Google is playing its "Android of robotics" card: providing the brains for humanoids, not building them

Google is making its Android move in robotics: supplying the brains of humanoids, not building them

🔎 Two strategies, one war

Two sets of news dropped just weeks apart, and they tell the same battle from two opposing angles. On one side, Tesla has converted a Model S/X-dedicated line at Fremont to assemble the Optimus Gen 3, and is now turning out several hundred robots per week. On the other, Google DeepMind unveiled Gemini Robotics 2 on July 30, 2026: the first model capable of controlling a humanoid's legs, torso, arms and hands under a single policy.

One is industrializing the body. The other is standardizing the mind. As WebProNews sums it up, Google's robotics bet deliberately mimics its Android play, while Tesla chases industrial scale.

This divergence is not an engineering detail. It will decide who captures the value of a market everyone announces as colossal, without knowing when it will truly take off. The question that should matter to you — whether you're a developer, an investor, or a mere observer — comes down to one sentence: will the humanoid robot of 2035 look like the smartphone (commoditized hardware, licensed brain) or the Model 3 (everything integrated, everything controlled)?

The Essentials

• Gemini Robotics 2 (July 30, 2026): DeepMind's first model to control an entire humanoid — legs, torso, arms, hands — under a single policy.
• Tesla Optimus: production rate multiplied by ~10 in a few months, from several dozen to several hundred units per week at Fremont; target of a line producing more than 1,000 robots/week by the end of 2026.
• Google plays horizontal: Intrinsic reintegrated into the heart of the group, Gemini Robotics powering Boston Dynamics' Atlas (DeepMind/Hyundai partnership).
• Tesla timeline: public sales "probable" by the end of 2027; Musk promises "thousands" of Optimus units by the end of 2026.
• Real bottlenecks: dexterous hands, suppliers, Chinese controls on rare earths (in effect since April 2025).
• The stakes: the Android model (commoditized hardware, monetized brain) versus the Apple model (end-to-end vertical integration).

Recommended tools

This news doesn't upend your everyday tools. But the "brain" layer of this future industry can be tested right now, and it won't cost you a dime.

My advice: stick to free tiers until your use case is validated. Models change too fast to justify a premature paid commitment.

Gemini Robotics 2: one model for the whole body

Direct answer: as of July 30, 2026, Gemini Robotics 2 is DeepMind's first model to control an entire humanoid — locomotion, torso, arms, hands — under a single policy.

That technical detail is in fact a turning point. Most current systems stack specialized controllers: one module for walking, another for manipulation, a third for balance. Gemini Robotics 2 merges these layers into a single model. Concretely, the robot coordinates its entire body in service of a task — which makes fine bimanual gestures possible, like carrying a crate while opening a door for itself.

RoboZaps, in its roundup of the 38 best humanoids of 2026, sees this as the shift from limb-by-limb control to whole-body control. A first for DeepMind, and a signal sent to the entire industry: general dexterity is no longer a lab dream, it's a roadmap.

Second piece of the puzzle: Google isn't staying in the realm of software abstraction. According to Biz Chosun, DeepMind now equips Boston Dynamics' Atlas through the DeepMind/Hyundai partnership, and Intrinsic — Alphabet's industrial robotics arm — has been reintegrated into the heart of the group. Google surrounds itself with hardware partners without ever breaking its implicit rule: it will not manufacture actuators.

This is no isolated first shot. Gemini Robotics was announced back in March 2025, with partnerships at Apptronik — in which Google invested in late 2024 — and Agile Robots. The logic has been consistent for two years: one model, dozens of bodies.

It's the Android playbook, to the letter. One OS, manufacturers, zero phone factories — sorry, robot factories.

The Android Analogy: Right on the Economics, Yet to Be Proven on the Technology

Direct answer: the analogy holds up remarkably well on the economic model, but it remains to be demonstrated on the technical front.

Let's start with what works. Google bought Android in 2005 for roughly $50 million, then gave it away to the world in November 2007 through the Open Handset Alliance, a coalition of some thirty manufacturers. The first Android phone, the HTC Dream, only arrived in October 2008. In the meantime, Google had convinced Samsung, HTC, LG, and dozens of others to build on its foundation.

The result: Android now powers about 70% of the world's smartphones (StatCounter, 2025). Google never mass-produced these phones, but it monetized the layer above — Play Store, services, advertising. The hardware got commoditized; the brain captured the margin.

This doctrine extends beyond robotics, for that matter: Google is pushing Android itself toward the status of an "intelligence system" with Gemini Intelligence and Googlebooks. Robotics is simply a logical extension of a twenty-year-old strategy: control the intelligence layer, let others manufacture the glass and aluminum.

But here's what the enthusiasts forget. A smartphone rests on a standardized hardware base: ARM chips, sensors, screens — components that are interchangeable from one manufacturer to the next. A humanoid is the opposite: every maker has its own actuators, its own kinematics, its own sensors, its own hands. A single policy that drives an Atlas just as well as an Apptronik robot doesn't happen by decree. It has to be proven robot by robot, and that will take years.

And there's a track record that partners haven't forgotten. Google bought Boston Dynamics in 2013, sold it to SoftBank in 2017, then disbanded its Everyday Robots team in 2023. Alphabet has already been burned in robotics, and manufacturers will sign a dependency agreement with the memory of those abandonments.

Analyst Linas Beliunas does indeed talk about an "Android moment" for robotics (LinkedIn). But an Android moment presupposes manufacturers that stay. That's the variable nobody can guarantee today.

Tesla in Fremont: vertical integration, with the numbers to back it up

Direct answer: Tesla is now producing several hundred Optimus units per week, a 10x increase in a few months — and unlike previous promises, this figure is documented.

The data comes from The Information, relayed by Gadgets Now: in the second quarter of 2026, output was in the tens of units per week. It has since risen to several hundred. Tesla is aiming for a line capable of more than 1,000 robots per week by the end of 2026.

The Fremont site in California was converted from the Model S/X lines to accommodate Optimus Gen 3 assembly, with production starting in late July–early August 2026. And RoboZaps indicates that public sales are "probable" by the end of 2027.

Let's stay clear-eyed on two points, because the topic calls for it.

First point: Elon Musk promised in June 2026 "thousands" of Optimus robots by the end of the year (ETC Journal). At a rate of several hundred per week, the cumulative target is achievable. But his track record of Optimus predictions, going back to 2021, warrants healthy skepticism: the deadlines have consistently slipped.

Second point: several hundred per week is tiny on an industrial scale. A car factory produces more vehicles in a single day. The 10x is real and deserves respect; the road ahead is still long.

What the vertical play offers Tesla, on the other hand, is a closed data loop: robots built in-house, deployed in its own factories, feeding its models with proprietary physical data. It's the Apple argument applied to robotics — and it's precisely what Google can't buy.

Two business models, one question: who captures the value?

Direct answer: when hardware becomes commoditized — and it will — the margin will go to the intelligence layer. The entire Google/Tesla duel boils down to anticipating this shift.

The history of technology offers precedents for both outcomes. The PC rewarded the software layer: Microsoft dominated for thirty years while assemblers fought over margins of a few percentage points. The smartphone rewarded both extremes: Apple captured most of the hardware profits through integration, while Google captured most of the usage through Android.

Robotics will likely be decided segment by segment. Industry first — where Intrinsic plays, with constrained robots and repetitive tasks. The general public next, assuming a domestic humanoid ever finds a massive use case beyond the demo.

There is, moreover, a third path, rejected by both giants: the full-stack approach. Genesis AI just unveiled GENE-26.5 and its own humanoid robotic hands: model and hardware designed together, without depending on either Google or a subcontractor. It's costly, but it avoids the classic trap of the middle layer being crushed between two integrators.

My opinion, without detours: Google is right about the long-term trajectory, Tesla is right about the next three years. The vertical wins the early battles; horizontality wins the standardization wars — provided it survives long enough to fight them.

The Underestimated Variable: China, Magnets, and Hands

Direct answer: the number one bottleneck in this industry isn't AI — it's the supply chain for magnets and dexterous hands.

Chinese controls on rare earths — indispensable to robot motors — are directly hampering Optimus production, as reported by ETC Journal. These restrictions, in place since April 2025, hit exactly the component a humanoid needs by the dozen: finger motors.

Because the other bottleneck, documented by The Information, is the hands. Manufacturing a five-fingered hand capable of manipulating delicate objects, at scale, at controlled cost: no one has solved this industrial problem. Tesla knows this better than anyone — it's its own bottleneck, cited in black and white by its suppliers.

Yet this is precisely where China is moving fast. Unitree deployed its G1 at Haneda Airport in Japan — a sign that Chinese physical robotics is scaling up and exporting to demanding markets. If Chinese hardware commoditizes robot bodies the way it commoditized drones, the Android scenario becomes mechanical: interchangeable bodies, a licensed brain.

A delicious paradox: Chinese pressure on hardware strengthens Google's thesis while weakening Tesla's, which must secure its magnets and suppliers to keep up its production pace.

The timeline: three deadlines that will settle the duel

Direct answer: three dates are enough to know who's right — end of 2026, end of 2026, and end of 2027. Yes, end of 2026 twice.

• End of 2026: Does Tesla reach a rate above 1,000 Optimus/week, as targeted according to The Information?
• End of 2026: Does Musk deliver "thousands" of Optimus units produced, as promised in June 2026?
• End of 2027: first public sales of Optimus, judged "likely" by RoboZaps?
• 2027: Does Gemini Robotics 2 equip Atlas and Apptronik's robots in real production, beyond the demos?

My recommendation: follow the production rates, not the keynotes. A verifiable weekly rate says more than a demo video polished in the edit. If Tesla crosses 1,000/week, the vertical takes an enormous psychological lead. If the bottlenecks persist — hands, magnets, suppliers — Google's window to install its OS at the manufacturers widens accordingly.

What this changes for developers — and how to position yourself

Direct answer: the layer that is standardizing first is the model layer, and it's already available for free. You don't need a robot to prepare for the Android moment.

Three skills will be worth their weight in gold when humanoids become widespread: multimodal reasoning (understanding a video scene and acting on it), sim-to-real (transferring behaviors from simulation to the real world), and orchestrating physical agents. The first two can be practiced today with models like Gemini 3.1 Pro, GPT-5.5, or Claude Opus 4.7 (Adaptive).

Good news: you can do this without spending a dime. Our guide to free AI APIs lists Groq, Google AI Studio, and OpenRouter, which give access to top-tier models with generous quotas. It's the best entry point for understanding what "giving a machine a brain" actually means in practice.

A practical tip: document everything. The robotics market will lack serious French-language references for years, and those who publish their experiments — with a simple self-hosted site on Hostinger — will capture the audience before the big media outlets arrive.

The Android moment of robotics, when it comes, will reward those who know the intelligence layer before everyone else. Not those who waited for the keynote.

❌ Common Mistakes

Mistake 1: Confusing production rate with sales

"Several hundred per week" does not mean "hundreds sold." Current Optimus units are working first in Tesla factories, where their opportunity cost is zero. Public sales remain "probable" by the end of 2027, not confirmed. Don't build a business plan on robots that don't yet exist on the market.

Mistake 2: Believing the Android analogy is a copy-paste

Android took hold on standardized hardware: ARM, screens, interchangeable sensors. Robotics has nothing of the sort — every actuator, every hand, every kinematic differs. A universal policy must prove itself robot by robot. The analogy is a solid working hypothesis, not an established fact.

Mistake 3: Neglecting dependence on critical components

Chinese controls on rare earths, in effect since April 2025, show that a business model can be blocked by a magnet. If you are investing or positioning yourself in this market, audit the supply chain before the demo. The best policy in the world is useless without motors.

❓ Frequently Asked Questions

Does Gemini Robotics 2 replace classical control code?

Not entirely. It unifies whole-body control under a single policy, which replaces some of the specialized controllers. But safety layers, hardware limits, and critical real-time systems remain necessary. Industrial robots won't switch over overnight: the transition will happen module by module, starting with manipulation.

Will Tesla sell Optimus to the general public?

Nothing is confirmed, but RoboZaps indicates that public sales are "probable" by the end of 2027. In the meantime, the robots are serving in Tesla factories. The price remains unknown; Musk has mentioned targets around $20,000 to $30,000 in the past, without ever officially confirming them. Price announcements rarely precede deliveries.

Why doesn't Google build its own robots?

Three reasons: a painful track record (Boston Dynamics sold off in 2017, Everyday Robots dissolved in 2023), a business model that is more profitable through licensing than through manufacturing, and a consistent doctrine since Android — capture the intelligence layer, leave the hardware to others. The reintegration of Intrinsic shows a presence on the industrial software side, without ever moving back up toward manufacturing.

Who wins between Google's strategy and Tesla's?

Over three years, Tesla: the product is vertical, will soon be for sale, and it controls its data loop. Over ten years, the advantage tilts toward Google if hardware becomes commoditized, as was the case with Android. Google's real risk isn't Tesla but abandonment — Alphabet has already left this field once. Consistency will be its main challenge.

How can you test "brain" models without a robot?

Start with the free APIs: Google AI Studio for Gemini 3.1 Pro, Groq for fast inference, OpenRouter to compare GPT-5.5 and Claude Opus 4.7. Add Hugging Face's open source LeRobot library to experiment with control policies in simulation. Our guide to free AI APIs details the quotas and limits of each option.

✅ Conclusion

Google doesn't want to build tomorrow's humanoids: it wants to be their Android, while Tesla proves in Fremont that vertical integration can scale — several hundred Optimus units per week, with a target of 1,000 by the end of 2026. Both can't be right at the same time within the same decade: watch which one — the brain or the body — captures the margin first.

To understand the intelligence layer today, start with our guide to free AI APIs — that's where the next standardization will play out.