For three months the gap between what AI could do and what organizations could absorb showed up where it had never been legible before — in distribution models, in labor markets, in supply chains, and in valuations. This is the quarterly read on what 357 new signals say, taken together, that no single signal says alone.
Q2 2026 was the quarter AI capability stopped being the constraint. The bottleneck became infrastructure — physical, financial, and human. Token spend outran the budgets built to measure it; a frontier model was recalled by export order three days after launch and replaced by an open-weight substitute at a tenth of the cost the next day; a profitable insurer detonated its own 19,000-agent distribution layer in a single decision; and a $965B raise priced an AI economy a Tufts dean argued does not yet exist at the organizational layer. Underneath all of it, the taxonomy itself couldn't hold still: the Stack stood up eight new categories in ninety days. The readiness gap is no longer an argument. It is showing up in the numbers.
Category growth is itself a trend indicator. When a tracked corpus has to invent new categories to hold the inflow, that genesis rate is a finding — not a footnote. Q2 added 357 signals and required eight new categories to file them.
The genesis rate is the quarter's quietest, loudest finding. Eight categories that did not exist on April 1 were carrying signals by June 30:
Eight categories in ninety days is not housekeeping. It is the measurement instrument straining against the rate of structural change. When the map needs redrawing this often, the territory is moving faster than the people standing on it can feel.
A movement is a pattern across many signals that no single signal demonstrates. Each below is named from the densest, most cross-validated clusters in the Q2 corpus, cited to its evidence, and read against the framework spine.
The densest mechanism in the entire quarter was sovereignty: who controls compute, models, and access, and whether that control can actually be held. The answer that emerged is that it increasingly cannot. A frontier model was pulled from the global market by emergency export order three days after launch — the first post-deployment recall of its kind — and an open-weight, MIT-licensed substitute at roughly a tenth of the cost, running on non-standard silicon, filled the vacuum the next day. The lab that red-teamed and disclosed got intervened against; the labs that look less hard ship freely. Meanwhile the physical floor hardened: 410,000 MW of data-center demand on file with one grid operator, GPUs financed like aircraft with a year-one depreciation clock, and a diagnostic framework (SAAFE-7) showing most "sovereign compute" claims fail on supply-chain integrity and jurisdiction.
For two years the gap between AI capability and organizational absorption was an argument practitioners made. This quarter it became a line item. A $44B fintech raised $750M naming tokens the third pillar of business cost — the fastest-growing cost in commercial history, invisible to the instruments every finance team was trained on. Enterprises that set 2026 AI budgets on 2024 usage rates blew through them in a single quarter; Uber, Microsoft, Amazon, and Walmart all capped spend. A formal arXiv proof established the structural Jevons paradox: per-token prices fell while total spend tripled, by design. And the academic-survey cluster — HBR's 67-point gap, McKinsey's 88/86 paradox, UST, the $252B-for-6%-impact finding — all measured the same thing from different angles. The gap is now denominated in dollars.
The place where economic value accrues moved up a level — from the model to the orchestration layer that runs fleets of agents across tools and data. MCP crossed to infrastructure grade (Linux Foundation, 97M monthly SDK downloads, first-class support across every major platform), and agentic execution shipped by default: Microsoft took Copilot Cowork to general availability worldwide, priced per task rather than per seat, running on a third-party frontier model. Analysts put $700B of SaaS displacement and $58B of productivity-suite reshuffle at stake. The pattern beneath all of it: whoever owns the orchestration and observability layer owns enterprise AI governance, regardless of which model wins — the Windows-above-hardware, AWS-above-servers pattern, repeating.
The quarter's most structural finding is that the firm as a unit of economic organization is eroding bidirectionally. Outward: a formal theorem (the Coordination Tax / CALM bridge) proved 42–74% of enterprise coordination spend is avoidable, and an Ismail survey found a majority of incumbents believe a two-person AI-native team could disrupt them. Inward: Microsoft's MAI family showed the world's largest software company re-integrating the model layer with a sub-10-engineer team. And on the supply side, an entirely new failure mode appeared — a profitable insurer compressing its own 19,000-agent captive distribution layer from the cost and distribution sides at once, the genesis event for Cat 22. The minimum viable size of a company is collapsing while the maximum reach of an incumbent's insourcing expands. The middle is where the firm used to live.
A report that only confirms itself is not authority. These are the strongest signals in the Q2 corpus that complicate, qualify, or push back on the movements above. They are logged with the same weight.
Five questions the Q2 corpus opened and did not close. These set the agenda for the next quarterly.
This report is derived from a tracked corpus, not assembled for the occasion. Every movement named above is backed by dated, sourced, cross-validated signals logged in the Signal Stack before they were known to form a trend. The work is reproducible: the full record is public — every signal referenced here is in the Signal Stack v10.5.
A named, dated development with a primary source, a category, a one-line thesis, and a set of cross-validation references to other signals. Signals carry a status — Confirmed, Formal, Watch, or Candidate — reflecting how independently corroborated they are. Watch-status items (e.g. the substrate-fragility counter-thesis) are explicitly flagged as unconfirmed.
A daily scan surveys the labs, the agentic frontier, regulators, capital markets, the academic literature, and the IBM i ecosystem, filters for durable mechanisms over transient news, and logs candidates against the existing record to avoid double-counting. The quarterly is the accumulation: 357 signals entered the Stack across Q2, taking it from v8.4 to v10.5.
Every signal is read against four lenses, applied consistently across quarters so trends can be tracked over time rather than re-invented each issue:
This is an instrument, not an argument. Where a finding supports a particular strategic thesis, that does not make it true; where the corpus contains its own counter-evidence, that counter-evidence is logged with equal weight (see Section 04). The report's value is that it is comprehensive, current, and disinterested — readers are expected to check the record and draw their own conclusions.
Linked where a canonical primary source is publicly reachable; named where the source is paywalled or proprietary. Each maps to one or more dated signals in the full record.