What happened
Figure AI has officially frozen the design for its next-generation humanoid robot, Figure 04, and is shifting its primary focus toward scaling its AI model, Helix, through massive investments in data and . During an all-hands meeting, CEO Brett Adcock announced a $3.5 billion partnership with Nscale to deploy up to 100,000 NVIDIA Vera Rubin-based GPUs in Texas by late 2027. The company is also scaling its 'Index' project, a global crowdsourced data-collection initiative that captures human manipulation behaviors to train its robots for zero-shot in unfamiliar environments.
Figure AI has completed the design lock for the Figure 04 humanoid robot. The new generation features significant upgrades to actuators, batteries, and a focus on a high-precision dexterous hand designed to mimic human manipulation. The company has not provided a specific launch date or full technical specifications for the unit.
The company is scaling its 'Index' project, which uses a global app to collect human behavior data from everyday settings. Adcock reported that the project is growing by approximately 20% weekly, with 210,000 unique users contributing data in the most recent week. This data is used to pre-train the Helix model, which has shown improved success rates in zero-shot tasks compared to models trained from scratch.
A $3.5 billion partnership with Nscale has been established to support future training needs. The infrastructure, to be located in Texas, is expected to go live in the second half of 2027. The company currently holds over $2 billion in cash to support these operations, though specific funding sources were not disclosed.
Figure AI reported that its previous model, Helix 2.5, achieved a 56% success rate in unfamiliar residential environments, compared to 9% for models trained from scratch. The company is currently testing a new, larger-scale model internally, which Adcock described as 'Baby AGI,' with further progress expected in the coming weeks.
Source details: finance.biggo.com
Why it matters
The announcement signals a strategic pivot in the humanoid robotics industry from basic mobility to the application of 'scaling laws'—the theory that increasing data and will yield predictable, exponential gains in robot intelligence. By committing billions to compute and crowdsourced data, Figure AI is attempting to build a 'data moat' that few competitors can replicate. The success of this approach is critical for the industry, as it tests whether humanoid robots can move beyond controlled lab environments into complex, real-world settings without requiring task-specific fine-tuning. If validated, this model could fundamentally change how robots are deployed in industrial and domestic sectors.
The shift toward 'robot scaling laws' suggests that the industry is moving toward a model where intelligence is a function of capital expenditure. By treating robotics as a -heavy problem similar to large language models, Figure AI is attempting to solve the 'long-tail' problem of real-world variability.
The reliance on crowdsourced data through the Index project represents a departure from traditional lab-based training. This approach allows the company to capture a wider variety of human behaviors, which is essential for robots intended to operate in diverse, unstructured environments like homes and warehouses.
The financial scale of these commitments—spending millions weekly on data and billions on —creates a high barrier to entry. This suggests that the future of general-purpose robotics may be dominated by a small number of well-capitalized firms capable of sustaining these massive infrastructure costs.
The transition from 'can it move' to 'can it learn' marks a maturation of the sector. The ability to perform tasks in new environments without retraining is the 'holy grail' of robotics, and Figure AI's focus on this capability is a direct attempt to prove the commercial viability of humanoid labor.
Interactive Mechanism: How It Actually Works
Explore the underlying technology behind this development interactively.
Which component of an AI application is the machine-learning model itself?
What to watch next
The primary milestones to monitor are the performance of the next-generation Helix model and the successful activation of the Texas cluster in 2027. Investors and industry observers should watch for whether the company can maintain its current trajectory of scaling data and compute to achieve higher task success rates in unseen environments. Additionally, the technical performance of the Figure 04's new dexterous hand—which aims for near-millimeter precision—will be a key indicator of whether the hardware can keep pace with the model's intelligence improvements.
Watch for the release of performance data for the next-generation Helix model. The company has promised new progress within weeks, which will serve as a test of whether their scaling approach is yielding the expected improvements in intelligence.
Monitor the development of the Texas facility. Any delays in power-supply permitting or hardware procurement could impact the company's ability to train larger models on the timeline they have projected for 2027.
Observe the real-world deployment of Figure 04. While the company has successfully tested previous models in logistics settings, the ability of the new hardware to handle the complexities of the dexterous hand and the demands of varied industrial environments remains to be proven at scale.
Watch for further disclosures regarding the company's financial runway. While Adcock stated they have $2 billion in cash, the massive ongoing costs for data and will require continued capital efficiency or further fundraising to maintain the current pace of development.