Or OpenAI's blog https://t.co/rnSrr68wy0 And attend today's Jalapeño Hot Chips session with Richard, Ravi, and Chris from OpenAI too. (7/7)
Who's Who of AI
SemiAnalysis
Why they matter
Tracked through public AI activity and peer connections inside the directory.
- AI signals
- 10 past 30d
- Sources
- 4 distinct domains
- Discussions
- 0 past 30d
- Latest signal
- 6d ago
What they're sharing
openai.com
OpenAI’ Jalapeño: Better Than Nvidia Blackwell newsletter.semianalysis.com
Gemini is Cooked but GCP is Cooking newsletter.semianalysis.com
Are Open Models Catching Up? newsletter.semianalysis.com
AgentX - InferenceXv3: Does CUDA Moat Hold up in Agentic Inferencing? newsletter.semianalysis.com
TokenBudgeting: Our Conversations with Enterprises on Token Spend newsletter.semianalysis.com
AgentX Methodology and Datasets | InferenceX inferencex.semianalysis.com
Cerebras's Next Generation CS-4: Fast Just Got Faster newsletter.semianalysis.com
SemiAnalysis Core Research semianalysis.com
AI Networking Model semianalysis.com
AI Cloud TCO Model semianalysis.com
Meta Compute: Everyone Wants To Be A Neocloud newsletter.semianalysis.com
Articles & links
Read more at our article👇️ (6/7) https://t.co/dbuAUrRSt1
AI Weekly's analysis
→
Read full analysis →
Gemini is Cooked but GCP is Cooking GCP YoY rev growth >100%, DeepMind's long term failure is Google Cloud's short term gain https://t.co/PqVUCF8qSg
AI Weekly's analysis
→
Read full analysis →
Are Open Models Catching Up? Comparing open vs. closed models across the eras of frontier models, Is the gap narrowing? https://t.co/TBZnvquwio
AI Weekly's analysis
→
Read full analysis →
AgentX - InferenceXv3: Does CUDA Moat Hold up in Agentic Inferencing? $3 Million USD dataset open sourced, 1 Mil+ Context Length, Multiturn, Sub Agents 95%+ KVCache HitRate, GB300 NVL72, MI355, B200 https://t.co/PgtKNYXdFr
AI Weekly's analysis
→
- SemiAnalysis's AgentX 1.0, built on 393 anonymized Claude Code traces at 1M+ context, cost more than $3M and used ~2MW across 1000+ chips.
- On Qwen3.5 SGLang the report puts Nvidia at 'over 20x better performance' at 90 tok/s/user; B300 FP4 shows '12x better performance per dollar' vs H100.
- AMD's ATOM stack shows strong single-GPU kernels but almost no production adoption — only one Alibaba ad unit runs it live, the authors say.
Read full analysis →
TokenBudgeting: Our Conversations with Enterprises on Token Spend Was Widespread TokenMaxxing Ever Really Here? https://t.co/tgNssBBGVk
AI Weekly's analysis
→
- Meta employees consumed over 60 trillion tokens in a 30-day window in early 2026, with one individual alone accounting for about 280 billion.
- Monthly per-employee caps now range from $250 at an aerospace and defense manufacturer to $2,000 at Workday and Stripe, with no cross-industry consensus.
- Ramp data cited by SemiAnalysis shows 99th percentile customers spend about $90,000 per employee per year while the median customer spends $136.
Read full analysis →
https://t.co/kmFqrIAzF2 (2/2)
AI Weekly's analysis
→
- AgentX v1.0 turns 393 opt-in Claude Code sessions into a replay benchmark, with a median 142k input tokens and 444 output tokens per request.
- To share traces safely, inputs are reduced to session-scoped chained hashes in 64-token blocks that preserve matching prefixes but not content.
- SemiAnalysis says the benchmark's biggest first-months output was 50+ upstream pull requests from partners including vLLM, SGLang and TensorRT-LLM.
Read full analysis →
Cerebras's Next Generation CS-4: Fast Just Got Faster Double the Performance, Double the Power, Double the Fun https://t.co/qUgaBrNOYA
AI Weekly's analysis
→
- Cerebras CS-4 delivers more than 4,400 tokens per second per user on GPT-OSS-120B, up to 30 times a GPU inference baseline.
- Each rack pairs three WSE-3 Turbo wafers for 750 PFLOPS, 129.6 PB/s of memory bandwidth, and two-microsecond wafer-to-wafer latency.
- AMD Instinct and AWS Trainium handle prompt processing while CS-4 acts as the decode accelerator; first shipments begin this quarter.
Read full analysis →
We discuss the potential TPU beneficiaries in our Core Research Exclusive Note, "TPU Long and Prosper: Don't Cry Because Gemini's Over, Smile Because TPU Happened." (3/3) https://t.co/Onejdf8iMG
AI Weekly's analysis
→
- SemiAnalysis projects more than $100 million in 2026 revenue, up from roughly $20 million a year earlier, according to The Information.
- Core Research is one of four institutional products, aimed at hedge funds, long-only asset managers, venture investors and corporate strategy teams.
- The Substack newsletter reaches 200,000-plus subscribers but is sold separately from the institutional models and Core Research.
Read full analysis →
To learn more about ICI networking topology, please refer to our networking model (8/8) https://t.co/kqt0IiSiFi
AI Weekly's analysis
→
- SemiAnalysis is offering device-level tracking of AI cluster networking across five fabric layers, with data running 2023 to 2026.
- Coverage spans 80+ hyperscaler configuration panels for Microsoft, Google, Meta, Amazon, Oracle, X.AI and neoclouds, tied to specific accelerator SKUs.
- 25+ suppliers are tracked, including Nvidia, Arista, Broadcom, Cisco, Coherent and Lumentum, across 200G to 1.6T transceiver speeds.
Read full analysis →
To learn more, check out our Datacenter Model, which covers BitDeer, and our AI TCO Model, which covers the economics of NeoClouds. 6/6🧵 https://t.co/wACjpkQLM8 https://t.co/EeBYF4i2ve
AI Weekly's analysis
→
- SemiAnalysis's AI Cloud TCO Model covers Nvidia, AMD, Intel, and custom accelerators across cost of ownership per hour, inference cost per million tokens, and training cost per FLOP.
- The $/hr calculation is built from upfront server capex, system power consumption, colocation and electricity costs, and cost of capital, wrapped in a three-statement financial model.
- Detailed install base projections run through 2028 and vendor unit shipment estimates through 2034, aimed at operators, procurement teams, and equity and debt investors.
Read full analysis →
Meta Compute: Everyone Wants To Be A Cloud Zuck Takes Plan B? SpaceX 2.0, Bedrock 2.0, MSL Isn't Giving Up, Scaling RecSys by 10x... ClusterMAX ranking coming soon? https://t.co/RKHoW1Du89
AI Weekly's analysis
→
- Meta has contracted over 5GW of capacity across cloud and colocation, per SemiAnalysis, after nearly 10GW of deals since early 2024.
- Meta is reportedly in final talks with Anthropic to host private Claude instances, akin to Bedrock or Vertex from rival hyperscalers.
- SemiAnalysis pegs Meta's strategy as four-track: frontier training, 10x-plus ads recsys, Claude hosting, and SpaceX-style external rentals.
Read full analysis →