Microsoft and Anthropic control 85% of Nscale's $103B data center backlog

SemiAnalysis maps MoE inference architecture across hardware layers, revealing KV cache and orchestration as critical differentiators

A comprehensive technical deep dive from SemiAnalysis details how Mixture of Experts models reshape inference infrastructure across prefill, midfill, and decode stages, with concrete implications for GPU clustering, memory hierarchy (HBM vs. DDR vs. NVMe storage), and networking topology. The analysis distinguishes four operating regimes with fundamentally different compute and memory demands: prefill reaches high arithmetic intensity but requires no prior context; midfill extends cached context with moderate intensity; decode attention and decode experts operate at low intensity with heavy data movement. The piece documents how KV cache management—spanning immutable blob-based storage systems and multi-tier memory (HBM → CPU DRAM → NVMe)—has emerged as a competitive advantage in long-context agentic workloads, where conversations can span thousands of turns across hours or days. This directly maps to infrastructure design decisions at companies competing on inference efficiency and cost-per-token economics.

Straight-line GPU depreciation artificially inflates hyperscaler margins during high-growth capex cycles

H100s generate approximately half their lifetime cash flow in the first two years but are depreciated evenly across six years, creating a timing mismatch between actual cash economics and reported profitability. At 50% annual fleet growth, this accounting effect adds roughly 10 percentage points of margin with zero change in underlying economics. This directly impacts how to interpret reported profitability at Nvidia's largest customers and becomes immaterial only when capex growth slows.

Warp's Wilson factory ships 2,000 PRs monthly with humans as primary bottleneck, measuring success by human interactions per PR

Warp CEO Zach Lloyd demonstrated how their agentic software factory (Wilson) moves from Slack request to merged pull request in 35 minutes on average, with automated QA verification via video recording now standard. The critical metric is "human interactions per PR"—counting every re-prompt, comment, or correction required—revealing that humans in code review are now the primary bottleneck rather than code generation. The factory replays real historical jobs against different model configurations on Pareto charts to optimize cost-quality tradeoffs, and uses failed runs to automatically propose code changes to the factory's own agent definitions. This represents a shift from standalone coding copilots to integrated, measurable systems that treat the entire development lifecycle as orchestrated and optimizable.

Chinese labs ship frontier-scale MoEs at 10-week cadence; Xiaomi's MiMo-V2.6-Pro claims top open-weights ranking at $2.6M training cost

A sustained wave of major releases from Chinese teams over 10 weeks—including Kimi K3 (2.8T), Qwen3.8-Max (2.4T), DeepSeek V4-Pro (~1.6T), Hy4 (770B), GLM-5.3 (~753B), and Atria Dawn (744B)—demonstrates frontier-scale capability shipping at open weights. Xiaomi's MiMo-V2.6-Pro (1.02T total / 42B active parameters) ranks as the top open-weights model on Artificial Analysis' Intelligence Index (46) with strong cost efficiency at $0.13 per task. The model is natively omnimodal and trained with reinforcement learning in just 130 hours for $2.6M, suggesting post-training has become a dramatically cheaper path to frontier-adjacent capability than previously assumed. This cadence and cost structure directly challenge the closed-model moat and signal a structural shift in competitive dynamics, with startups increasingly building custom models on open weights to reduce dependence on OpenAI/Anthropic at 10–50x lower cost.

Amazon blocks Meta's Muse from shopping on its platform; Shopify embraces integration in critical e-commerce agent battleground

Amazon is preventing Meta's new AI agent from accessing Amazon.com to complete purchases on users' behalf, citing terms of service violations and security risks, while also filing an amended complaint against Perplexity's Comet browser over credential handling. The move signals major e-commerce platforms are willing to restrict agent access to protect customer relationships and data, even as Shopify embraces Muse integration—a critical competitive dynamic for how AI agents reshape retail infrastructure and platform control.

TypeSafe AI launches Jev, a "System One" model class optimized for structured decisions and software integration over chat

Diogo Almeida's TypeSafe AI released Jev, a new model class built for fast, structured decisions that software can consume directly, prioritizing calibrated decision-making and "intelligence per dollar" over RLHF-style human preference alignment. Jev achieves similar intelligence on System One tasks compared to existing LLMs while being two orders of magnitude faster and more efficient, with input tokens at $0.042/MTok and output tokens free. The model outputs all probabilities in parallel rather than autoregressively generating by token, eliminating hallucination and type errors—table stakes for automation. The launch video reached ~40M views and drove 100K+ Discord community members, representing a fundamental shift in how frontier models are positioned: away from general-purpose assistants toward programmable, machine-native systems designed for code consumption and automation workflows.

H100 rental prices reverse sharply upward after 60% decline; CoreWeave reports capacity sold out across GPU fleet

Spot rates for H100 GPUs fell from $8+/hour in 2023 to $1.70/hour by October 2025 but rebounded to $2.75 by September 2026. CoreWeave reported capacity sold out and rising prices across H100, H200, and A100 fleets in Q1 2026. The underlying asset depreciates slowly; what decays is the rental premium, which is now recovering as demand outpaces supply.

Nscale's $103 Billion in Contracts: Microsoft and Anthropic Dominate British Data Center Developer's IPO Disclosure

Bloomberg reported on Nscale's IPO filing disclosing $103 billion in total contract value, with concrete customer concentration metrics. Main takeaways:

• Microsoft and Anthropic represent 85% of Nscale's $103B contract backlog, reflecting extreme customer concentration in frontier AI infrastructure
• Nscale's contracts grew 1,000x over three years, driven almost entirely by two AI companies' compute demand
• Anthropic's deal with Nscale remains unfunded, creating execution risk for the British data center developer's growth plans

Tokenizers v1: Performance Improvements Across Encoding, Caching, and Multi-Language Support

Hugging Face published a technical report on tokenizers v1, a major performance upgrade to its widely-used tokenization library. Main takeaways:

• Tokenizers v1 achieves 3-30x faster encoding than v0.23 on single-threaded benchmarks (M4 Max), with particularly large gains on non-Latin languages (14.6x on Amharic, 12.5x on Hindi, 11.9x on Bengali)
• Time-to-first-token latency reduced from 110µs to 17µs (median p99 on 512-byte English documents), directly impacting inference performance for long-context agent workloads
• Word cache optimization eliminates redundant merge computation for repeated pre-tokens, with cache hits saving 34% of uncached work on typical corpora
• Bitcannon SIMD-based splitting replaces regex engines for BPE models, processing 64 bytes per register operation instead of character-by-character scanning
• Scales to 76% linear efficiency across 8 worker threads; maintains identical token ID output to v0.23 for backward compatibility

Building Eval Systems That Improve Your AI Product

Lenny's Newsletter published an essay on systematic evaluation methodology for AI products. Main takeaways:

• Error analysis via open coding (free-form critique) and axial coding (pattern grouping) should precede metric selection; most teams fail by measuring fashionable metrics like hallucination without understanding actual product failures
• Binary pass/fail judgments with detailed critiques outperform Likert scales; nuance is captured in explanations, not rating granularity
• LLM-as-a-judge requires ground truth calibration against human-labeled data; trust is established by measuring judge agreement with domain experts on held-out test sets
• Code-based evaluators handle objective, rule-based failures; LLM judges handle subjective judgment; choosing the right tool per failure mode is critical
• Operationalization requires continuous improvement flywheel: error analysis → evaluation suite → production monitoring → new error discovery

Notes from Inside China's AI Labs

Nathan Lambert in Interconnects published an essay titled "Notes from Inside China's AI Labs". Main takeaways:

• Chinese labs benefit from student-heavy teams with fresh perspectives on modern techniques, less ego-driven organizational politics, and willingness to do non-flashy work to maximize final model quality—contrasting with US labs' emphasis on individual researcher visibility and career advancement.
• Most Chinese AI developers actively use Claude despite nominal bans, suggesting strong latent demand for inference services that contradicts assumptions about limited Chinese software spending; demand may track cloud markets rather than SaaS.
• Chinese AI industry lacks a centralized master plan; it reflects practical equilibrium across multiple companies (ByteDance, Alibaba, DeepSeek, Meituan, Ant Group) with DeepSeek respected as technical leader but not positioned as economic winner.
• Chinese researchers show less engagement with philosophical debates on AI economics, safety, and societal impact compared to US counterparts, reflecting different educational and incentive systems oriented toward engineering execution.

Kraken Parent Plans Permissioned Hyperliquid Perps for US Traders

The Defiant reported on Payward's plan to offer U.S. clients perpetual futures on Hyperliquid through a permissioned structure. Main takeaways:

• Payward will deploy HIP-3* markets on Hyperliquid with dual allowlists (NinjaTrader and Bitnomial) restricting access to approved U.S. accounts
• Bitnomial (CFTC-regulated exchange/clearinghouse) deploys and clears markets; NinjaTrader Clearing (FCM) carries customer accounts; trades settle onchain on Hyperliquid
• Structure separates regulatory obligations (with named entities) from order book/settlement (on Hyperliquid blockchain), creating a regulated on-ramp to builder-deployed markets
• Payward identifies Hyperliquid as first protocol for U.S. client onchain perpetual futures, subject to regulatory approval