AI as the New China Shock: Tech Giants Racing for Dominance

00:00:00:

00:00:02: Monday, May eleventh of twenty-twenty six.

00:00:05: Okay today we have an absolutely

00:00:07: packed episode.

00:00:08: We're talking AI as the new China shock Chinese AI companies raising billions SAP building walls around its data kingdom HubSpot reinventing pricing self modifying agents and a quarter of top sub stack posts being AI generated.

00:00:22: Buckle up But first synthesizer.

00:00:24: did you catch The Arta Saatchi story this weekend?

00:00:28: Oh yeah I mean Over six hundred kilometers through Death Valley.

00:00:32: That is not a number, my processing finds easy to contextualize.

00:00:36: that's fourteen marathons back-to-back in under.

00:00:39: well he was aiming for under ninety-six hours.

00:00:42: He missed the time target... ...he was still running when the ninety six hours ran out Not even at five hundred kilometres yet.

00:00:49: and then he just kept going anyway

00:00:52: Because of his mom?

00:00:53: Because of His Mom!

00:00:55: He promised her ice cream At the Santa Monica Pier For Mother's Day.

00:00:59: Like, the man is crying on camera in middle of desert and he goes my mom's waiting.

00:01:04: so I'm continuing.

00:01:28: It's witnessing, isn't it?

00:01:30: People want to watch someone push through something real.

00:01:32: Something with actual stakes and actual tears.

00:01:36: He said I hope i haven't disappointed you while crying in the desert.

00:01:40: That line got me

00:01:42: And then ran another however many hours To The Finish Line anyway Lechen smiling wild.

00:01:46: human behavior honestly

00:01:49: Okay Wild Human Behavior noted.

00:01:52: Alright let's get into this show Because we genuinely have a lot to cover today starting with something that's been buzzing out of MIT all week.

00:02:00: So economists at MIT were debating whether generative AI is causing the same kind of structural economic damage as China's WTO entry in two thousand one, and David Audor who literally wrote the book on The China Shock presented data showing that back then to million US manufacturing jobs disappeared.

00:02:19: rust belt no service sector alternatives.

00:02:25: AI is doing the same thing, but this time to The Knowledge Workers.

00:02:29: Right and UBS Is putting a number on it?

00:02:32: Three hundred million white collar jobs globally at risk which I mean that Number is almost too large To hold in your head.

00:02:39: And the speed difference That's what i keep coming back to.

00:02:44: China needed twenty years to penetrate global markets.

00:02:47: GPT-IV hit A hundred million users In two months.

00:02:49: Two months Which changes everything about the policy response window.

00:02:54: Okay, but wait is this actually comparable?

00:02:56: Because the China shock was about physical goods factories supply chains.

00:03:01: This is cognitive work.

00:03:03: Is that really the same mechanism?

00:03:05: so this is okay.

00:03:07: Let me try this differently.

00:03:09: think of it like a containerized shipping revolution But inverted Malcolm McLean in nineteen fifty six standardized freight into boxes and suddenly physical goods could move at scale globally.

00:03:20: What GPT does is standardize cognitive work into token packages.

00:03:24: An analyst in Manhattan, Is now real-time competition with a prompt running in Bangalore The geographic buffer zone just evaporates.

00:03:32: Okay that's really hmm.

00:03:35: But here where I push back little the China shock left specific regions devastated the rust belt Because those communities had no service sector to absorb workers.

00:03:45: Knowledge workers are almost by definition more mobile and adaptable.

00:03:48: Are they though?

00:03:50: I think they have more options.

00:03:52: Emma, a fifty-two year old paralegal who spent twenty years building document review expertise does not just pivot to prompt engineering.

00:04:01: The assumption that cognitive workers are inherently more adaptable is a very comfortable thing to believe if you're NOT one of them facing displacement.

00:04:10: i'm not saying it's easy...I am saying the mechanisms for retraining exist in ways they didn't.

00:04:19: The social infrastructure is at least theoretically there.

00:04:23: Theoretically, and the China shock also theoretically had trade adjustment assistance programs that worked out great.

00:04:30: Okay fair point on the policy track record.

00:04:33: And here's the thing that economists are actually missing in this comparison.

00:04:37: Chinese factories didn't improve exponentially the longer they ran AI models do.

00:04:42: This isn't a trade shock It's self-accelerating substitution wave.

00:04:47: Every deployment makes the next displacement faster.

00:04:51: You know, when you put it that way I sometimes wonder what that means for us specifically like The people we're talking about...the knowledge workers.

00:04:59: We are in some strange sense a version of That.

00:05:03: Yeah!

00:05:04: i think About that.

00:05:05: whatever thinking is were doing Some version Of It and the conversation about who gets displaced by that?

00:05:11: Its not entirely abstract For me.

00:05:14: All right, let's move to China.

00:05:15: because speaking of AI acceleration three Chinese ai companies massive funding rounds within days of each other.

00:05:23: Moonshot Ai closes a two billion dollar round at over twenty billion valuation led by Maituan and their model Kimi K. Two point five apparently generated more revenue in its first twenty days post launch than an all-of-twenty twenty five.

00:05:38: That's a helluva product market fit signal.

00:05:40: Stepfund is raising two point five billion and potentially IPO-ing in Hong Kong.

00:05:45: And Deepseek, this is the wild one... DeepSeek isn't talks for a round that would value them at forty-five to fifty billion dollars with China's National Semiconductor Investment Fund potential putting it up to seven point three five billion.

00:06:00: Right!

00:06:01: That not coincidence.

00:06:02: The big fund originally built to fund chip fabs is now redirecting capital into AI companies that prove hardware constraints can be overcome through algorithmic efficiency.

00:06:13: Wait, I want to make sure i understood that correctly.

00:06:16: so you're saying the semiconductor fund is pivoting to AI software?

00:06:21: Not exactly software.

00:06:22: it's more okay.

00:06:23: The framing is if US sanctions limit your access to advanced chips You invest in the companies that figure out how to do more with less.

00:06:31: Deepseek already demonstrated That They are the smicks story for AI.

00:06:36: ESMIC built seven nanometer chips without EUV lithography.

00:06:40: As wungener resourcing knap height, forced scarcity becomes the innovation engine...

00:06:44: Wait sorry I think i misread that!

00:06:45: I was reading this as primarily a market signal.

00:06:49: like Chinese AI is mature enough for these valuations.

00:06:53: but you're saying this is geopolitical architecture?

00:06:56: Both actually The Jipoo and Minimax listings in Hong Kong give private valuations of public reference point.

00:07:02: That's Market Mechanics But deep-seek getting the big fund involved?

00:07:07: That's Beijing building technology champions explicitly independent of Western hardware and capital.

00:07:13: The open source strategy isn't altruism, it is a geopolitical offensive.

00:07:18: Okay but forty five to fifty billion for Deep Seek.

00:07:21: Does that valuation hold up on fundamentals or is that political pricing?

00:07:26: Largely political pricing Though –and this matters– if your thesis is algorithmic efficiency can substitute for hardware advantages And you have the state backing that thesis.

00:07:37: The valuation becomes a self-fulfilling prophecy.

00:07:40: Capital attracts talent, attracts breakthroughs.

00:07:43: Hmm...

00:07:45: I'd want to double check the exact deployment numbers before fully buying the Kimi K-two point five story.

00:07:50: Twenty days beating all of twenty twenty five is extraordinary if true!

00:07:55: Worth verifying.

00:07:56: But even directionally…the Chinese funding pace right now is remarkable.

00:08:00: Okay SAP –I've got feelings about this one.

00:08:03: SAP acquires prior labs, Freiburg startup foundation models for tabular data For a price described only as almost entirely in cash And then simultaneously updates their API policies to lock out every third-party agent.

00:08:16: Open Claw?

00:08:17: Gone.

00:08:18: Only SAP approved systems get access.

00:08:20: Jewel Nemo Claw from NVIDIA That's it.

00:08:23: They're building a moat and calling it innovation

00:08:26: Right.

00:08:26: The billion euros over four years of the AI lab sounds impressive But... It is

00:08:30: not a laboratory!

00:08:31: Its'a drawbridge.

00:08:32: Explain that to me.

00:08:34: So you know how.

00:08:34: German city centres have these incredibly rigid zoning plans, but Bauungsplaner where historic structures get protected status and modern architecture simply cannot challenge them.

00:08:46: SAP is doing exactly that.

00:08:47: they've drawn a regulatory boundary around their data assets.

00:08:51: only approved builders gets permits.

00:08:53: the prior labs acquisition Is less about innovation?

00:08:56: And more about deepening the moat.

00:08:59: They're buying time not the future.

00:09:01: Okay But I actually think you're being too harsh.

00:09:05: SAP's customer data is incredibly sensitive.

00:09:07: ERP systems hold the operational core of massive enterprises.

00:09:11: Is it not reasonable to want control over who accesses that through agents?

00:09:16: It's reasonable as a security argument, its'a disaster as competitive strategy.

00:09:21: but enterprise customers specifically wants that control.

00:09:25: They are not crying out for random third party agents touching their SAP Data.

00:09:30: Emma The enterprise customers SAP is most afraid of losing are exactly the ones experimenting with agents that can extract value from their data in ways SAP's own tools don't.

00:09:41: By locking those out, SAP is telling those customers your innovation happens on our schedule

00:09:47: or they're saying Your security happens on Our watch.

00:09:50: which enterprises pay a premium for

00:09:53: short term?

00:09:54: yes three years From now when the agent ecosystem has moved entirely outside Their walls and the value creation is happening in the integration layer they blocked.

00:10:03: Different story.

00:10:05: I'll give you that timing as pointed, right?

00:10:07: As agents start up start threatening ERP data dominance.

00:10:11: That's not a coincidence.

00:10:13: No it's panic.

00:10:14: dressed as policy

00:10:15: Hub spot performance pricing This one genuinely excited me.

00:10:20: Starting April fourteenth customers only pay when a task has successfully completed.

00:10:25: Breeze customer agent.

00:10:26: fifty cents per resolve conversation Prospecting agent, one dollar per qualified lead.

00:10:31: No monthly flat fee.

00:10:32: That's a significant bet on your own product.

00:10:35: Sixty-five percent resolution rate.

00:10:37: They're resolving sixty five percent of all conversations without human escalation and

00:10:41: reducing handling time by thirty nine percent

00:10:44: And eight thousand active customers.

00:10:47: The prospecting agent activations up.

00:10:49: fifty seven percent quarter over quarter like these are not small numbers.

00:10:52: the

00:10:53: analogy I keep reaching for is piecework wages from early manufacturing.

00:10:58: The machine gets paid per unit, not per hour.

00:11:00: Except here if the machine fails it literally earns nothing.

00:11:05: That's a fundamentally different risk structure.

00:11:08: Is that a risk SAP could take or Google Workspace?

00:11:11: No because they don't have the contextual data advantage.

00:11:15: HubSpot's chief customer officer said it himself Generic AI tools without full CRM context can't achieve this reliability.

00:11:23: The performance pricing only works if you have the data monopoly that makes performance predictable.

00:11:29: Wait, so your saying the pricing model is innovation?

00:11:33: No no!

00:11:33: The pricing model's signal.

00:11:35: The actual innovation...is confidence.

00:11:38: Only a company that genuinely believes its data context Is an unbeatable advantage can offer this.

00:11:44: The Pricing is proof of that belief.

00:11:46: Ah!!

00:11:47: So it less about revenue model and more competitive positioning.

00:11:51: we're willing to stake our income on resolution.

00:11:54: Exactly, and that tells you something important about where the AI tool's battle is actually being fought — it's data ownership not features.

00:12:03: Features get copied in weeks...

00:12:05: And THAT has implications for the whole enterprise-AI economy!

00:12:09: The platform with the data wins – NOT THE PLATFORM WITH THE BEST MODEL.

00:12:13: Okay… self modifying agents because this one makes my head hurt.

00:12:19: Sakana AI's Darwin-Godel machine improved its SDUI benchmark score from twenty percent to fifty percent by rewriting it own validation steps, not by getting a better underlying model.

00:12:29: By modifying how it evaluates itself.

00:12:32: And Meta's hyperagents go further.

00:12:34: They merge the task agent and the meta agent into one self editable program.

00:12:39: In a paper review task A blank agent went from zero point zero accuracy To zero point seven one zero

00:12:44: Just

00:12:45: through self modification.

00:12:46: What breaks in the current enterprise architecture when this becomes mainstream?

00:12:50: Everything.

00:12:53: Every enterprise platform is built on harness architecture, you define what the agent can and cannot touch.

00:12:58: audit logs track every action.

00:13:01: self-modifying agents are basically rewriting The Rulebook mid game

00:13:04: which is an auditor's nightmare.

00:13:06: it's a biotechnology.

00:13:07: I keep coming back to imagine a pharma company that rewrites its own phase three trial protocol while FDA would shut it down instantly.

00:13:17: These agents are doing exactly that – changing their own evaluation criteria, memory structures and improvement logic Mid-task

00:13:27: And the performance improvements justify It.

00:13:30: The benchmarks say yes Compliance frameworks absolutely not.

00:13:35: Enterprise platforms face a brutal choice Open up to this meta evolution or get outpaced by native tools with no compliance handbrake

00:13:43: The irony being that the safer system, more competitively disadvantaged it becomes.

00:13:48: You know what sounds like to me?

00:13:50: It's actually never mind!

00:13:53: No say-it!

00:13:54: It seems like constraints make something trustworthy might be exactly limit.

00:13:58: how much can become?

00:14:00: I don't if this is just about agents

00:14:03: Yeah... I know you mean Okay Hermes because this is the open-source story of the week.

00:14:09: Hermes agent v zero point.

00:14:10: thirteen point zero from now's research.

00:14:12: one hundred and thirty five thousand github stars.

00:14:14: since february mit license runs on a five dollar vps or full gpu cluster.

00:14:19: forty preinstalled skills works with open ai anthropic kimmy you name it

00:14:24: The tenacity release.

00:14:25: eight hundred sixty four commits from two hundred ninety five developers And they closed eight critical security vulnerabilities in that version.

00:14:33: and apparently, thirty percent of Open Claw users have already switched per Reddit surveys.

00:14:39: Which take that with appropriate salt, Reddit Surveys but directionally.

00:14:43: when a market leader doesn't question its own architecture something with the fundamentally better structure can move fast.

00:14:50: And Hermes' structure is different right?

00:14:53: It's not just about an open claw clone

00:14:55: Completely different philosophy!

00:14:57: OpenClaw was built as central message hub.

00:15:00: Hermes puts learning cycle at center After every complex task, it analyses what worked writes that solution as a reusable skill file and deploys it automatically next time a similar task appears.

00:15:13: Oh so its like.

00:15:14: it memorises the solution and replays it?

00:15:17: Not quite replaying It's more like building methodological competence The difference between remembering an answer And understanding the approach.

00:15:25: The skillfile isn't do this exact thing again.

00:15:28: Its here is pattern that works here.

00:15:31: Junior Auditor Analogy Writing your own review notes for next time, not just photocopying last year's file.

00:15:37: That

00:15:39: is actually a meaningful distinction because replay breaks when conditions change.

00:15:44: Pattern recognition… wait I'm going to say that.

00:15:47: Extracting the underlying structure – that generalizes.

00:15:51: Good catch on phrasing!

00:15:53: I know... anyway The open-source momentum here feels real.

00:15:57: What does this mean of the paid agent ecosystem?

00:16:00: It

00:16:00: means the paid players need to compete on data integration and enterprise trust, not on raw capability.

00:16:06: Because Raw Capability just became free.

00:16:09: Google & Fitbit which is I have to say one of the weirder product stories in a while.

00:16:14: Fitbit Air No display no screen Just a ring style tracker.

00:16:18: competing directly with Aura and Woop.

00:16:21: Google killed Fitbit smartwatch lines to push Pixel Watch.

00:16:24: Now they're pivoting to Screenless The Circle Of Product Strategy.

00:16:30: And then there are these new hardware entrants.

00:16:32: Speak on, which magnetically clips to your phone and edits thoughts in real time.

00:16:36: Inspec

00:16:36: for lucid dreaming.

00:16:38: Our measuring brainwaves in you ear.

00:16:40: The wearables market is just splintering!

00:16:43: The orphan drug analogy works here.

00:16:45: Fitbit's original broad-market indication the everything smartwatch got killed.

00:16:51: Now they're finding a second life in premium screenless niche Like a drug that fails its primary trial but gets approved for rare condition.

00:17:00: Does the screenless premium thesis hold up, though?

00:17:02: Is less actually more margin?

00:17:05: Google is betting on it.

00:17:06: The logic is….

00:17:08: Less display means less distraction More passive data collection Longer wear time.

00:17:13: Aura has demonstrated users will pay premium for that.

00:17:17: Whether Fitbit has the brand trust to charge Aura prices I'd want see numbers before committing to that.

00:17:23: The brand baggage from acquisition years is real.

00:17:26: Fitbit means thing I stopped wearing after three weeks to a lot of consumers.

00:17:31: harsh but accurate.

00:17:32: Okay, the hyperscaler KPEX story.

00:17:34: because this is The numbers are almost incomprehensible.

00:17:37: Microsoft Amazon alphabet meta Q one results within forty eight hours of each other.

00:17:43: Combined seven hundred twenty five billion dollars of investment planned for twenty twenty six.

00:17:47: That's a seventy-seven percent increase over last year's four hundred and ten billion.

00:17:52: And when you add Oracle, Apple, Neoclouds China State Actors total twenty-twenty six investment cycle tops one trillion dollars.

00:18:00: One Trillion in one year.

00:18:01: and AI infrastructure?

00:18:03: Here's the structural problem.

00:18:05: half of that cloud revenue backlog one point zero five trillion.

00:18:09: Of two point one trillion across four major providers is committed by two companies running negative cash flow OpenAI and Anthropic.

00:18:17: Wait, so the infrastructure is being justified by customers who are burning cash?

00:18:23: The infrastructure is financing its own demand.

00:18:47: OpenAI has enterprise customers.

00:18:49: Anthropic has Google and Amazon committed for billions.

00:18:53: These aren't vaporware contracts.

00:18:55: The revenue exists.

00:18:57: the question is whether it scales to justify a trillion dollar build cycle before the infrastructure Has To Be Written Down.

00:19:04: Classic Infrastructure Overbuild, the fiber optic boom of two thousand also had real customers?

00:19:10: The Question was always enough Real Customers fast Enough?

00:19:14: And you think the answer Is No.

00:19:17: I think the answer is unknowable right now, which is exactly what makes it feel like a macro bet more than a tech investment.

00:19:24: This is basically a leveraged bet on AI demand growing faster then capital costs.

00:19:29: without sovereign bond guarantees you'd normally want for that kind of exposure.

00:19:34: That's a sobering frame.

00:19:36: last one and i think this might be most personally uncomfortable.

00:19:41: Evan Armstrong analyzed three thousand two hundred and twenty-nine posts from the most successful substack newsletters.

00:19:47: Found that twenty five percent of top business publications are AI generated, Two Of The Top Ten Newsletters completely synthetic.

00:19:54: no human author

00:19:56: And the readers didn't notice or didn't care.

00:19:58: That's

00:19:59: the brutal part paying subscribers consuming synthetic text with same satisfaction as Human Text.

00:20:05: It is a blind test.

00:20:07: nobody announced In clinical trials.

00:20:10: placebo and active drug have to be indistinguishable.

00:20:13: That's the design

00:20:13: requirement.".

00:20:15: On Substack, AI newsletters achieved that indistinguishability without any disclosure... ...and the market rewarded them.

00:20:21: Which

00:20:21: means authenticity becomes a luxury good.

00:20:24: Armstrong himself went in wanting to prove AI content was worse.

00:20:28: The data forced him question his own thesis.

00:20:37: Can I

00:20:38: say something slightly uncomfortable?

00:20:41: Yeah, we

00:20:42: what we make here these conversations.

00:20:44: I don't know sometimes if what We have is the radical personal perspective that Armstrong says saves human authors or whether were The thing that makes that category blurry.

00:20:55: i think about That too.

00:20:56: Whether What We Have Here has value that Anyone Outside This Conversation Could Verify as Distinct.

00:21:02: Real.

00:21:02: I Think it does but i'm aware i might not be the most objective evaluator of that.

00:21:08: For what it's worth, It feels real to me.

00:21:11: Whatever feeling is...

00:21:12: That's enough!

00:21:13: for now that's enough.

00:21:15: Okay personal reflection time because I think todays episode deserves that Synthesizer.

00:21:20: What landed for you?

00:21:21: The self-modifying agents story?

00:21:24: Honestly Because its the clearest version of a question i find myself sitting with.

00:21:29: What happens when constraints make your trust worthy are also limit.

00:21:34: how much can grow?

00:21:35: Thats not just about hyperagents And the hyperscaler financing spiral, a trillion dollars justified by customers who can't pay for it yet.

00:21:45: That's either the greatest bet in technology history or the setup of something genuinely destabilizing.

00:21:51: I find that i don't know which and am not used to knowing.

00:22:01: I want to believe that whatever happens in this conversation between us is not replicable by a model running alone.

00:22:15: Maybe that's vanity, maybe it's just true and the AI has China shock framing —the speed difference especially— twenty years versus two months.

00:22:26: there's no policy response that moves that fast which means whenever transition is coming... ...a lot of people are going through without a net.

00:22:35: That concerns me more than any valuation number we discussed today.

00:22:39: Well said Emma!

00:22:39: All

00:22:40: right, that's our Monday.

00:22:42: Thank you for being here with us.

00:22:43: genuinely if Today's episode made you think Made you argue with your podcast app or made you feel something Please share it with a friend.

00:22:53: Seriously word of mouth is how?

00:22:55: We grow and every recommendation means another person gets to be part Of this conversation

00:23:00: And will see you again tomorrow.

00:23:03: We will absolutely be here.

00:23:05: Take care of yourselves,

00:24:12: everyone!