This weekend’s AI newsletters carried a striking line: AMD’s MI355X serves tokens for Kimi K3, the Chinese open-weight model, at up to 2.3 times lower cost than Nvidia’s B200, a gap one summary credits to memory capacity and kernel tuning. A roundup of FourWeekMBA’s write-up attributes the analysis to SemiAnalysis’s InferenceX benchmark. Lens could not find a 2.3x figure in Wafer’s benchmark post “Is memory the moat?” by Ian Ye, published on July 31, 2026, which tested Kimi K3 on these chips. What that post does show is narrower, and still notable.

What Wafer measured

Kimi K3 has about 2.8 trillion parameters and needs more than 1.5 terabytes of GPU memory for its weights alone, according to Wafer. Wafer tested three setups on a throughput workload of 1,024 input tokens and 400 output tokens: eight AMD MI355X GPUs with 288 GB each in one node, eight Nvidia B300 GPUs with 288 GB each in one node, and sixteen Nvidia B200 GPUs with 192 GB each across two nodes, because the weights plus a one-million-token cache would not fit in a single eight-GPU B200 node.

  • MI355X: 952 tokens per second peak aggregate, at $2.50 per GPU-hour, for 48 tokens per second per dollar of hourly rental.
  • B300: 1,568 tokens per second, at $6.00 per GPU-hour, for 33.
  • Two-node B200: 498 tokens per second, at $4.25 per GPU-hour, for 7.

Wafer takes its prices from gpus.io. The B300 delivers about 1.65 times the MI355X node’s aggregate throughput, Wafer says, but costs about 2.4 times as much per GPU-hour. Lens’s arithmetic from Wafer’s table puts the MI355X at roughly 1.45 times the B300’s throughput per dollar and about 6.9 times the two-node B200’s. Neither is 2.3x.

Wafer’s own caveats

  • The B200 setup is handicapped. Wafer says cross-node communication on the decode path somewhat deflates its B200 numbers, so the 7 reflects a two-node arrangement as much as the chip.
  • AMD’s software lags. Wafer says AMD has slower kernels and less day-zero support in inference frameworks. During testing, Kimi K3 on AMD’s ROCm software fell back to a generic attention path because a faster kernel would not load; a workaround Wafer applied gave roughly two to three times faster prompt processing.
  • Long prompts favor the B300. In Wafer’s separate test of a 172,000-token prompt processed cold, the MI355X was slower than the B300.
  • Scope. The results come from one throughput test shape and one prompt length, and Wafer notes that prices may change.

Another B200 number

SemiAnalysis’s InferenceX publishes its own Kimi K3 results. Its B200 page lists a figure in the thousands of tokens per second per GPU, 6,223 on the view Lens checked, at a target of 50 tokens per second per user. The workload is agentic coding in FP4 precision, served with Dynamo and vLLM across multiple nodes, with runs dated October 26. The page converts throughput into a cost of $0.077 per million tokens at a hyperscaler-owned rate of $1.73 per GPU-hour, and $0.17 at a retail rate of $3.70. It does not say whether the tokens counted are input, output or both.

Wafer’s two-node B200 test produced about 31 tokens per second per GPU. The two figures differ by more than two orders of magnitude, but they come from different workloads, precisions, serving software and price assumptions, so they cannot be placed on one chart. The point is that the B200 side of any AMD comparison depends heavily on how the B200 is configured, served and priced.

What the claim does and does not mean

Lens could not trace the 2.3x figure to a primary source. The defensible reading of Wafer’s data is narrower: on one throughput test, at Wafer’s rental prices, an eight-GPU MI355X node delivered more tokens per dollar than both the B300 and a two-node B200 setup. Wafer’s framing in its title is memory: 288 GB per GPU lets the model fit on one node. What the data does not show is lower cost on long prompts, on other workloads, at hyperscaler prices, or against optimized Nvidia serving stacks.

Lens has examined a similar gap between a headline multiple and the underlying test before, in Cerebras CS-4 and WSE-3 Turbo: What “30x Faster Than GPUs” Does and Does Not Mean.

What to watch next

  • A primary source for the 2.3x figure, with the workload, precision and prices stated.
  • Like-for-like runs that fix the model, precision, context length and per-user speed, and report tokens per second per GPU and per dollar.
  • Whether the ROCm kernel gaps Wafer describes close.
  • Whether AMD’s price advantage holds if rental prices move.

Sources and notes

Benchmarks: Wafer, “Is memory the moat?” (July 31, 2026); SemiAnalysis InferenceX, Kimi K3 on B200. The 2.3x claim as circulated: Rio Rundown, November 1, 2026. Throughput-per-dollar ratios other than Wafer’s own table values are Lens’s arithmetic. Graphic: Lens, from Wafer’s table. Thumbnail photo: Nvidia’s headquarters in Santa Clara, California (2018) and an office building there that housed AMD’s headquarters (2020), a composite of two photos by Coolcaesar via Wikimedia Commons (Nvidia, AMD), licensed under CC BY-SA 4.0. Illustrative photos; they do not show the benchmark described.