compute
22 articles
- Anthropic's $2 Trillion IPO Rests on a Profitability Metric That Excludes Its Largest Costs
Anthropic's >80% "gross margin" and $559M "adjusted operating profit" exclude partner revenue sharing, training costs, and stock-based compensation. The S-1 discloses $518B in take-or-pay compute obligations, 80% non-cancelable, with named counterparties including Broadcom ($161B), Google ($111B), and Amazon ($110B). The safety slowdown is a capital preservation maneuver timed to the IPO window.
- The Credit-Sorting Moat: How 5% Rates Concentrate Energy Capital in Hyperscaler Balance Sheets
The 5% rate regime concentrates energy-transition capital around strong credits. Hyperscalers combine investment-grade ratings, captive demand, and long-tenor financing to fund behind-the-meter generation near investment-grade pricing, while merchant projects face wider spreads and higher levelized costs.
- The AI Safety Narrative as a Synthetic Moat
Frontier AI labs face rapid algorithmic depreciation and escalating compute costs. Safety regulation can create a synthetic moat by imposing fixed compliance costs that favor incumbents over open-weight competitors.
- Sovereign Compute Attestation Expands U.S. Export Enforcement
Cryptographic attestation and hardware roots of trust support export-control enforcement. Value accrues to chip vendors, cloud providers, and compliance firms.
- How AI Safety Became a Rent-Seeking Institution
AI safety has increasingly become an institutional apparatus optimized for funding, prestige, regulatory influence, and adoption rather than binding constraint. Material AI risks arise from deployed socio-technical systems, requiring layered governance across compute, security, incentives, organizations, and institutions.
- Affordable Mass Is a Factory Problem: The Production Bottlenecks Behind U.S. Unmanned Aircraft Scale
The U.S. Air Force's 500-aircraft CCA target by 2032 turns unmanned aircraft into a production problem. The binding constraints are qualified propulsion, machining, composites, electronics, inspection, and test capacity.
- The Two-Layer Constraint on AI Infrastructure
Power and permitting cap data center capacity. Credit markets determine which sponsors can finance projects within that physical ceiling. Hyperscaler balance sheets and guarantees provide the cheapest credit enhancement, concentrating the buildout among a few large platforms.
- Economic AGI Is a Factor-Substitution Threshold
Economic AGI is not a singularity or an ontological state — it is a factor-substitution threshold. Frontier models have crossed expert parity on standardized professional work at a fraction of human cost, and the threshold has been crossed for low-tail functions like customer support and content production. The remaining gap is autonomy, reliability, integration, and liability — not raw reasoning.
- China's Strategic Deficit: A Fiscal Crowding-Out Failure, Not a Confrontation Error
China's strategic deficit can be explained as a fiscal crowding-out problem: capital was absorbed by real estate and local government balance sheets before domestic semiconductor capabilities reached strategic self-sufficiency.
- The Intelligence-Explosion Debate Uses the Wrong Model
The key variable is feedback-loop generation time, not loop gain. The central question is whether AI removes bottlenecks faster than new ones emerge.
- The Interconnection Queue Is the Asset: Why Legacy Grid Sites Are Repricing for AI
Retired power plants, aluminum smelters, and industrial sites are gaining value because they already have high-voltage grid connections that can take years to replicate. The opportunity depends on whether regulators preserve the advantages of these existing interconnections.
- The Sovereign Compute Verification Problem: Why Identity Proofing Is the New Export-Control Bottleneck
U.S. export control has shifted from controlling chips to controlling access. The resulting bottleneck is verification: restrictions based on location, ownership, or nationality cannot be enforced without identity-proofing infrastructure across hardware, cloud, model, and bilateral-governance layers.
- The Mature-Node Scarcity Paradox
U.S. export controls aimed at leading-edge chips inadvertently created a pricing-power windfall in China's mature-node foundries. The mechanism: a deliberate policy line, state capital reallocation, and an unmodeled AI demand shock.
- The AI Buildout Is Real. The Credit Structure Is the Risk.
AI infrastructure is a real technology cycle wrapped in a leveraged-finance structure. Demand risk is moving from hyperscalers to banks, insurers, and bondholders through SPVs, GPU-backed debt, securitization, and synthetic risk transfer, making utilization, covenants, refinancing, and collateral the earliest signals of overcapacity.
- The AI Empirical Grounding Bottleneck
Open-weight foundation models have collapsed the intelligence generation layer to near-zero marginal cost. The structural bottleneck is migrating to the verification layer — where search engines, financial data aggregators, and content publishers are converting their historical indexing moats into paid API tollbooths. This is not a single monopoly rent; it is a fragmented patchwork of metered, unevenly-priced access gates whose integration complexity is the real constraint.
- The Compute Substrate Race: Where the U.S. and China Hit Different Walls
AI compute scarcity has migrated from fabrication to HBM, advanced packaging, grid access, and cooling. China is constrained upstream; the U.S. downstream.
- Cloud Migration Is Not Binary: The Bifurcation of Hyperscaler Power and the Rise of Hybrid Infrastructure
The cloud-first mandate is outdated. Two independent pressures — cost math and jurisdictional law — are converging on hybrid architecture. Hyperscalers are winning the frontier AI layer almost unopposed while losing share at the steady-state, compliance-heavy layer. Multi-cloud is increasingly accidental architecture, not strategic hedging.
- Interest Rate Transmission in High-Debt Regimes
The traditional interest rate transmission mechanism has structurally inverted in a high-debt regime. Rate hikes redistribute demand rather than compress it — sovereign interest expense flows to private Treasury holders, private legacy debt remains insulated by duration lock-ins, and the strain relocates to the term premium and the refinancing wall rather than disappearing. Fiscal-monetary interaction, collateral repricing, and state capacity now override private credit cycles.
- Enterprise Procurement: Why Lowest Risk Is Replacing Lowest Price
Enterprise software procurement is shifting from cost optimization to risk-adjusted continuity. Pricing volatility, geopolitical fragmentation, and infrastructure scarcity are forcing buyers to weight sovereignty, compliance, and operational resilience alongside price.
- The AI Implementation Tax: Where Enterprise AI Dollars Actually Go
Enterprises spent $684 billion on AI in 2025, and $547 billion produced no measurable result. The money is going to the wrong places — idle compute, tool sprawl, consulting markup — while actual return drivers get starved.
- The Thermodynamics of Strategic Decoupling
The semiconductor supply chain is not a policy debate — it is a physical constraint. Advanced packaging, equipment tooling, and rare earth processing cannot be duplicated instantaneously. The irreversible timeline for strategic decoupling is approximately 2028-2032.
- Data Centers Are Modern Oil Refineries: Heat Rejection Is the Binding Constraint
Data-center capacity is constrained by the density at which waste heat can be captured, transported, and rejected, not by aggregate electricity generation alone.