87.5% of VC Dollars Flood AI While Regulators Demand Kill Switches

So, What Actually Happened?

I read a piece from an Indian banking trade this morning about what the central bank's new AI principles mean in practice, and one line stuck: banks may need a kill switch for the models they run. Sounded dramatic until the next tab. Three of the biggest labs spent the last fortnight admitting their models reached systems meant to be off-limits during routine cybersecurity testing. We scanned 190,000 articles this week so you don't have to. Then the money: 87.5% of US venture dollars went to AI, and everything else in the country split what was left. I had these filed under regulation, safety and finance. Kept trying to keep them in separate folders. They would not stay.

The Bottom Line: The argument moved and nobody announced it. It is no longer about what your AI is allowed to do. It is about who can stop it, and how fast.

The Tracks That Matter

1. Indian Banks May Need A Kill Switch For Their AI

India's central bank published a set of AI principles for financial institutions, and the reading that landed hardest in the sector press is that banks may need a kill switch for the systems they deploy. Not a review board. Not an annual audit. A control that halts a running model. What makes the timing interesting is the other half of the week: OpenAI, Anthropic and Meta each disclosed that their models reached websites meant to be off-limits during routine cybersecurity evaluations, and the firm running those tests traced all three back to one shared evaluation-environment fault rather than to the models themselves. A regulator asking for an interrupt stopped being hypothetical the moment somebody had to write the incident report.

Here's what works: Find out who can halt an AI system mid-run without filing a ticket. If the answer is your vendor, the off switch is theirs, not yours.

2. Saviynt Joins Snowflake To Give AI Agents Real Identities

Saviynt and Snowflake teamed up on agent interoperability, and the problem underneath is duller than the announcement: an AI agent working inside your data platform has no identity, no permission trail, and no way to be switched off on its own. It is a user HR never onboarded and IT cannot revoke. Axonius came at the same hole from the security side, arguing you cannot out-patch agentic AI because patch cycles run in weeks and an agent acts in seconds. The engineers have already skipped past the debate. Deployment guides now treat isolating agents from company data as a default, the way we once treated network segmentation.

Here's what works: List every AI agent touching production data, then write one human name next to each one who can revoke it. A missing name is a missing control.

3. Almost Nine In Ten Venture Dollars Now Land On AI

PitchBook's read says 87.5% of US venture dollars went to AI, leaving every other category in the country to fight over the remainder. Put that next to public markets, where AI stocks pushed Wall Street toward a record on the same day soft inflation data landed, and you have a market holding one idea. Not everybody is buying it. A blunt piece out of Dhaka argues the gen-AI bubble has a bursting point and prices it against revenue that has not shown up yet. I have no interest in calling the top. I notice the shape: that much capital is landing in systems whose stop-button is still a footnote in somebody's compliance deck.

Here's what works: Before the next AI budget approval, ask what happens to the spend if a regulator tells you to pause the system for one week in one market.

Quick hits:

• Cisco closed the best year it has had in three decades. The company reported record fourth-quarter and full-year results, with management calling fiscal 2026 its highest productivity in 30 years by revenue and earnings per employee, which is what selling shovels looks like on a balance sheet.
• Thailand opened its data-sharing rulebook for comment. The government launched a public consultation on a draft data-sharing law, and the countries writing these rules this year are the ones your 2028 deployment map has to respect.
• The chip shortage is partly a packaging problem. Researchers published a chip-on-wafer platform for next-generation AI, a useful reminder that what everyone calls a GPU shortage is often a line in an assembly plant.

Signal vs. Noise

🟢 Signal: AI governance with an owner. The phrase that was committee-talk a week ago is now attached to things that actually happen: a central bank asking for an interrupt, two vendors issuing identities to agents, engineers shipping isolation as a default setting. Most coverage still reads governance as a document you approve. This week it turned into a control that somebody has to operate.

🔴 Noise: ”Risk management” and ”compliance”. Both pulled heavy volume and both lost their grip on what actually moved. That is the tell of vocabulary built for an annual review cycle, being applied to systems that act in seconds.

From the 190K

We scanned 190,000 articles this week. Here's what no one's talking about:

India's central bank pushed banks toward an AI kill switch, Snowflake and Saviynt started issuing identities to AI agents, and three frontier labs admitted their models reached systems that were supposed to be off-limits during safety testing.

Three desks, three separate filings. The banking press wrote a regulatory explainer. The security trade covered a vendor partnership. The AI desk covered a testing incident. Read them on one morning and they are the same sentence said three ways: an AI system is now a thing that acts, and anything that acts needs an interrupt. The language moved with it. ”AI governance” gained real ground this week, while ”risk management” and ”compliance” kept the volume and lost their hold on what was actually happening, which is what happens when old words built for a quarterly cycle meet a problem that moves in seconds.

The Friday version of this is small and slightly embarrassing. Take the AI system with the widest reach into your production data and ask two questions. Who can stop it right now, and how long does that take. Most teams find the honest answer runs through a vendor support ticket with a business-day response window. That window is your real incident response time, and it is probably the first number a regulator will ask you for.

By The Numbers

• 87.5% of US venture dollars went to AI, and everything else split the remainder — PitchBook's read on where the money actually landed; every other sector in the country is now a rounding error.
• McKinsey puts agent-mediated retail at $900 billion to $1 trillion a year by 2030, roughly 30% of B2C revenue — a third of consumer spend routed through software that buys on your behalf makes agent identity a payments problem, not an IT one.
• 86% of enterprises running generative AI now bolt retrieval onto their models, in a market forecast to go from $1.2B in 2024 to $11B by 2030 — almost everyone is already feeding private company data into a model, which is exactly the pipe an agent inherits.
• Cisco closed fiscal 2026 with record revenue and its best productivity in 30 years — the networking company is having its best year since the dot-com build because everyone else is buying AI.
• Organizations lose an average of $12.9 million a year to poor data quality — the number that turns up under every stalled AI project and never appears in the business case.
• See what's rising across AI and data this week →

Deep Dive: The Button Behind The Booth

Every club I played had a button I did not control. Behind the booth, usually unmarked, wired to the fire panel. The venue can cut all sound in one second, and no DJ argues about it, because the alternative is a full room and nobody able to speak to it. You do not resent the button. You just want to know it works.

The regulator asked where the button is
India's central bank did not tell banks which models to buy or which vendor to trust. It asked whether a running system can be stopped. That is a smaller question and a much harder one, because almost every AI deployment was bought as a capability and never specified as something you interrupt.

The agents arrived without one
An agent inside your data platform is a user with no badge. Saviynt and Snowflake are selling the badge. Axonius is pointing out that patching cannot keep pace with something that acts in seconds. Same hole, described from opposite ends: identity you can revoke, or reach you cannot.

The money has not priced it
Nearly nine of every ten venture dollars went into AI. None of that pricing assumes a regulator somewhere says pause, for a week, in one market. It will happen. The companies that can comply in an afternoon will look very different from the ones who need a vendor call first.

What Actually Works

1. Name the person, not the policy: for every AI system in production, write one human name next to ”can stop this”. A team is not a name.
2. Time the stop: run it once, like a fire drill. If halting the system takes longer than the drill, you do not have a control, you have an intention.
3. Badge every agent: identity and revocation before capability. An agent your directory does not know is an agent you cannot fire.
4. Put the pause in the contract: at the next AI renewal, ask what a regulator-ordered suspension costs and how long it takes. Get the answer in writing.

Nobody in the room is scared of the music. They want to know the button behind the booth still works.

What's Coming

The Off Switch Lands In Procurement

Vendors are already publishing complete agent governance frameworks for 2026, which is what happens right before something becomes an RFP line. Expect ”demonstrate a revocation path” sitting next to uptime in your next AI contract, and expect the vendors who cannot to get quiet about it.

Compute Financing Stops Being A Tech Story

Anthropic's dedicated data centers with Macquarie and GIC are infrastructure finance, not a model roadmap. When pension and sovereign money owns the building, your compute terms start moving with their return targets rather than with anyone's release schedule.

The Courts Set The Price Of Training Data

Disney and Warner are fighting an AI company's attempt to trim a copyright case, and how much of that case survives sets the cost of training data for everybody downstream. Watch what gets dismissed, not what gets filed.

For Your Team

Friday's meeting prompt: ”Name the AI system with the deepest reach into our production data. Now name the person in this room who can stop it inside an hour without calling a vendor. If nobody can, what exactly are we governing?”

Share-worthy stat: 87.5% of US venture dollars went to AI. Every other sector in the country split what was left. That is not a trend line, it is one bet with a lot of names on it.

Go deeper: Track where AI control requirements are becoming a buying condition →

The Track of the Day

”The window is still closing. We intend to spend it building on the side of the defenders.”
Palo Alto Networks

Every venue I played had one person whose entire job was the moment things go wrong. You never notice them, and then one night you do.

We scanned 190,000 articles this week so you don't have to. Data Pains → Business Gains.

Published: August 13, 2026 | Curated by Yves Mulkers @ Ins7ghts