Nvidia's agent-workload lead scales with interactivity, and the second-source budget line does not

The mechanism worth arguing about is where reusable state sits between turns. On DeepSeek V4, SemiAnalysis measured a 91% cache hit rate in high-bandwidth memory on a B300 configuration with 3 terabytes of DRAM and eight-way expert parallel decode at 384 concurrent sessions, with another 1.36% served from host memory ^[13]. The same engine on B200, at concurrency 196, hit 73% in HBM and pulled a further 20% off the host ^[14]. Total coverage is almost the same, roughly 92.4% against 93% ^[6]. What differs is which tier answers, and SemiAnalysis ties the heavier offload to sharply worse latency as user counts scale ^[14]. That is why the reported gap is not a constant: on the same model, B200's advantage over MI355X sits near 57% at 108 output tokens per second per user and 247% at 141 ^[4]. Thirty-three tokens per second of extra responsiveness multiplies the advantage by more than four ^[3].

Two other things only appear when a benchmark has a previous turn. Median input of 142,000 tokens against median output of 444 works out to roughly 320 tokens in for every token out ^[1], and 45% of the public sessions spawn at least one subagent ^[4]. Nvidia's TensorRT-LLM added boundary-aware incremental tokenization so a conversation resending its history does not re-tokenize it; on a Qwen 3.5 trace it matched full tokenization across all 1,087 transitions while mean per-turn processing time fell from 185.1 milliseconds to 11.3 ^[15], a cut of about 94% ^[2]. A fixed 8,000-in, 1,000-out sequence has no prior turn to reuse and therefore cannot show any of it ^[12].

The procurement question this reframes is not which vendor is faster in the abstract. It is what interactivity number goes into the service level target, because that number now selects the answer. Lowering it does not flip the ordering in the published data, only narrows it ^[4]. And the ledger against AMD gets worse on like-for-like open stacks: Qwen 3.5 served through SGLang on both sides puts Nvidia more than 20 times ahead at 90 tokens per second per user, according to SemiAnalysis ^[19].

Two caveats belong in the same breath. Forbes reads the five-times figure as a claim about one configuration rather than a market-wide ratio ^[5], and the public v1.0 subset is 393 Claude Code sessions under Apache 2.0 ^[8], transformed into session-scoped chained hash blocks of 64 tokens each so that prefix relationships survive but the original prompts and code do not ^[9]. That preserves the cache behaviour the argument rests on. It also means the evidence describes coding agents, replayed from one firm's own developers, and anyone buying on it is extrapolating from that to their own traffic.