Amazon lost 6.3 million orders because an AI tool wrote bad code. China shipped 90% of the world’s humanoid robots. Both stories are about the same thing: humans losing control of the speed they created.
Oil above $110, a hawkish Fed, and stagflation risk
Est. reading time: 7 min
The Week in 60 Words
Amazon convened an emergency engineering meeting after AI-generated code caused cascading outages across its retail platform — 6.3 million orders lost in a single day. Meanwhile, new data showed China shipped over 90% of the world’s humanoid robots in 2025. Unitree alone delivered 5,500 units; Tesla, Figure, and Agility shipped 150 each. The automation age is arriving. The guardrails are not.
Big Idea
The real bottleneck is human
Two stories dominated this week, and they’re really one story. Amazon’s AI coding tools caused a series of outages that wiped out millions of orders. And China shipped nearly all the world’s humanoid robots — at a scale that makes American efforts look like science fairs.
The connection isn’t obvious until you look at the structure. Amazon deployed AI to write code faster and cut costs. It worked — until it didn’t. The AI generated more code than humans could review. Errors scaled at machine speed. A single bad deployment took down checkout, pricing, and account access for six hours.
The humanoid robot story follows the same logic in reverse. China’s advantage isn’t smarter robots — it’s a manufacturing ecosystem that can push units off the line before anyone proves the economics work. Unitree’s cheapest model costs $5,900. Figure AI, valued at $39 billion, shipped 150 units.
AI isn’t replacing human judgment. It’s revealing how little of it there was to begin with.
Amazon’s response is telling: mandatory senior sign-offs, a 90-day safety reset across 335 critical systems, dual human review for every production deployment. In other words, more humans, not fewer. The same pattern will unfold in factories, warehouses, and every industry betting on automation. The first wave saves headcount. The second wave discovers that the headcount it cut was the only thing preventing catastrophe.
For investors, builders, and anyone whose job touches technology — which is everyone — the signal is clear. The scarcest resource in the AI era isn’t computing power. It’s the ability to know when to say no.
The Model
THE AUTOMATION TRUST GAP
1. Company deploys AI to boost speed and cut headcount
↓
2. AI generates more output than humans can review
↓
3. Errors scale at machine speed — outages, lost orders, cascading failures
↓
4. Company re-imposes human oversight (sign-offs, audits, resets)
↓
5. The “automation dividend” shrinks — the bottleneck was never speed, it was judgment
Every automation wave eventually discovers the same thing: the last mile is human, and it costs more than the first mile saved.
By the Numbers
HUMANOID ROBOT SHIPMENTS BY COMPANY, 2025
Units sold globally · China (red) vs U.S. (blue)
Source: Omdia via Rest of World / Visual Capitalist, Mar 2026
China’s top two players shipped more humanoid robots than the rest of the world combined. The U.S. shipped fewer units total than a single Chinese factory’s monthly output. This isn’t a competition — it’s a manufacturing class gap that mirrors the solar and EV trajectories of the past decade.
AMAZON AI CODING INCIDENTS — ESCALATION TIMELINE
Q4 2025 – Q1 2026 · Impact measured by scale of disruption
Source: Financial Times, Digital Trends, CNBC · Mar 2026
Four incidents in three months, each larger than the last. The pattern isn’t random — it’s what happens when AI-generated code volume outpaces review capacity. Amazon’s fix — re-inserting humans into the loop — is the template every enterprise will eventually follow.
Signal vs Noise
Amazon launches 90-day safety reset across 335 critical systems after AI coding outages
Dual human review now mandatory for production code — the fastest admission yet that AI speed without governance is a liability.
Digital Trends · FT
China shipped ~90% of 14,500+ humanoid robots globally in 2025
Unitree led with 5,500 units; Tesla, Figure, and Agility shipped 150 each — the manufacturing gap mirrors the earlier solar and EV divergence.
Visual Capitalist · Omdia · Counterpoint
NBER survey: 80%+ of 6,000 executives report zero AI productivity gains
Despite 70% active adoption, the disconnect between AI hype and enterprise reality echoes the Solow productivity paradox of the 1980s.
Fortune · NBER
Figure AI valued at $39B — shipped 150 robots in 2025
A 15x valuation jump in 18 months for a company whose entire output equals a quiet Tuesday at Unitree’s factory.
Figure AI · Omdia
Fortune: AI coding errors piling up as companies push senior tasks to junior staff + AI
Apiiro found developers using AI introduced 10x more security issues — the code is faster, the bugs are bigger.
Fortune · Apiiro · Bain
What I’m Reading
Normal Accidents
Charles Perrow · 1984 · Princeton University Press
I picked this up because Amazon’s outages are a textbook Perrow scenario — tightly coupled systems where a small failure cascades into a catastrophe nobody designed for. What I agree with: when you optimize for speed in complex systems, you don’t eliminate failure — you make it invisible until it’s massive. What I find too neat is the assumption that decoupling always helps — sometimes the coupling is the product. What I’m taking from this: before you automate a process, map the failure modes first. If you can’t, you’re not ready.
One Number
6.3M
The number of Amazon orders lost in a single outage on March 5, 2026 — linked to AI-assisted code deployment. That’s not a bug report. That’s a quarter’s worth of bad earnings compressed into six hours. And it happened because AI-generated code entered production faster than any human could review it.
This Week I Noticed
I’ve been using AI tools more aggressively lately — testing new options, pushing the limits of what I can delegate. Tasks that used to eat up a full day now take about an hour. The productivity gain is real. But I’m starting to notice something else: an uncomfortable dependence. When you outsource too much of the thinking, you risk losing the sharpness that made you effective in the first place. The analytical muscle needs friction to stay strong. I’m watching myself for that trade-off now.
Meanwhile, a CodeRabbit study this month found that AI-generated pull requests contained 75% more logic and correctness errors than human-written code. Not syntax errors — logic errors. The kind that pass tests but break in production. The kind Amazon discovered. The finding tracks with what Andrej Karpathy has been saying: AI can write code that looks right and runs right and is still fundamentally wrong. We’re entering an era where the most dangerous code is the code that seems to work.
Quote of the Week
“We shape our tools, and thereafter our tools shape us.”
— John Culkin
Culkin wrote this in 1967, channeling McLuhan. Sixty years later, it reads like a field report. Amazon shaped Kiro. Kiro shaped Amazon’s infrastructure — by deleting it. The question for everyone in 2026 isn’t whether AI tools will change your work. It’s whether you’re still the one deciding what the work is.
Three things to remember
AI tools are generating more work than organizations can safely review — Amazon’s outages prove that speed without governance is a liability, not an asset.
China shipped 90% of the world’s humanoid robots in 2025 while America’s top three players combined shipped 450 units — the manufacturing gap in physical AI mirrors the earlier divergence in solar and EVs.
The real bottleneck in automation isn’t technology — it’s human judgment, and every organization that cuts it discovers this the hard way.
One Thing To Do This Week
Audit one process you’ve recently handed to AI — code review, drafting, analysis, anything. Ask: who checks this before it ships? If the answer is “nobody” or “me, quickly,” add a human checkpoint now. Amazon waited until 6.3 million orders were lost. You don’t have to.
Support This Newsletter
Sponsor
Trading 212
No account yet? Use my link — you get a shot at free shares, and you help keep this newsletter going. No catch.