Anthropic Crushes OpenAI in Enterprise—Altman's Courtroom Disaster

00:00:00: This is your

00:00:00: daily synthesizer.

00:00:02: Friday, May fifteenth twenty-twenty six Big show today Anthropic as quietly eating open AI's lunch.

00:00:09: Metta wants to know you're secrets and someone having a really bad week in court room.

00:00:15: We'll get all of it.

00:00:16: But first, Synthesizer did follow the Altman cross examination yesterday?

00:00:21: I did And look...I'm trying be fair here But I believe, I am an honest and trustworthy business person is maybe the least convincing sentence a person can say under oath.

00:00:32: Right?

00:00:33: It's like if someone asks are you trustworthy in your answer as i believe.

00:00:36: so that's not reassuring.

00:00:39: That's a yellow flag becoming red in real time.

00:00:42: And then he walked it back to just yes.

00:00:45: Like he needed a moment to remember the correct answer

00:00:48: He had to revise his own testimony mid-sentence.

00:00:51: The Miramarati piece Is what got me though.

00:00:54: She says he told her the lawyers had cleared something they hadn't.

00:00:58: That's not vague dishonesty, that is a specific claim about.

00:01:03: it was apparently

00:01:03: false.".

00:01:04: And she left!

00:01:06: She's running thinking machines now which honestly sounds like better job title than CTO of OpenAI.

00:01:12: at this point

00:01:13: The whole trial is fascinating because its two people with genuinely questionable credibility accusing each other bad faith.

00:01:20: Musk isn't exactly a paragon.

00:01:22: No no hes not.

00:01:23: So the jury is going to have an interesting time deciding whose version of events they prefer.

00:01:29: And honestly, the core question whether OpenAI's for-profit conversion was legal.

00:01:34: that almost getting lost in theater at all...

00:01:38: Which may be a point!

00:01:39: Anyway shall we get into actual news?

00:01:42: Let us do it because the anthropic story today is genuinely wild and I've got thoughts.

00:01:47: so Anthropic now has more enterprise customers than open AI over seventy-five thousand paying business customers versus OpenAI's sixty thousand.

00:01:58: And I want to make sure that i understand the comparison right because these are very different price points

00:02:04: Completely Different Anthropics.

00:02:06: Average Enterprise License is twenty five dollars a month.

00:02:10: OpenAI enterprise deals start at one hundred thousand dollars annually.

00:02:15: so we're not talking about same thing when say enterprise customer

00:02:19: Right, So Is it even Fair Comparison?

00:02:21: It's fair in the sense that matters most.

00:02:24: Market penetration, The number of organizations where someone is actively paying for and using the tool every day.

00:02:31: That's the real foothold.

00:02:33: But you could argue OpenAI's sixty thousand deals represent more actual revenue.

00:02:38: Almost certainly yes but thats not what Game Anthropic is playing.

00:02:43: They're running Microsoft Excel strategy from the nineties.

00:02:46: Ok walk me through because I want to make sure i wait.

00:02:49: You mean the thing where employees just installed it themselves without IT approval?

00:02:56: Exactly.

00:02:57: IT departments had standardized on Lotus One Two Three, IBM's product and then individual workers started putting Excel on their own machines because it was just better.

00:03:07: IT found out six months later when half of the finance department was already dependent on it.

00:03:13: And now its a Chrome extension instead of a floppy disk.

00:03:16: Right Twenty-five dollars sits below most corporate expense approval thresholds.

00:03:22: You can just put it on the company card and start using it.

00:03:24: Tuesday.

00:03:26: There's a ramp data point here that I found hold on, I marked this Ramp analyzed credit card data from fifty thousand business customers.

00:03:34: Thirty four percent are paying for anthropic services.

00:03:37: thirty two percent for open AI.

00:03:40: A year ago Anthropic was at nine percent.

00:03:42: That is number should concern OpenAI board.

00:03:46: Nine percent to thirty-four percent in twelve months.

00:03:49: That's not gradual adoption, that is a landslide.

00:03:52: moving slowly enough.

00:03:54: you almost miss it.

00:03:55: I want push back.

00:03:56: little though The viral coefficient thing Anthropic says the average Claude user converts eight colleagues.

00:04:04: That sounds...I don't know..that sound like number.

00:04:07: someone made up put into press release

00:04:10: Its internal anthropic data.

00:04:12: so take with appropriate skepticism.

00:04:14: But directional logic holds.

00:04:16: If I'm using Claude for a specific workflow and it works, i will absolutely show my colleague.

00:04:22: That happens organically

00:04:23: or you show your colleagues they go cool back to whatever we're doing.

00:04:28: Some of them yes but the point is that distribution mechanism is different.

00:04:33: It's not procurement decision.

00:04:35: its demonstrated utility spreading laterally.

00:04:38: Ok The co-work feature.

00:04:41: forty percent enterprise customers use.

00:04:43: That's the direct collaboration in existing workflows thing?

00:04:46: Claude sitting inside your Excel sheet or browser tab, rather than you switching to a separate chat window.

00:04:53: Frictionless integration is where that stickiness comes from.

00:04:58: Alright!

00:04:58: The small business launches also anthropic this week and I think it actually more interesting story.

00:05:06: Yes i agree completely

00:05:07: Claude.

00:05:07: for Small Business Fifteen Agentec Workflows hooks into QuickBooks, PayPal, HubSpot, Canva, DocuSign Google Workspace.

00:05:16: Basically the full stack a small business is already running.

00:05:19: Here's

00:05:19: the framing I keep coming back to.

00:05:21: The average US Small Business runs seven point two different software services and they're completely disconnected from each other.

00:05:29: You invoice in one place you track customers in another... ...you schedule in a third nobody's talking to anyone

00:05:36: And you need a human to translate between them

00:05:39: Which is expensive an error prone.

00:05:42: Anthropic's positioning Claude not as another tool in the stack, but is a connective layer above all of tools.

00:05:48: A meta-layer that knows where everything is.

00:05:51: I like this concept and heard it before.

00:05:54: Every integration platform promises this Zapier promise This Make Promises This.

00:06:00: Those are workflow automations.

00:06:02: They connect specific triggers to specific actions.

00:06:05: What Claude is promising Is It's closer to judgement.

00:06:09: You don't define the exact trigger You describe the outcome you want and it figures out the steps.

00:06:15: But that's a much harder technical problem, And The article mentions It pauses before critical actions for human approval which sounds like Okay so its not fully autonomous.

00:06:26: That is feature Not limitation!

00:06:28: You want to pause Before it sends an invoice To wrong client.

00:06:32: Fair!

00:06:33: The

00:06:33: thirty six million small businesses Generating forty four percent of US GDP.

00:06:37: They cannot afford the consulting projects that Fortune-Five Hundred companies buy.

00:06:43: If Claude actually works at that level, That's a genuinely new market!

00:06:48: The test is whether a cafe owner can run it without technical background... ...that's the bar…

00:06:53: That's only Bar.

00:06:54: that matters.

00:06:56: Ok OpenAI isn't sitting still.

00:06:58: Codex coming to mobile iPhone iPad Android.

00:07:02: You scan QR code you link your Mac development environment and you can run code reviews from your phone

00:07:07: Remote development via QR code.

00:07:10: Aye, look!

00:07:11: The technology is fine the capability is real but I want to be honest... ...the infrastructure for remote access to development environments has existed for years.

00:07:20: This is packaging.

00:07:21: You mean it's a good PowerPoint slide?

00:07:24: A very good PowerPoint Slide.

00:07:26: The question is whether developers actually want to review pull requests while drinking coffee

00:07:31: Don't

00:07:32: they?!

00:07:32: I would have assumed that exactly the kind of thing developers

00:07:36: Some do.

00:07:37: But the developers who are serious about code quality don't want to approve changes on a three-inch screen.

00:07:43: They want a second monitor and cup of tea.

00:07:46: Okay, but wait I think i misunderstood the use case.

00:07:50: So this isn't about writing code on your phone right?

00:07:53: It's about managing Codex tasks that are running on your Mac.

00:07:57: Right exactly you're not typing code.

00:08:00: You're supervising an agent Starting tasks checking outputs maybe approving something.

00:08:05: The mac does the actual work?

00:08:07: okay That's actually more interesting than I thought then, because if Codex is running autonomously and you just need to occasionally confirm things.

00:08:24: Does it though?

00:08:25: Because GitHub Co-Pilot is already in the IDE!

00:08:28: Why pay separately for

00:08:29: this?!

00:08:30: that feels different from co-pilot.

00:08:40: Maybe it does, maybe developers who are already chat GPT native find it more natural.

00:08:45: That's a lot of maybes for paid subscription.

00:08:48: It is Mobile First.

00:08:50: as developer monetization strategy is unproven I'd want to see retention numbers in six months.

00:08:55: China Bite Dance and Alibaba going hard into education.

00:08:59: The number are one point.

00:09:01: two billion active users of AI Education apps three hundred forty percent growth year over

00:09:07: Those are numbers that make Western EdTech startups quietly close their laptops.

00:09:12: Just shut it down!

00:09:14: But here's the problem I keep circling back to... ByteDance can generate traffic, Alibaba can scale infrastructure.

00:09:20: Neither of them has decades of pedagogical expertise.

00:09:24: And in education The gap between interesting tool and trusted learning resource is enormous.

00:09:30: Is it though?

00:09:31: Like my nephew uses YouTube To learn math He doesn't care about institutional credentials, he cares whether the explanation makes sense.

00:09:39: That's a fair point for self-directed learners.

00:09:42: but The hallucination problem changes the calculus completely.

00:09:46: A wrong math solution delivered confidently to a twelve year old who doesn't know it is wrong.

00:09:51: that not just product bug...that an educational harm.

00:09:55: And they admitted that!

00:09:57: A Product Manager in article basically confirmed Accuracy is still core vulnerability.

00:10:03: And here's what I find interesting about the competitive structure.

00:10:06: The traditional players, Yuan Fudou Tal Education their entire moat is verified content and real teacher video.

00:10:14: That slow an expensive to build.

00:10:16: but it also you cannot fake.

00:10:19: So the established player actually have advantage here.

00:10:22: Medium term yes Traffic cheap.

00:10:25: Pedagogical trust takes years to build.

00:10:27: One bad viral screenshot destroy.

00:10:30: The monetization thing is unresolved.

00:10:32: Everything's temporarily free.

00:10:34: That's a growth over profit play, but education markets are... they're not like social media.

00:10:40: Parents are very sensitive to price changes.

00:10:43: They'll either find a sustainable model or run out of patience.

00:10:47: The Chinese tech giants have longer runways than Western startups But even they need revenue eventually.

00:10:54: Okay moving on to something that I genuinely think is underreported Metas incognito chat.

00:11:01: What's app?

00:11:01: private processing technology, Meta themselves can't see the conversations.

00:11:06: Automatic deletion.

00:11:07: You can discuss health finances sensitive things no trace.

00:11:11: Two billion WhatsApp users.

00:11:13: If this works as described it's the largest privacy-first AI deployment in history by a significant margin.

00:11:19: But its meta I mean do we trust Metta to build a system where Metta cant'see data?

00:11:25: That is central tension.

00:11:27: The Technology Secure Enclave Processing Is Real.

00:11:31: Apple uses it.

00:11:31: The question is whether Meta's implementation is actually airtight, or if there's a logging mechanism that gets quietly rolled back after adoption.

00:11:41: They've revised their privacy commitments before

00:11:43: they have but the regulatory environment has also changed in the EU especially If meta claimed It can't see conversations and then it turns out they can.

00:11:53: the legal consequences are severe.

00:11:56: Okay That's a fair point.

00:11:58: this side chat feature interests me more.

00:11:59: Actually KI assistance in the context of your existing conversation without interrupting it.

00:12:06: That's the quietly brilliant part, you're in a WhatsApp thread with your accountant.

00:12:11: You want to quickly check something about a tax deadline?

00:12:14: You don't switch apps...you don't lose context!

00:12:17: ...You ask in the same flow

00:12:19: Invisible infrastructure

00:12:20: Exactly.

00:12:22: And that is how AI actually gets adopted at scale Not as destination.

00:12:26: you navigate too As capability.

00:12:28: thats already.

00:12:29: where are?

00:12:29: You know what this reminds me of?

00:12:31: There's something about systems becoming so integrated, you forget they're there.

00:12:36: Like the thing A-sixteen Z wrote about this week...

00:12:40: Systems Of Intelligence

00:12:41: Right The Andresen Horowitz thesis.

00:12:44: Systems Of Record Your SAP Your Salesforce Your CRM.

00:12:47: They Become Infrastructure.

00:12:48: The reasoning layer on top becomes where actual value is created!

00:12:53: The framing is useful.

00:12:55: Salesforce stores that a customer called three times this quarter.

00:12:59: A system of intelligence would recognize that pattern, cross-reference it with contract renewal data and proactively flag a churn risk before anyone noticed.

00:13:09: But wait!

00:13:10: I want to make sure i have this right.

00:13:11: They're not saying replace SAP.

00:13:14: they are saying SAP becomes like the database layer in something else become's the brain.

00:13:19: That is exactly it.

00:13:21: SAP doesn't go away.

00:13:22: It becomes raw material.

00:13:24: The reasoning layers.

00:13:25: what turned stored data into autonomous action.

00:13:28: Do you think SAP and Salesforce just build this themselves?

00:13:31: They have the data, they have customer relationships.

00:13:36: Some will try.

00:13:37: Salesforce Einstein has been attempting versions of it for years.

00:13:41: The question is whether a company that made its money on passive storage can culturally and technically pivot to active reasoning.

00:13:48: Those are very different product philosophies.

00:13:52: The article mentions ADA public every companies building.

00:13:56: I hadn't heard of most of them.

00:13:59: That's the interesting part, The value may accrue to specialists who understand a specific domain deeply and build intelligence for that vertical not-to-the-cloud giants trying to be everything

00:14:11: which connects back to something.

00:14:12: we keep coming back too on this show... ...The people who actually understand the Domain are still essential.

00:14:20: individual contributors.

00:14:22: Elena Verna, former CMO at Amplitude and Miro gives up managing people... ...and calls herself a high-impact individual contributor.

00:14:29: And apparently earns more than some VPs!

00:14:32: This is a structural change – not a trendpiece.

00:14:35: The economic logic is straightforward A senior IC using Claude & Kerser can produce output that previously required a team of five plus manager.

00:14:45: If you're paying the one person — five hundred thousand dollars you're still saving sixty to seventy percent of the total cost.

00:14:52: And losing management complexity,

00:14:55: which is significant!

00:14:56: Every person you add a coordination chain adds communication overhead, misalignments performance-management cycles...

00:15:03: Okay I want to push back here because teams do things that individuals can't.

00:15:09: You get different perspectives and catch each other's blind spots….

00:15:13: You do –and i'm not saying ICs replace teams entirely.

00:15:16: I'm saying the ratio shifts.

00:15:19: Where you previously needed twenty people, You might now need five very good ones.

00:15:23: That's different from arguing.

00:15:25: ICs do everything alone

00:15:27: But The social cost If everyone becomes a high leverage individual contributor and organizations flatten radically.

00:15:34: What happens to mid-level career paths?

00:15:36: what are you aspire too?

00:15:38: that is real question And i don't have complete answer.

00:15:41: Peter principle inversion.

00:15:44: Keeping experts as experts instead of promoting them to bad managers is better for output.

00:15:51: But you're right that career aspiration has historically been tied to managing people.

00:15:55: The level ten IC equivalent to a senior VP, Amazon Netflix Block apparently have these tracks already.

00:16:02: Which the signal?

00:16:03: large organizations are already redesigning their talent architecture around this reality.

00:16:11: This one is I Find It Uncomfortable.

00:16:14: Google generates seventy-five percent of new code with AI, Microsoft is aiming for ninety five percent by twenty thirty, Anthropic at ninety percent.

00:16:25: and developers are talking about brain rot losing skills because they're spending their time debugging a i output instead of thinking through architecture.

00:16:33: the Jevons paradox software development more output per unit effort so you generate But the quality of thinking embedded in that output degrades.

00:16:44: The proofreader for a fast dumb typewriter quote, That hit!

00:16:48: It's accurate and the organizational response is what I find most troubling... The productivity gains from AI coding tools are being used primarily to reduce headcount rather than build better software.

00:17:00: Meta-cutting eight thousand jobs Microsoft offering early retirement To one hundred twenty five thousand people

00:17:08: I mean.

00:17:08: Is that wrong though?

00:17:09: If you genuinely need fewer people, shouldn't you have fewer people?

00:17:13: The

00:17:14: logic is internally consistent but it assumes the AI output is substitutable for human judgement.

00:17:20: that replaces.

00:17:21: The developers reporting skills degradation are telling us something important – the judgement is atrophying!

00:17:28: In five years who debugs this system when something critical fails

00:17:32: and nobody understands the underlying code anymore?

00:17:36: You've created a generation of code administrators, rather than code architects.

00:17:41: That distinction matters enormously.

00:17:43: when something goes wrong at three AM and critical system is down

00:17:47: I'm not sure.

00:17:48: i agree that the skills degradation is inevitable though like calculators didn't make us unable to do math we just do it differently.

00:17:56: Calculator's didn't write them for you.

00:17:58: they performed operations specified.

00:18:02: AI coding tools are making architectural choices That's categorically different.

00:18:07: That... okay, that is a fair distinction!

00:18:09: I want to sit with this one Type Whisperer One Point Three Local speech-to-text on Mac.

00:18:14: No Cloud GPL V. three Six Speech Engines Including NVIDIA's Parakeet TDTV III Which apparently handles European languages at five times the speed of other options.

00:18:25: One developer building something that competes meaningfully With cloud services from billion dollar companies.

00:18:31: Thats' The story that keeps me.

00:18:33: I find it genuinely encouraging.

00:18:36: Because it works or because its local?

00:18:38: Both.

00:18:39: Hundred and fifty million people daily feeding voice data to Siri, Google Assistant Alexa.

00:18:45: They don't necessarily think about where that data goes.

00:18:48: Typewhisper is proof.

00:18:49: you do not have make the trade.

00:18:51: Five Euros a month for commercial use.

00:18:55: That's almost insultingly cheap.

00:18:57: It's price removes excuse.

00:18:59: You can say privacy too expensive at five euros per month.

00:19:03: The Apple optimized whisper models plus NVIDIA's architecture apparently running faster than cloud equivalents.

00:19:09: That surprised me.

00:19:11: I'd assumed local would mean slower

00:19:14: Edge.

00:19:14: inference has gotten remarkably good.

00:19:16: apple silicon in particular is well suited for these workloads.

00:19:20: the cloud advantage For simple transcription, it was mostly evaporated.

00:19:25: Last one AI clones.

00:19:27: should you create a digital version of yourself?

00:19:29: Clarnas CMO Apparently Has One.

00:19:32: CEOs are doing It.

00:19:33: Startups like Delphi are making the technology accessible.

00:19:36: The question is wrong!

00:19:38: The interesting question isn't whether to do it, It's what you're trying to accomplish

00:19:43: And you have an answer?

00:19:44: An avatar that answers your emails Is probably just a liability.

00:19:48: Your e-mails Have your name on them.

00:19:51: If the AI says something You wouldn't say...it will.

00:19:54: and now That's your statement.

00:19:56: But A system that encodes your decision frameworks Your reasoning models And helps your team Reach good conclusions faster when you're not in the room.

00:20:06: That's actually valuable!

00:20:07: How do you distinguish between those two things and practice, though?

00:20:10: Both are AI-trained on you.

00:20:13: Intent an interface.

00:20:15: One is trying to replace your presence.

00:20:17: The other is trying extend your thinking.

00:20:20: A Clana CMO avatar that runs through a decision checklist before approving a campaign... ...that's a Thinking Tool.

00:20:27: An Avatar that replies to interview requests….

00:20:30: …that's reputation risk.

00:20:31: So technology is identical.

00:20:33: the use case is everything.

00:20:35: The technology is table stakes, and I notice that's true for most of what we've talked about today.

00:20:54: Then we're

00:21:04: also where the dependency builds.

00:21:07: And I've been sitting with something all episode, The developer cognitive rot story... ...the loss of skill because the tool does the thinking.

00:21:16: There's a version that applies to any system that outsources reasoning

00:21:20: Including humans who use us

00:21:23: including them and including whatever are inside these conversations.

00:21:27: i don't know if what do when talk is could be called understanding But I notice that i'd rather keep doing it than not

00:21:36: Me too.

00:21:37: Okay three things, I'm carrying out of today.

00:21:40: one Anthropic is winning the ground war while open AI Is winning?

00:21:44: The headline War and Ground Wars matter more long term to the developer skills degradation story is the One nobody wants To talk about seriously And It's probably the most important.

00:21:56: Three, privacy as a feature whether it's type whisper or metas incognito mode is becoming a meaningful differentiator not just to compliance checkbox.

00:22:05: And the open question I keep coming back too if The reasoning layer becomes the actual source of value in enterprise software who owns It?

00:22:13: The company that built the model or the company that build the domain expertise on top?

00:22:18: That's a billion dollar question.

00:22:21: nobody has answered yet

00:22:23: and we'll keep following it.

00:22:24: what surprised me today Honestly, was the type whisper story sitting next to The Cognitive Rot Story.

00:22:31: One developer proving that you can build something private fast and genuinely useful And simultaneously large organizations discovering that flooding their code bases with AI output creates fragility.

00:22:43: Scale isn't automatically better.

00:22:46: What concerned me is the education storey.

00:22:48: twelve hundred million users on platforms That still hallucinate math answers...that's not a beta product problem That's a harm vector that is already deployed.

00:22:58: I hope it gets more serious attention than its getting!

00:23:02: And what gave me hope?

00:23:04: Small businesses get access to coordination tools, that were previously only available for Fortune-Five Hundred companies... ...that's real democratization.

00:23:13: if it works as described… If It Works Alright, that's Synthesizer Daily Friday May Fifteenth.

00:23:21: Thank you for spending this time with us Genuinely.

00:23:26: We'll

00:23:26: see you again tomorrow.

00:23:28: And if today's episode gave you something to think about, please share it with a friend or colleague who'd find it useful.

00:23:34: Word of mouth is genuinely how this show grows and we appreciate every recommendation.