Dataconomy’s claim that AMD’s Instinct MI355X beats Nvidia’s B300 on Kimi K3 inference cost is not supported by the public InferenceX record it cites. As of August 3, SemiAnalysis’s InferenceX performance-per-dollar index lists exactly one Kimi K3 hardware pairing: Nvidia B200 versus Nvidia B300. It does not publish a Kimi K3 MI355X-versus-B300 cost comparison, and its B200/B300 page says no interpolated cost-per-token data is available for the default configuration.
The discrepancy is more than a missing chart point. Dataconomy appears to have combined real MI355X results for the older 1-trillion-parameter Kimi K2.5/K2.6/K2.7-Code family with Kimi K3’s new hardware requirements. The resulting headline converts a plausible broader AMD price-performance narrative into a Kimi K3 result that the benchmark’s own public dashboard does not show.
For enterprises deciding whether Kimi K3 can be self-hosted on AMD infrastructure, the meaningful news is narrower but still consequential: AMD has a documented, validated path to load and serve Kimi K3 on an eight-GPU MI355X node. It has not yet released independently comparable K3 throughput or cost-per-token figures against B300.
SemiAnalysis describes InferenceX as an open-source, continuously updated benchmark whose recipes, run logs, artifacts, and resulting database entries are publicly auditable. Its Kimi K3 landing page lists the model as a 2.8-trillion-parameter workload, but its performance-per-dollar directory identifies only one K3 GPU pair: B200 versus B300.
Opening that K3 comparison reveals an important limitation. The page provides estimated “owning hyperscaler” GPU costs of $1.73 per B200 GPU-hour and $2.26 per B300 GPU-hour, but it supplies no interpolated dollars-per-million-token result for the default K3 FP4 workload. There is no MI355X entry on that page, no AMD-versus-Nvidia K3 comparison, and no disclosed result that could substantiate a claim that MI355X delivered lower K3 token cost than B300.
That is sharply different from the InferenceX page for Kimi K2.5/K2.6/K2.7-Code. There, SemiAnalysis does publish a B300-versus-MI355X comparison. At one 46-token-per-second-per-user operating point, the site estimates $0.19 per million output tokens for B300 and $0.12 for MI355X. At 72 tokens per second per user, the gap narrows to $0.21 versus $0.18. At 97 tokens per second per user, B300 becomes the more cost-efficient option, at $0.23 versus $0.25.
In other words, the older Kimi benchmark does support a qualified conclusion: MI355X can be cheaper for some Kimi K2-class serving targets, while B300 can pull ahead at others. It does not establish a blanket AMD win, and it does not establish any Kimi K3 result.
The $2.50-per-hour MI355X and roughly $6-per-hour B300 figures in the submitted report are also not the economics used by InferenceX’s K2-family comparison. SemiAnalysis uses its own “owning hyperscaler” total-cost model, pegging MI355X at $1.50 per GPU-hour and B300 at $2.26 per GPU-hour in that specific chart. Public cloud pricing is a different measurement with different assumptions around availability, tenancy, region, commitment, networking, and whether a provider sells a single GPU or only an eight-GPU node.
That footprint is why the K3 story is about memory capacity and system topology before it is about chip-level token speed. Nvidia’s B200 carries 180 GB of HBM3E per GPU, putting eight B200s at 1.44 TB—below the K3 checkpoint size. Nvidia’s B300 and AMD’s MI355X each provide 288 GB per GPU, giving an eight-GPU server 2.304 TB of aggregate HBM.
SemiAnalysis has made the same distinction publicly: a single eight-GPU B200 system cannot hold Kimi K3 at FP4, while a B300, GB300 NVL72, or MI355X deployment has the per-GPU memory capacity needed for the model’s practical single-node configurations. vLLM separately says K3 can barely fit in one Nvidia DGX B300 and needs at least 16 B200 or GB200 GPUs on that generation.
This is a material correction to the common shorthand that K3 “requires 288 GB per GPU.” The model requires a cluster-scale memory pool and a viable distributed-inference configuration; it does not run on one 288 GB accelerator. For the practical minimum configurations being discussed, organizations are renting or buying eight MI355X or B300 accelerators at once, not shopping single-GPU hourly prices in isolation.
At the submitted article’s nominal $2.50 per MI355X GPU-hour, an eight-GPU node would be $20 per hour before host CPUs, local storage, networking, management, and any provider markup for the full server. At $6 per B300 GPU-hour, the equivalent eight-GPU node would be $48 per hour. Those figures illustrate why price matters, but they cannot produce a cost-per-token verdict without measured throughput, latency, concurrency, cache behavior, and actual deployment topology.
AMD also says it completed a full 1,319-sample GSM8K five-shot correctness validation on the MI355X configuration, using a 16,384-token maximum model length. That is legitimate deployment evidence: the model loads, starts an OpenAI-compatible service, and completes a basic correctness test on AMD’s latest accelerator.
But AMD explicitly excludes the missing metrics from its article. It makes no throughput, time-to-first-token, time-per-output-token, or kernel-efficiency claims. It says K3-specific tuning for tensor-parallel collectives, MXFP4 grouped MoE, Kimi Delta Attention, prefill, decode, and long-context serving will appear in later work.
That omission is understandable. K3 is a complicated mixture-of-experts model, and serving it efficiently requires more than having enough HBM. The workload’s expert routing, long-context cache management, speculative decoding, inter-GPU communication, and framework maturity can change the answer substantially. A system that loads K3 with useful headroom is not automatically the system with the lowest production cost per completed request.
vLLM’s K3 launch coverage shows the same split. The project announced day-one support across Nvidia Hopper and Blackwell platforms plus AMD MI355X, but its published performance headline was for 16 Nvidia GB300 NVL72 GPUs: 118 tokens per second per user without speculative decoding and up to 370 tokens per second with DSpark. It did not publish an equivalent MI355X K3 throughput figure in that announcement.
The reported MI355X victory on K3 cost, however, should not guide capacity planning. SemiAnalysis’s public K3 data does not show that matchup, does not provide default interpolated cost-per-token results even for its B200/B300 K3 pair, and cannot support a claim about AMD’s K3 token economics. The public result being cited belongs to the smaller Kimi K2.5/K2.6/K2.7-Code model family, where the outcome varies with the operating point.
For now, the defensible procurement position is simple: treat MI355X as a validated Kimi K3 hosting platform, treat B300 as the corresponding Nvidia high-memory option, and demand a reproducible K3 run with matched context length, concurrency, precision, interactivity target, software stack, and full-node price before assigning either vendor a cost-per-token crown.
For enterprises deciding whether Kimi K3 can be self-hosted on AMD infrastructure, the meaningful news is narrower but still consequential: AMD has a documented, validated path to load and serve Kimi K3 on an eight-GPU MI355X node. It has not yet released independently comparable K3 throughput or cost-per-token figures against B300.
The Kimi K3 comparison is missing from InferenceX
SemiAnalysis describes InferenceX as an open-source, continuously updated benchmark whose recipes, run logs, artifacts, and resulting database entries are publicly auditable. Its Kimi K3 landing page lists the model as a 2.8-trillion-parameter workload, but its performance-per-dollar directory identifies only one K3 GPU pair: B200 versus B300.Opening that K3 comparison reveals an important limitation. The page provides estimated “owning hyperscaler” GPU costs of $1.73 per B200 GPU-hour and $2.26 per B300 GPU-hour, but it supplies no interpolated dollars-per-million-token result for the default K3 FP4 workload. There is no MI355X entry on that page, no AMD-versus-Nvidia K3 comparison, and no disclosed result that could substantiate a claim that MI355X delivered lower K3 token cost than B300.
That is sharply different from the InferenceX page for Kimi K2.5/K2.6/K2.7-Code. There, SemiAnalysis does publish a B300-versus-MI355X comparison. At one 46-token-per-second-per-user operating point, the site estimates $0.19 per million output tokens for B300 and $0.12 for MI355X. At 72 tokens per second per user, the gap narrows to $0.21 versus $0.18. At 97 tokens per second per user, B300 becomes the more cost-efficient option, at $0.23 versus $0.25.
In other words, the older Kimi benchmark does support a qualified conclusion: MI355X can be cheaper for some Kimi K2-class serving targets, while B300 can pull ahead at others. It does not establish a blanket AMD win, and it does not establish any Kimi K3 result.
The $2.50-per-hour MI355X and roughly $6-per-hour B300 figures in the submitted report are also not the economics used by InferenceX’s K2-family comparison. SemiAnalysis uses its own “owning hyperscaler” total-cost model, pegging MI355X at $1.50 per GPU-hour and B300 at $2.26 per GPU-hour in that specific chart. Public cloud pricing is a different measurement with different assumptions around availability, tenancy, region, commitment, networking, and whether a provider sells a single GPU or only an eight-GPU node.
Kimi K3 changes the hardware floor
Kimi K3 is not simply a somewhat larger Kimi K2. Moonshot AI’s technical report puts the model at 2.8 trillion total parameters, with 104 billion activated per token, native vision support, and a 1-million-token context window. Its released mixed MXFP4 and BF16 checkpoint occupies roughly 1.56 TB before the memory required for serving state, cache, communication buffers, and runtime overhead.That footprint is why the K3 story is about memory capacity and system topology before it is about chip-level token speed. Nvidia’s B200 carries 180 GB of HBM3E per GPU, putting eight B200s at 1.44 TB—below the K3 checkpoint size. Nvidia’s B300 and AMD’s MI355X each provide 288 GB per GPU, giving an eight-GPU server 2.304 TB of aggregate HBM.
SemiAnalysis has made the same distinction publicly: a single eight-GPU B200 system cannot hold Kimi K3 at FP4, while a B300, GB300 NVL72, or MI355X deployment has the per-GPU memory capacity needed for the model’s practical single-node configurations. vLLM separately says K3 can barely fit in one Nvidia DGX B300 and needs at least 16 B200 or GB200 GPUs on that generation.
This is a material correction to the common shorthand that K3 “requires 288 GB per GPU.” The model requires a cluster-scale memory pool and a viable distributed-inference configuration; it does not run on one 288 GB accelerator. For the practical minimum configurations being discussed, organizations are renting or buying eight MI355X or B300 accelerators at once, not shopping single-GPU hourly prices in isolation.
At the submitted article’s nominal $2.50 per MI355X GPU-hour, an eight-GPU node would be $20 per hour before host CPUs, local storage, networking, management, and any provider markup for the full server. At $6 per B300 GPU-hour, the equivalent eight-GPU node would be $48 per hour. Those figures illustrate why price matters, but they cannot produce a cost-per-token verdict without measured throughput, latency, concurrency, cache behavior, and actual deployment topology.
AMD has validated loading Kimi K3, not benchmarked a K3 win
AMD’s own July 27 technical article is more precise than the cost-performance claim. It says Kimi K3 can be deployed on eight MI355X GPUs using tensor parallelism across eight accelerators, with the ATOM serving stack and an FP8 latent KV cache. AMD calculates roughly 190.974 GiB of model weights per GPU and approximately 205.401 GiB after adding known state for one 1-million-token sequence. That leaves about 82.6 GiB per GPU before accounting for remaining runtime overhead.AMD also says it completed a full 1,319-sample GSM8K five-shot correctness validation on the MI355X configuration, using a 16,384-token maximum model length. That is legitimate deployment evidence: the model loads, starts an OpenAI-compatible service, and completes a basic correctness test on AMD’s latest accelerator.
But AMD explicitly excludes the missing metrics from its article. It makes no throughput, time-to-first-token, time-per-output-token, or kernel-efficiency claims. It says K3-specific tuning for tensor-parallel collectives, MXFP4 grouped MoE, Kimi Delta Attention, prefill, decode, and long-context serving will appear in later work.
That omission is understandable. K3 is a complicated mixture-of-experts model, and serving it efficiently requires more than having enough HBM. The workload’s expert routing, long-context cache management, speculative decoding, inter-GPU communication, and framework maturity can change the answer substantially. A system that loads K3 with useful headroom is not automatically the system with the lowest production cost per completed request.
vLLM’s K3 launch coverage shows the same split. The project announced day-one support across Nvidia Hopper and Blackwell platforms plus AMD MI355X, but its published performance headline was for 16 Nvidia GB300 NVL72 GPUs: 118 tokens per second per user without speculative decoding and up to 370 tokens per second with DSpark. It did not publish an equivalent MI355X K3 throughput figure in that announcement.
The useful conclusion for Windows and infrastructure teams
AMD now has a credible Kimi K3 deployment option for organizations building on ROCm: an eight-MI355X system has the memory capacity to host the 1.56 TB checkpoint, and AMD has documented a working TP8 serving recipe. That is strategically important because it prevents K3’s unusually large checkpoint from becoming a B300-only procurement decision.The reported MI355X victory on K3 cost, however, should not guide capacity planning. SemiAnalysis’s public K3 data does not show that matchup, does not provide default interpolated cost-per-token results even for its B200/B300 K3 pair, and cannot support a claim about AMD’s K3 token economics. The public result being cited belongs to the smaller Kimi K2.5/K2.6/K2.7-Code model family, where the outcome varies with the operating point.
For now, the defensible procurement position is simple: treat MI355X as a validated Kimi K3 hosting platform, treat B300 as the corresponding Nvidia high-memory option, and demand a reproducible K3 run with matched context length, concurrency, precision, interactivity target, software stack, and full-node price before assigning either vendor a cost-per-token crown.
References
- Primary source: dataconomy.com
Published: 2026-08-03T08:34:37+00:00
AMD’s MI355X beats Nvidia B300 on Kimi K3 inference cost
AMD’s Instinct MI355X GPU ran Moonshot AI’s 2.8-trillion-parameter Kimi K3 model at less than half the hourly cloud rental costdataconomy.com - Related coverage: tomsguide.com
What is Kimi K3? Everything you need to know about the new ChatGPT rival | Tom's Guide
What is Kimi K3? Discover Moonshot AI’s new open-weight ChatGPT rival, its key features for coding and agents, pricing, and the latest news.www.tomsguide.com - Related coverage: tomshardware.com
China's Moonshot AI reportedly used Nvidia Blackwell chips for training Kimi K3 — company circumvented both U.S. export and Chinese import controls to acquire compute | Tom's Hardware
The US gov't says Moonshot has purchased Blackwell systems and rented time on foreign clouds to help it train models.www.tomshardware.com - Related coverage: amd.com
AMD Instinct™ MI355X GPUs
Built on the 4th Gen AMD CDNA™ architecture, AMD Instinct™ MI355X GPUs deliver leadership AI and HPC performance enabling high density infrastructures with 288GB HBM3E memory, 8TB/s bandwidth, and expanded MXFP6 and MXFP4 datatype support.www.amd.com - Related coverage: kie.ai
What Is Kimi K3? Moonshot's 2.8T, 1M-Context Flagship
Kimi K3 is Moonshot AI's next-gen model — a ~2.8T MoE with a 1M-token context window, launched July 16, 2026 on Kimi Code and the Kimi app.kie.ai - Related coverage: ctxwire.com
Kimi K3: A 2.8T Open-Weight Bid That Isn't Downloadable Yet | Context Wire
Moonshot AI announced Kimi K3, a 2.8-trillion-parameter model with strong agentic benchmarks. The open weights are promised, not published. What we verified.
ctxwire.com
- Related coverage: axios.com
China's open-weight Kimi model stuns AI world with frontier-level results
Kimi K3's early performance is fueling alarm in Silicon Valley and Washington.www.axios.com
- Related coverage: implicator.ai
Moonshot Launches 2.8-Trillion-Parameter Kimi K3
Moonshot has put Kimi K3 online with 2.8 trillion parameters, native vision and a one-million-token context window, but the model's most important test cannot begin yet. Its benchmark table puts K3 near proprietary rivals even though the comparisons used different agent harnesses, and one early...
www.implicator.ai
- Related coverage: todayforai.com
Moonshot AI Launches Kimi K3: 2.8T Parameters, 1M-Token Context, and Full Open Weights Due by July 27 | Today for AI
Moonshot AI launched its flagship Kimi K3 model on July 16, 2026. The 2.8-trillion-parameter MoE model natively supports vision and a 1-million-token context window. It is live across Kimi, Kimi Work, Kimi Code, and the API, with full model weights scheduled for release by July 27.todayforai.com
- Related coverage: presenc.ai
Moonshot Kimi K3, 2.8T Open Weights, 1M Context, Benchmarks, Brand-Visibility Brief | Presenc AI
Release brief for Moonshot AI Kimi K3: July 16 2026 launch, 2.8 trillion parameters, the largest open-weight model released to date, 1M context, and...presenc.ai - Related coverage: kimi-ai.chat
Kimi K3: Specs, API Pricing, 1M Context & Weight Status
Kimi K3 specs, API pricing, 1M context, model IDs, downloadable Hugging Face weights and the custom Kimi K3 License. Verified July 29, 2026.kimi-ai.chat - Related coverage: amd.com
Kimi-K3 on AMD Instinct GPUs
Day 0 support for Kimi-K3 on AMD Instinct MI355X GPUs with validated TP8 setups.www.amd.com - Related coverage: linkedin.com
Kimi K3 GPU requires large system for memory | SemiAnalysis posted on the topic | LinkedIn
Kimi K3 2.8T is so large that it will not fit on a single NVIDIA DGX B200, even at FP4. A GB300 NVL72, B300, or MI355X system is required, as each GPU has 288 GB of memory. One optimization that could make Kimi K3 fit on B200 is to gang multiple nodes together and use a technique called WideEP...www.linkedin.com
- Related coverage: blockchain.news
Kimi-K3 Scales to 2.8T Parameters on AMD Instinct GPUs
Kimi-K3, a 2.8T-parameter AI model, deployed on AMD Instinct MI355X GPUs with Day 0 support for advanced inference capabilities.blockchain.news - Related coverage: cometapi.com
Kimi K3 Self-Hosting vs API in 2026: Hardware & Cost - CometAPI
Kimi K3 self-hosting needs 8+ GB300 or MI350X/MI355X GPUs and about 1.56 TB of weights. Compare API pricing, license terms, and break-even costs.www.cometapi.com - Related coverage: lightseek.org
Kimi K3 at Day 0: Frontier Model Enablement on Leading Platforms with TokenSpeed | LightSeek Foundation
TokenSpeed enables Day 0 support for Moonshot AI's 2.8T-parameter Kimi K3 on NVIDIA Blackwell and AMD CDNA4 within one week, with a unified flat KV cache, disaggregated serving, and specialized CuteDSL, TensorRT-LLM, and Gluon kernels.
lightseek.org
- Related coverage: runpod.io
Deploy Kimi K3 on Runpod
Deploy Kimi K3, Moonshot AI's 2.8T-parameter open-weight model, on Runpod. Call the OpenAI-compatible Public Endpoint or self-host on multi-node GPUs.
www.runpod.io
- Related coverage: tokencost.app
Kimi K3 Self-Hosting Cost vs the $3/$15 API | TokenCost
Kimi K3's open weights are 1.56TB and need 8x B300 minimum. Every host charges $3/$15. Break-even math, GPU rental rates, and the license thresholds.tokencost.app
- Related coverage: gpufinder.dev
B300 Cloud Pricing from $6.94/hr (2026) — 5 Providers
B300 cloud GPU pricing from $6.94/hr across 5+ providers in 2026. Compare on-demand and spot rates with real-time availability, updated daily.gpufinder.dev - Related coverage: docs.nvidia.com
Components — NVIDIA HGX AI Factory
docs.nvidia.com
- Related coverage: aiwiki.ai
- Related coverage: docs.nvidia.com
System Hardware & Components — NVIDIA NVL72 AI Factory
docs.nvidia.com
- Related coverage: cyfuture.ai
NVIDIA B300 GPU Explained: Features, Specifications & Performance (2026)
NVIDIA B300 GPU: explore full specifications, performance benchmarks, cloud pricing, and why AI teams are rushing to rent NVIDIA B300 GPU servers for LLM training and inference in 2026.
cyfuture.ai
- Related coverage: computeprices.com
HGX B300 GPU Price: $2.63/hr | Compare 10+ Providers | ComputePrices.com
Compare HGX B300 GPU prices across 10+ cloud providers from $2.63/hr. 288GB VRAM, Blackwell architecture. Find the best rates for AI training and inference.computeprices.com - Related coverage: cyfuture.cloud
Rent NVIDIA B300 GPU | NVIDIA B300 GPU Cloud | Cyfuture Cloud
Rent NVIDIA B300 GPU Server with enterprise-grade infrastructure. Get scalable B300 GPU as a Service, transparent B300 GPU price, and high-performance AI compute.cyfuture.cloud - Related coverage: investor.nvidia.com
NVIDIA Corporation - NVIDIA Blackwell Ultra AI Factory Platform Paves Way for Age of AI Reasoning
Top Computer Makers, Cloud Service Providers and GPU Cloud Providers to Boost Training and Test-Time Scaling Inference, From Reasoning to Agentic and Physical AI New Open-Source NVIDIA Dynamo Inference Software to Scale Up Reasoning AI Services With Leaps in Throughput, Faster Response Time and...investor.nvidia.com - Related coverage: oracle.com
- Related coverage: nvidianews.nvidia.com
- Related coverage: nvidianews.nvidia.com
- Related coverage: tomshardware.com
Nvidia could be readying B30A accelerator for Chinese market — new Blackwell chip reportedly beats H20 and even H100 while complying with U.S. export controls | Tom's Hardware
Blackwell for China is in the works.www.tomshardware.com