Figure AI's Brett Adcock Bets Gigawatt NVIDIA Deal on Hark Handoff

Hark, the personal AI startup founded by Figure AI's Brett Adcock, has signed a multi-year strategic partnership with NVIDIA that includes gigawatt-scale compute on the upcoming Vera Rubin platform. The deal follows the template NVIDIA used with Mira Murati's Thinking Machines Lab earlier this year, which committed to deploying at least one gigawatt of next-generation Vera Rubin systems for frontier model training. Hark says the capacity will underwrite both training of its own foundation models and inference for a consumer platform launching within weeks.

The announcement lands ahead of what Hark describes as its first public product release. According to the company's announcement, NVIDIA infrastructure has already powered development of Hark Handoff, the computer-use model the startup unveiled earlier this summer, and will handle distributed inference at launch plus future foundation model training.

Who is Hark, and why does NVIDIA care

Hark is Adcock's second act. The Figure AI founder is self-funding the company with $100 million of his own money, and Figure itself was valued at $39 billion in late 2025. Hark debuted with a promise to build a vertically integrated system that can see, listen, speak, and "touch and influence the world," developing everything from foundation models to bespoke native hardware devices in-house. The stated goal is an intelligence that offloads mental workload by thinking like you, and sometimes ahead of you.

The design and engineering bench is unusually deep for a stealth company. Hark's design lead is Abidur Chowdhury, formerly an industrial designer at Apple credited with the iPhone Air. The engineering team includes researchers with backgrounds in large-model training, tokenization, and safety, plus hardware engineers from Apple, Tesla, and Meta. NVIDIA, AMD, Intel Capital, and Qualcomm Ventures all participated in an earlier round, which is unusual because chip companies do not typically back consumer AI product bets. Their involvement suggests they see Handoff as a workload that will consume substantial inference compute at scale, which aligns with the architecture choice of spinning up a dedicated virtual machine per user session.

The Handoff proof point

Handoff is the technical calling card that helped Hark land this deal. The company says the model recorded the top-ever score on Online-Mind2Web, a third-party human-evaluated leaderboard for web agents, posting a 97.7 against 92.8 for OpenAI's GPT 5.4, 84.1 for Anthropic's Claude Opus 4.8, and 69 for Google's Gemini 2.5 Pro. Hark says it can serve the model at less than one-tenth the token price of competing frontier models.

The architecture explains why gigawatt-scale inference matters. For every task, Handoff spins up a dedicated virtual computer with its own browser, file system, and terminal. The agent connects to a user's existing accounts and logs in on their behalf, with access to saved addresses, payment methods, and preferences. It then reads the visual structure of each page and makes click-by-click, keystroke-by-keystroke decisions in real time. That is orders of magnitude more compute per user than a typical chatbot session.

The benchmark claims deserve an asterisk. The current leaders, OpenAI's GPT-5.6 and Anthropic's Opus 5, are absent from Hark's comparison, as are strong open-source contenders like DeepSeek V4, Kimi K3, and Qwen3.8-Max. That means Hark's "top-ever" claim cannot currently be checked against the strongest available systems. Two of the three benchmarks in Hark's results table were also run inside Hark's own harness with pass rates computed by Hark's internal LLM judge.

What Hark is actually running on NVIDIA's stack

Beyond raw GPUs, the partnership pulls in a substantial software stack. Hark says it is using:

• Megatron for model training
• Dynamo for inference serving
• Nemotron permissively licensed pretraining data
• NVSentinel for GPU cluster resiliency

The target hardware is NVIDIA's next-generation platform. Vera Rubin NVL72 production is ramping at partners including CoreWeave, Google Cloud, Microsoft Azure, Oracle Cloud Infrastructure, and Nebius. The rack design features energy-efficient 45 degree Celsius liquid cooling, a new liquid-cooled busbar for higher performance, and 20x more energy storage to keep power steady.

A pattern of gigawatt deals

Hark joins a growing list of AI labs whose forward roadmaps are effectively bolted to NVIDIA silicon. Alongside Thinking Machines Lab, xAI announced it will deploy Vera CPUs to accelerate the next generation of agentic AI applications powering Grok, and formalized that its entire forward compute roadmap is being built on NVIDIA's next-generation architecture with no dual-sourcing strategy in sight. These deals typically bundle deep engineering collaboration with capacity guarantees, and in Thinking Machines' case NVIDIA also made a significant equity investment. Hark's announcement did not disclose whether NVIDIA is taking a stake, though its earlier participation in the startup's cap table is on the record.

What this means for anyone building agents

For teams building browser-driving or computer-use agents, Hark's push signals a few concrete shifts:

1. Per-session VMs are becoming a serious architecture. If Handoff's model of a full virtual computer per user request wins, inference costs and infra design assumptions for agent products change substantially.
2. Personalized memory plus multimodal is the next wedge. Adcock has said the underlying LLM will focus on speech and memory, with improved capabilities to anticipate user needs. Expect competition to shift from raw reasoning benchmarks to personal-state retention.
3. Consumer AI hardware is heating up again. Adcock has launched Hark to build a family of AI devices, jumping into an emerging segment where OpenAI, Apple, Meta, and Google are all planning to launch hardware specifically for AI usage.

The immediate risk for Hark is execution. The company has no shipping consumer product, its benchmark story omits the newest frontier models, and Vera Rubin capacity is being courted by every well-funded lab on the planet. Adcock now has to convert compute commitments and hired talent into a platform that ships before end of summer, then a hardware device that gives people a reason to hand off their computer to an agent that keeps a memory of them. Gigawatts are the easy part.