The Early Scale: AI Is Deployed Everywhere. It's Working Almost Nowhere.

Good morning

AI is everywhere, but producing results almost nowhere. Three separate reports dropped this week showing the same brutal gap: deployment is up, outcomes are down, and the culprit is almost always the same thing, strategy (or the total lack of one). Meanwhile, the hardware race just got more expensive, and utilities are learning that building big doesn't mean getting paid. Grab your coffee. Here's what matters this morning.

The Big Three

AI Is Deployed Everywhere. It's Working Almost Nowhere.

Two major studies landed within days of each other and they tell the same uncomfortable story. Kyndryl's 2026 People Readiness Report found that AI has been deployed in 57% of enterprises, yet only 11% have hit both of their top objectives. Info-Tech Research Group's survey of 551 senior leaders found that enterprises with a formal AI strategy are 3x more likely to report measurable impact than those just running pilots and hoping. The variable separating winners from everyone else is not the model, the vendor, or the budget. It is whether someone owns the strategy, the data, and the workforce readiness behind it.

The B2B angle: Stop buying more AI tools and audit the ones you already have: assign clear ownership, define two measurable objectives, and run a data-readiness check before the next renewal cycle.

Automation Capital Is Back, but Humanoid Hype Is Getting Ahead of the Factory Floor

Roland Berger is forecasting 6-9% annual growth in industrial automation capital spending through 2030, a significant acceleration from recent years. The firm's analysts and executives at Automate 2026 both agree: the cycle is real, but the headlines are running well ahead of deployable reality. Humanoid robots are drawing the crowds and the venture dollars, while the actual near-term gains are coming from intelligent software platforms layered on top of existing equipment. Manufacturers who over-rotate toward robotics demos risk missing the unglamorous but profitable integration work happening right now.

The B2B angle: If you sell into manufacturing or industrial, position around integration and ROI proof points now, before humanoid hype resets buyer expectations upward to a level your product cannot meet.

Fitch Calls Utilities 'Deteriorating' as a $240B Capex Wave Meets Affordability Walls

Fitch downgraded its outlook for the utility sector to 'deteriorating' in June 2026, and the math is not hard to follow. Utilities are committing to a historic $240 billion capital spending wave, largely to support grid modernization and data-center load growth. But regulators and consumers are pushing back hard on rate increases, threatening cost recovery on investments that are already being made. The squeeze is tightening: capital is going out the door faster than it can be recouped, and Fitch says the affordability pressure is structural, not temporary.

The B2B angle: Energy infrastructure vendors and project developers should front-load rate-case documentation and ROI modeling now, because utility procurement teams are about to face intense internal scrutiny on every new commitment.

Also worth knowing

Anthropic launched Claude Opus 5 at the same price as Opus 4.8, with performance approaching GPT-4.1-level benchmarks at roughly half the cost, and introduced an 'effort dial' that lets enterprise teams trade response depth for predictable billing. For B2B teams tired of unpredictable AI spend, that dial is the real product.

Etched closed a $300M Series C at a $10.3B valuation led by Sequoia, doubling its valuation in seven months, with $1B in inference chip orders already booked. The inference chip market is no longer a future bet: it is a present supply constraint.

Rockwell Automation surveyed 1,560 manufacturers and found 93% have a Manufacturing Execution System, but only 23% have fully integrated it enterprise-wide. Owning the software is not the same as using it. The integration gap is where margin is being left on the table.

Smart plays for the week

Before your next AI vendor renewal, map every active AI tool to a named internal owner and two specific, measurable objectives, then kill any tool with neither. Kyndryl found only 11% of enterprises hit their AI goals; Info-Tech confirmed that ownership and strategy, not deployment volume, are the deciding variables.

If you market to manufacturers, lead your next campaign with integration ROI data rather than technology capability, and use the 93-vs-23 MES gap as your opening hook. Rockwell's survey proves nearly every manufacturer already has the software; the selling conversation has shifted entirely to 'are you actually using it across the enterprise.'

Test Anthropic's Claude Opus 5 effort dial this week by setting a lower effort tier on high-volume, lower-stakes tasks such as email drafts or CRM summaries, and track whether your AI spend drops without a quality hit. Opus 5 launched at the same price as its predecessor with a new cost-control lever, making this a zero-risk experiment to convert unpredictable AI bills into a manageable variable.

Something to think about

With 57% of enterprises running AI and only 11% hitting their goals, this framing cuts through the noise: the bottleneck is not the model, it is the management system around it.

Teach me something: Inference Chip

Most people have heard of training AI, which is the expensive, months-long process of teaching a model on massive datasets. Inference is what happens after that: every time you ask an AI a question or run a prediction, the model is 'inferring' an answer in real time. Inference chips are purpose-built silicon optimized for that task at scale, trading the general flexibility of a GPU for dramatic gains in speed and cost per query. As enterprises shift from experimenting with AI to running it constantly in production, inference costs become the dominant expense, which is exactly why Etched's $1B in chip orders and Anthropic's new 'effort dial' both matter so much right now.