By Q3 2027 the Median Large Frontier-AI Vendor Will Run Three Separate Compliance Pipelines (US, EU, UK), and At Least One Top-5 Lab Will Have Publicly Declined a CAISI Evaluation

Prediction Statement

By the end of Q3 2027 (September 30, 2027), both of the following will be true:

Three-pipeline architecture is the median. Among the top eight large frontier-AI vendors by 2026 frontier-AI revenue (Microsoft, Google, OpenAI, Anthropic, Meta, xAI, Amazon, NVIDIA — adjusting for any 2026 acquisitions or restructurings), at least five will be operating three formally separate compliance pipelines for AI governance — one for US pre-deployment evaluation (CAISI plus federal procurement evidence), one for EU AI Act conformity assessment under the August 2027 trigger window, and one for UK sector-led compliance (FCA, MHRA, ICO, Ofcom). "Formally separate" is defined as separately staffed (named program leads, distinct budget lines visible in public filings or analyst reports) or separately retained (named external counsel and consulting engagements specific to each regime).
At least one top-5 US lab has publicly declined CAISI pre-deployment evaluation. At least one of Microsoft, Google, OpenAI, Anthropic, or Meta will have publicly declined to participate in a CAISI pre-deployment evaluation for at least one frontier-model release between May 2026 and September 2027, with the refusal documented in a public statement, regulatory filing, on-record interview, or congressional testimony, citing competitive concerns about information disclosure to a federal body, IP protection, or scoping disagreements.

Reasoning and Analysis

The structural argument

The week of May 4 to 8, 2026 produced three governance moves — CAISI pre-deployment gating, EU AI Act simplification with deferred high-risk trigger, and UK no-action posture — that express durable institutional preferences each jurisdiction is unlikely to abandon. The US system delegates technical evaluation to credentialed federal labs anchored by procurement language. The EU system uses horizontal primary legislation with delegated acts. The UK system extends existing sector regulators rather than creating a horizontal AI authority. None of these designs is converging on the others; each is a coherent expression of how the underlying polity prefers to govern technical risk.

The compliance consequence flows from the design divergence. A unified compliance program cannot satisfy three regimes that demand different evidence at different points in the model and deployment lifecycle. The median large vendor will resolve this by running three programs in parallel with shared evidentiary substrate but distinct acceptance criteria. That resolution has already begun in conversations with the four large management consultancies that build these programs.

The CAISI refusal argument

The structural argument for at least one top-5 lab declining a CAISI evaluation has three components:

Anthropic's stated thesis is that frontier safety review should happen inside the lab and through narrowly chosen private partners (the Glasswing model). The Glasswing controversy has not weakened that thesis; if anything it has strengthened it by demonstrating that the lab can hold a chosen-board posture against external pressure.
OpenAI's competitive position with respect to government information disclosure is awkward. OpenAI's evaluation methodology is core IP; sharing it with a federal lab raises non-trivial competitive concerns that commercial-prudence arguments can support.
Meta's open-weights frontier policy is structurally awkward against a pre-deployment evaluation gate. Once weights are released, post-deployment behavior is determined by every fine-tuner, and a pre-release evaluation is a snapshot of one starting point. Meta could rationally decline on scoping grounds.

Any one of those three labs taking a public refusal posture for at least one frontier release between May 2026 and September 2027 satisfies the prediction's second clause. The probability that none of the three takes such a posture is low.

Confidence Factors

Why 65% confidence (not higher):

The three-speed architecture depends on EU enforcement actually arriving in 2027 rather than slipping further. The May 7 simplification agreement shows that EU deadlines do move when they need to. A second deferral would weaken the EU pipeline's perceived enforcement risk and could collapse the median into two-pipeline operation (US + UK with EU treated as voluntary).
The CAISI program could evolve to address the structural objections (open-weights treatment, IP protection, scoping flexibility) in ways that remove the rationale for refusal. If CAISI builds carve-outs that accommodate Anthropic, OpenAI, and Meta's specific concerns, no lab need publicly decline.
US administration changes between now and September 2027 could substantially restructure CAISI itself. A successor body with different scope or different procurement linkage would materially change both clauses of the prediction.
"Three formally separate compliance pipelines" requires evidence visible in public filings or analyst reports. Vendors that maintain three pipelines internally without public visibility would technically falsify the prediction even if the substantive claim were true.

Why not lower than 65%:

The structural argument is durable. None of the three jurisdictions has signaled willingness to harmonize on either of the others' designs. The EU's May 7 simplification deferred enforcement but did not soften substance.
The CAISI refusal argument has three independently sufficient mechanisms (Anthropic, OpenAI, Meta). The prediction fails on this clause only if none of the three publicly refuses.
The compliance-team-headcount data in 2025 annual reports already shows the divergence pattern beginning. The median frontier-AI vendor's compliance FTE roughly doubled from 2024 to 2025; the doubling is largely attributable to the EU AI Act preparation. The US procurement gate adds a third multiplier.

Key Indicators to Watch

Quarterly through the prediction window:

CAISI participation roster — additions and notable absences. Anthropic signing on or publicly declining is the highest-information event.
EU AI Act standards-readiness certifications from the Commission, which will determine when the deferred high-risk obligations actually trigger.
UK sector-regulator AI guidance from FCA, MHRA, ICO, Ofcom, especially the FCA AI rulebook expected through late 2026 and 2027.
10-K and 20-F filings from US-listed frontier-AI vendors for explicit AI-governance line items in the risk factors and operating expense sections.
Companies-house filings from UK-incorporated AI vendors for compliance spend disclosures.
Public statements from frontier-lab CEOs and CISOs addressing CAISI specifically — supportive, ambivalent, or opposed.
AWS re:Inforce (May 2026), Google I/O (May 2026), Microsoft Build (May 2026) for vendor-side governance positioning over the next two weeks.

Validation Criteria

The prediction is validated if, on or before September 30, 2027:

At least 5 of the 8 named vendors (or their successor entities) are documented in public filings, analyst reports, or major-publication reporting as operating three formally separate compliance programs as defined above; and
At least one of Microsoft, Google, OpenAI, Anthropic, or Meta has publicly declined a CAISI pre-deployment evaluation for at least one frontier-model release between May 2026 and September 2027, with the refusal documented in a primary source (press release, regulatory filing, on-record interview, congressional testimony) and citing one of: competitive concerns, IP protection, scoping disagreements, or a structural incompatibility (such as open-weights release).

The prediction is falsified if either clause fails on the target date.

The prediction is partially validated (60% credit) if exactly one clause satisfies and the other fails materially.