Hark Has 70 Employees and a $6 Billion Valuation. Figure AI Had Working Hardware and Took Three Years to Get There.

Six Months From Stealth to $6B

Brett Adcock founded Hark in late 2025 with $100 million of his own money. The company surfaced publicly on March 24, 2026, and by May 21 had closed a $700 million Series A at a $6 billion post-money valuation, as reported by TechCrunch. Founder-funded stealth to a $6 billion price tag took roughly six months. Public launch to a closed mega-round took less than two.

That pace is the story. A normal Series A signals early evidence of product-market fit (some usage data, maybe a few paying customers). Hark had none of that: no revenue, no released product, no hardware form factor, no disclosed customer count. TechCrunch reported the round was oversubscribed. The investor syndicate — Parkway Venture Capital leading, with Nvidia, Intel Capital, AMD Ventures, Qualcomm Ventures, Salesforce Ventures, ARK Invest, Brookfield, Greycroft, Align Ventures, Prime Movers Lab, and Tamarack Global participating — reads closer to a strategic consortium than a typical first round.

Stack Hark's velocity against the best-funded robotics startups of the past two years and the gap becomes clearer.

Funding velocity: Hark vs. peers

Figure AI, Adcock's own humanoid-robot startup, took three years to reach a comparable valuation tier — and it had working hardware, BMW deployment pilots, and an estimated $60 million in annual recurring revenue by the time it hit $2.6 billion in 2024. Hark reached unicorn status with 70 employees, a data center full of Nvidia B200 GPUs, and a public timeline that amounts to "models this summer, hardware later."

The closest analog is Safe Superintelligence, which reportedly reached a $32 billion valuation with no product and no revenue — a bet placed almost entirely on founder credibility and the scarcity of elite AI research talent. Hark occupies a similar category, except it adds a hardware thesis on top of the AI-lab pitch.

The timeline tells you what the investors are underwriting: not a product, not revenue, not even a public demo. They are underwriting Adcock's track record (Figure and Archer both went from concept to commercial pilots inside three years) and the bet that a vertically integrated personal-AI-device company needs to be funded at scale before the first model ships. Whether that bet is rational or recursive depends on what Hark actually releases this summer.

What Hark Is Actually Building

Hark is not building a chatbot with better hardware. It is building a vertically integrated personal-AI platform (foundation models, software, and native hardware designed as a single system) that aims to function as what Adcock calls a "universal interface between humans and machines."

The distinction matters because it separates Hark from three categories that already exist: humanoid robotics, smart speakers, and AI wearables.

Not a humanoid robot. Adcock's other company, Figure AI, builds humanoid robots for labor automation. Hark is a separate entity with a separate thesis. Where Figure's machines replace physical work, Hark's devices sit between the user and the digital world — listening, seeing, remembering, and acting on behalf of one person. The hardware is purpose-built for a personal assistant, not a general-purpose body.

Not a smart speaker. Amazon's Echo and Google's Nest run cloud-based models on general-purpose silicon. Hark is building its own multimodal foundation models (integrating speech, text, vision, and contextual awareness) alongside custom hardware designed specifically for those models. Adcock has said today's AI feels "basic" because it lacks persistent memory and is constrained by decades-old consumer devices. Hark's bet is that co-designing the model and the chip eliminates that constraint.

Not an AI wearable. Meta's Ray-Ban glasses and similar devices bolt AI onto existing form factors. Hark's approach, led by design director Abidur Chowdhury — the former Apple industrial designer who worked on the iPhone and Mac, including the recent iPhone Air — is to build bespoke native hardware from the ground up. Chowdhury has said the goal is a system that "understands you, intelligently anticipates your needs, and loves doing tasks that you don't want to do."

The core technical thesis is persistent, proactive personalization. Hark's models are designed to build memory over time (preferences, goals, context) and to act before being prompted. That requires running inference locally or on dedicated infrastructure rather than routing every query through a third-party API. The company has signed a deal to bring a large cluster of NVIDIA B200 GPUs online for multimodal pre-training and post-training, and Jensen Huang said NVIDIA is "excited to support Hark's work with NVIDIA accelerated computing."

The product rollout follows a software-first sequence: Hark plans to release its first agentic, multimodal AI models in summer 2026, giving early users access to the personal intelligence platform before the native hardware ships. That sequencing lets the models train on real usage patterns while the hardware finishes development — a practical advantage that pure-software competitors building on top of existing devices do not have.

The full-stack ownership (models, OS, silicon, industrial design under one roof) is the structural bet. Whether it works depends on whether Hark can recruit the talent to execute across all four layers simultaneously.

The Investor Stack: What the Cap Table Signals

Parkway Venture Capital led Hark's $700 million Series A, but the lead investor is not the story here. The story is that four rival semiconductor companies (NVIDIA, AMD Ventures, Intel Capital, and Qualcomm Ventures) all wrote checks into the same round for a 70-person company with no shipping product. That does not happen by accident.

When competitors pile into the same deal, they are not betting on equity returns alone. They are buying a seat at the design table. Whose silicon ends up inside Hark's AI-native hardware secures a flagship design win for the personal-device category. The cap table is a pre-positioning move for a chip-supply contract that has not been awarded yet.

Reading the strategic bloc

The investor list breaks cleanly into four camps, each signaling a different bet on what Hark becomes.

NVIDIA's participation carries dual weight. Hark already operates a dedicated B200 GPU data center for model training, as the company confirmed in its announcement. NVIDIA is effectively subsidizing its own compute-revenue pipeline: fund the model trainer today, sell the GPUs tomorrow, and potentially power the edge device the day after. AMD Ventures and Intel Capital making concurrent bets means they see the same trajectory and refuse to let NVIDIA lock up another high-profile customer unchallenged.

Qualcomm Ventures points in a different direction. Qualcomm's design wins live in mobile and edge devices (smartphones, headsets, automotive). Their presence suggests Hark's hardware may target power-constrained, always-on form factors rather than a desk-bound hub. You do not bring Qualcomm into a round if you are building a server rack.

Why Salesforce Ventures belongs on this cap table

Salesforce Ventures looks like the odd one out among chip makers and deep-tech funds. But Salesforce invests where AI meets the enterprise user. If Hark's personal AI platform works as described (a persistent, context-aware assistant that operates across services) the enterprise play is obvious. An AI that "remembers who you are and what you say," as Hark's own announcement puts it, and manages digital workflows is a productivity tool, not just a consumer gadget. Salesforce's check signals that Hark's addressable market extends beyond the home.

The cap table's verdict

Hark is not being financed as a pure software play, a research lab, or an AI-infrastructure company. The investor stack treats it as a vertically integrated consumer-device company with enterprise reach — one that controls its own models, its own silicon selection, and its own hardware. The four-way chip-investor presence is the clearest signal: the strategic money expects Hark to ship a physical product, and the fight over which architecture powers that product has already started inside the cap table.

San Jose Hiring Blitz — Which Roles Hit the Payroll First

Hark hasn't posted roles to Zero G Talent's board yet. But Adcock's track record and the company's San Jose headquarters tell you plenty about which disciplines hit the payroll first.

When Adcock built Figure AI, he staffed mechanical engineering and integration roles in San Jose alongside ML inference and graphics engineers — the same split Zero G Talent now tracks at Figure, where eight open roles include Staff AI Inference and Acceleration Engineer at $180,000–$275,000, Sr/Staff Graphics Engineer at $150,000–$275,000, and Software Engineer for Manufacturing Systems at $160,000–$250,000. That pattern — ML-infrastructure and rendering talent at senior bands, with manufacturing software alongside mechanical interns — reveals the build sequence for a physical product: get the model and visual pipeline running, then stand up the line.

Hark's thesis demands a different ratio. A humanoid robot needs actuators, joint controllers, and balance systems before it walks. A personal-AI device (something closer in physical complexity to a laptop than a biped) shifts weight from mechanical engineering toward silicon, on-device inference, and supply-chain ops. The hardware has to be simple enough to manufacture at consumer volumes, and the model has to run without round-trips to a cloud API.

That means the first Hark hires likely cluster in three buckets:

Mechanical engineering doesn't disappear (the device still needs an enclosure, thermal management, and actuators for whatever physical interaction Hark intends). But the headcount tilt runs toward ML-systems and hardware-silicon integration, not toward the mechatronics-heavy orgs at Figure or Tesla Optimus.

San Jose as the base reinforces the read. The Valley corridor gives Hark access to the same inference-optimization and embedded-systems pool that Figure draws from, but Adcock isn't building a lab adjacent to automotive-plant talent in Austin or Detroit. He's staffing for a consumer-electronics launch cadence — the kind where supply-chain engineers earn their keep months before the first unit ships.

Why Edge AI Gets Stronger Every Time Cloud API Costs Rise

The case for putting AI on a person, rather than routing every query through a cloud API, is mostly an economics argument. And that argument gets stronger every time the cloud side gets more expensive.

Right now, most consumer AI products (chatbots, copilot assistants, image tools) run on large language models hosted in data centers. The user sends a prompt to a server, the model runs on rented GPU time, and the result comes back over the network. That architecture works. It also means the product's operating cost scales directly with usage: every query costs the company real money in compute, and the company passes that cost along through subscriptions, rate limits, or ads.

Hark's thesis inverts that model. If the intelligence runs on the device itself (on custom silicon, with a proprietary model optimized for that silicon) the per-query marginal cost drops toward zero after the hardware is sold. The user pays once for the device. The company pays for the chip, the battery, and the engineering to make the model small enough to run locally. After that, usage doesn't compound the cost.

This is the same logic that pushed Apple to build neural engines into iPhones and that drove Qualcomm and MediaTek to add on-device AI accelerators to mobile SoCs. The difference is that Hark is building the full stack (model, silicon, device, and the integration layer between them) rather than bolting AI onto an existing product category.

The timing matters. When a frontier model's token costs stay flat or rise, every startup building on top of that API watches its margins compress. A company that owns its own model and its own inference hardware doesn't face that squeeze. It faces a different problem (the enormous upfront cost of training and silicon design) but that's a capital problem, not a recurring operating one. And capital is exactly what Hark just raised $700 million to spend.

The workforce implications follow directly. If Hark is serious about on-device inference, it needs engineers who can compress models, write firmware for custom accelerators, and tune power budgets — not just ML researchers who train large models in the cloud. That's a different hiring profile than a pure-software AI company, and it overlaps more with semiconductor and embedded-systems talent than with the typical OpenAI or Anthropic job description. Watch the San Jose postings for roles in model quantization, silicon bring-up, and power optimization. Those will tell you whether the edge thesis is real or just a pitch deck slide.

How Hark's Workforce Blueprint Compares to Figure, Tesla Optimus, and Anduril

Hark sits at a different organizational intersection than any of the three companies most often mentioned alongside it. Figure AI, Tesla Optimus, and Anduril each have a clear dominant identity — humanoid logistics robot, automotive-scale manufacturing line, autonomous defense systems — and their hiring reflects it. Hark's structure, by contrast, has to fuse four disciplines that rarely live under one roof: consumer hardware design, proprietary ML model development, silicon-aware inference engineering, and supply-chain ops for a physical device sold to individuals.

Figure AI, the most direct comparison among humanoid plays, is running a focused two-front build. Zero G Talent's board lists 8 Figure roles added in the past week — a Staff AI Inference and Acceleration Engineer, a Sr/Staff Graphics Engineer, a Manufacturing Systems Software Engineer, and a cluster of mechanical and validation internships, all in San Jose. Half the open roles are on the inference and silicon-aware software side, half on mechanical integration and production. Figure is building one robot (the Figure 02) for one use case (warehouse and manufacturing logistics), so its hiring can stay tight around a single hardware-software feedback loop. The internships suggest a deliberate pipeline strategy — Figure is seeding its production floor with early-career mechanical and validation engineers who can grow into a scaled manufacturing org.

Tesla Optimus is a different beast entirely. Reports from March 30 put Tesla's Optimus job listings at 100-plus production-specific roles — manufacturing, hardware, and robotics functions tied to Fremont repurposing and a Giga Texas line Musk has said targets 10 million units per year. The scale dwarfs anything Hark or Figure is doing. Tesla is hiring to staff a full automotive-grade production line, which means process engineers, quality technicians, and assembly roles far outnumber the research positions. The talent profile skews heavily toward manufacturing execution — people who know how to run high-volume lines, not people designing a first-generation personal device from scratch.

Anduril occupies the third point of the triangle. The defense-tech company, valued at $14 billion after a $1.5 billion raise in June 2024, posted 450 engineering positions across autonomous systems, software, and manufacturing as of late 2024, with roughly 40% of its total workforce in engineering roles. Zero G Talent's board shows 242 Anduril roles added in the past 7 days alone — Sustainment Leads, Supply Chain Directors, Camera Test Engineers, Space Modeling and Simulation Engineers — spread across Costa Mesa, Hudson, Lexington, and El Segundo. Anduril's hiring is mission-driven: every role connects to a defense program (Lattice AI, Fury, counter-drone systems) and requires security-cleared or clearance-adjacent talent. The org is deep in manufacturing ops and field sustainment, not consumer product design.

Hark's workforce has to do something none of these three are optimized for: build a consumer device that runs proprietary AI models on-device, at a price point individuals will pay, with the reliability expectations of a phone and the safety requirements of a robot. Figure's engineers can optimize for a single warehouse environment. Tesla's can lean on a century of automotive production playbook. Anduril's can rely on defense contracts that tolerate higher unit costs and longer development cycles. Hark can't borrow any of those crutches.

That means the roles Adcock staffs first will reveal whether he's betting on vertical integration the way Apple did (owning silicon, model, and device simultaneously) or whether Hark is closer to a hardware-shell-plus-cloud-model play that would make it more akin to a smart-speaker company with legs. If the early roles are silicon engineers, on-device inference specialists, and supply-chain leads who've shipped consumer electronics at volume, Hark is building something structurally unlike anything in the current robotics talent market. If the roles lean heavily on cloud ML researchers and generic mechanical engineers, the "proprietary model on personal hardware" thesis is mostly branding — and the workforce will look a lot like a smaller, less focused version of Figure.

What the Hark Bet Means for Robotics Talent in 2026

The robotics labor market in 2026 is a paradox: companies can't hire fast enough, yet the skills they need are shifting under the feet of the engineers already in the field. Global direct robotics employment hit roughly 2.1 million in 2026, up from about 1.6 million in 2023, according to Robotomated's workforce analysis. The fastest-growing category isn't mechanical design or even general software — it's AI/ML for robotics, up 35% year-over-year, with base salaries ranging from $140,000 to $220,000. For engineers deciding where to place their next two years, the Hark bet sharpens the choice into three distinct paths.

The humanoid track (Figure AI, Tesla Optimus, Agility) is a robot learning engineer's market. The specialization is VLA models, imitation learning, sim-to-real transfer. The ceiling is high: senior robot learning engineers at top startups report total compensation of $280,000 including equity, according to the Robotics Center of Silicon Valley's 2026 salary guide. The risk is concentration — these programs are capital-intensive, milestone-dependent, and still proving whether a general-purpose humanoid is a product or a science project.

The defense-tech track (Anduril, the primes' robotics divisions) is hiring at volume right now. Anduril added 242 roles in the past week on Zero G Talent's board, spanning supply chain directors at six figures to micro-electronics technicians at $27–$37 per hour. The work is hardware-heavy, security-cleared, and tied to government contract timelines. Compensation is strong but more defense-industrial in structure: solid base, modest equity, location-locked to places like Costa Mesa, El Segundo, and Lexington.

The personal-AI hardware track Hark is opening sits between these. It needs people who can do embedded ML deployment, real-time perception on constrained hardware, and consumer-grade systems integration — a stack closer to Apple's silicon play than to Figure's robot learning pipeline. The AI-fluent salary premium is stark: engineers combining robotics with computer vision and deep learning skills earn $200,000–$210,000, roughly 40–55% above the $133,000–$142,000 base for traditional robotics roles, according to 2026 salary data aggregated from Glassdoor and the Qubit Labs AI Talent Report.

The practical takeaway is about skill stacking. An engineer with ROS 2 fluency and PyTorch-based perception experience can work at any of the three. But the person who also knows how to quantize a model for edge inference, optimize latency on embedded silicon, and integrate with a consumer supply chain is the one Hark will compete with Apple and Meta to hire — and that profile is scarce enough to command the top of any of these bands. If you're in robotics right now and haven't added an AI skill layer, the window to do so while demand dramatically exceeds supply is open, but it won't stay open for long.

Working in robotics? Zero G Talent tracks the openings: browse robotics jobs, openings at Anduril Industries, OpenAI and Figure AI, and the people building the field.