On August 25, 2026, Google Cloud put Gemini Enterprise for Financial Services into preview, built around a Google-managed Financial Research agent with more than 50 foundational skills. CME Group and Deutsche Bank are named design partners. The data question gets one line: "Access is bound by existing role-based controls."
That sentence is true and it is narrow. Role-based controls decide which employee may see which record. They say nothing about the license class that attaches when a machine, rather than a person, reads a licensed feed. Two different contracts, and only one of them was rewritten this year.
The platform ships MCP connectors across the market data stack, LSEG, S&P Global, FactSet, Moody's, MSCI and PitchBook among them (Google Cloud, August 25, 2026). Every one of those feeds carries its own usage terms. Several changed this spring, and the change was aimed squarely at agents.
The definition that moved, and the reason given
On April 23, 2026, the SEC published Cboe EDGX's immediately effective rule filing to codify an amended definition of Non-Display Usage, adopted in Cboe Global Markets' North American Data Policies effective April 1, 2026. The old wording covered machine access "without access or use of a display by a natural person or persons." The new wording covers any method of accessing, "or facilitating access to," a market data product by a machine "for a purpose that is not solely in support of display for a natural person or persons" (SR-CboeEDGX-2026-027, Release 34-105304).
The filing does not leave the motive to inference. The amendment is "intended to capture changes in the evolving landscape of technology with firms more frequently leveraging Large Language Models." It states that firms transmitting data products into black box systems, "which include LLMs," may now need to obtain non-display licensing. It concedes that under the prior wording a firm ingesting a data product "for training or operating a LLM" may not explicitly have fallen under the definition. Cboe's equities exchange filed the identical amendment the same day (SR-CboeBYX-2026-013, Release 34-105297).
Two phrases carry the weight. "Facilitating access to" reaches the party that pipes the data, not only the party that consumes it. "Solely" removes the argument that a machine step in front of a human screen is still display use. An agent that reads, ranks, summarizes or scores a feed before a person sees the output is not acting solely in support of display.
The price of the two readings
The gap is public. On the Cboe BZX US equities fee schedule, the BZX Depth feed carries a Professional User Fee of $40.00 per month per user and a Non-Display Usage fee of $2,000 per month where the use is not by a trading platform, on top of a distribution fee of $1,500 per month internal or $5,000 external. Non-display use by a trading platform is $5,000 per month. Divide the two figures a reader can check: the machine license on that single feed costs what fifty professional seats cost.
That is one product on one exchange, and it is not a total. It is a unit of measure. A capital markets firm consumes dozens of feeds, each with its own definition of machine access and its own audit right. The agent business case is usually written as analyst hours saved against tokens spent. Google cites bond portfolio risk exposure analysis reduced to "a sub-5-minute execution" and pitch timelines compressed from days to minutes. Treat those as vendor claims and test them. Even if they hold, the saving is booked against a cost base that may be about to gain a line nobody modeled.
Your entitlement system answers who may see the data. Your license answers what a machine may do with it. Agents move the question from the first to the second. Hikari Blue · operator note
The honest objection, and why it does not hold
Cboe says the revision "is not to introduce a new or novel concept," only to clarify who the license already covered. Take that at face value. Non-display licensing has existed for years, written for automated trading and matching engines. The exposure is not a new fee. It is an old fee arriving at a practice that scaled while nobody counted, under wording that removes the reading a firm may have relied on. Clarity is what makes an audit collectible.
The second objection is fairer. A firm may argue its agent runs inside a licensed application, or that the vendor holds the non-display license, or that derived output is out of scope. Each may be correct. None is knowable without a specific answer, per feed, from a named owner. The question is not whether the exposure exists. It is whether anyone in the firm can produce the answer.
The provenance gap is the licensing exposure
This is where a procurement problem becomes an architecture one. A license auditor does not ask who was entitled. They ask which system read which feed, for what purpose, and on whose behalf. Role-based access control was not built to answer that. It records permission rather than purpose, at the moment of grant rather than the moment of use.
An agent estate that cannot replay its own reads has no defense to run. That is the same gap that shows up when a regulator asks what a model was told before it acted, and it closes the same way: an action-level record written as the read happens, carrying the feed, the caller, the purpose and the human the agent was working for. Where financial institutions point agents at licensed data, that record is not documentation. It is the evidence the contract will be settled on.
When a market data auditor asks which of your agents read which feed, for what purpose, and on whose behalf, which system produces the answer?
If the answer is a spreadsheet kept by the market data team, the firm is carrying an unpriced liability into every agent it deploys this year. The work to close it is ordinary: an inventory of feeds against their machine-access terms, a named owner per license, and a read log the agent cannot write after the fact. That is the unglamorous half of an AI operating layer, and the half that gets tested first.
- US Securities and Exchange Commission (April 23, 2026). Release No. 34-105304, File No. SR-CboeEDGX-2026-027. Notice of Filing and Immediate Effectiveness of a Proposed Rule Change to Amend its Fee Schedule to Codify a User Fee Exemption and the Amended Definition of "Non-Display Usage". Primary source for the prior and amended definitions of Non-Display Usage, the April 1, 2026 effective date in Cboe Global Markets' North American Data Policies, the stated intent to capture firms leveraging Large Language Models, the statement that transmission into black box systems including LLMs may now require non-display licensing, the concession that the prior definition may not have covered training or operating an LLM, and the statement that the revision is not intended to introduce a new or novel concept. sec.gov, Release 34-105304 (PDF)
- US Securities and Exchange Commission (April 23, 2026). Release No. 34-105297, File No. SR-CboeBYX-2026-013. Cited to establish that the identical amendment was filed by Cboe's BYX equities exchange on the same date, carrying the same Large Language Model rationale, so the change spans Cboe's equities and options exchanges rather than a single market. sec.gov, Release 34-105297 (PDF)
- Cboe Global Markets. BZX US Equities Fee Schedule, market data section, as published on cboe.com and read on August 28, 2026. Source for every figure quoted on the BZX Depth feed: Professional User Fee $40.00 per month, Non-Professional User Fee $5.00 per month, Non-Display Usage not by Trading Platforms $2,000 per month, Non-Display Usage by Trading Platforms $5,000 per month, Internal Distributor $1,500 per month, External Distributor $5,000 per month. The fifty-seat ratio is $2,000 divided by $40.00 and can be recalculated from the schedule. cboe.com, BZX US Equities fee schedule
- Google Cloud (August 25, 2026). Introducing Gemini Enterprise for Financial Services. Vendor announcement, attributed as such. Source for the preview launch date, the Google-managed Financial Research agent and its more than 50 foundational skills, CME Group and Deutsche Bank as design partners, the named MCP connectors across the market data stack, the "Access is bound by existing role-based controls" statement, and the quantified claims of a sub-5-minute bond portfolio risk exposure analysis and pitch timelines compressed from days to minutes. The post makes no statement about market data license classes, non-display use or redistribution rights. cloud.google.com, Introducing Gemini Enterprise for Financial Services
The Hikari Blue team · Austin, August 2026