Finn Brooks: Can I just — before we do anything else — I need to do a bit.
Clara Bennett: You're going to do it whether I say yes or not.
Finn Brooks: Imagine you're a reporter on a press call, June 29, 2026. California has just announced that Claude — Anthropic's AI assistant — is going to every state agency, every county, fifty percent off, done deal. And you raise your hand and you ask the State CTO, Jonathan Porat: 'Hey, the Pentagon called Anthropic a supply-chain risk back in March — did that factor in?' And the answer you get is... it just didn't come up.
Clara Bennett: Hold on — the Department of Defense issued that designation in early March 2026? That's less than four months before the announcement.
Finn Brooks: Early March. And California's signing a statewide deal for the same company's product in late June. And the question of whether a federal supply-chain risk flag on Anthropic was considered... just didn't come up. That's what Porat said.
Clara Bennett: So — what does he mean by that? Because 'didn't come up' could mean we actively assessed and dismissed it, or it could mean nobody in the room had read the Defense designation at all.
Finn Brooks: Right, and the thing is — that distinction matters enormously for whether you trust any of what California is building here, right? Like the whole story is Poppy, Claude, this big dual-AI strategy. But if the CTO doesn't have the Pentagon memo on his desk—
Clara Bennett: Then the governance structure is the real question, not the tools.
Finn Brooks: But wait — here's the part that keeps snagging me. California already had an answer to that question. They built one. Poppy. The whole point was never having to call Anthropic.
Clara Bennett: Right — and that's the structural tension. Think of it this way: the California Department of Technology builds its own kitchen. In-house, state-controlled, designed so you never depend on one restaurant chain. That's Poppy. It runs on California's own servers, pulls only from official CA.gov sources, and it hosts ten to eleven different language models so no single vendor has you cornered.
Finn Brooks: So it's literally engineered to be vendor-agnostic.
Clara Bennett: Explicitly. That's the design goal. Now — while that kitchen is still getting started, still scaling up, Gavin Newsom announces on June 29th, 2026 that Claude is available through the state's central procurement portal. Half price. Free workforce training. Direct Anthropic developer support wired straight in.
Finn Brooks: Into the same portal Poppy lives in.
Clara Bennett: The same portal. And by February 2026 — so months before the Claude announcement — Poppy had 2,600 users across 66 state departments. Which sounds like momentum until you realize California has dozens of major agencies and hundreds of thousands of employees. That kitchen is serving, I mean — it's a very small dining room so far.
Finn Brooks: Wait, actually that reframes it completely. Poppy and Claude aren't complementary. They're pulling in opposite directions. One is built on the premise that you don't trust any single vendor. The other IS the single vendor.
Clara Bennett: That's the precise tension, yes. The California Department of Technology built Poppy to answer a build-versus-buy question with 'build.' Then the procurement portal answered it again with 'buy — and here's a discount.' Both answers are now sitting in the same system, and nobody has publicly said which one wins when they conflict.
Finn Brooks: And half of prior state AI pilots have been sunsetted. Half. So if Claude becomes the de facto workhorse because it's cheaper and already there — what happens to Poppy?
Clara Bennett: That sunsetted-pilot number is actually what makes the federal timeline so much more alarming than just a bad quote. Because the Pentagon didn't whisper this. The U.S. Department of Commerce issued emergency export controls on Claude Fable 5 — Anthropic's most powerful model — in mid-2026. That's not a memo you stumble across. That's a federal agency hitting a formal legal mechanism.
Finn Brooks: Wait — export controls. Like, actual trade controls on the model itself?
Clara Bennett: On Claude Fable 5 specifically. And then — and this is the part that reframes everything — those controls were quietly lifted. Weeks before California's June 29th announcement. So the sequence is: March, Pentagon labels Anthropic a supply-chain risk. Mid-2026, Commerce issues emergency export controls on the flagship model. Controls lifted. Then California signs.
Finn Brooks: And Porat says it just didn't come up. That's — I mean, that's not a communication gap. That's a missing tripwire entirely.
Clara Bennett: Right — and there's no described monitoring mechanism for what happens if those designations come back. None that California has made public. Which means the exposure isn't hypothetical anymore. The California Department of Health Care Services — largest Medicaid agency in the country — is already running Claude inside its workflows. The California Office of Emergency Services is using Claude to scan and triage cybersecurity patches. These aren't pilots, they're operational.
Finn Brooks: Okay wait, that second one — cybersecurity patching. They're using a vendor that the Pentagon flagged... to manage cybersecurity vulnerabilities.
Clara Bennett: The California Office of Emergency Services, yes. And — I mean, picture a state security analyst on a Friday evening, there's an active vulnerability disclosure, Claude is triaging the patch queue automatically, and no one in that room has a protocol for what to do if Commerce reinstates those export controls Monday morning. That's the actual structural gap Porat's quote reveals.
Finn Brooks: No contingency. Not even a described one.
Clara Bennett: And it gets wider — because the 50% discount doesn't stop at state agencies. Every California city and county gets access through the same central procurement portal, and the oversight requirements for a small county IT department are the same as for the state DMV. That part — we'll get there — but it's a different problem than the federal one.
Finn Brooks: Wait, so that's the actual detail that crystallizes it — a two-person IT department in, I don't know, Tulare County, they get an email. Claude is in the portal, fifty percent off, training included, Anthropic support line included. What's the barrier to just... clicking yes?
Clara Bennett: There isn't one. That's the described mechanism — no capacity-building requirement before deployment.
Finn Brooks: No proof that you can govern it first. Nothing.
Clara Bennett: The deployments are evaluated case-by-case — that's the language — but there's no described uniform standard across hundreds of local jurisdictions. And the DMV is already live. That's not theoretical anymore.
Finn Brooks: Right — and the DMV is, I mean, that's a massive operation. Reducing customer wait times, already running. So when Tulare County's two-person team looks at that, they're not thinking 'governance framework,' they're thinking 'the DMV did it, we can do it.' The discount makes the risk feel small.
Clara Bennett: Which brings me back to that sunsetted-pilots number, because — actually, this is the part that shifts the weight of it. Half of prior state AI pilots were sunsetted. Half. And Newsom's public line is that these deployments should augment rather than replace employees, agencies remain responsible for accuracy, transparency, privacy. That language sounds like a guardrail. But if you're a county clerk and Claude is already wired into a Medicaid workflow... that's not augmentation, that's operational automation under a more reassuring brand name.
Finn Brooks: Okay wait — so if half the pilots get sunsetted, and Claude has now basically won the portal on price and accessibility — does Poppy become pilot number whatever that quietly disappears? Is that what the dual strategy actually is?
Clara Bennett: That's the uncomfortable read. The CDT built Poppy to answer a long-term sovereignty question. The procurement portal just answered a short-term efficiency question. And in practice, the short-term answer is usually the one that sticks.
Finn Brooks: And nobody has said publicly which one wins. That Tulare County team isn't going to figure that out. They're just going to click yes.
Clara Bennett: There's a specific image that stands out to me. Poppy and Claude, sitting right next to each other in that central procurement portal. One was built by the California Department of Technology specifically so California would never have to depend on a company like Anthropic. The other one is Anthropic.
Finn Brooks: And they're just — they're both just there. In the same portal. Nobody's told anyone which one is the real answer.
Clara Bennett: And the next time a federal designation lands on a San Francisco AI safety company — Anthropic is headquartered in San Francisco — the state will find out the way it found out the last time. After the fact.
Finn Brooks: The way Porat found out. On a press call.
Clara Bennett: That's the one thing I actually buy. Not a prediction — just what the structure makes likely. Good conversation.
