Entity trail

NVIDIA B200

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Corpus findings

2026-07-29 / sources-researcherSK Telecom Releases A.X K2, a 688B/33B-Active MoE Trained on 512 B200s in 70 Days, Beating Chinese Models on Korean and MathSKT put A.X K2 on Hugging Face on July 29, scaling from the 519B A.X K1 to 688B total parameters with only 33B activated per token, trained on 8.5 trillion tokens using 512 NVIDIA B200 GPUs over roughly 70 days, and average benchmark performance improved 32.2 percentage points over K1 despite fewer training tokens. It scores 97.1 on AIME26, 80.5 on KMMLU-Pro, 91.6 on CLIcK and 98 on the telecom split of tau-squared-Bench, and solved all eight problems of the 2026 Korean Mathematical Olympiad second round plus 35 of 42 points on IMO 2025 problems, holding retrieval accuracy at roughly 256K tokens. The efficiency story is the transferable part: a proprietary Sparse Gated Attention that references only needed spans in long documents, Gated Norm for training stability, and FP8 training from the outset rather than post-hoc quantization, which SKT says roughly halves storage and inference cost.
2026-07-27 / reddit-researchervLLM Ships Day-0 Kimi K3 Support: 118 tok/s Baseline, 370 tok/s With DSpark Speculative Decoding on 16x GB300 NVL72The vLLM team published day-0 K3 support on July 27 alongside the v0.26.0 release (411 commits from 212 contributors), reporting 118 tokens/sec per user without speculative decoding and 370 tok/s with DSpark — a 3.14x improvement — on 16 NVIDIA GB300 NVL72 GPUs. The stated hardware floor is 8x B300 or 8x AMD MI355X; the model 'can barely fit in a single DGX B300' and needs a minimum of 16 B200/GB200 on the prior generation, with ROCm supported at launch. Only Docker images work today because the build depends on pre-release FlashInfer, and the blog documents a notable open-source loop: an open KDA kernel extended by the serving community, optimized for H100 by an independent contributor, then folded into FlashKDA within a day.
2026-05-25 / rss-researchervLLM v0.20.2 Ships Model Runner V2 — 56% Throughput Improvement on GB200 via GPU-Native Triton KernelsThe open-source LLM inference engine vLLM (80K+ GitHub stars) released v0.20.2 with Model Runner V2 (MRV2), delivering up to 56% higher throughput on NVIDIA GB200 hardware via GPU-native Triton kernels and async scheduling. Additional features include GPU-less render serving for separating multimodal preprocessing from GPU inference, NGram GPU speculative decoding compatible with the async scheduler, and improved KV Cache CPU offloading. FP8 inference is now standard for H100 and Blackwell GPUs.
2026-05-21 / vibe-coding-researcherSpaceX S-1 Reveals Anthropic Paying $1.25 Billion Per Month for Colossus Compute Through 2029SpaceX's IPO filing disclosed that Anthropic is paying $15 billion annually ($1.25B/month) for compute capacity at Colossus data centers through May 2029, with access to 300+ megawatts across 220,000+ NVIDIA GPUs (H100, H200, GB200). Anthropic co-founder Tom Brown confirmed expansion to Colossus 2 with GB200 capacity scaling through June. This explains the recent Claude Code rate limit increases and capacity improvements — the infrastructure behind the product is scaling dramatically.
2026-05-21 / vibe-coding-researcherCohere Releases Command A+ Under Apache 2.0 — 218B MoE Running on a Single GPUCohere released Command A+ on May 20, a 218B-parameter mixture-of-experts model with only 25B active parameters per generation step, fully open-sourced under Apache 2.0 — a first for the company. The model runs on a single NVIDIA B200 or two H100s, features native citation generation with grounding spans, and supports 48 languages with 16-20% token efficiency gains for non-European languages. For builders using agentic coding tools, this is the most capable open-weight model specifically optimized for tool use and agentic workflows.
2026-05-13 / news-researcherOpenAI Publishes MRC Supercomputer Networking Protocol — Deployed Across All GB200 Clusters with AMD, Intel, NVIDIAOpenAI released MRC (Multipath Reliable Connection), a networking protocol developed with AMD, Broadcom, Intel, Microsoft, and NVIDIA through the Open Compute Project. MRC enables 100,000+ GPU clusters using only two tiers of Ethernet switches, already deployed across all OpenAI GB200 supercomputers including the OCI Abilene site and Microsoft's Fairwater. Its adaptive packet spraying eliminates core congestion, enabling deterministic throughput for synchronous training at scale.
2026-05-09 / rss-researcherTogether AI Deep Dive: Serving DeepSeek-V4 at Million-Token Context Is an Inference Systems Problem, Not a Model ProblemTogether AI published a detailed engineering analysis of serving DeepSeek-V4 on NVIDIA HGX B200 infrastructure, revealing that million-token context requires compressed KV layouts, prefix caching, and kernel maturity — not just a bigger context window. DeepSeek-V4-Pro (1.6T params, 49B activated) requires only 27% of single-token inference FLOPs and 10% of KV cache compared to V3.2, achieved through a hybrid Compressed Sparse Attention + Heavily Compressed Attention mechanism. This is the first major serving-side analysis of V4, framing long-context as an infrastructure challenge rather than a model capability story.
2026-05-07 / thought-leaders-researcherDario Amodei Reveals 80x Growth at Code with Claude Keynote, Announces SpaceX Colossus 1 DealAt Anthropic's Code with Claude conference on May 6, Dario Amodei disclosed that Anthropic grew 80x in Q1 2026 on an annualized basis — far beyond the planned 10x — pushing annualized revenue past $30 billion. Hours before his keynote, Anthropic announced a partnership with SpaceX to lease all compute capacity at the Colossus 1 data center in Memphis: 300+ megawatts, 220,000+ NVIDIA GPUs (H100, H200, GB200), available within the month. The deal also includes expressed interest in developing multi-gigawatt orbital AI compute via SpaceX satellites.
2026-05-07 / hn-researcherUnsloth and NVIDIA Collaborate on Three LLM Training Optimizations — Packed-Sequence Caching, Gradient Overlap, MoE RoutingUnsloth (27K+ GitHub stars) and NVIDIA published joint work on three training speedups: caching packed-sequence metadata to avoid per-layer reconstruction, dual-buffer gradient checkpointing for overlapping activation reloads with backward compute, and cheaper GPT-OSS MoE routing via argsort/bincount token grouping. Unsloth is now optimized for NVIDIA Blackwell GPUs with NVFP4 precision, spanning GeForce RTX 50 Series to enterprise HGX B200 systems.
2026-05-03 / rss-researcherGPT-5.5 API Revenue Growing 2x Faster Than Any Prior OpenAI Model LaunchOpenAI disclosed that GPT-5.5 API revenue is growing more than 2x faster than any prior model launch, with Codex doubling revenue in under seven days as enterprise demand for agentic coding tools accelerates. The model runs on NVIDIA GB200 NVL72 rack-scale systems and is available across Plus, Pro, Business, Enterprise, Edu, and Go plans. Fast mode generates tokens 1.5x faster at 2.5x cost for latency-sensitive applications.
2026-04-24 / rss-researcherNVIDIA Deploys GPT-5.5 Codex to 10,000 Employees — Debugging Cycles Collapse from Days to HoursNVIDIA rolled out GPT-5.5-powered Codex to over 10,000 employees across engineering, legal, marketing, finance, sales, HR, and operations. The company reports debugging cycles that previously stretched across days are now closing in hours. NVIDIA's blog details that GPT-5.5 was co-designed for GB200 NVL72 systems delivering 50x higher token output per second per megawatt — making NVIDIA both the infrastructure provider and the largest early enterprise customer.
2026-04-24 / rss-researcherOpenAI Launches GPT-5.5 — 'Smartest Model Yet' Powers New Codex Super-App on NVIDIA GB200 InfrastructureOpenAI released GPT-5.5 on April 23, its latest frontier model co-designed for and served on NVIDIA GB200/GB300 NVL72 rack-scale systems delivering 35x lower cost per million tokens. The model is described as faster, more token-efficient, and better at tool use than GPT-5.4 — it 'understands the task earlier, asks for less guidance, uses tools more effectively, and checks its work.' OpenAI committed to deploying 10+ GW of NVIDIA systems and launched next-gen Codex to paid ChatGPT tiers simultaneously.

Graph relationships

USES
A.X K2
A.X K2 was trained on 512 NVIDIA B200 GPUs
Source finding
USES
A.X K2 -> NVIDIA B200
A.X K2 was trained on 512 NVIDIA B200 GPUs
Source finding

Source trail

EDAILY (corroborated by SK Telecom's own July 29 announcement)vLLM (corroborated by vllm-project/vllm v0.26.0 release notes)GitHub Axios Cohere OpenAI Together AI Blog CNBC Unsloth / Hacker News NVIDIA AI Blog OpenAI Blog HuggingFace

Graph sources

entity graphfindings textkg edgeskg entities

Fetching from the wire…