Hark Bets on NVIDIA Vera Rubin for Personalized AI at Gigawatt Scale

Hark has signed a multiyear NVIDIA partnership that promises gigawatt-scale compute, even though its first public AI platform has not launched. The agreement centers on Vera Rubin, NVIDIA’s next-generation system for training and serving reasoning models and AI agents. A brief rsshub 36kr item surfaced the news for Chinese readers, but the underlying announcement came directly from Hark.

That order matters. Hark is committing to infrastructure designed for the largest AI workloads before outside users can evaluate its product. The company says the capacity will support personalized agents that combine speech, vision, persistent memory, and computer control.

Hark is not alone in making that bet. Thinking Machines Lab, OpenAI, and other well-funded developers have also tied future products to large Vera Rubin deployments. Hark must now prove that infrastructure at this scale creates a better personal agent, rather than a larger operating bill.

The Partnership Commits Hark to More Than GPUs

Hark is buying into NVIDIA’s full computing system, not simply reserving a collection of accelerators.

The companies announced their agreement on August 27, 2026. According to the partnership announcement, it covers technical collaboration and gigawatt-scale capacity based on Vera Rubin.

Neither company disclosed a deployment schedule, exact capacity, data-center location, or financial commitment. “Gigawatt-scale” describes the intended infrastructure class, not a verified amount of live computing capacity. Hark also did not specify whether it will own systems, lease capacity, or use several infrastructure providers.

That missing detail separates the announcement from an operating milestone. A power target does not reveal how many systems Hark has secured, when those systems become available, or how efficiently it can use them.

The agreement nevertheless reaches across Hark’s planned AI stack. Hark says NVIDIA infrastructure has already supported Hark Handoff, a computer-use model designed to operate software interfaces. The same infrastructure will reportedly serve Hark’s initial platform and help train its future foundation models.

Computer-use models interpret a screen and take actions through software interfaces. They can open applications, enter information, and complete multistep workflows, subject to permissions and reliability limits.

Hark also named four NVIDIA technologies involved in the collaboration. Megatron supports large-model training, while Dynamo coordinates distributed inference, meaning the process that runs trained models for users. Nemotron provides permissively licensed training data, and NVSentinel monitors GPU clusters for failures.

These components reveal the broader commitment. Hark is aligning its training, inference, data, networking, and cluster operations with NVIDIA’s software environment. That approach can reduce integration work, but it also makes NVIDIA central to several technical layers.

The platform itself is built around rack-scale computing. NVIDIA says a Vera Rubin NVL72 system combines 72 Rubin GPUs and 36 Vera CPUs. It also includes ConnectX-9 networking, BlueField-4 data-processing units, and sixth-generation NVLink.

NVLink creates a high-bandwidth connection among processors, allowing one rack to handle large models as a coordinated system. That architecture targets workloads that would otherwise spend considerable time moving information between separate accelerators.

Hark’s announcement links this infrastructure to an upcoming public platform. The company said that product would arrive before summer ends, creating an unusually short gap between its infrastructure story and its first major market test.

The first release will therefore carry more weight than a normal product preview. It must demonstrate why Hark needs an infrastructure commitment commonly associated with frontier model laboratories and hyperscale services.

Why Personalized Agents Consume So Much Compute

A personal agent becomes expensive when it must remember context, perceive several media types, reason through tasks, and respond without noticeable delays.

A conventional chatbot processes a prompt and generates text. Hark describes something more persistent. Its planned system is supposed to understand speech and vision, remember the user, control computers, and operate through existing devices or dedicated hardware.

Each layer increases the workload. Speech must be transcribed and interpreted. Images or video require visual processing. Memory requires relevant records to be retrieved, ranked, and inserted into the model’s context.

An agent then needs to plan actions, call tools, inspect results, and revise its approach. One request can trigger several model passes instead of one response. Serving millions of such sessions would demand substantial inference capacity, even without continuous model training.

Inference is particularly important for Hark because personalization happens during use. A system cannot feel personal if every interaction waits behind a long processing queue. It also cannot act reliably if context disappears whenever a session ends.

NVIDIA positions Vera Rubin around these constraints. Its platform combines processors, networking, memory systems, and software to reduce the cost and latency of reasoning workloads. These are company claims, and Hark has not published independent benchmarks for its implementation.

The Vera CPU handles orchestration and data movement around accelerated workloads. Rubin GPUs perform the dense computation required by models, while BlueField processors and high-speed networking manage traffic across larger installations.

This architecture addresses a real bottleneck. Agentic systems do more than generate tokens. They move context among storage, processors, tools, and external applications throughout a multistep task.

Hark’s proposed hardware adds another challenge. A personal device cannot carry data-center hardware, so the product must divide work between local components and remote infrastructure. The company has not explained that boundary.

Local processing can reduce latency and limit the information sent to servers. Cloud processing can run larger models and provide more frequent improvements. A practical system will probably use both, but the balance affects privacy, performance, and cost.

Hark says its AI will work through existing devices and bespoke hardware. It has not revealed the device design, operating system, sensors, or local computing capabilities. It also has not explained whether its first launch includes hardware.

That uncertainty matters because Hark’s product vision depends on access to the user’s environment. Speech, vision, computer control, and memory are most useful when they connect across applications and devices.

Consider a routine work scenario. A user could ask an agent to review meeting notes, compare them with project documents, prepare a status update, and schedule follow-up tasks. The agent would need permissioned access, reliable retrieval, tool execution, and a record of previous decisions.

People already build narrower versions of that workflow with a personal knowledge base. Hark’s ambition is broader because it wants the agent to coordinate perception, memory, and actions through one interface.

The infrastructure commitment suggests Hark expects those interactions to generate intensive, recurring workloads. The unanswered question is whether customers will value that integration enough to support its cost.

Hark Faces a Personal AI Race, Not an Infrastructure Race

Hark’s primary competition comes from established assistants that already possess users, distribution, and growing memory features.

OpenAI, Google, Anthropic, Meta, Microsoft, and Apple are all developing systems that can retain context, use tools, or operate software. Their approaches differ, but they compete for the same valuable position between users and their digital lives.

Hark enters without that installed base. It must persuade people to trust another service with conversations, files, preferences, and possibly device access. Infrastructure can improve a product, but it cannot create distribution or trust by itself.

The company’s answer appears to be vertical integration. Hark says it is developing foundation models, software interfaces, and native hardware together. That strategy can produce tighter coordination than a product assembled from unrelated components.

It can also increase execution risk. Hark must train models, build an agent platform, design consumer hardware, secure personal data, and deliver reliable computer control. Each task would support a substantial company on its own.

The identity of Hark’s founder raises expectations. Brett Adcock previously founded Archer Aviation and Figure AI. Hark presents his new company as an AI laboratory focused on personal intelligence rather than industrial robotics.

The company has disclosed limited information about its staff, funding structure, or product development. Its public website says it is building “personal intelligence,” but provides few operational details beyond that vision.

That secrecy gives Hark room to develop before launch. It also makes independent comparison difficult. There are no public evaluations showing that Hark Handoff performs better than competing computer-use systems.

The Vera Rubin agreement creates another comparison. In March, NVIDIA and Thinking Machines Lab announced a one-gigawatt partnership for customizable AI and frontier model development.

Thinking Machines disclosed a minimum capacity target and an intended deployment period. NVIDIA also invested in that company. Hark’s announcement describes gigawatt-scale capacity but does not identify a minimum commitment or an NVIDIA investment.

The similar language shows how NVIDIA is expanding its role beyond hardware supply. It is forming long-term technical relationships with laboratories before their products reach broad adoption.

For NVIDIA, those agreements can lock future training and inference demand into its architecture. For the startups, they offer access to systems, software, and engineering support that would be difficult to assemble independently.

The trade is not automatically favorable for every startup. Deep integration can accelerate development, but it can also constrain future choices. Moving workloads to another processor platform becomes harder after models and operations depend on vendor-specific tooling.

Hark therefore competes on two fronts. It must match the assistant capabilities offered by much larger technology companies. It must also use its infrastructure more effectively than startups pursuing similar personalized or customizable AI products.

Raw model performance will not settle that contest. Personal agents need consistent memory, understandable permissions, low latency, and dependable actions. A smaller model with better product design can outperform a larger model in a specific workflow.

Hark’s advantage, if it emerges, will come from the coordination of those pieces. Its partnership provides a foundation for that approach, but the product must supply the evidence.

The rsshub 36kr Headline Leaves Key Risks Unanswered

The central risk is not whether Hark can acquire compute, but whether it can make a memory-rich agent safe, reliable, and economically sustainable.

The original rsshub 36kr summary accurately captures the companies, the multiyear agreement, and the Vera Rubin plan. It cannot answer questions that Hark omitted from the announcement.

First, persistent memory changes the privacy stakes. A useful personal agent may retain relationships, routines, work history, locations, communications, and preferences. That collection can reveal more than any single file or conversation.

Users need to know what the system remembers, where that information resides, and how they can delete it. They also need clear controls for separating personal, professional, and shared contexts.

Hark has not publicly detailed those controls. Its announcement emphasizes an AI that “knows you,” but does not describe consent, retention periods, encryption, data portability, or model-training policies.

Those omissions are not proof of weak safeguards. They are unresolved product questions. Hark’s public launch should answer them before users grant the agent broad access.

Second, agents introduce security risks beyond ordinary chatbots. The NIST agent inquiry highlights indirect prompt injection, poisoned models, harmful actions, and specification gaming.

Indirect prompt injection occurs when an agent encounters hostile instructions hidden inside external content. A compromised document or webpage can attempt to redirect the agent while it acts for the user.

That threat becomes more serious when an agent can read private information or control applications. A mistaken text response is inconvenient. An unauthorized message, purchase, file transfer, or configuration change can cause lasting harm.

Hark has not published security evaluations for Handoff or its broader platform. It has not stated which actions require confirmation, how permissions expire, or whether users can audit completed steps.

Third, computer control remains difficult to evaluate. Interfaces change, visual elements move, and websites present ambiguous states. Agents can appear competent during demonstrations while failing on long or unfamiliar tasks.

A credible release needs more than selected examples. Hark should disclose task completion rates, intervention frequency, latency, failure categories, and the conditions used for testing.

Independent evaluations will matter because Hark calls Handoff an advanced computer-use model. Until comparable results appear, that description remains the company’s characterization.

Fourth, gigawatt-scale infrastructure introduces financial and operational exposure. Data centers at this level require power, cooling, networking, facilities, and long-term supply agreements. None of those resources becomes valuable merely because a company reserves them.

Hark must keep expensive systems occupied with useful training or paid inference. Underused capacity would weaken the economics, while heavy usage by unprofitable users would create a different problem.

The company has not disclosed its business model. A consumer subscription, hardware sale, enterprise service, or blended model would produce different requirements for margins and customer support.

Dedicated hardware raises further questions. Consumer devices involve manufacturing, inventory, repairs, distribution, and replacement cycles. These pressures differ from operating a cloud software service.

Finally, the environmental footprint deserves direct treatment. A gigawatt is a measure of power, not intelligence. The real impact depends on utilization, grid supply, cooling efficiency, hardware performance, and the duration of the deployment.

NVIDIA says Vera Rubin delivers more tokens per watt than Blackwell. That is a vendor claim until workload-specific measurements and operating data become available.

Efficiency improvements can also lower usage costs and increase total demand. Hark should therefore report both performance efficiency and absolute resource consumption if its deployment reaches the announced scale.

These issues do not invalidate the partnership. They define the evidence required to judge it.

What Hark Must Prove After the Vera Rubin Deal

Three signals will determine whether Hark’s infrastructure commitment represents product conviction or capacity secured ahead of evidence.

The first signal is the public platform launch. Hark said its initial product would become available before summer ends, leaving little time between announcement and delivery.

That launch should reveal the product’s actual scope. Readers should watch whether Hark offers open access, a limited preview, or a tightly controlled demonstration.

The most important tests will involve complete workflows. Can the agent remember relevant context, operate common software, recover from errors, and explain actions without constant supervision?

A launch with measurable reliability would strengthen Hark’s argument that integrated models and infrastructure create a differentiated assistant. A delayed or narrowly staged release would weaken the connection between the deal and near-term user value.

The second signal is a concrete deployment schedule. Hark needs to clarify how much Vera Rubin capacity it has committed to, where that capacity will operate, and when it becomes available.

“Gigawatt-scale” could describe a firm deployment, a phased goal, or access through partners. Those arrangements carry different levels of financial commitment and execution risk.

A schedule with named infrastructure partners and delivery milestones would make the announcement easier to evaluate. Continued ambiguity would suggest that the headline runs ahead of the operational plan.

The third signal is evidence covering safety, reliability, and adoption. Hark should publish computer-use benchmarks, memory controls, permission rules, and results from real users.

Adoption must also extend beyond sign-ups. Useful indicators include returning users, completed tasks, intervention rates, and the cost of serving each active user.

Those measures would connect compute to outcomes. Without them, a large deployment says more about capital availability than product quality.

NVIDIA also has something to prove. Vera Rubin is designed around agentic inference, but customer workloads will determine whether its architectural gains translate into lower operating costs.

Hark can become an important test because its proposed product combines several demanding workloads. Speech, vision, memory, reasoning, and computer control must work together in real time.

Success would strengthen NVIDIA’s claim that tightly integrated rack-scale systems support the next generation of agents. It would also encourage more AI startups to secure infrastructure earlier in their development cycles.

Failure would not necessarily indict the chips. Product design, model quality, privacy choices, distribution, and customer demand can overwhelm infrastructure advantages.

For developers, the lesson is to separate capacity from capability. Hardware determines what can run at scale. It does not determine whether an agent chooses the correct action or earns user trust.

Enterprise buyers should apply the same distinction. A vendor’s processor commitment does not replace security documentation, audit logs, deployment controls, or workload-specific performance results.

Knowledge workers should focus on the permission model. An agent becomes valuable when it can use relevant context, but broad context also expands the consequences of a mistake.

Hark’s bet is therefore clearer than its product. The company believes personalized agents will require frontier-scale infrastructure, persistent memory, and interfaces designed around natural interaction.

The next few months will show whether that thesis survives contact with users. Watch the launch, the deployment schedule, and the safety evidence in that order.

If Hark delivers all three, the Vera Rubin partnership will look like preparation for measurable demand. If those signals remain absent, the rsshub 36kr headline will stand as an infrastructure promise awaiting a product case.