Cheaper tokens excite me less when the work stays stuck. NVIDIA announced Rubin in January 2026, promising substantially lower inference costs than Blackwell, with partner products expected in the second half of the year. Good. I want that improvement to reach a task I can actually finish.
I am writing this retrospective in October 2026, after encountering a much closer source of frustration. In June, I reported agents stalled by a Claude usage limit and asked for work to resume. In July, I questioned whether I was using the capacity available across Cursor, Codex, and Claude Code. I wanted the work moving and was unsure whether the available tools were being put to use.
Those episodes tell us nothing about Rubin's performance. They explain why I distinguish purchased capacity from capacity I can use. A hardware announcement describes part of the bill; the service determines how that capacity reaches users. My own organization still comes after that. I can have tools sitting idle while a task waits for my decision.
The first thing I want when an agent stops is the reason. Did its allowance run out? Is accessible information missing? Does another task need to finish? Each answer changes the next step. Calling everything slowness makes complaining easy and leaves the queue untouched. Apparently even a useful complaint needs to establish who is waiting for whom.
Several tools give me options, but moving work requires context. A task tied to a particular integration may take more effort to move than to resume later. I want to examine the prepared task, establish what it needs, and then choose where to execute it. Collecting subscriptions without assigning work just gives me more icons.
I also need to watch the urge to fill every spare bit of capacity. Requesting more investigations is easy; reading the results still takes time. If a price reduction produces only a larger pile of answers, I have created another queue myself. How much of that material helps with a pending decision?
Better hardware deserves attention, including the conditions behind NVIDIA's comparison. To assess its effect on my work, I want a closer measure: how long a prepared task sits idle and how much effort restarting it takes. My next step is to record the cause of the wait before looking for more capacity to buy.