Three things happened in the last 36 hours that, taken as a unit, tell you exactly where the council pattern is in May 2026. They look like separate stories. They aren't.
One. Capgemini announced its investment in the OpenAI Deployment Company on May 12, becoming the first named follow-on consulting investor 24 hours after the entity was unveiled. Two structural details surfaced that the launch coverage didn't have. The entity-level valuation is being reported at $14 billion by Axios, NextWeb, and Constellation Research — the $4B widely quoted is committed funding, not the entity valuation. And the Nifty IT index fell 3.6% in a single day: TCS down ~4%, Infosys 4%, HCL and Wipro 2.5–4%, Persistent Systems ~5%. That is the cleanest equity-market admission that the orchestration tier above the model is now where the enterprise consulting margin lives.
Two. Today, May 13, is the overlap day. AI Council 2026 is in Day 2 at the SF Marriott Marquis — vendor-neutral, ten tracks. Interrupt 2026 opened this morning at The Midway in Dogpatch — Harrison Chase keynote at 9:30 AM PT, Day 1 production case studies from Apple, Lyft, LinkedIn, Toyota, Coinbase, Clay, Rippling, Workday, Uber, and Honeywell. Andrew Ng and Jensen Huang are on the Interrupt program across the two days. Roughly 3,000 practitioners are physically in San Francisco today, and both stages are live-tweeting the same architectural question.
Three. On May 11, Cristiano De Nobili at Critiqality in Milan submitted arXiv:2605.10528 — "Collective Alignment in LLM Multi-Agent Systems: Disentangling Bias from Cooperation via Statistical Physics." It's the sharpest contrarian paper on the multi-agent consensus pattern we've seen this year. The setup: a 2D lattice of identical LLM agents, each holding a binary state, updating by querying the model conditioned on four nearest-neighbor states. Three open-weight models tested — llama3.1:8b, phi4-mini:3.8b, mistral:7b. The headline finding: for current open-weight models under minimal prompting, multi-agent consensus should be treated as "amplified single-agent opinion" rather than "deliberated group judgment."
Three signals. Three different reads. Read them together.
What the $14B SKU actually says
The OpenAI Deployment Company is being framed in most coverage as a consulting arm. That framing misses what's happening. McKinsey, Bain, Capgemini, and Goldman are inside the alliance. TCS, Infosys, HCL, Wipro, and the Accenture stack are outside. The firms that didn't invest took the equity hit. The firms that did bought the upside.
This is consolidation, not competition. The Deployment Company is a roll-up vehicle for forward-deployed engineering capacity across enterprise verticals — McKinsey-style strategy paired with OpenAI-stack integration paired with Capgemini-style implementation muscle. The $14B valuation against $4B committed funding is the giveaway. The 3.5x multiple is pricing in the consulting-bench acquisition wave that hasn't happened yet.
Stacked with the rest of the month: Anthropic shipped Claude Managed Agents and multi-agent orchestration to developers on May 7. Microsoft shipped Agent 365 at $99/user to enterprises on May 1. OpenAI put a $14B price tag on the human-deployed orchestration tier on May 11. Apple previews iOS 27 Extensions on June 8. The orchestration tier is now in production at every layer from consumer OS to enterprise SaaS to consulting bench.
That is also what the conferences in SF this week are mostly about. The Interrupt theme is literally "agents at enterprise scale — what does the team, the tooling, and the infrastructure look like when agents aren't a proof of concept anymore?" That's the routing-and-orchestration question. It has an enormous institutional answer.
It is not the only question.
The seam above the orchestration tier
There's a primitive that sits above routing. When two or more models reason about the same question and disagree — not because they were given different jobs, but because they read the same problem and produced different reasoning — something has to extract the decision. Averaging the outputs throws away the signal. Picking the highest-confidence one rewards the model that was wrong most loudly. Letting them argue forever produces what we've been calling permanent contradiction.
That primitive is deliberation. It's a different primitive than routing. And every time the orchestration tier ships another GA SKU, the seam above it gets sharper, not blurrier.
Which brings us to the De Nobili paper.
The honest critique deserves an honest answer
Most practitioner writing this week will either ignore arXiv:2605.10528 or dismiss it. The differentiated move is to engage it on its merits.
De Nobili is right about the setup he measured. Three small open-weight models in the 3–8B range, sharing roughly the same training-data class, running on minimal prompting. Under those conditions, the consensus is an echo chamber. Same model class plus minimal scaffolding produces neighbor-influenced agreement that looks like deliberation but isn't. Naming that failure mode — amplified single-agent opinion — is useful work.
The vocabulary stack of named failure modes the literature has produced this year now reads, roughly in order: consensus trap, artificial consensus, Debate Trap, homogeneous-ensemble identity bias, anchoring, sycophantic conformity, contextual fragility, consensus collapse, epistemic herding, and now amplified single-agent opinion. Every name is real. Every failure mode is what happens when you build a council wrong.
What the paper doesn't measure is the inverse setup. Heterogeneous frontier models across distinct training lineages. Structured disagreement protocols that force divergence rather than allowing convergence. Chairman synthesis that audits where the panel agreed and where it didn't, rather than averaging the disagreement away.
That inverse is the architectural answer. Heterogeneity at the model layer breaks the same-model-class echo. Strategy at the protocol layer produces real disagreement to synthesize over. Synthesis at the Chairman layer surfaces the structure of the disagreement instead of flattening it.
So when we run a council, we don't run three quantized open-weight models on minimal prompting and call it deliberation. We run Claude, GPT, Gemini, and Grok — four frontier models from four distinct training lineages — and we pick a protocol that matches the question. Red Team vs. Blue Team forces adversarial divergence. Survivor eliminates the weakest reasoning by structured jury vote. Round Robin iterates refinement chains. Quick Take is the don't-deliberate-when-you-don't-need-to primitive — because the honest answer to the echo-chamber critique includes admitting not every question deserves a council in the first place.
That's the engineering form of what the De Nobili paper implicitly endorses. And it's the engineering form of what Karpathy said at Sequoia Ascent this week — "Even for writing, you can imagine having a council of LLM judges and getting something reasonable." The verifiability bottleneck is real. The council is the way through it. As long as the council can actually disagree.
The point
The orchestration tier is institutionally consolidated. $14B in the consulting bench. $99/user in the enterprise SaaS bundle. GA in the developer platform. Previewing on the consumer OS in 26 days. The room for that work is in San Francisco this week, two stages, three thousand practitioners.
The deliberation layer is the open seam. It is the primitive that extracts a decision when heterogeneous models reason about the same question. It is what the De Nobili critique is actually about — not whether to deliberate, but how to make sure deliberation isn't just amplified single-agent opinion in a costume.
The honest answer to the echo-chamber critique isn't don't deliberate. It's make sure your council can actually disagree. Heterogeneous frontier models, structured disagreement protocols, Chairman synthesis that audits rather than averages. Try one on a real decision — the kind expensive to get wrong — and see what the output looks like.
Try it free. shingik.ai — no signup.