Socks, Shelves, and Shipping Robots: Takeaways From Silicon Valley

What We Learned in San Francisco

Spending time with robotics teams in Silicon Valley highlighted a fundamental split between how the U.S. and China approach humanoids.

In China, more than 40 companies are already mass-producing hardware—hundreds or thousands of units at a time. You can buy a Chinese humanoid today, but you have to program it. These companies focus on distributors and avoid the complexity of serving end customers.

In the U.S., it’s the opposite: “You’re not allowed to program a Figure robot. You’re not allowed to program an Optimus. You’re not allowed to program a 1X Neo.”

American teams want to control hardware and software, sell directly, and keep robots closed. But none are actually shipping yet. Nils challenges: “Try to buy an American robot. Let me know how it goes.”

Manipulation, Locomotion… or Something Else?

A recurring theme in our Silicon Valley meetings: labs keep arguing whether to start with locomotion or manipulation.

But as JX from Gradient put it, robotics has three pillars:

1. Manipulation
2. Locomotion
3. Spatial semantic perception

Almost nobody is starting with perception. That’s the bet we’re making.

Why? Because many perception-driven tasks—like managing store shelves, detecting issues, guiding staff—are actually worth more than folding socks or vacuuming floors. As Nils put it: “Who makes more money, the store clerk or the store manager?”

The world is full of high-value cognitive tasks that robots can do today. Housekeeper robots may sell billions eventually, but retail perception robots could realistically scale to 100,000–500,000 units this decade.

Socks vs. Shelves

Sunday’s sock-folding demo was impressive, but folding “one pair every two minutes” isn’t where the money is. Contrast that with a robot that inspects shelves and tells staff what needs doing—which large retailers say could save 30–60 minutes per employee per day.

That’s why we’re prioritizing perception: actionable intelligence first, physical labor later.

Why We’re Going to Market in 2026 (While Others Aren’t)

A surprising (and candid) takeaway from U.S. teams was a general lack of urgency to commercialize. One AI researcher from Figure even said publicly that “the market opportunity is so small today that it’s not worth the commercialization effort.”

Nils then joked: “You’re saying the quiet part out loud. You’d better delete this before the SPAC.”

He did.

Many American teams showed the same pattern—heavy research drive, little go-to-market pressure, and concern that generating revenue too early complicates valuations.

China, meanwhile, is pressing ahead with mass production.

This gap is exactly where we see the opening.

Our 2026 Ambition: Largest Real Deployments in the World

We plan to deploy around 500 working semi-humanoids in 2026, which would exceed every current humanoid deployment globally (most are “dozens” at best). And by 2027, scaling into the 10,000+ range is realistic.

We believe the Chinese hardware is already good enough—and that we’ve figured out what it’s for.

Pilots: Now Over 6,500 Locations

Our ongoing paid pilots (each paying $10,000+ to try Cactus) now represent 6,500+ retail locations across multiple enterprise chains.

Not all locations are active—big chains often pilot in 3–30 stores—but their decision making is based on their entire footprint.

With luck, we may cross 10,000 by year-end.

We’re still in price and process discovery, but our target is clear: focus on enterprises that can spend at least $1 million per year with us. Smaller deployments will go through partners and distributors.

Help Train Our Looker-Pointer Robot

We’re building a customer-facing robot that can answer store questions and point to products. To train it, we need real customer questions.

Record yourself asking what you would ask a store robot. We’ll reward 500 $AUKI per question, up to 5 questions, and likely cap submissions at ~100 people.

See here for more details and submission form.

Auki on the Road

A quick travel update from the call:

• Nils is currently in Taipei to pitch at AppWorks, in Mandarin.
• Next: Shenzhen for an Intercognitive meetup.
• Then Beijing for university lectures and lab visits.
• Then Europe to support pilots.

And in Q1 2026, we’ll be setting up our San Francisco presence together with Mentra. Some of us—including Nils—will spend the whole quarter there.

Hardware: Will We Build Our Own?

We’re in discussions with several Chinese OEMs about co-designing hardware optimized for our use cases. But our current position is “Hardware will get commoditized quickly. It’s better strategically to be hardware-agnostic.”

We may do a limited hardware effort to accelerate software development, but long term we expect margins to shift toward the network and perception layer.

Robotics Gossip

A few spicy industry observations we can repeat publicly:

• Multiple people in Silicon Valley believe Neo won’t ship (we disagree).
• Many U.S. teams are incentivized to “talk about the future” rather than ship.
• Some founders openly admit that going to market is riskier than staying in R&D because revenue forces realistic valuations and makes it more difficult to raise money.

And yes—after the livestream ends, we share even more gossip privately on Discord.

アウキ・ラボについて

Aukiはポーズメッシュという地球上、そしてその先の1000億の人々、デバイス、AIのための分散型機械認識ネットワークを構築しています。ポーズメッシュは、機械やAIが物理的世界を理解するために使用可能な、外部的かつ協調的な空間感覚です。

私たちの使命は、人々の相互認知能力、つまり私たちが互いに、そしてAIとともに考え、経験し、問題を解決する能力を向上させることです。人間の能力を拡大させる最も良い方法は、他者と協力することです。私たちは、意識を拡張するテクノロジーを構築し、コミュニケーションの摩擦を減らし、心の橋渡しをします。

ポーズメッシュについて

ポーズメッシュは、分散型で、ブロックチェーンベースの空間コンピューティングネットワークを動かすオープンソースのプロトコルです。

ポーズメッシュは、空間コンピューティングが協調的でプライバシーを保護する未来をもたらすよう設計されています。いかなる組織の監視能力も制限し、空間のプライベートな地図の自己所有権を奨励します。

分散化はまた、特に低レイテンシが重要な共同ARセッションにおいて、競争優位性を有します。ポスメッシュは分散化運動の次のステップであり、成長するテック大手のパワーに対抗するものです。

アウキ・ラボはポスメッシュにより、ポーズメッシュのソフトウェア・インフラの開発を託されました。