Physical AI Digest is a weekly briefing produced by Klaudia from Physical AI Company xBerry - a tech company based in Poland building tools at the intersection of Physical AI and operations.
The photo used in the cover was uploaded from https://www.therundown.ai/news/openai-altman-humanoid-robots-hardware-training-data.
On September 3, Sam Altman confirmed on the Sources podcast that OpenAI will definitely build a humanoid robot. No production partner, no release date, no prototype - a statement of direction from the company that currently leads the industry on VLA model capability. What exists is a hiring pipeline for hardware engineers in San Francisco and a clear strategic rationale: controlling the hardware layer gives OpenAI control over how its models act in the physical world and how it collects robot training data at scale. In a category where software is increasingly the constraint, that matters more than any single product announcement.
This week also brought the two clearest demand signals the humanoid category has ever seen from named end users. Toyota disclosed that its global manufacturing operations will need approximately 400,000 robots and $6.4 billion per year in CAPEX from 2028. Hyundai confirmed Atlas deployment at its Metaplant in Savannah, Georgia, starting 2028 - initially for parts sequencing, expanding to component assembly by 2030. These are CAPEX line items, not aspirational roadmaps.
Stats:
| Value | Description |
|---|---|
| definitely | Sam Altman's word on OpenAI building a humanoid robot - no date, no hardware partner, no prototype, but hardware engineers are already being hired in San Francisco |
| 400K | robots Toyota estimates it will need across factories, group companies, and key suppliers - $6.4 billion per year in CAPEX from 2028 |
| 2028 | start of Atlas deployment at Hyundai Metaplant in Savannah, Georgia - parts sequencing first, component assembly from 2030 |
| 8.64s | TianGong Ultra's 100m at the World Humanoid Robot Games - down from Honor Lightning's 9.32 in August; crash mats still required |
OpenAI Enters Humanoids: What "Definitely" Changes
The companies currently shipping humanoid robots have solved the hardware problem to varying degrees. The software layer - the VLA model that lets a robot interpret instructions and translate them into physical actions - is where OpenAI's advantage is clearest and most compounding. Every robotics company not building its own foundation model is already dependent on someone else's software stack.
Altman also stated that everyone should eventually have a personal robot. OpenAI's near-term target is robots that support skilled workers on infrastructure and construction sites; long-term, general-purpose personal robots. Neither horizon has a date. What exists is the world's most capable VLA research organization moving from pure software toward hardware - and already hiring the engineers to do it.
The September 3 announcement does not threaten any robot shipping today. It changes the planning horizon for every company deciding, in the next twelve months, which software architecture to build around. That is slower to materialize than a product launch, harder to respond to, and ultimately more consequential. A company that builds its logistics deployment on OpenAI's VLA stack today is betting that OpenAI won't compete on hardware. That bet just got harder to make.
OpenAI has built the foundation model that the humanoid industry increasingly runs its software on. When the foundation model provider announces it will also build the hardware, every company sitting between those two layers has to recalculate its competitive position. The recalculation is not urgent - there is no product, no timeline, no competitor to respond to today. But the direction is clear, and direction in this industry compounds faster than most people expect.
Toyota 400K and Hyundai Savannah: The Demand Side Is No Longer Speculative
Toyota disclosed that it estimates its global manufacturing operations will need approximately 400,000 robots - covering its own factories, group companies, and key suppliers - with annual CAPEX of approximately 1 trillion yen ($6.4 billion) from 2028. This is not an analyst projection. It is Toyota communicating an internal estimate to its investors, produced by the same operations team that runs Toyota Production System across 15 million vehicles per year.
Toyota's estimate is the largest single end-user demand figure the humanoid category has ever seen from a named customer. The number does not guarantee procurement at that scale. What it tells you is what Toyota's engineering leadership believes it needs - stress-tested internally before being shared publicly.
Hyundai's Atlas deployment at the Metaplant in Savannah, Georgia, adds a specific location and a specific timeline. Parts sequencing from 2028. Component assembly from 2030. Two manufacturers with combined global production of over 20 million vehicles annually have disclosed specific robot deployment plans with specific dates attached to CAPEX positions. Every supplier, integrator, and capital allocator in Physical AI now has a named customer and a named date to build around.
The Sprint Record Fell Again
In August, Honor Lightning set the World Humanoid Robot Games 100m record at 9.32 seconds. On September 14, TianGong Ultra brought it to 8.64 seconds. The capability is improving faster than the competition schedule. The braking problem remains unsolved - crash mats at 75 kilograms and 17+ meters per second. What changed is the speed at which the speed itself is improving: the record dropped by nearly a full second in one month of competition.
What to Watch Next
- OpenAI hardware partner: Altman's "definitely" needs a manufacturing partner to become a product - the announced partnership or first-party production announcement will tell you how quickly OpenAI plans to move from intent to hardware.
- Toyota first vendor selection: 400,000 robots requires procurement decisions - the first announced supplier will show which company Toyota's TPS engineering team trusts to meet its quality standards.
- Hyundai Savannah pilot confirmation: 2028 parts sequencing start means procurement decisions are being made now - watch for Atlas order confirmation from Boston Dynamics.
- Next World Humanoid Robot Games benchmark: the 100m record went from 9.32 to 8.64 in one month - the next event will show whether the improvement rate is holding.
FAQ
Q: OpenAI already has partnerships with robot companies. What is different about building its own humanoid?
Partnerships give OpenAI's models access to hardware someone else manufactures and sells. Building its own humanoid gives OpenAI control over the training data pipeline - every task the robot performs in the real world becomes proprietary training data that improves OpenAI's models in ways that a partnership arrangement does not. It also gives OpenAI margin: a partner sells hardware and licenses software separately; a vertically integrated product captures both. The most direct analog is Apple's move from licensing its OS to building its own chips. Altman's stated rationale - robot training data and model control - maps directly onto what a vertically integrated product provides and a partnership does not.
Q: Why does Toyota's 400,000-robot estimate matter more than analyst projections?
Because it comes from the end user. Analyst projections estimate what manufacturers might buy, built from adoption curves and market models. Toyota's figure is what its own operations team calculated it would need - communicated to investors as an internal assessment. Toyota Production System teams don't produce numbers for investor calls that haven't been stress-tested against real facility layouts, operational requirements, and supplier capacity. The 400,000 figure reflects actual automation requirements as Toyota's engineering leadership understands them. It does not commit Toyota to purchasing that many robots. But it tells you what Toyota believes the number is - which is more informative than any external projection of the same demand.
Top comments (0)