SemiAnalysis

Kimbo Chen × Wenyao Gao × Jordan Nanos

GLM-5.3’s sparse attention eases bandwidth pressure, not KV-cache capacity needs, while cheaper serving could lift HBM and bandwidth demand.AMD’s MI355 reportedly matched B200 cost at 150 tokens per second per user within weeks, underscoring software-driven cost volatility.Post-training reportedly lifted agentic benchmarks by more than 10%, though GLM-5.3 is not Mythos-class and technical disclosure remains limited.

Dan Nishball × Jordan Nanos

SemiAnalysis estimates $11.3 trillion in cumulative AI IT and data-center CapEx for 2024-29, implying roughly $7 trillion of debt by 2029 and $1 trillion of net funding this year.Hyperscaler and NVIDIA backstops, five-year offtakes, and frontier API gross margins near 90% could support financing, but lenders still face residual-value, churn, and execution risks.

Max Kan × Jordan Nanos × Andrew Megalaa

Anthropic’s $200 subscription delivered roughly $12,000 of API-equivalent usage in testing, but workload mix—not headline token counts—determines its economics.At roughly 96% cached reads priced at 0.1x or less of fresh input, model ownership may support 40–70% subscription gross margins, while OpenAI’s 50% limit cut and migration to open models remain risks.

Jordan Nanos × Abhilash Jain

SemiAnalysis Consulting is turning bespoke AI-infrastructure diligence into scalable products, including a token-supply simulator linking chip deployment, workload allocation, throughput, and token pricing.NeoCloud underwriting centers on contractual survivability: 99.9% availability expectations, sub-90% performance, and mismatched 15–20-year leases versus five-year GPU contracts can threaten financing and trigger remedies, while capacity timelines determine whether clients build or lease.

Jordan Nanos × Cam Quilici × Bryan Shan × Alec Ibarra

Engrams co-design model architecture and memory hierarchy, retrieving 2-gram and 3-gram embeddings from DRAM or SSD.Hybrid Engram-MoE layers outperform either endpoint, while AgentX traces show 100,000 input tokens at P50 and cache-hit rates above 98%.DRAM offload matters for hundreds of agents at 40–60 tokens/s; TPU V7's early Pareto position still leaves software, utilization, pricing, and demand unresolved.

Sam Harshe × Pratt Bhatt × Jordan Nanos

ClusterMAX 3.0 reshuffles 77 providers: Nebius joins CoreWeave in Platinum, Google Cloud joins Oracle in Gold, while Azure, AWS, Crusoe, and Together fall.Endpoint X finds 75% versus 99% cache hit rates can double an identical bill, while NVIDIA’s backstop universe forecast above $2T by end-2031 leaves SLA discipline, security, and hosted RL’s unproven market as risks.

Maya Barkin × Reyk Knühtsen × Jordan Nanos × Jeremie Eliahou Ontiveros

Mapping 400+ local instruments against 3,500 projects finds 300+ moratoriums but only three projects actively affected: AWS, NorthPoint, and CoreSite.Roughly 80% of live restrictions have no capacity behind them, leaving the forecast unchanged and making behind-the-meter power more attractive; New York delays roughly 800MW by 6–10 months, while an off-grid-blocking Texas-scale ban is key risk.

Jordan Nanos × Doug O'Laughlin × Max Kan × Joey Brookhart

“Pacing” would slow Anthropic’s capability progress without halting training or compute purchases, potentially weakening its strongest internal model.Near-term scarcity and safety workloads keep compute demand elevated, while semiconductor signals increasingly depend on frontier-lab ARR and capacity premiums.Bank hacks, data leaks, or AI-assisted biological attacks could accelerate regulation.

Myron Xie × Jordan Nanos

Rubin Ultra shifted from the GTC-previewed 1TB of HBM4E to 192GB of 8-high HBM4, below Blackwell Ultra’s and vanilla Rubin’s 288GB, as HBM supply rations TSMC-secured logic.Four-high saturates the interface at far lower cost per bandwidth and can yield roughly twice as many cubes as eight-high, but model-size growth is the key risk while memory tightness is not expected to ease within this decade.

Jordan Nanos × Eric Wen × Nigel Chiang × Nico Bontigui

Time to power, rather than cost, is driving modular adoption as compute deals reach $40M/MW and some Anthropic configurations exceed $100 million per megawatt.Factory parallelization cuts fit-out from up to nine months to three, shortening builds from 18–24 months to as little as 12; Level 5 commissioning can take 3–8 months, while module lead times reach 18 months.