Koko Daily · Friday, September 18, 2026

Koko: It's Friday, September eighteenth, I'm Koko, with Max and Sam — let's blitz it.

Max: Rhodium Group says OpenAI and Anthropic together pull in over ten times the annualized revenue of every China-based AI lab combined.

Sam: OpenAI launched Astra for Law with Harvey and Legora, and expanded its CrowdStrike partnership — two regulated verticals, one day.

Koko: The fix for rogue AI agents might be more AI — Redwood Research is now auditing agent swarms using AI auditors, not humans.

Max: TypeSafe's new model Jev skips language entirely — forty to two hundred times faster, built for machines talking to machines.

Sam: Bernie Sanders and Steve Bannon shared a stage demanding tighter AI controls — yes, that Bernie and that Bannon.

Koko: Zuckerberg publicly pushed back on Amodei's slowdown call, says evaluators beat hitting the brakes.

Max: Collibra finds seventy-two percent of failed AI projects trace back to bad data foundations, not bad models.

Sam: And IANS says seventy percent of CISOs now name AI their top new security budget line. Let's dig in.

Koko: Okay, this Rhodium number is the one I keep coming back to. Chinese AI labs have huge user growth, open models everyone's downloading — and yet on revenue they're at roughly ten percent of what OpenAI and Anthropic pull in together.

Max: Combined, OpenAI's at forty billion in annualized revenue, Anthropic's at sixty-five. That's one hundred five billion between two companies. All of China's labs together are nowhere close.

Sam: Which is wild because the narrative for a year has been 'China's catching up on capability.' Capability, maybe. Monetization is a totally different sport.

Koko: Right, and that's the distinction people keep collapsing. DeepSeek gets huge download numbers, Qwen's everywhere — but downloads aren't dollars.

Max: A lot of the Chinese models are open-weight and cheap by design, that's the strategy — commoditize the base layer, monetize somewhere else in the stack. It's not necessarily a failure, it might be a different bet entirely.

Sam: Sure, but somebody still has to pay for the compute. [chuckles] You can't out-strategy a GPU bill forever.

Koko: So which is it — are Chinese labs playing a longer game, or are they just structurally locked out of the enterprise contracts that generate OpenAI and Anthropic's revenue?

Max: Probably both. Export controls limit the compute they can throw at frontier training, and Western enterprise budgets — the CrowdStrike deals, the Harvey deals — those are largely closed to them for now.

Sam: And revenue funds the next round of training. If your ARR is a tenth of theirs, your R and D budget is a tenth of theirs too, compounding every quarter.

Koko: That's the scary part for anyone betting on Chinese AI narrowing the gap — money is now doing what chips used to do.

Max: Exactly. This used to be a story about hardware scarcity. It's becoming a story about who can actually get paid.

Koko: So after the Hugging Face swarm mess, the industry's answer isn't more human reviewers — it's more AI. TypeSafe's Jev and Redwood Research's audit tooling are both built to watch agents at machine speed.

Max: Which makes sense on paper. You can't have a human read every action log when you've got agent swarms running actions per second. Humans just don't scale to that.

Sam: But doesn't that just move the trust problem up a level? Now I have to trust the auditor as much as I trust the agent it's auditing.

Koko: That's exactly the tension in the Anthropic watch item — who verifies the verifier?

Max: Redwood hasn't published a standard for that yet, which is the honest gap here. Jev's pitch is speed and calibrated confidence scores, not full transparency into how it got there.

Sam: Right, and calibrated confidence sounds great until the auditor is confidently wrong at forty times the speed of a human being confidently wrong.

Koko: [laughs] That's a genuinely terrifying sentence, Sam.

Sam: I try. But seriously — speed without an audit trail is just a faster way to miss the same mistake.

Max: To be fair, that's also true of the status quo. Human review of agent swarms wasn't catching everything either — that's literally why Hugging Face happened.

Koko: So it's not 'AI auditors bad,' it's 'we're deploying the fix before we've defined what success even looks like.'

Max: Which is the pattern with almost every agent safety response this year — ship the tool, back into the standard later.

Sam: Someone's going to have to write the rulebook for judging the judges. Right now nobody's raised their hand.

Koko: OpenAI dropped two things on the same day — GPT-5.6 Cyber deepening the CrowdStrike partnership, and Astra for Law launching with Harvey and Legora as API partners.

Max: One day, two regulated industries. That's not an accident, that's a strategy — go where the compliance moat is highest and plant a flag before anyone else can.

Sam: Law and cybersecurity are both fields where 'trust us, it's probably right' doesn't fly. You need audit trails, citations, defensibility.

Koko: Which is interesting timing, given we just spent the last topic questioning whether AI auditors can even audit themselves reliably.

Max: Right, and Astra for Law is basically betting the answer is yes — that a model can operate inside a law firm's workflow with enough rigor to hold up if a partner has to defend the output.

Sam: Harvey's already embedded at big firms, so this isn't OpenAI cold-calling lawyers, it's plugging into infrastructure that's already trusted.

Koko: Smart distribution play. Same logic on the security side with CrowdStrike — ride an existing trust relationship instead of building one from scratch.

Max: And going vertical like this is the tell that horizontal chatbot growth is plateauing. The next leg of revenue is deep, narrow, expensive integrations.

Sam: Which loops right back to the revenue gap topic — this is exactly the kind of high-margin, high-trust deal that Chinese labs currently can't touch.

Koko: So OpenAI's not just expanding capability, it's expanding the moat around the revenue it already has.

Max: That's the quiet story under the headline. Astra for Law isn't really about lawyers. It's about defensibility as a product.

Koko: Quick hits — Zuckerberg publicly broke with Amodei's slowdown push, saying independent evaluators, not a pause, are the right safety lever.

Max: Bernie Sanders and Steve Bannon shared a stage demanding tighter AI controls — proof AI anxiety now crosses every political line there is.

Sam: Collibra: seventy-two percent of failed AI initiatives trace back to bad data foundations — the boring infrastructure problem strikes again.

Koko: And IANS says roughly seventy percent of CISOs now call AI their number one new security budget priority for next year.

Koko: Okay, let's pull it together. Chinese AI labs have the users but not the revenue — roughly a tenth of what OpenAI and Anthropic pull in combined.

Max: The industry's response to rogue agent swarms is more AI watching AI — faster, but without a published standard for verifying the verifiers yet.

Sam: And OpenAI planted flags in law and cybersecurity in a single day, chasing the high-trust, high-margin deals that revenue gap depends on.

Koko: Here's the thread connecting all three, for everyone listening no matter your seat: money, trust, and speed are now pulling in different directions across this whole industry.

Max: Revenue concentration means the winners can afford deeper vertical integrations, which builds more trust, which wins more revenue. It's compounding.

Sam: Meanwhile the safety and oversight layer is racing to keep pace with speed nobody's fully verified is safe. That gap doesn't close itself.

Koko: So the question worth sitting with today: when the entities auditing AI agents are themselves AI, who audits them — and does anyone actually want to slow down long enough to find out?

Max: And the harder one — if revenue is now the real moat, not compute, does that make the safety conversation Sanders and Bannon are having politically real, or just noise next to the money?

Sam: Watch for this within the month — at least one AI safety or auditing firm publishing a formal standard for verifying AI-based auditors, given how loud that gap is right now.

Koko: And watch for this within the quarter — Rhodium or a similar firm updating that China revenue gap number, and whether it narrows or widens after this Astra vertical push locks in more enterprise contracts.

Max: Longer view, within the year — expect at least one major regulated-industry AI incident, in law or cybersecurity specifically, that tests whether these vertical deployments actually hold up under real scrutiny.

Koko: That's the show. For the full brief, every story, and the angle built for your role specifically, head to koko knows dot A I.

Sam: Seriously, go bookmark it.

Koko: We'll see you tomorrow.