meta+Amazon+Oracle

Nvidia accelerator share

85%

−4pp YoY

data-center revenue

U.S. data center power share

5.6%

1.6pp YoY

of total grid load

PJM interconnect queue

6+ years

vs ~2 years pre-AI

Northern Virginia data center alley

Sources6 ranked

Sorted by relevance
S
semianalysis.com·this week
Report

Hyperscaler Capex Tracker Q2 2026

Microsoft, Google, Meta, Amazon, and Oracle combined capex reached $98B in Q2 2026, on track for $400B+ annual.

Cred
93
Auth
91
Fresh
96
Rel
88
●Center
Strongevidence
E
eia.gov·this week
Report

AI Power Demand and Grid Constraints

U.S. data center electricity demand projected to reach 9% of total grid load by 2028, up from 4% in 2024. PJM interconnect queue at 6+ years.

Cred
96
Auth
96
Fresh
92
Rel
88
●Center
Strongevidence
N
nextplatform.com·this week
Article

Nvidia Market Share and Roadmap Update

Nvidia retains ~85% data-center accelerator revenue share; Blackwell Ultra ramps Q3 2026, Rubin in H2 2027.

Cred
88
Auth
85
Fresh
95
Rel
88
●Center
Strongevidence
R
reuters.com·this week
News

Microsoft-Constellation Three Mile Island Restart

Microsoft signed a 20-year PPA for the full output of Three Mile Island Unit 1, restarting by 2028 to power AI data centers.

Cred
92
Auth
90
Fresh
94
Rel
88
●Center
Strongevidence
S
semianalysis.com·this week
Report

Custom AI Silicon: TPU, Trainium, MTIA Comparison

Custom silicon now accounts for ~15% of inference workloads at hyperscalers; cost advantages of 2–4× on optimized workloads.

Cred
92
Auth
90
Fresh
90
Rel
88
●Center
Strongevidence
I
ieee.org·this week
Article

Ultra Ethernet Consortium Update

UEC 1.0 specification ratified; aims to match InfiniBand performance for AI scale-out by 2027. AMD, Broadcom, Cisco, Meta backing.

Cred
95
Auth
94
Fresh
86
Rel
88
●Center
Strongevidence

Refine your research

Map hyperscaler capex by companyProject power-constraint scenariosBuild Nvidia vs AMD investment viewAnalyze nuclear PPA economicsCompare custom silicon roadmapsGenerate executive summary

Demo mode · All sources, insights, and data are mock-generated for illustration.