
After GPT-6 Astra, Cheap Tokens Are Not Enough: Price the Work That Gets Done
Acceptance rates, human review, retries and workload expansion reshape agent economics and the value captured by workflows, models and infrastructure.
Read the column →Source-linked analysis of data-center contracts, cloud capital spending, power availability and the path from announced capacity to billable infrastructure.

Acceptance rates, human review, retries and workload expansion reshape agent economics and the value captured by workflows, models and infrastructure.
Read the column →
Coherent transport and line systems take center stage. Strong revenue, lower next-quarter margin guidance and concentration reveal the economics of networks beyond the data center.
Read the column →
Use the reported Lambda deal and a separately documented Hut 8 financing case to distinguish cloud purchases, property leases and the project-debt cash waterfall.
Read the column →
Separate AI revenue, custom accelerators and networking to examine growth, margins, customer-designed systems and cash conversion.
Read the column →
Separate purchase consideration from equity retention, then test whether discovery and inference routing create durable distribution power without sacrificing openness.
Read the column →
Both companies shifted growth toward APIs and open platforms. Z.ai disclosed 100,000-plus domestic-chip inference and positive API margin; MiniMax showed global enterprise expansion. Adjusted losses and funding needs remain substantial.
Read the column →
Owned campuses, partner-owned assets, GPU-backed debt and behind-the-meter power show how Nebius separates capacity growth from balance-sheet burden.
Read the column →
Anthropic's $47B run-rate and AWS, Google, Broadcom and SpaceX capacity are mapped across demand, contract, delivery and cash burn.
Read the column →
Data Center revenue reached $89.0B, while $366B of commitments, $108.5B of guarantee exposure, receivables and inventory move NVIDIA toward infrastructure orchestration.
Read the column →
Z.ai served GLM-5.3-Flash across tens of thousands of Chinese accelerators. The important shift is model, engine, interconnect and scheduling co-design turning domestic compute into a serviceable cluster.
Read the column →
The 8 IT-GW plan separated into NVIDIA credit support, SB Energy land-power-shell, OpenAI demand and unfinished energization.
Read the column →
Four operators received provisional 50MW allocations. The real gate combines green power, IT efficiency, AI value and project execution.
Read the column →
Separating NVIDIA's third-party capital objective from project finance, energization, collateral and customer cash flow.
Read the column →
China Mobile's interim results unpacked across AIDC, compute services, data centers, utilization and revenue quality.
Read the column →
Tracing supplier ramps through packaging, server qualification and customer configuration to map synchronized delivery of next-generation AI platforms.
Read the column →
CoreWeave, Nebius and Bitdeer compared by revenue, ARR, backlog, service-ready capacity and cash conversion.
Read the column →
ERCOT's Batch Zero puts eligibility, project evidence and system reliability inside large-load interconnection.
Read the column →
GLM-5.3, DeepSeek Harness and Anthropic commercial signals separated into capability, runtime and an unproven causal chain.
Read the column →
Meta and BlackRock place the under-construction El Paso campus in a venture: outside capital owns most assets while Meta retains the lease, management and residual-value support. The 1GW target begins online in 2028; energization, acceptance and use remain to be verified.
Read the column →
Volta's official $10B compute agreement names no customer; media report Anthropic, unconfirmed. The three-layer capital stack and delivery verification boundaries, separated.
Read the column →
Private capital packages generation, storage, distribution and operations into EaaS. The deal is open and the 2GW pipeline is intention until four gates close.
Read the column →
A Morgan Stanley-led group plans to turn ~$15B of Hubbard campus loans into speculative-grade bonds; Google support activates only after completion.
Read the column →
TrendForce confirms undersupply into 2027; NVIDIA's three HBM options remain undecided and CSPs weigh lower ASIC content.
Read the column →
Cloud revenue absorbs capacity while the supply gap persists; hyperscaler free cash flow diverges. Vertiv and Arm stay not-determined.
Read the column →
Q2 power-generation sales reached $3.098B, led by large reciprocating engines and turbines for data centers. AI spending is spilling into firm power, storage and service.
Read the column →
EU support can cover up to 17% of capex or pre-purchase equivalent access time. The consortium still carries the remaining capital, opex, silicon and utilization burden.
Read the column →
The Broadcom-Apollo-Blackstone AI XPV platform links silicon, sites, long-term demand and private credit. Residual value and utilization matter more than headline size.
Read the column →
EPA guidance, PJM capacity policy and state permitting are moving new generation, grid upgrades and air compliance onto data-center economics.
Read the column →
A buyer can diversify at portfolio level while standardizing each large cluster. GPU, network, software and operations turn chip choice into infrastructure architecture.
Read the column →
Palantir Q2 and sovereign AI separated into reported contracts, management claims and the still-unproven compute, network, power and cooling read-through.
Read the column →
Seven tests across external cloud growth, model-service ARR, the RMB380B buildout, proprietary silicon, margin and cash conversion.
Read the column →
One gigawatt, faster GPU activation and shifting lease accounting reveal how Microsoft must convert infrastructure into cloud revenue.
Read the column →
GPT-5.6 and Anthropic's Sonnet 5, Opus 4.8 and Fable 5 turn capability, speed, price and reasoning depth into separate infrastructure variables.
Read the column →
Beyond the record print, customer agreements, HBM4 execution and disciplined capacity are the three signals that matter for the memory cycle.
Read the column →
A ranked read-through across MU, AMAT, LRCX, KLAC, NVDA, AMD and RMBS through conventional DRAM, HBM, equipment controls and China exposure.
Read the column →
Azure, Google Cloud and AWS accelerated together; capacity gaps and free-cash-flow conversion now define the hyperscaler AI buildout.
Read the column →
Four disclosed agreements approach 1.8GW. Contract, energization and billing reveal how campus-scale orders convert into cash flow.
Read the column →