NextFin News - Brett Adcock's AI startup Hark has launched Hark Pro, a "proactive" personal AI assistant available on the web, iOS and iPad, betting that the software layer — memory, anticipation, and the ability to act across apps — will be ready well before the hardware device meant to replace the phone and computer, which is not due until 2027. The sequencing is the story: Hark is shipping the agent first, giving away paid subscriptions to its first 100,000 sign-ups, and treating the coming months as a public stress test of the intelligence layer that its yet-unrevealed device will eventually live inside.
The launch arrives five days after OpenAI used its September 29 DevDay to unveil "dots," always-on agents inside ChatGPT and Codex that can execute tasks and answer phone calls, and a new $500-per-month Pro 500 plan offering 25 times the usage of a Plus account. The race for the proactive-assistant category is no longer theoretical — and Hark is entering it as a well-funded private company rather than a public lab with an installed user base.
The Launch: An Agent Before a Device
Adcock announced the launch on October 4, 2026:
"Hark launches this week. The first 100,000 sign-ups get a paid plan for free. I don't do anything without Hark anymore. The team has obsessed over every detail, and it shows."The company, which emerged from stealth in March 2026 after Adcock put in $100 million of his own money in late 2025, has since raised more than $700 million in a Series A that valued it at $6 billion. The round was led by Parkway Venture Capital and included Nvidia, AMD Ventures, Intel Capital, Qualcomm Ventures, Salesforce Ventures, Brookfield, ARK Invest, Greycroft, Prime Movers Lab, Align Ventures and Tamarack Global.
Hark Pro is positioned as the opposite of a chatbot. It carries what the company describes as persistent "Memory and Soul," supports Projects, widgets and scheduled tasks, and — critically — includes computer use: the agent can operate a virtual machine, look at a screen, move a cursor and type, rather than depending on APIs. That matters because, as Adcock put it on the My First Million podcast in August 2026, only about one in 1,000 websites exposes an API, so an assistant that waits for integrations can never be general-purpose. The paid tiers — Hark Pro, Pro², Pro³ and Pro⁴ — carry usage multipliers of 5× and 20×, with one tier offering twice the allowance of Pro³. Hark has not published dollar pricing.
The model stack, according to code references identified by third-party observers, points to Opus 5.5 for general tasks with Gemini 3.1 Flash Image, Qwen3.6-27B and GPT-6 Astra routed for specific workloads — a multi-model routing strategy that mirrors a broader 2026 industry shift away from single-provider dependence. Hark has not officially confirmed the stack.
Why Software First: The Moat Is Not the Device
The decision to launch an app roughly a year ahead of hardware is a deliberate inversion of the playbook that produced the two most conspicuous failures in the category. Humane's AI Pin launched at $699 plus a $24 monthly subscription, raised more than $230 million, and was acquired by HP for $116 million in 2025 — roughly half the capital raised — after devices were discontinued and cloud service terminated on February 28, 2025. The Rabbit R1, at $199 with no subscription, is still shipping and receiving software updates, but never approached mass adoption either. Both shipped purpose-built devices before the underlying agent was reliable, and both discovered that a new form factor cannot compensate for an intelligence layer that cannot execute. Hark is running that sequence backwards — prove the agent in the wild, on devices people already own, and only then give it a body.
There is a data-flywheel logic underneath. A proactive assistant that remembers a user's life, anticipates tasks and acts across apps generates a proprietary stream of interaction data — what users actually ask for, where the agent fails, which workflows recur. That feedback is exactly what a reinforcement-learning computer-use agent needs to improve, and it is unavailable to a company that keeps its product sealed inside unreleased hardware. By the time Hark's device arrives in 2027, the company wants the intelligence to have already logged millions of real interactions. The hardware then becomes a distribution event for a product that already works, rather than a research prototype with a launch party.
The 100,000 free paid plans are not generosity; they are a data-acquisition and lock-in strategy. A user who has trained a personal agent on months of their own memory, projects and scheduled tasks faces real switching costs — the kind that turn a free trial into a durable subscription base. Adcock's own claim that he now routes his Slack, to-do list and morning coffee order through Hark is the prototype for the retention argument: the assistant is sticky when it lives inside the user's actual workflow, not when it sits in a separate app waiting to be prompted.
The Hard Part: Computer Use at Scale
Hark's technical claim rests on a reinforcement-learning post-training loop in which the model practices computer tasks and improves against reward signals — a process Adcock says may be unique.
"We have a reinforcement learning process that we think is maybe nobody else in the world has done," he said.The company's first hardware lead previously designed several generations of iPhone, MacBook and MacBook Pro at Apple, and Hark has a fabrication facility and design studio in place.
Compute is not the constraint it would have been two years ago. In March 2026 Hark announced a deal to bring a large cluster of thousands of NVIDIA B200 GPUs online in April to support multimodal pre-training and post-training. Nvidia's founder and CEO Jensen Huang endorsed the effort:
"The new era of personal AI will be defined by intelligent agents that understand context, reason across modalities, and act on our behalf. Bringing that vision to life requires enormous compute to build powerful multimodal foundation models, and we're excited to support Hark's work with NVIDIA accelerated computing."With both Nvidia and AMD on the cap table, Hark also sidesteps the GPU supply allocation that has strangled other AI-hardware startups in 2026.
But the benchmark is moving. OpenAI's "dot," an always-on personal agent inside ChatGPT and Codex, and Anthropic's work on agentic systems mean Hark is not alone in chasing the proactive-assistant category. OpenAI's DevDay on September 29, 2026, also introduced GPT-6.1 Sol, positioned close to GPT-6 Astra at roughly a fifth of the price, and an Ultrafast processing tier delivering up to eight times the token generation in Codex. The difference Hark is betting on is vertical integration: it is building foundation models, software, interfaces and hardware as one stack, the way Apple once built silicon, software and devices together.
"We believe the next computing platform will be personal AI—intelligence that understands you and works alongside you every day," Adcock said in the company's March announcement. "But that future only becomes possible when the entire stack is built together."
The former Apple designer Hark recruited to lead its interface work, Chowdhury, framed the ambition in stark terms:
"It just feels like there's an opportunity for better, and I've not felt like that since the iPhone came up."He also drew a line against the wearables approach that defined the previous hardware cycle:
"I don't think it's appropriate to put a layer between humanity and the interfaces we use in the world. I have similar discomfort with pins, or that kind of stuff that is going around with cameras."
Cyclical Hype, Structural Shift
Is this another AI-hardware bubble, or a genuine regime change? The answer splits by layer. The hardware bet is cyclical and mean-reverting: venture capital is flooding AI-device startups, valuations are ahead of shipped products, and the category's recent history — Humane, Rabbit — shows that capital alone cannot manufacture product-market fit. Hark's $6 billion valuation, reached before shipping a single device, sits squarely in that speculative leg. If the agent disappoints, the hardware will follow it down.
The software shift, however, is structural and will not revert. The transition from chatbots that answer questions to agents that execute tasks — reading screens, clicking buttons, remembering context across sessions — is a change in what the product fundamentally is, not its price or its marketing. Once users hand an agent persistent access to their accounts and workflows, the relationship changes from transactional ("ask, get an answer") to relational ("delegate, get it done"). That does not unwind when sentiment cools. The structural question is not whether agentic AI arrives, but which company owns the memory layer that sits between the user and every other application.
This is why the sequencing matters more than the hardware. If Hark nails the agent first, the 2027 device becomes the iPhone moment for a product that already has a constituency. If it does not, the device becomes an expensive shell around an undifferentiated chatbot — the Humane outcome dressed in better industrial design.
The cyclical leg has a visible tell: the talent war. Adcock estimates only 20 to 30 people in California can build frontier models, and that Meta is paying packages around $36 million to win them — "guaranteed cash" beating startup equity. That is a cyclical constraint in the sense that compensation premiums normalize once the labor market clears, but it is also a real execution risk: if the people who can actually improve the models keep leaving for incumbents, Hark's technical edge erodes regardless of how elegant the hardware becomes.
The Counter-Thesis: Humane 2.0
The strongest case against Hark is that it is repeating the same mistake with better press. The counter-argument runs: Hark has raised $700 million at a $6 billion valuation, recruited engineers from Google, Meta, Apple and Tesla, and secured GPU supply — and still has no shipped hardware, no published pricing, and no independently verified proof that its computer-use agent works outside Adcock's own workflow. Humane and Rabbit were also lavishly funded, also staffed by alumni of the industry's best companies, and also certain they had identified the next platform. Funding and pedigree predicted nothing then; they may predict nothing now.
There is also the distribution problem that Humane could not solve. A new category requires users to change behavior, not just buy a device. OpenAI does not face that problem: ChatGPT already has 1.2 billion weekly active users, and a dot agent lands inside an app they already use. Hark must first persuade users to download a new app, train it on their life, and trust it with accounts and tasks. That is a steeper adoption curve than any of its well-funded rivals face.
The answer to the counter-thesis is empirical, not rhetorical, and the test is near. Hark's own framework sets the bar: the consumer app landing in October 2026 shows whether the computer-use agent works at scale outside the founder's personal workflow; the end of 2026 tests whether Figure's robots can perform unsupervised multi-day work; and 2027 is the hardware deadline. The falsifying signal is specific and observable: if fewer than roughly 30 percent of the 100,000 free users are still actively using Hark Pro three months after sign-up, or if the company cannot demonstrate computer-use reliability on routine tasks — booking travel, reconciling expenses, managing calendars — without human intervention, the "proactive agent" claim collapses into marketing and the hardware launch should be read as a pivot, not a product line.
What to Watch
Short term (weeks): whether the waitlist converts, whether the app functions reliably across iOS and web, and whether the computer-use agent can complete multi-step tasks end to end. Early user reports will matter more than the launch announcement.
Medium term (through 2027): retention of the first 100,000 users, the announcement of dollar pricing, and whether Hark discloses a concrete hardware form factor, target price and launch market — all of which remain undisclosed. Figure AI's parallel progress — thousands of Figure 3 units deployed, a BMW partnership, package sorting at 2.9 seconds apiece for 200 continuous hours — is a proxy for whether Adcock's teams can ship real-world AI at all.
Long term (structural): who owns the personal memory layer. If Hark's agent becomes the default interface through which users access other services, the company captures a toll on the entire digital economy. If it remains a feature inside existing apps, it competes on price against OpenAI, Anthropic, Google and a dozen aggregators — a race to the bottom on margin.
The base case is that Hark ships a competent, genuinely proactive assistant that finds a loyal but narrow power-user base by 2027, giving the hardware launch a real audience but not a mass market. The upside case is that the computer-use agent proves reliably general, the memory layer locks in tens of millions of users, and the 2027 device becomes the first AI-native hardware product that people keep. The downside case is the Humane replay: a beautiful device, an agent that cannot be trusted to act alone, and a $6 billion valuation written down.
Adcock's own summary of the bet is the clearest risk disclosure available:
"It's either going to be a very big thing or very bad."Hark Pro is the first public data point. The hardware, whenever it arrives, will be judged not on its design but on whether the agent inside it has already earned the right to act on your behalf.
Explore more exclusive insights at nextfin.ai.
