Hark needs task-level unit economics

The core issue is that Hark is stacking several expensive businesses on top of each other before proving that consumers will regularly pay enough to cover each completed task. Every Handoff request spins up model inference plus a dedicated virtual computer, while the broader product also carries hardware, connectivity, and model training costs. That makes completed task economics, not downloads or device sales, the real test of whether the integrated model can become durable.

• Hark was still pre revenue as of September 2026, with Handoff in research preview and hardware not yet shipping. That means willingness to pay is still unproven at the moment when the company has already committed to large fixed costs in compute, hardware, and carrier distribution.
  1 sacra
• The likely pricing problem is simple. A user may want a flat monthly subscription, but Hark incurs variable cost every time the agent browses, clicks, retries, and runs for minutes on a cloud computer. That is why credits, task caps, or paid tiers matter more here than for ordinary chatbots.
  4 sacra
• Recent AI device history shows why this matters. Humane sold a dedicated AI device but ended in a $116M asset sale to HP, while Meta scaled smart glasses through existing distribution and daily utility. Hark needs the Meta outcome, not the Humane one, and that requires repeat paid usage, not novelty.
  5 hp 6 fb 7 fb

The path forward is to turn Handoff from an impressive demo into a habit where users trust it with enough high frequency jobs to justify recurring spend. If Hark can make travel booking, shopping, scheduling, and returns feel reliably faster than doing them by hand, the hardware and carrier layer become leverage instead of overhead.