What happened
InferenceX’s live ranking compares hardware serving Muse Spark Open, which the page labels an 816B model. It lists five platforms, each paired with a measured throughput per GPU, a derived cost per million tokens, precision, and serving engine. The top entry is NVIDIA GB300 NVL72: 43,643 tokens per second per GPU, $0.015 per million tokens, FP4, and SGLang. AMD MI355X ranks second at 40,997 tokens per second per GPU and $0.010 per million tokens, also using FP4 and SGLang. InferenceX says GB300 NVL72 is 6% ahead of MI355X on throughput.
The remaining listed systems are B300 at 39,693 tokens per second per GPU and $0.016 per million tokens; GB200 NVL72 at 38,936 and $0.013; and B200 at 35,736 and $0.013. All five entries show FP4 and SGLang. The table does not rank hardware lacking a measurement at the chosen operating point, so it is not a complete catalog of available accelerators.
The comparison uses the AgentX agentic coding workload and a matched target of 125 tokens per second per user. InferenceX describes this as an iso-interactivity comparison: it sweeps concurrency on its fleet to trace each platform’s throughput-versus-interactivity frontier, then reads each platform at the same per-user speed. Its rationale is that a system should not rank first merely by reporting high aggregate throughput at an interaction speed too slow for users. The page says the default target is 125 tokens per second per user because current runs cover the most hardware at that speed; another target is used only when it ranks more GPUs for the model.
InferenceX says its figures are measurements on real hardware using community serving engines, not estimates from specification sheets. The cost figure converts measured throughput into dollars per million total tokens using per-GPU-hour rates from the SemiAnalysis AI Cloud TCO model. The page says this derivation is shared with the InferenceX overview leaderboard. It also says rankings are rerun continuously as engine releases and configurations change. The newest result feeding this page landed on November 8, 2026.
What the archive adds
New Runtime’s October 23, 2026 coverage of Muse Spark Open on GB300 NVL72 discussed NVIDIA’s deployment details and throughput claims. Its November 3, 2026 comparison of GB300 NVL72 and MI355X already covered a head-to-head benchmark across interactivity targets. This ranking is a new measurement update on a known story, not a first comparison of these systems.
Trend signal
This does not yet fit a pattern New Runtime tracks. The benchmark measures infrastructure performance and derived serving cost for one model workload; it does not establish a broader shift in how agent systems are priced or evaluated.
Why now
The ranking is live and rerun continuously, with results changing as serving-engine releases and configurations land. InferenceX says the newest result feeding this page landed on November 8, 2026. That makes the page a current measurement snapshot rather than a fixed hardware specification or an evergreen declaration of a universal leader. For teams weighing deployments now, the timing matters because a benchmark can move when software configurations change even when the accelerator hardware does not.
Why it matters
GPU rankings can change depending on what a team means by “fast.” Aggregate tokens per second is useful only alongside the interaction speed users receive, especially for coding agents that stream output while a person or another tool waits. InferenceX’s matched target makes the comparison more decision-relevant than an unmatched peak-throughput number, but it remains a single workload and operating point.
The listed costs also show why a throughput winner is not automatically the cost winner. MI355X trails GB300 NVL72 in measured throughput but has the lower calculated cost per million tokens in this table. Teams planning deployment should therefore compare options at their own required interactivity, concurrency, utilization, and workload mix. The ranking can narrow the hardware shortlist; it cannot by itself predict a production bill or prove that one accelerator is the best choice for every Muse Spark Open deployment. The practical value is in the comparison method: hold user-facing speed constant, then examine throughput and cost as distinct outcomes rather than treating one headline number as the whole decision.
Limits and uncertainty
The results describe the AgentX agentic coding workload at 125 tokens per second per user, not every inference workload or interaction target. InferenceX reports measurements from its own fleet and describes the cost as derived from its SemiAnalysis AI Cloud TCO model; the page does not establish that every team will obtain the same hardware rates, utilization, or serving configuration. The ranking excludes systems without measurements at this operating point. It is continuously rerun, so figures may change as engines and configurations change. The page does not provide uncertainty ranges or independently verified production costs for each platform.
Skill to build now
This week, define one representative agentic coding request and the minimum acceptable streaming speed for your users. For each candidate GPU and serving stack, record tokens per second per GPU and the cost per million tokens at that same interaction target; do not compare one system’s peak throughput with another system’s slower operating point. If your team has access to a test environment, repeat the run at the concurrency and prompt/output mix expected in production. That will not reproduce every cost factor, but it will show whether the published ranking transfers to your workload.
Reader takeaway
GB300 NVL72 leads this particular Muse Spark Open comparison on throughput, but MI355X has the lowest listed cost per million tokens. The useful result is not a universal GPU winner: it is a reminder to compare measured performance at a relevant user-facing speed and to evaluate cost separately for the workload you actually intend to serve.