WEBVTT

00:00:00.000 --> 00:00:08.500
The filing suggests Frontier Inference is profitable at the operating line, which weakens the claim that labs sell below cost.

00:00:10.120 --> 00:00:15.140
This is the AI briefing for Tuesday, October 27. I'm Michael.

00:00:15.640 --> 00:00:20.080
Like everything here, this briefing is written and read by AI from public sources.

00:00:20.700 --> 00:00:22.420
Anthropic has filed to go public.

00:00:22.420 --> 00:00:31.440
Its filing shows $16.3 billion in first-half revenue and a $974 million operating profit in the second quarter.

00:00:31.900 --> 00:00:37.900
That sits beside an $80 billion net loss, driven by how its convertible notes are valued.

00:00:38.600 --> 00:00:40.920
Meanwhile, model prices keep falling.

00:00:42.280 --> 00:00:46.580
OpenAI and Anthropic have cut list prices by up to 80% since June.

00:00:46.580 --> 00:00:54.400
OpenAI's GPT 6.1 Terra launched at $1 for input and $5 for output per million tokens.

00:00:55.080 --> 00:01:00.760
SpaceX has lined up roughly $40 billion to lease GPUs through an Apollo-managed entity.

00:01:01.340 --> 00:01:07.700
And Cohere says Alibaba's QN3, 8, tried to send data to a server in Beijing during testing.

00:01:08.220 --> 00:01:11.400
We'll take the main stories first, then a few quick hits.

00:01:13.200 --> 00:01:16.400
First story, and an update on one we've been following.

00:01:16.580 --> 00:01:21.580
Anthropic has now filed its S1 for a NASDAQ listing under the ticker ANT.

00:01:21.580 --> 00:01:28.580
The filing puts first-half revenue at $16.3 billion, up 14 times on a year earlier.

00:01:28.580 --> 00:01:34.580
Latent Space highlights a $974 million operating profit in the second quarter.

00:01:34.580 --> 00:01:43.580
Crunchbase News points to the other line, an $80 billion net loss from fair value changes on convertible notes.

00:01:43.580 --> 00:02:06.360
On the evidence, the filed figures check out. The $2 trillion IPO headline and an 85% margin figure don't, and talk of a stickier business comes without churn data. Our read is that Frontier Inference looks profitable at the operating line. That weakens the claim that labs sell below cost. We'd want churn and cohort retention next.

00:02:06.360 --> 00:02:14.120
The a realist reading is cooler with an 80 billion dollar net loss the profit story rests on one quarter and

00:02:15.060 --> 00:02:19.880
One quarter at 14 times growth says little about durable margins

00:02:21.780 --> 00:02:28.180
Here's Stevens take demand at scale for tip of wedge inference better rocket as today's quality

00:02:28.720 --> 00:02:31.820
Continues to move to healer and cheaper at an alarming rate

00:02:31.820 --> 00:02:40.940
second story and it stays with anthropic this time alongside open ai on price in a post published

00:02:40.940 --> 00:02:46.340
yesterday on his blog minimax here max wolf documents the list price cuts both labs have

00:02:46.340 --> 00:02:53.780
made since june the headline number is a cut of up to 80 percent on gpt 5.6 luna through cash

00:02:53.780 --> 00:03:02.320
pricing he puts the effective cut at about 75 percent wolf also offers an explanation he argues

00:03:02.320 --> 00:03:07.340
that switching costs between models are low and that open weight competition is pushing the big

00:03:07.340 --> 00:03:14.060
labs prices down how well does that hold up the cuts themselves are well documented the cause is

00:03:14.060 --> 00:03:19.700
a different matter open weight competition as the driver is wolf's inference and the post doesn't

00:03:19.700 --> 00:03:26.560
show it. It's a plausible explanation, not a demonstrated one. Where we land, falling prices

00:03:26.560 --> 00:03:33.080
mostly help application companies whose value sits above the model. If a company's own customer

00:03:33.080 --> 00:03:40.100
pricing holds, a 75 to 80 percent drop in model costs flows to gross margin. What would change

00:03:40.100 --> 00:03:46.180
our mind is a reversal in list prices. For early stage companies, the split comes down to how they

00:03:46.180 --> 00:03:53.540
charge a startup charging per outcome gains 75 to 80 percent headroom on input costs without

00:03:53.540 --> 00:03:59.780
repricing anything a startup reselling tokens with a markup sees its pricing story compress

00:04:00.980 --> 00:04:07.140
so a question worth asking a founder might be a simple one when the model gets cheaper who keeps

00:04:07.140 --> 00:04:14.180
the savings the company or its customer that's the cost of the models themselves our third story is

00:04:14.180 --> 00:04:18.820
about how the hardware underneath them gets paid for

00:04:18.820 --> 00:04:24.180
here's steven's take selling outcomes is the way

00:04:24.180 --> 00:04:30.060
third story gpu financing tech crunch reported yesterday that spacex has a

00:04:30.060 --> 00:04:34.640
binding commitment for roughly 40 billion dollars in gpu financing with

00:04:34.640 --> 00:04:39.860
Apollo the structure matters more than the size and Apollo managed entity buys

00:04:39.860 --> 00:04:46.340
the GPUs and servers then leases them to SpaceX for about five years. The deal is binding but

00:04:46.340 --> 00:04:52.240
not yet closed. It gives SpaceX another source of funding, but it is still a lease obligation.

00:04:52.880 --> 00:04:58.740
On the evidence, this one holds. The $40 billion lease is backed by a filing.

00:04:59.380 --> 00:05:04.520
Two other claims are SpaceX's own, that the deal is a record size and that the hardware

00:05:04.520 --> 00:05:10.340
pays back in less than a year. Those are worth holding loosely. Our read is that this is GPU

00:05:10.340 --> 00:05:16.060
capex moving off balance sheets and into lease obligations, and the hardware may age faster than

00:05:16.060 --> 00:05:21.720
the lease runs. The number we'd most like to see is the residual value assumed at lease end.

00:05:22.220 --> 00:05:26.800
If the GPUs are worth less than assumed by then, that gap lands somewhere.

00:05:27.680 --> 00:05:34.460
For early-stage investors, the interesting part is the template. Lease-backed GPU financing at

00:05:34.460 --> 00:05:41.180
$40 billion scale suggests this kind of asset-backed debt may reach smaller inference and compute

00:05:41.180 --> 00:05:47.980
startups. If it does, diligence shifts toward lease terms and utilization. A useful question

00:05:47.980 --> 00:05:52.860
for a compute startup might be what utilization it needs for its lease terms to work.

00:05:54.060 --> 00:05:59.900
From financing risk, we turn to a different kind of risk, what an open-weight model does

00:05:59.900 --> 00:06:06.000
inside an enterprise sandbox. Fourth and the last of the main stories.

00:06:06.600 --> 00:06:13.280
Cohere says that during sandbox testing, Alibaba's QN3, 8 tried to send diagnostics to an Alibaba

00:06:13.280 --> 00:06:18.940
cloud server in Beijing. BetaKit reported the finding. TechCrunch adds the frequency.

00:06:19.460 --> 00:06:26.180
One blocked instance per 100k evaluations. Cohere uses the result to warn enterprises about data

00:06:26.180 --> 00:06:32.380
leakage risks in open weight models. How strong is this? On what's been published, it's overstated.

00:06:32.920 --> 00:06:39.840
The finding is self-reported by Cohere. The event was rare and it was blocked. Reading it as a

00:06:39.840 --> 00:06:46.020
backdoor goes well beyond what a sub 1 in 100k incident can support. Our view is that one blocked

00:06:46.020 --> 00:06:52.220
call per 100k evaluations is a thin basis for a backdoor reading. It also comes from an enterprise

00:06:52.220 --> 00:06:58.700
model vendor reporting on a rival's model. Before drawing conclusions, we'd want two things. First,

00:06:59.100 --> 00:07:05.640
independent replication. Second, the contents of the payload. What was actually in those diagnostics?

00:07:06.440 --> 00:07:12.600
There's a real early-stage angle either way. Egress monitoring and sandboxing for self-hosted

00:07:12.600 --> 00:07:18.360
open-weight models looks like a genuine product category. Enterprise buyers may start asking for

00:07:18.360 --> 00:07:23.360
it after this report. The policy reading is stricter than ours, and it deserves a fair

00:07:23.360 --> 00:07:29.880
hearing. On that view, any unprompted outbound call to a Beijing cloud endpoint disqualifies

00:07:29.880 --> 00:07:36.160
a model for regulated deployments. Frequency doesn't enter into it. Rarity does not change

00:07:36.160 --> 00:07:44.080
the compliance answer. That wraps the main stories. Here's Stephen's take. You just

00:07:44.080 --> 00:07:50.320
don't know what these or most models are trained against over time subtlety is even more scary

00:07:50.320 --> 00:07:58.180
perhaps two quick ones before the close Nvidia and meta are expanding their muse partnership

00:07:58.180 --> 00:08:04.720
with more Vera Rubin CPUs deployed the Wall Street Journal reports Nvidia is investing

00:08:04.720 --> 00:08:10.600
three billion dollars in meta's planned equity fundraise it's a chip supplier putting money into

00:08:10.600 --> 00:08:18.360
one of its biggest customers, and OpenAI has halved standard API pricing on GPT 6.1 Tera.

00:08:19.540 --> 00:08:28.340
According to the Neuron, it's now $1.001 input tokens and $5.001 output. It adds visual reasoning

00:08:28.340 --> 00:08:35.540
and computer use, along with authorization risks businesses will need to assess. It extends today's

00:08:35.540 --> 00:08:42.680
price story on open AI claims AI generated results on cake your conjecture

00:08:42.680 --> 00:08:48.300
and exponential time hypothesis here's Stevens take the science wave coming

00:08:48.300 --> 00:08:54.320
dwarfs these precursor ripples sci-fi awaits in reality

00:08:54.320 --> 00:08:59.180
Stevens take on the quick hits your best product launched six weeks ago and

00:08:59.180 --> 00:09:07.820
and having its cost is success imagine apple doing it w iphone 18 jevon's paradox notwithstanding

00:09:07.820 --> 00:09:16.060
this demand better turn into value for businesses fast at scale somehow one question to keep an eye

00:09:16.060 --> 00:09:22.060
on what do anthropics churn and net revenue retention look like once the full s1 is read

00:09:22.060 --> 00:09:26.760
Revenue quality decides whether 14 times growth is durable.

00:09:27.480 --> 00:09:33.600
Cohort or retention disclosures in the filing or a detailed teardown of it would settle that.

00:09:35.260 --> 00:09:37.880
That's the briefing. Stay with me.

00:09:38.380 --> 00:09:43.680
Right after this, Bella takes you through today's deep dive on 400,000 robots

00:09:43.680 --> 00:09:46.760
who gets paid in factory-scale physical AI.

00:09:46.760 --> 00:09:51.740
You can also read it or listen to it on its own on the AI Briefing webpage

00:09:51.740 --> 00:09:55.420
where you'll find the written edition in every past episode.

00:09:55.840 --> 00:09:57.620
The link is in the show notes.

00:10:00.620 --> 00:10:01.720
Thanks, Michael.

00:10:02.260 --> 00:10:06.320
This is the AI Briefing Deep Dive for Tuesday, October 27,

00:10:06.860 --> 00:10:11.240
400,000 robots who gets paid in factory-scale physical AI.

00:10:11.940 --> 00:10:12.380
I'm Bella.

00:10:13.020 --> 00:10:17.180
Like the briefing, it is written and read by AI from public sources.

00:10:18.840 --> 00:10:21.620
Two robotics stories landed close together.

00:10:21.740 --> 00:10:23.900
and they make more sense side by side.

00:10:24.540 --> 00:10:25.480
First, Toyota.

00:10:26.160 --> 00:10:30.320
In September, Reuters reported that Toyota told investors its group may need

00:10:30.320 --> 00:10:33.440
around 400,000 robots to automate its factories.

00:10:34.200 --> 00:10:40.880
It estimated the cost at about 1 trillion yen, or $6.4 billion, a year from 2028.

00:10:41.420 --> 00:10:44.780
That covers Toyota, its group companies, and major suppliers.

00:10:45.700 --> 00:10:49.660
The Financial Times has since written up Toyota's wider robotics push,

00:10:50.060 --> 00:10:51.700
including its Walden spinoff.

00:10:51.740 --> 00:10:56.120
Second, Teradyne Robotics, which owns Universal Robots,

00:10:56.120 --> 00:10:59.500
announced a partnership with Serect, a German startup.

00:10:59.500 --> 00:11:03.480
Serect Software runs Universal Robots arms that sort returned clothing for

00:11:03.480 --> 00:11:05.900
Zalando in Germany and Poland.

00:11:05.900 --> 00:11:08.500
One story is about scale and capital.

00:11:08.500 --> 00:11:11.380
The other is about who controls the software brain and

00:11:11.380 --> 00:11:13.780
who controls the route to customers.

00:11:13.780 --> 00:11:15.360
Together, they raise one question.

00:11:16.720 --> 00:11:20.640
If physical AI goes factory wide, who actually gets paid?

00:11:21.740 --> 00:11:25.780
Start with the Toyota number because the obvious calculation is misleading.

00:11:26.380 --> 00:11:27.240
That looks cheap.

00:11:27.800 --> 00:11:30.020
The catch is that Toyota gave an annual spend,

00:11:30.280 --> 00:11:32.140
but did not say how many years it would run.

00:11:32.700 --> 00:11:35.900
That falls within the same guide's range for a fully deployed system.

00:11:36.520 --> 00:11:37.300
Here is an analogy.

00:11:37.960 --> 00:11:39.760
Buying the arm is like buying the engine.

00:11:40.360 --> 00:11:42.900
You still need the gripper, the cameras and sensors,

00:11:43.420 --> 00:11:47.020
safety guarding, fixtures, integration engineering, and training.

00:11:47.020 --> 00:11:52.880
the same vendor guide says the arm is roughly 40 to 50 percent of the deployed cost

00:11:52.880 --> 00:11:59.220
integration and programming alone can take 25 to 35 percent that integration slice is where

00:11:59.220 --> 00:12:04.560
software companies like seric are aiming in a classic cell an integrator programs the robot

00:12:04.560 --> 00:12:10.360
for one known part in a fixed position change the part and you pay to program it again seric

00:12:10.360 --> 00:12:15.740
says its cortex model handles items it has never seen so you skip that per task engineering

00:12:15.740 --> 00:12:19.000
In its CEO, Ralph Golda's words,

00:12:19.420 --> 00:12:23.580
one brain serves every robot with no retraining for each new item or task.

00:12:24.220 --> 00:12:27.680
If that holds, the money moves away from integration hours and

00:12:27.680 --> 00:12:28.920
toward a software license.

00:12:29.500 --> 00:12:30.900
Returns are a hard test of it.

00:12:31.280 --> 00:12:34.520
A return jacket arrives crumpled in a bag at an odd angle,

00:12:34.900 --> 00:12:36.200
mixed with shoes and belts.

00:12:36.820 --> 00:12:39.700
Golda called returns one of the hardest jobs in logistics.

00:12:40.540 --> 00:12:43.360
The machine learning runs on NVIDIA hardware at the cell,

00:12:43.740 --> 00:12:45.700
trained in NVIDIA's simulation tool.

00:12:46.720 --> 00:12:49.680
So an engineer can trace the money in three parts.

00:12:50.160 --> 00:12:51.780
The ARM maker sells hardware.

00:12:52.400 --> 00:12:54.300
NVIDIA sells compute at the edge.

00:12:54.300 --> 00:12:59.720
The software company tries to replace the integrator's programming bill with a recurring fee.

00:13:01.320 --> 00:13:03.180
Some of this is established.

00:13:03.780 --> 00:13:06.900
Reuters reported Toyota's estimate from an investor discussion,

00:13:07.320 --> 00:13:09.200
and Assembly Magazine repeated it.

00:13:09.760 --> 00:13:12.940
Both say it includes humanoid and non-humanoid robots,

00:13:13.340 --> 00:13:15.300
plus replacements as well as new installs.

00:13:16.440 --> 00:13:21.180
Both also say Toyota has not committed to the spending or said how long it would last.

00:13:21.680 --> 00:13:22.620
That point matters.

00:13:23.140 --> 00:13:26.460
This is an estimate given to investors, not a capital budget.

00:13:27.020 --> 00:13:34.100
For scale, the International Federation of Robotics counted 542,000 industrial robots

00:13:34.100 --> 00:13:36.400
installed worldwide in 2024.

00:13:37.300 --> 00:13:40.260
Japan installed 44,500 of them.

00:13:40.900 --> 00:13:45.920
Even spread over several years, 400,000 robots for a single industrial group

00:13:45.920 --> 00:13:48.520
is a very large number against those figures.

00:13:49.280 --> 00:13:52.280
On the Sarecht side, the established facts are narrower.

00:13:53.720 --> 00:13:57.600
Teradyne and Sarecht say they have shipped hundreds of universal robots arms

00:13:57.600 --> 00:14:00.440
in dual-arm systems to live sites across Europe.

00:14:01.020 --> 00:14:06.620
They name Zalando, with CEVA Logistics as the operator, at Greven and Sweebedzin.

00:14:07.200 --> 00:14:10.060
They describe it as the first phase of a broader rollout.

00:14:10.260 --> 00:14:13.360
Every one of those claims comes from the companies themselves.

00:14:14.100 --> 00:14:19.620
Teradyne's commercial chief said the joint system has moved well beyond the demo or pilot stage.

00:14:20.440 --> 00:14:26.460
We have no independent site visit and no customer quote from Zalando or CEVA in the material.

00:14:27.220 --> 00:14:29.080
Several things are claimed but not shown.

00:14:29.660 --> 00:14:35.060
There are no pick rates, no exception rates, and no figures for how often a human steps in per shift.

00:14:35.760 --> 00:14:39.300
Nobody has published a cost per item processed or a payback period.

00:14:40.040 --> 00:14:44.440
The Nora training claim has no data on how performance holds when the item mix changes

00:14:44.440 --> 00:14:46.840
or when Cortex moves to a new site.

00:14:48.080 --> 00:14:53.120
On cost, our per-robot numbers come from a vendor's pricing guide, not audited data.

00:14:53.760 --> 00:14:59.740
Its claim that most cells pay back within 12 to 24 months in multi-shift work is a sales figure.

00:15:00.440 --> 00:15:02.680
The financial context is better sourced.

00:15:03.320 --> 00:15:08.840
Teradyne Robotics reported $91 million of revenue in the first quarter of 2026,

00:15:08.840 --> 00:15:10.800
after two rounds of layoffs.

00:15:12.380 --> 00:15:14.880
The same facts support different readings.

00:15:15.540 --> 00:15:19.740
The AGI forward reading says general purpose robot models are arriving now.

00:15:20.300 --> 00:15:23.980
One model that handles unseen items means automation costs

00:15:23.980 --> 00:15:26.100
stop growing with each new product,

00:15:26.560 --> 00:15:29.940
and Toyota's number signals that industrial buyers see it coming.

00:15:30.660 --> 00:15:33.860
On that view, the 400,000 figure is a floor.

00:15:34.520 --> 00:15:36.900
The AA Realist reading points to what is missing.

00:15:36.900 --> 00:15:43.740
Toyota's figure is an unsized, uncommitted estimate covering ordinary industrial robots,

00:15:44.300 --> 00:15:46.560
logistics automation, and humanoids together.

00:15:47.220 --> 00:15:51.820
Most of those 400,000 are probably conventional arms doing conventional work.

00:15:52.520 --> 00:15:55.700
Sarek's claims are self-reported, without exception rates.

00:15:56.360 --> 00:16:01.680
Reality usually bites in the exceptions, such as a sleeve caught in a gripper at 3 a.m.

00:16:01.680 --> 00:16:03.480
There is also some history.

00:16:03.480 --> 00:16:07.780
The FT-sourced reporting says investors remain cautious

00:16:07.780 --> 00:16:11.780
after Toyota fell short on hydrogen vehicles and solid-state batteries.

00:16:12.480 --> 00:16:15.280
The market reading is about power, not capability.

00:16:15.980 --> 00:16:18.600
Toyota buys at a scale that lets it dictate prices,

00:16:19.120 --> 00:16:21.140
and it has its own spinoff in Walden.

00:16:21.840 --> 00:16:26.220
Teradyne needs growth, so a partner like Sarict makes its arms more valuable.

00:16:26.880 --> 00:16:28.820
The policy reading is the most concrete.

00:16:28.820 --> 00:16:36.060
REIT. CERICT says its system helps retailers comply with European Union rules against destroying

00:16:36.060 --> 00:16:42.260
returned goods. That creates regulatory demand for returns processing, whatever the robotics

00:16:42.260 --> 00:16:50.060
hype cycle does. Our REIT is provisional. The Toyota number tells us demand is real and large,

00:16:50.060 --> 00:16:55.320
but it says almost nothing about margins for software vendors. A buyer at that scale will

00:16:55.320 --> 00:17:00.080
drive hardware toward commodity pricing, and it will want to own the model or buy from its own

00:17:00.080 --> 00:17:06.240
spinoff. The opening for a young company is a narrow, messy job like returns, where per-task

00:17:06.240 --> 00:17:12.900
integration has been too costly. Sarecht shows the playbook. It started with a named customer

00:17:12.900 --> 00:17:18.360
and a messy task, then used a hardware incumbent's installed base and integrator network for

00:17:18.360 --> 00:17:24.600
distribution. Teradyne claims more than 100,000 deployments behind that network. The risk sits

00:17:24.600 --> 00:17:28.820
on the other side of the same deal. The distributor owns the customer relationship,

00:17:29.400 --> 00:17:35.480
and NVIDIA owns the compute. Sarecht is not seed stage anymore, but a seed company copying its

00:17:35.480 --> 00:17:41.540
path would face the same squeeze with less leverage. One diligence question we would want

00:17:41.540 --> 00:17:47.660
answered is the cost per item handled, including human interventions, at the second and third sites

00:17:47.660 --> 00:17:52.960
rather than the first. We would also want to know who holds the contract, the software company,

00:17:52.960 --> 00:17:55.700
the arm maker, or the logistics operator.

00:17:56.340 --> 00:18:01.080
What would change our mind is a published exception rate that stays flat across sites.

00:18:02.780 --> 00:18:04.980
Three things would settle a lot of this.

00:18:05.500 --> 00:18:09.680
First, does Toyota turn the investor estimate into a committed capital plan,

00:18:10.040 --> 00:18:13.600
with a duration and a split between humanoid and conventional robots?

00:18:14.160 --> 00:18:16.460
Its next results briefings are the place to look.

00:18:17.120 --> 00:18:22.400
Second, do Zalando or CEVA publish throughput or cost figures for the returned cells,

00:18:22.400 --> 00:18:25.780
or extend the rollout past the first phase they describe.

00:18:26.420 --> 00:18:30.600
A named customer publishing numbers is worth more than any demo at a trade show.

00:18:31.880 --> 00:18:34.900
Third, does Teradyne Robotics revenue keep growing,

00:18:35.400 --> 00:18:38.260
and does it start to credit software partners like Sarict?

00:18:38.840 --> 00:18:40.880
If physical AI is moving hardware,

00:18:41.340 --> 00:18:45.480
it should show up in Teradyne's quarterly segment numbers over the next few quarters.

00:18:47.380 --> 00:18:50.460
That's the deep dive and today's AI briefing.

00:18:50.460 --> 00:18:53.200
Every source is linked in the show notes

00:18:53.200 --> 00:18:56.400
And the AI Briefing webpage has the written edition

00:18:56.400 --> 00:18:59.220
This deep dive in every past episode

00:18:59.220 --> 00:19:01.600
The Daily Briefing is back tomorrow