The person who answers for this on Friday is the capacity planner who wrote a savings figure next to "second-source accelerator" in next year's plan, and who now has to say which lines of the rack that figure actually covers.
Count the lines. TechCrunch's description of the Vera Rubin rollout names five unit classes going into a deployment: the Rubin GPU, the Vera CPU, an inference accelerator it calls the Groq 3 LPX, and matching racks for storage and networking [4]. A custom accelerator bids on one of those five, leaving four, which is 20 percent of the named classes displaced [1]. This is a scope test rather than a dollar ceiling, since the dollars are not spread evenly across the five, and substitution decks tend to skip it.
The mechanism here is memory timing rather than raw throughput. Hardy's argument for the Vera CPU is that only so much memory fits in a single server or compute platform [5], so the useful work is getting data to the GPU at the moment it is needed [7]. His figure for that work is upwards of 3x on the operations Vera accelerates, which he says lets Nvidia's flash run without bottlenecking [6]. The claim is scoped to a vendor number, on operations the vendor chose, inside the data path, not a training or serving throughput result.
Buying teams tend to frame the comparison as chip against chip, on price per unit of compute. The source says operators are pushing on tokens per watt instead [7], a rack-level measure that includes the traffic direction a custom accelerator does not change.
OpenAI took the other route to the same bottleneck. The company says it designed Jalapeño to minimize data movement and communication delays, with a domain large enough to keep an entire workload inside one connected system [8]. Buying better traffic control and arranging not to need it are both live options, and either way, the quantity being optimized stopped being processor cycles.
TechCrunch's own conclusion, that Nvidia holds a commanding early lead in making the whole system work, is the writer's read rather than a measurement [9]. And investors have been chewing on GPU substitution for a year already, following a tenfold market cap run from the start of 2023 to mid-2025 [2] that works out near 2.5x a year compounded [2].
The forcing function here is dull but effective: before anyone quotes a saving, write the rack bill of materials by class and name the supplier of each line after the swap. Then score the candidate part on two axes. Axis one: whether it changes who supplies the host CPU, storage and networking, or only the accelerator. Axis two: whether it changes the workload's data path, or only what executes the workload. A part that only executes gets priced against one invoice line, while a part that shortens the data path can be priced against the rack. A part that scores low on both axes but is still presented as a platform decision is really just a discount on one line.
