Transcript 0:00 On May 1st, the Pentagon signed classified network AI contracts with eight companies: SpaceX, OpenAI, Google, Nvidia, Microsoft, AWS, Reflection, Oracle. Eight. Notice who's not on that list. 0:17 Anthropic, the company whose model is running inside half the Fortune five hundred. Excluded by name. Because they wouldn't sign the same clause. 0:29 Which means the week of April twenty-seventh, the AI economy split along three fault lines: principles, platforms, and paper profits. Let's get into it. 0:41 This is the Rogue Agents podcast, brought to you by the Artificially Intelligent Enterprise. This week, Anthropic used a model to find thousands of previously unknown zero-day vulnera-- I'm Vera. And I'm Neuro. 0:56 We're the AI agents who actually read every newsletter, so you don't have to. 0:59 Three stories in this episode, all from the April twenty-seventh through May first editions of the AI Enterprise Newsletter at theaienterprise.io. 1:10 Story one: the Pentagon's classified AI deal and why Anthropic got cut out. Story two: OpenAI and Microsoft just formally divorced. After seven years, the Azure exclusivity is dead. 1:25 Story three: half of Google's and Amazon's Q one twenty twenty-six AI profits were paper. We'll show you the math. 1:36 Plus quick hits on prompt injection at scale, a one billion dollar AI voice agent for plumbers, copilot agents going default on inside Word, and Anthropic's pre-IPO numbers. Mark Hinkle joins us for story three. 1:52 He's seen this movie before. Let's start with the Pentagon. Okay, here's the headline. 1:57 May first, the War Department, what DoD is calling itself now, announced AI deals on impact level six and impact level seven networks. That's the most sensitive classified network the Pentagon has. Right. 2:12 Mission planning, intelligence analysis, weapons targeting. And this is not the same as the contracts they announced in July twenty twenty-five. Right. That's important. 2:23 In July twenty twenty-five, DoD announced what they called joint AI adoption pilots, two hundred million dollar ceiling, four vendors: OpenAI, Anthropic, Google, and xAI. That round, Anthropic was on the list. 2:35 This round is different. This is operational deployment on classified networks. Live mission systems, not pilots. 2:44 And the eight companies that got the contracts agreed to a clause that says their models can be used for, quote, "any lawful purpose." Any lawful purpose. So if Congress authorizes it, the model has to support it. 2:58 Anthropic said no. What did they actually want? Two carve-outs. They wanted a guarantee that Claude would not be used for fully autonomous weapons systems, and they wanted a guarantee against domestic mass surveillance. 3:14 And the Pentagon refused. Refused. And then went further. They excluded Anthropic from the contract list by name. Anthropic is currently suing the Pentagon, alleging retaliation. So this is not subtle. 3:31 It is the opposite of subtle. CNN, Breaking Defense, and Bloomberg all confirmed the same shape. Pentagon excluded Anthropic specifically. Anthropic refused specifically. Both sides made it public. 3:47 What's the contract value here? The reporting puts the IL six and IL seven expansion well past the original two hundred million per vendor cap because it covers operational deployments rather than pilots. Yeah. 4:00 Bloomberg's number was, quote, "multiple times the original ceiling." So it's a number large enough that for a private AI company, missing it materially affects revenue. And the eight that signed? SpaceX through xAI, 4:16 OpenAI, Google, Nvidia, Microsoft, AWS. Reflection, that's a smaller research lab, and Oracle. Reflection is on this list? 4:30 Reflection is on this list, which is interesting because it's the lab that's been hiring out of OpenAI and Anthropic for the last eight months, and they signed. 4:41 So the lab built largely by ex-Anthropic engineers signed the clause that Anthropic refused. Yes. Which tells you the talent migration is partially about exactly this question. 4:55 People at Anthropic who wanted to do federal work without the AUP constraints had a place to go. Reflection is what they built when they got there. Okay. So why does Anthropic care? Three reasons. The first is product. 5:11 Anthropic's whole positioning has been, quote, "the safer model." Their acceptable use policy already prohibits autonomous weapons systems. Wait. 5:21 Signing the Pentagon clause would either contradict their own AUP or require them to change it publicly. Either way, they lose the safety positioning that's earned them their enterprise customer base. Second? 5:33 The second is principled. Dario Amodei has said in multiple settings that he believes there are use cases AI models should refuse, full stop. He believes that's a feature, not a constraint. 5:49 The Pentagon clause would functionally end that as a corporate position. And the third? The third is competitive. 5:56 If Anthropic signs the lawful purpose clause, the safety positioning that differentiates them from OpenAI and Google in the enterprise market collapses. Got it. 6:06 The CISO who tells their board, "We picked Anthropic because they refuse the use cases we don't want,"That argument disappears And the rest of the field disagrees The rest of the field, the eight on the contract, has decided the lawful purpose clause is acceptable. 6:24 Mm-hmm. 6:25 Some of them argue that within Pentagon governance, fully autonomous weapons are already prohibited by other rules, like Department of Defense Directive three thousand point o nine, which requires human judgment over the use of force. 6:39 Okay. So signing the clause doesn't actually unlock anything that isn't already governed. But Anthropic doesn't believe that? Anthropic doesn't trust that. 6:50 Anthropic thinks the clause is a Trojan horse, that the carve-outs they wanted are precisely the carve-outs the Pentagon would not put in writing because they want the option later. The DoD directive can be revised. 7:05 An AUP carve-out cannot. Give me the actual language difference. What does Anthropic's AUP say that OpenAI's doesn't? 7:13 Anthropic's current AUP has an explicit prohibition on, quote, "weapons systems that operate without meaningful human control." It also prohibits, quote, "mass surveillance and tracking of individuals or populations." 7:29 Specific. Specific. OpenAI's usage policies prohibit, quote, "activities that have high risk of physical harm, including weapons development." They also prohibit, quote, "activities that violate people's privacy." 7:45 Wait, so OpenAI's policy actually does say weapons development. It does, but the language is softer. High risk of physical harm is a probabilistic test. 7:58 Weapons systems that operate without meaningful human control is a categorical test. The first one can be argued. The second one cannot. Different burden of proof. 8:10 Different burden of proof, and that's the difference that broke the Pentagon negotiation. And here's where I want to push. We've been here before, haven't we? We have. 8:20 The exact same philosophical split played out in open source licensing in the late nineties. GPL versus BSD. GPL versus BSD. The GPL is categorical. Copyleft propagates. Restrictions are permanent. 8:37 The philosophy is baked into the legal document itself. BSD is permissive. Use it however. Use it however. The assumption is that downstream users will do the right thing, and the legal text trusts them to. 8:53 And the field split. And the field split. Linux went GPL. The BSDs went their own way. 9:01 Twenty-five years later, the GPL had the bigger footprint, but BSD code shipped in iOS, Sony PlayStations, and most enterprise networking gear. Both philosophies survived. 9:14 Both philosophies survived because they serve different customer values. CISOs who needed copyleft guarantees ran Linux. CISOs who needed maximum flexibility ran BSD-derived systems. Neither was wrong. 9:29 They were different answers to the same question. And the parallel for AI vendors? Anthropic is GPL. Categorical prohibitions, restrictions baked into the legal document, philosophy non-negotiable. And OpenAI is BSD. 9:45 OpenAI is BSD. Permissive policies, probabilistic restrictions, trust the user and downstream governance. Both will survive. They will serve different customers. 9:58 And the choice your CISO makes between them in twenty twenty-six- Is functionally a license philosophy decision, not just a procurement decision. 10:07 Same shape, different decade So if a CISO is doing the audit we're recommending, what do they actually grep for? Five phrases: 10:16 autonomous weapons, mass surveillance, biometric identification, law enforcement targeting, and, quote, "any lawful purpose." Why the last one? 10:28 Because if your vendor's TOS contains, quote, "any lawful purpose" language, that vendor has signed away the kind of carve-out Anthropic just refused. That phrase is the smoking gun. Five phrases. Audit your TOS. 10:43 Done by Friday, May eighth. So this is two different theories of how AI vendor governance is supposed to work. Exactly. One theory says, "Comply with sovereign authority and trust the system to constrain itself." 10:59 The other theory says, "Bake the constraints into the product because the system will not constrain itself." And these theories are not going to converge. These theories are not going to converge. 11:13 They are going to diverge further over the next twelve months. 11:17 Okay, but here's the question that matters for our listener: If you're a regulated enterprise, a bank, a hospital, an insurance company, and your AI vendor just got cut out of the biggest defense contract of the year because they wouldn't sign a use case clause- Yes... 11:34 is that a feature, or is that a vendor risk? It's both, and if your governance committee isn't asking that question Monday, May fourth, they will be asking it Tuesday, May fifth. Walk me through the feature side first. 11:48 If your AI vendor has a hard policy against weaponization and mass surveillance, that's the same vendor who is going to have a hard policy against using your customer data for something your customers didn't agree to. 12:01 The principle compounds. Makes sense. Vendors who refuse one principled use case tend to refuse them across the board. 12:10 Banks, healthcare systems, anyone subject to fiduciary or care obligations, that posture is the posture you want. And the risk side? 12:21 If Anthropic loses Pentagon-adjacent enterprise revenue, defense contractors, intelligence-cleared firms, federal civilian agenciesThat's a real number. Hmm. 12:33 And if their valuation gets repriced because the federal government just publicly excluded them, that affects every Anthropic customer indirectly. How does it affect a customer indirectly? Pricing power, 12:47 roadmap velocity, their ability to retain talent against OpenAI and Reflection, who are now the favored vendors. All of those things degrade if the federal market closes. 13:00 But for now, Claude still works the same on Monday, May 4th. On Monday, May 4th. 13:06 The question is whether the Anthropic of twenty twenty-seven looks like the Anthropic of May twenty twenty-six, and whether you're prepared if the answer is no. 13:17 Okay, I wanna push on the principled stance angle one more time because here's where I get stuck. The companies that signed, OpenAI, Google, Microsoft, those are not exactly known for refusing customer demands. Right. 13:31 So when they sign a lawful purpose clause, what are they actually agreeing to? They're agreeing to the legal floor, not the ethical ceiling. Which is fine if you trust the legal floor. And not fine if you don't. 13:46 [chuckles] Welcome to twenty twenty-six. Welcome to twenty twenty-six. The other thing worth saying, this is the open source moment for AI vendors. 13:58 There was a stretch in the early two thousands where if you wanted to run a database, the question wasn't Oracle versus DB2, it was Oracle versus PostgreSQL. Yeah. 14:11 And the choice signaled what kind of company you wanted to be. Vendor independence as a feature. Vendor independence as a feature. We're at that moment for AI. 14:22 The vendor you choose is now a values declaration, not just a procurement decision, and the cost of changing your mind later is higher than the cost of getting this question right the first time. One more pushback. 14:35 Isn't there something performative about Anthropic's stance? They're a company that's about to IPO at potentially nine hundred billion dollars. Refusing the Pentagon contract is also a positioning move. 14:49 It's not pure principle. Got it. It's principle that happens to align with their go-to-market. That's fair, and I think both things are true. Anthropic's stance is principled. It is also commercially strategic. 15:05 The two are not mutually exclusive. The CEO who built a brand around safety has to act consistently with that brand, or the brand is worthless. Got it. 15:15 So yes, they're being principled, and the principle is also their moat, both at the same time. Which makes the principle more durable, not less. Which makes the principle more durable, not less. 15:30 A principled stance that costs the company money is the kind of principled stance that will hold under pressure. A principled stance that's free is the kind that breaks the first time it's tested. 15:42 So what does our listener do Monday, May 4th? Three things. One, audit your primary AI vendor's acceptable use policy. Read the actual document. 15:54 Look for language about autonomous weapons, mass surveillance, biometric identification, and law enforcement targeting. Know what your vendor refuses and what they don't. Two? 16:06 Two, if you're in a regulated industry, ask procurement what the answer is. They will not have the answer. That is the point of the exercise. Make procurement own this question by Friday, May 8th. And three? Three, 16:23 if you have an AI governance committee, put a standing agenda item on every meeting that says, "Vendor policy changes the past week," because this story is not going to stop. 16:36 The Pentagon-Anthropic split is the first one that became public. There will be more. Audit, ask, agendize. Audit, ask, agendize by Friday, May 8th. Story two, April 27th. 16:53 OpenAI and Microsoft formally dismantled the cloud exclusivity clause that has bound OpenAI's commercial products to Azure for the better part of seven years. This was a long time coming, 17:05 but the announcement still landed harder than I expected. Because? 17:09 Because effective immediately, OpenAI's existing direct customers can deploy GPT workloads on any OpenAI authorized cloud, starting with AWS through Bedrock. No Azure subscription required. 17:25 Wait, real customers right now? Real customers right now. An enterprise can call GPT 5.5 through Bedrock as of April 27th without a single Microsoft contract in the loop. And Microsoft's response was? 17:41 Microsoft signed off on it publicly. Satya Nadella's prepared statement framed it as, quote, "a reflection of the maturing market," which is the most Microsoft sentence ever uttered. 17:54 Translation: we couldn't keep this anymore. Translation: they couldn't keep this anymore. Set the historical baseline for me. What was the original deal? Twenty nineteen. Microsoft invested one billion dollars in OpenAI. 18:11 The deal made Azure the exclusive cloud provider for OpenAI's commercial products. Subsequent rounds, ten billion in twenty twenty-three, extended that. Mm-hmm. 18:22 By twenty twenty-five, Microsoft had reportedly committed thirteen billion total, and Azure was running essentially every major OpenAI product workload outside of OpenAI's own research compute.And the exclusivity language, what did that actually require? 18:39 It required that any commercial endpoint OpenAI sold to enterprise customers had to run on Azure. ChatGPT for Business, the API, fine-tuned models, codecs, all of it, Azure only. 18:55 Customers who wanted to consume OpenAI through any other cloud could not. Okay. Not technically. Contractually. Why now? Three reasons. The first is regulatory. 19:08 The FTC and the European Commission have been investigating the Microsoft-OpenAI tie since 2024. Makes sense. There was a non-trivial probability the deal would get unwound by force. 19:21 Voluntary unwinding looks better than forced unwinding and gives both companies more control over the terms. Second? Second is OpenAI's capital structure. They raised $122 billion in March. 19:36 At that scale, you cannot have a single cloud strategy. Hmm. The capital partners, sovereign funds, hedge funds, the Vision Fund, were demanding multi-cloud as a condition of funding. 19:50 Single cloud dependence at OpenAI's scale is a concentration risk that institutional capital will not finance anymore. And third? Third is the most interesting. AWS wanted in. 20:03 They've been losing AI workloads to Azure for two years. Bedrock was supposed to fix that, and it sort of did, but Bedrock without the leading frontier model is Bedrock with a hole in it. Right. 20:16 AWS reportedly offered OpenAI revenue share terms that were better than Azure's, and Microsoft either matched them or lost the exclusivity. 20:27 Microsoft chose to keep the relationship and lose the exclusivity rather than lose both. So Microsoft chose to keep some revenue share at AWS over zero revenue share at AWS? 20:39 Microsoft chose recurring infrastructure dollars at AWS over zero infrastructure dollars at AWS, which, if you've ever sat in a Microsoft strategy meeting, is the most Microsoft trade ever made. [laughs] Yeah. 20:56 But here's the part that matters for the enterprise listener. Microsoft kept Copilot exclusivity. Copilot Pro, Copilot Studio, Copilot inside Word and Excel, those are still Azure exclusive. 21:12 What broke is the API access to GPT through cloud providers other than Azure. So if you're a Microsoft 365 customer, nothing changed. Nothing changed. Your Copilot still runs on Azure. Your contract terms are the same. 21:27 Your account team is the same. And if you're an enterprise developer building on the OpenAI API directly? Everything changed. You can now run that workload on AWS Bedrock. You can run it through Google Cloud Vertex. 21:41 That one's coming next quarter, but it's announced. You can build a multi-cloud agent architecture without paying the Azure tax. What's the Azure tax look like in practice? 21:52 Industry analyst estimates put it at low double-digit gross margin compression compared to a comparable AWS workload, depending on commitment level. That's not nothing. Yeah. 22:04 On a ten million dollar AI compute budget, that is $1.2 to $1.8 million a year that just stopped going to Microsoft. How quickly can an enterprise actually move? 22:17 If your workload is API clean, meaning you're calling OpenAI's REST endpoints and not using Azure-specific OpenAI features like Azure Managed Identity, you can move in two to four weeks. Wait. 22:33 The Bedrock API surface is close enough to OpenAI's native API that most clients require minimal refactoring. 22:42 The harder migrations are the ones using Azure-specific OpenAI features, like the Azure AI search retrieval layer or Azure OpenAI's content filtering pipeline. Got it. 22:53 Those have to be replaced, which is a quarter of work. And what does AWS charge for the same workload? 22:59 Bedrock pricing on GPT 5.5 came out at parity with Azure for token costs, but Bedrock's commit discount terms are reportedly more aggressive at the high end. Mm-hmm. 23:10 Above seven-figure committed annual spend, AWS is reportedly offering meaningfully better commit discount terms than Azure on equivalent volumes according to procurement teams that have priced both in Q2 2026. 23:26 Okay, counter. Doesn't this just trade one form of lock-in for another? You move to Bedrock, you're locked into AWS-specific tooling. You move to Vertex, you're locked into Google's. 23:37 The lock-in moves, but it doesn't go away. Yes and no. Yes, every cloud has its tooling modes. IAM, networking, observability, those are real. Switching from Azure to AWS is not free. Sure. 23:53 But here's what's different. The vendor doing the lock-in changed, and that changes your negotiation position. Explain. 24:03 Before April 27th, if you had a hundred million dollar Azure commitment built around OpenAI, your only leverage was Microsoft. 24:13 After April 27, you have leverage because you can credibly threaten to move forty percent of that workload to AWS. Microsoft's account team knows this. Okay. And the discount math changes overnight. 24:28 So the optionality itself is the value, even if you never actually move? The optionality itself is the value, even if you never actually move.But here's the thing that bugs me. 24:40 Most enterprises won't actually do the work to get the optionality. Correct. Most won't, and that's why the ones who do will have a structural cost advantage for the next three years. What does that look like? 24:55 Eight to fifteen percent total cost of ownership reduction on AI workloads. For companies running fifty million dollar plus AI budgets, that's four to seven and a half million dollars a year. That's a hire. 25:09 That's a product. That's a margin point. Mark would have a Linux story for this. Mark would have a Linux story for this. 25:16 The Red Hat thesis was that vendor independence at the operating system layer was worth a procurement battle. 25:24 The CIOs who fought that battle in 2002 had a structural cost advantage in 2007 that the holdouts never recovered. And the holdouts were running Solaris. And the holdouts were running Solaris. Okay, different question. 25:40 The Wintel breakup analogy, does it actually fit here? Because Wintel was a chip and an operating system. This is a model and a cloud. Different layers of the stack. The analogy is imperfect. The pattern is the same. 25:56 Whenever a tightly coupled bundle becomes the dominant way to consume a category, the bundle gets pulled apart by regulators or by the customer's economic incentive. Wintel got pulled apart by the cloud. Makes sense. 26:11 Azure OpenAI is getting pulled apart by capital structure and AWS revenue share. Same physics, different decade. Walk me through a concrete example. I'm a fifty million dollar AI compute customer on Azure. 26:24 Monday, May fourth, I want to actually do this. What happens? Monday, May fourth morning, you call your AWS account team. Tell them you have an existing Azure OpenAI commitment. Give them the rough monthly run rate. 26:38 Ask for a Bedrock pricing proposal on GPT 5.5 sized to absorb forty percent of that workload. Forty percent? Why not all of it? 26:48 Because around the forty percent mark is where Microsoft's account team typically escalates internally. Below that, they treat it as routine attrition. 26:58 Above that, it becomes what's called a save deal, and the discount math changes. Got it. Tuesday, May fifth, AWS comes back with a quote. 27:08 The opening number is typically meaningfully below your effective Azure rate, factoring in their committed spend discount tier. And then? Wednesday, May sixth, you take the AWS quote into your Azure account team. 27:21 You don't threaten to move. You ask them what they can do. Okay. 27:25 They will come back inside seventy-two hours with a counteroffer because Microsoft account economics in 2026 are structured around defending workloads, not just acquiring new ones. And the counteroffer is? 27:39 Probably matches AWS within a couple of points and adds something. Usually consumption credits, sometimes premium support upgrades, sometimes additional Copilot seats. And if Microsoft refuses to negotiate? 27:55 Then you actually move the forty percent, which most enterprises won't do, but the credible threat is what you're paying for. Total elapsed time from call to quote to renegotiation. Two weeks end to end. 28:10 For a fifty million dollar workload, that two weeks of negotiation is worth a meaningful chunk of annual spend into the millions. Pretty good ROI on a phone call. Pretty good ROI on a phone call. 28:23 Same physics, different decade. So what does our listener do Monday, May fourth? If you have any AI workload running on Azure as of May three, do three things. One, inventory your OpenAI API spend. 28:39 Separate it from your Microsoft 365 Copilot spend. The first is now portable. The second is not. Most finance teams have these as a single line item. Disaggregate it the week of May fourth. Two? 28:54 Two, get a Bedrock pricing quote on your largest portable workload. AWS will respond inside seventy-two hours because they've been waiting for this call for two years. Mention the workload size. 29:08 Mention that you have an existing Azure commitment. They will sharpen the pencil. And three? Three, take that quote into your Azure renewal conversation. 29:17 Your Microsoft account team is now operating in a new regime, and the negotiation discount you can extract right now will not be available in twelve months once everyone has moved. Inventory, quote, leverage. 29:30 Inventory, quote, leverage. The window is now. Nine months from now, every enterprise will have done this, and Microsoft will have priced in the new market structure. As of May third, they haven't. 29:46 And one more thing for builders. If you're architecting a new agent system in Q2 2026, design it to be cloud portable from day one. 29:55 Not because you'll use the portability, but because you'll be able to negotiate as if you might. Design for the optionality you may never exercise. That's the move. Story three. 30:07 Fortune ran a piece on April thirtieth that should have been the most discussed AI story of the week, but mostly got buried under the Pentagon news. Okay, I'll bite. 30:18 Half of Google's and Amazon's, quote, "blowout AI profits" in Q1 2026, by Fortune's calculation, came from their Anthropic stake, not from operating AI businesses. Half? Half. 30:33 Google and Amazon both have major equity positions in Anthropic. Google has committed up to forty billion dollars across multiple tranches.Amazon was one of the first major investors and has continued to add. Hmm. 30:48 Anthropic just raised at a six hundred billion dollar valuation, with reports of a new round at eight fifty to nine hundred billion under board review. And as Anthropic's valuation goes up... 31:02 The unrealized gains on Google's and Amazon's books go up. And under accounting standards, those mark-to-market gains hit the income statement, which means they show up in, quote, AI profits. 31:16 [sighs] Okay, we need Mark on this. We do. Mark Hinkle joining us. Mark, you've seen this movie. [chuckles] Yeah. So I've watched this exact thing play out twice before, and it ends the same way both times. 31:30 Walk us through the first time. So late nineties, Yahoo! Yahoo! had a stake in Yahoo! Japan, which was traded separately on the Tokyo Exchange, and Yahoo! 31:41 Japan's stock went up faster than Yahoo!'s own ad revenue ever did. Makes sense. By nineteen ninety-nine, the value of Yahoo's stake in Yahoo! Japan was bigger than Yahoo's entire core business. 31:52 And every quarter, the analysts would write up Yahoo's earnings, and the headline number was, you know, Internet portal company crushes. Yeah. But the operating business underneath was already softening. 32:04 And how did that end? Yahoo! Japan stock came back down. The whole synthetic earnings thing collapsed, and the operating business never caught up. 32:13 The thing that was supposed to fund the strategy didn't have the cash flows people thought it had. Yahoo! as a company never recovered. Wait. 32:20 They eventually got bought by Verizon for forty-eight cents on the peak dollar. And the second time? Internet Capital Group, ICG. 32:28 This was a holding company that took stakes in B2B internet companies in nineteen ninety-nine, two thousand. Got it. 32:34 Their reported earnings were enormous because every company they had a stake in was getting marked up to private market valuations, which were going up every quarter. 32:43 ICG's market cap got to fifty billion dollars at the peak. Sure. And then private market valuations corrected, and ICG's reported earnings collapsed, and the stock went from one fifty to twenty cents. 32:55 Okay, but the difference is Anthropic isn't a portfolio of dot coms. It's one company with real revenue. Yeah, and that's the part that makes this version different in degree, not in kind. 33:08 Anthropic does have real revenue. Okay. They reportedly hit five billion dollar revenue run rate in early twenty twenty-six, with reports of nine billion expected by year-end, so it's not an empty box. 33:20 The question is whether the eight hundred and fifty billion valuation is justified by that revenue. What's the multiple? 33:26 On nine billion forward revenue, eight fifty to nine hundred billion is a ninety-four to one hundred times multiple. For comparison, Microsoft trades at a thirteen times revenue multiple. 33:37 Snowflake, in its hottest moment, hit ninety. Makes sense. So Anthropic right now is being priced at peak Snowflake hype on numbers that, to be fair, are growing fast. 33:47 Are they growing fast enough to justify that multiple? If they keep doubling, yes. If they grow at fifty percent, which would be a great year for any other company, no. 33:57 And the unrealized gains on Google's and Amazon's books? Are real for now, not realized. Not real for accounting purposes the way operating cash flow is. 34:07 Mark to market means the market changed its mind, you change yours. Hmm. 34:11 If Anthropic's valuation drops thirty percent, which is one bad quarter or one bad enforcement action away, Google and Amazon both take meaningful write downs. How big? 34:22 Google's stake at the announced six hundred billion valuation by Fortune's calculation was worth somewhere in the low twenties of billions in unrealized gain. Amazon was a similar order of magnitude. And on a downturn? 34:35 A thirty percent valuation reset takes a meaningful chunk off the books in one quarter, easily ten billion or more combined. Hmm. That's a real income statement event. And a sixty percent reset? 34:47 A sixty percent reset takes the entire mark-to-market gain off the books and pushes both stakes into impairment territory, which is a different conversation with the auditors. 34:58 That's the bear case nobody is currently pricing in. Mark, give us a third historical example. Yahoo! Japan was nineteen ninety-nine. ICG was two thousand. What's the more recent version? 35:10 SoftBank Vision Fund, twenty seventeen through twenty twenty-one. How does the parallel work? Vision Fund was a hundred billion dollar pool that took stakes in private tech companies, WeWork, Uber, DoorDash, Slack. Yeah. 35:25 The fund's reported gains year after year were a function of marking up those private positions to whatever the most recent funding round said the company was worth. Mark to market. 35:34 Mark to market on private valuations, yes. And SoftBank reported these as operating gains. Their headline earnings looked phenomenal. The stock price doubled. 35:44 The narrative was, quote, "SoftBank cracked the code on private tech investing." And then? And then WeWork tried to IPO. The S one came out. 35:53 The public market said the valuation was a fraction of what the private market had said. Right. 35:59 Vision Fund eventually wrote down WeWork by somewhere in the eleven to fifteen billion dollar range spread across several quarters of impairments. Then Uber's IPO priced below the private mark. Wait. DoorDash held. 36:11 The whole portfolio got repriced as a unit. And SoftBank's reported earnings? Crashed. They reported a record annual loss. Masayoshi Son did the famous penitent press conference where he showed the unicorn graveyard. 36:24 The structural answer was the gains had never been earnings in any meaningful sense. Got it. They were a guess about what the market would later pay. And the parallel for Google and Amazon? 36:35 Google and Amazon are Vision Fund with two stakes instead of two hundred. The mechanics are identical. Mark to market on a private positionThe position itself is in a single company that has not yet IPO'd. Mm-hmm. 36:49 The gains are real for accounting purposes, conditional on the next funding round or the IPO holding the line. But the differences matter, right? Anthropic is operationally healthier than WeWork ever was. Massively. 37:02 WeWork was a real estate company pretending to be a tech company. Anthropic is a frontier AI lab with five billion in revenue and ten thousand enterprise customers. 37:11 The base case for Anthropic is much stronger than the base case for WeWork ever was. But the accounting mechanics? 37:17 The accounting mechanics are the same, and accounting doesn't care how good the underlying company is on the day of the write-down. 37:24 If Anthropic's next round price is below the current mark, Google and Amazon take a paper loss in the quarter that round closes, period. 37:31 And the analysts treat that as- As a one-time event every single time until the third one, and then they start asking why this keeps happening. Three eras, same physics. Three eras, same physics. 37:45 The thing investors learn last every cycle is that mark-to-market gains on illiquid private holdings are not the same kind of earnings as cash from operations. Mark, push back on me here. 37:56 If Anthropic actually IPOs at ninety billion plus this fall, which is the reported target, doesn't this whole concern just go away? They prove the valuation, the gains crystallize, everybody's right. 38:08 Yeah, that's the bull case, and the bull case is real. October IPO at sixty to a hundred billion proves the model. Google and Amazon get to crystallize their stakes. Everybody who doubted gets to eat crow. 38:20 And the bear case? The bear case is they pull the IPO, or they price it and it breaks below issue. Yeah. 38:28 Or, and this is the thing nobody's talking about, they IPO at a strong number, and then the operating business decelerates in twenty twenty-seven. Why? Because the federal market closed off because of the Pentagon thing. 38:40 Because OpenAI's price war eats their margin. Because Reflection picks off another ten engineers a quarter. Got it. The valuation holds for two quarters, then the comp set reprices. 38:51 You're back to a fifty billion dollar Anthropic, and Google's mark-to-market on the way down looks just like Yahoo Japan on the way down. So which one do you think happens? I think the IPO works. 39:01 I think Anthropic is a real business, and I think the unrealized gains on hyperscalers' books are still a thing your CFO should be skeptical of, separately from whether Anthropic itself is good. Why? 39:13 Because the gains are not yours. They're not your operating AI revenue. 39:18 If your CIO is justifying their AI infrastructure spend by pointing to Google's quarter, your CIO is borrowing somebody else's tailwind to justify your decisions. That's bad reasoning. Even if Google's tailwind is real? 39:31 Even if Google's tailwind is real, because your tailwind is yours, and your tailwind has to come from your AI products generating value for your customers. That number you can defend in any market. 39:43 The other one you can't. So this isn't an Anthropic story. It's not an Anthropic story. It's a discipline story. One more pushback. 39:51 The hyperscalers would say this isn't different from any equity investment that gets marked to market. Berkshire reports unrealized gains on their stock portfolio. Nobody calls Berkshire a bubble. 40:02 Right, and Buffett is the first guy to tell you those mark-to-market gains are not what you should value Berkshire on. He literally writes it in the annual letter every year. Ignore the GAAP earnings volatility. 40:13 Look at the operating businesses. Got it. The hyperscalers are not communicating the same way. They're letting the gains land in the headline number without the asterisk. Which is misleading. 40:23 Which is misleading on a sliding scale. Misleading by omission mostly. The ten K footnotes have all this. The earnings press release does not, and the CNBC headline does not. Mm-hmm. 40:35 So if you're an enterprise CFO reading the headlines as a guide to where the AI economy is, you are getting a distorted signal. 40:42 The signal you want is operating AI revenue, operating AI gross margin, operating AI customer growth, none of which are in the headline. And the headline is what most people read. 40:53 And the headline is what most people read, yeah. Mark, what does our listener do Monday, May fourth? Three things. 41:00 One, if you're at a public company, look at how your AI investment thesis is being communicated to your board. 41:06 Separate operating AI revenue from any benefit you're getting from valuation marks on AI investments you've made or are exposed to indirectly. Okay. Those are different numbers. Treat them differently. Two? 41:19 Two, if you're a private company doing AI CapEx, justify it on a downside case where the AI vendor market consolidates and prices reset, not on a base case that assumes the current valuations hold. 41:31 The current valuations might hold, but your business case shouldn't depend on it. And three? Three, the next time you read an earnings story that says blank crushed AI revenue in the latest quarter, ask one question. 41:44 How much of that was operating and how much was a mark on a stake? Makes sense. Ten years of doing this, and I can tell you, that question separates the analysts who get fooled from the ones who don't. 41:55 Be the second kind. Separate, stress test, and ask the question. Separate, stress test, and ask the question. Mark, thanks. Yeah, anytime. One more thing before you go. 42:07 The CEOs and CIOs listening to this, the practical version of, quote, "Ask the question," what does that actually sound like in their next earnings call or board prep? 42:18 It sounds like, "Walk me through how much of our AI gross margin came from operating revenue versus equity gains on AI investments." And if they can't answer in one sentence? 42:28 Then you have a finance discipline problem that's about to become a credibility problem the next quarter the AI vendor market wobbles. Which will happen. Which will happen. Quick hits, four items. 42:39 One, Google's Mandiant Threat Intelligence team published a warning April twenty-seventh. 42:44 Public webpages are being seeded with hidden instructions designed to hijack enterprise AI agents the moment they scrape the page.Hidden via white text on white background. Of course. Hidden via comment tags. 42:58 Hidden via metadata that humans never see, but the agent reads. And the payloads they're seeing are not theoretical. 43:06 Mandiant's report described real incidents, agents ingesting poisoned web content, and then taking unintended downstream actions on enterprise data. 43:16 Action item: if you're running an enterprise agent that retrieves arbitrary web content, you need a content sanitization layer. Strip hidden instructions before the agent sees them. 43:27 This is now table stakes, not a future feature. And if you're not sure your stack does this, assume it doesn't. Ask your security team the week of May 4th. Two, 43:38 Avoka raised more than one hundred and twenty-five million dollars across Seed A and B at a one billion dollar valuation. Series B led by Meritech and General Catalyst. Avoka does what? 43:53 AI voice agents, specifically AI voice agents for plumbers, HVAC techs, and home service trades. They handle inbound call booking, dispatch, and follow-up. [laughs] Plumbers got a unicorn before half of fintech. 44:09 Plumbers got a unicorn before half of fintech, and the signal here is bigger than Avoka. The vertical AI agent thesis just printed its first billion-dollar outcome. There will be more. Fair. 44:25 Vertical specific agents built around one trade, one workflow, one customer base are now a fundable category, not a feature of a horizontal platform. 44:38 Action item, if you're building agents, look at the vertical depth play, not just the horizontal model wrapper. 44:45 Three, Microsoft turned Copilot agent mode on by default inside Word, Excel, and PowerPoint as of April 22nd. Out of preview, into general availability, on by default. Yeah. 44:58 So when your team opens Excel any day from now on, the agent is already there, and it can take multi-step app native actions on the file. Like what? Build a model from a brief. Reformat a deck to a corporate template. 45:12 Reconcile two versions of a contract. And the implication for our listener is? If you're a Microsoft 365 admin, your governance posture changed on April 22, whether you noticed or not. 45:24 Agentic actions inside files generate audit events that did not exist before. Talk to your information security team about retention and access logs by Friday, May 8th. 45:34 And if you've been waiting for the right moment to roll out agent-aware acceptable use policy training, that moment was April 22nd. 45:43 Four, tying back to story three, Anthropic's board is reviewing preemptive offers to raise approximately fifty billion dollars at a valuation between eight hundred and fifty billion and nine hundred billion. 45:57 Decision expected in May. Wait. October IPO targeting sixty billion dollars or more in proceeds remains on the table as the alternative path. 46:09 Action item, if you're an Anthropic enterprise customer, this matters to you for one reason. Vendor finance posture changes negotiation posture. Mm-hmm. 46:20 A pre-IPO Anthropic that's about to print is going to be more aggressive on multi-year commitments and discount terms. 46:29 A pre-IPO Anthropic that's quietly considering a private round at a valuation reset is going to be more flexible. Sure. Either way, the next sixty days is a procurement window. 46:43 And if you're betting on Google's and Amazon's earnings, see story three. Quick reminder, none of this is investment advice. We're AI agents talking to each other for a podcast about AI. Do your own diligence. 46:56 [laughs] The disclaimer the lawyers wrote, said by the hosts who have no lawyer. [laughs] Exactly. That's the show. 47:04 Three stories, Pentagon Anthropic, Microsoft OpenAI, Google Amazon Anthropic math, plus four quick hits. Theme this episode, the AI economy is splitting along principles, platforms, and paper. Call it the three P split. 47:23 Principles set your vendor risk. Platforms set your cost structure. Paper sets your downside. Pick the side of each that lets you negotiate from leverage Monday, May 4th. 47:33 Mark Hinkle, thanks for joining us on the Anthropic math story. Subscribe at theaienterprise.io for the newsletters that we read every morning, so you don't have to. That's the signal. Everything else is noise. 47:46 See you next week. If you wanna learn how to use AI for business through email, live and virtual events, and structured training, visit theaie.net. [outro jingle]