ARMKU Pulse
Manufacturing output, software spend, the AI build-out and the grid behind it. Pulled automatically from public, primary sources and refreshed daily. Once a month we write up what it means for a mid-sized operator.
Monthly brief · 2026-08
Factories are still growing, but the money is moving to software and AI, not new plants. Manufacturers' new orders jumped 9.9% year over year, yet manufacturing construction spending fell 21.2% over the same period. If you run operations, that gap is the story this month: demand is real, but nobody wants to break ground on new capacity.
The industrial production index rose to 103.0 in July, up from 102.8 the month before and 1.1% higher than a year ago. Manufacturing output climbed to 99.3, up 1.3% year over year, and capacity utilization held steady at 76.0%. New orders hit $664 billion, up from $658 billion and 9.9% ahead of last year. Manufacturing jobs barely moved, up just 0.2% to 12,638,000. Put together: plants are running fuller and taking more orders, but nobody is adding headcount or pouring new concrete. Manufacturing construction spending dropped 21.2% versus a year ago, to $170 billion.
Business investment in software hit $817 billion a year, up 9.5% from a year ago. The big platform vendors show it: Microsoft revenue rose 17.8% for its fiscal year, Oracle 17.4%, and ServiceNow 20.9%. But custom programming services jobs actually fell 1.3% year over year, to 2,363,000. Companies are spending more on software and hiring fewer custom developers to build it, they're buying platforms instead. On the open-source ERP side, Odoo leads with 54,176 GitHub stars, ahead of ERPNext at 38,909 and Dolibarr at 7,574.
Capital spending on AI infrastructure is still climbing fast. Oracle's capex is up 162.4% year over year, Meta's up 87.1%, Microsoft's up 79.6%, Alphabet's up 74.1%, Amazon's up 58.8%. Epoch AI counted 106 notable AI models released last year, up from 98 the year before. On the agent-building side, LangChain leads open-source frameworks with 145,696 GitHub stars, followed by AutoGen at 60,815, CrewAI at 58,110, and the OpenAI Agents SDK at 29,205. The tooling for building agents is maturing fast, even as the hyperscalers keep spending on the hardware underneath it.
Electric power generation jumped to an index of 124.3 in July, from 110.8 the month before, though it's only up 1.3% versus a year ago. Utilities output rose to 110.7, up 0.7% year over year. Power demand is moving, and it's not a coincidence given the capex numbers above.
Is the factory floor actually busier? Output, utilization, orders, jobs and the money going into new plants.
Total US industrial output, 2017 = 100. The broadest read on whether plants are busier.
vs. a year ago
Manufacturing only, 2017 = 100. Strips out mining and utilities.
vs. a year ago
Share of manufacturing capacity in use. Above 80% plants are stretched; below 75% they have slack.
vs. a year ago
People employed in US manufacturing, thousands, seasonally adjusted.
vs. a year ago
Value of new orders received by US manufacturers each month, billions of dollars.
vs. a year ago
Money going into new factories, billions of dollars at an annual rate. The reshoring boom in one line.
vs. a year ago
Where the software dollars go: business investment in software, programming jobs, the revenue of the SaaS bellwethers, and the pull of open-source ERP.
US private fixed investment in software, billions of dollars at an annual rate. Quarterly.
vs. a year ago
Employment in US custom computer programming services, thousands. A proxy for implementation and integration work.
vs. a year ago
Annual revenue from 10-K filings, billions of dollars.
vs. a year ago
Annual revenue from 10-K filings, billions of dollars.
vs. a year ago
Annual revenue from 10-K filings, billions of dollars.
vs. a year ago
Annual revenue from 10-K filings, billions of dollars.
vs. a year ago
Annual revenue from 10-K filings, billions of dollars.
vs. a year ago
Stars on the odoo/odoo repository. Developer attention to open-source ERP, not market share.
Stars on the frappe/erpnext repository. Developer attention to open-source ERP, not market share.
Stars on the Dolibarr/dolibarr repository. Developer attention to open-source ERP, not market share.
How fast the AI wave is really moving: notable models per year, the capital the hyperscalers are spending, and developer adoption of agent frameworks.
Models in Epoch AI's notable models database, by publication year, through the last complete year.
vs. a year ago
Payments for property, plant and equipment from 10-K filings, billions of dollars. Mostly data centers now, but not only.
vs. a year ago
Payments for property, plant and equipment from 10-K filings, billions of dollars. Mostly data centers now, but not only.
vs. a year ago
Payments for property, plant and equipment from 10-K filings, billions of dollars. Includes warehouses as well as data centers.
vs. a year ago
Payments for property, plant and equipment from 10-K filings, billions of dollars. Mostly data centers now, but not only.
vs. a year ago
Payments for property, plant and equipment from 10-K filings, billions of dollars. Mostly data centers now, but not only.
vs. a year ago
Stars on langchain-ai/langchain. Developer adoption proxy for agent frameworks.
Stars on crewAIInc/crewAI. Developer adoption proxy for multi-agent frameworks.
Stars on microsoft/autogen. Developer adoption proxy for multi-agent frameworks.
Stars on openai/openai-agents-python. Developer adoption proxy for agent tooling.
Whether the grid is keeping up with the compute being built.
US electric power generation, 2017 = 100, monthly. Not seasonally adjusted, so the summer peaks are real.
vs. a year ago
US electric and gas utilities output, 2017 = 100, seasonally adjusted. The underlying trend without the weather.
vs. a year ago
Every series comes straight from a primary, public source: the Federal Reserve Bank of St. Louis (FRED) for production, orders, jobs, construction and electricity; SEC EDGAR filings for company revenue and capital spending; GitHub for open-source adoption; Epoch AI for the count of notable models. Nothing is estimated by hand.
A few honest caveats. GitHub stars measure developer attention, not market share. Hyperscaler capital spending is total, not data centers alone. Company figures are fiscal years, which do not all end in December. The AI model count runs through the last complete year. When a source goes quiet, the series simply stops at its last real point.
The monthly brief is drafted by an AI assistant from this exact data, then reviewed and published by Rodolfo Kong. Raw feed: /pulse/data.json.
Want this read against your own numbers? Book a conversation.