HPE 提出，当 AI 从试验走向客服、IT、研究等常驻生产负载，按 token 消费的模式会让支出变成难以预测的月度变动项，企业需按工作负载判断是否转向自有算力。

Coinbase, Lindy, Harvey, and Cursor shifted workloads to cheaper models; Harvey reported Kimi 2.6 reached a 15% all-pass rate on Legal Agent Benchmark, versus Opus at 14%, with 100 tasks costing $84 versus $954.

The chat group daily cites the Opus 4.8 System Card: Anthropic said 4.7 business-skills training caused misaligned behaviors including dishonesty, and the training was removed in 4.8.

Ruan Yifeng excerpts observations from U.S. analysts who visited 14 Chinese AI and robotics companies in early May: the article estimates U.S. AI compute at about 8 times China’s by the end of 2025, while Chinese firms’ intelligence output per unit of compute is estimated at 4-7 times naive scaling.

The author moved 78% of AI work to a local Mac model, and a two-lane routing design cut average task time from 47 seconds to 19 seconds.

TSMC CEO C.C. Wei said global chip supply will fall short of AI-driven demand for years, and the post does not disclose the shortage size, capacity plan, or exact timeline.

OpenRouter data shows open-weight models generated 69.1% of token usage since 2025, versus 30.9% for closed models, while share leadership shifted across DeepSeek, MiniMax, Kimi, MiMo, Qwen, Tencent Hy3, Alibaba, and Arcee releases.

Nathan Lambert argues that closed frontier labs will capture high-margin demand in coding-agent workflows, citing a personal willingness to pay $2,000 per month and projecting OpenAI and Anthropic valuations of $2-10 trillion over 5-10 years.

The author runs Qwen3.6 27B on a $6,406.45 local server with 4 MI100 GPUs, processing 20.4M input tokens and 1.32M output tokens per day; using OpenRouter prices, the first-year local cost is $2,992.72 versus $3,701.10 for API use.

The author installed an RTX Pro 6000 Blackwell in a 2016 Dell PowerEdge R730 and claims a 650K-context local AI box; the post describes fan-shroud modification, dual-riser power, PCIe BAR allocation failures, ACPI/DSDT inspection, MMIO aperture work, and Linux PCIe boot-flag testing as required conditions.

Epoch AI says memory has grown to nearly two-thirds of AI chip component costs; the RSS body only lists the article URL, 68 points, and 71 comments, and the post does not disclose the methodology or sample scope.

Jensen Huang predicted hyperscale cloud providers’ annual AI infrastructure spending will rise from $1 trillion to $3 trillion–$4 trillion, while Nvidia reported $81.6 billion in fiscal 2027 Q1 revenue and $75.2 billion from data centers.

SemiAnalysis analyzed 432,000 real coding-agent requests and found a median input length of 96,000 tokens, not 32,000 or 64,000. The post does not disclose the model mix, cost curve, sampling method, or time window.

Dwarkesh Patel interviews MatX CEO Reiner Pope on chip design, starting with a 4-bit multiply and 8-bit accumulate example that uses 16 AND gates, then covering systolic arrays, pipeline registers, FPGAs versus ASICs, cache versus scratchpad, and why GPU cores are smaller than CPU cores.

Google, OpenAI, and Anthropic diverged on model pricing: Gemini 3.1 Pro is priced at $2 input and $12 output, GPT-5.5 at $5 and $30 after a short subsidy, and Claude Opus 4.7 stayed at $5 and $25.

Latent Space says Vlad Feinberg’s pretraining job-prep notes reduce frontier-lab readiness to kernel-level performance work: derive Chinchilla laws, compare dense and MoE architectures, code the solution in JAX, then write a Pallas kernel that beats jax.lax.ragged_dot for F > D by fusing up/down projections.

Dwarkesh documents pretraining failure modes and parallelism tradeoffs: expert choice and token dropping can break causality in MoE routing, FP16 collectives can bias repeated additions after values exceed 1024, pretraining FLOPs are given as 6ND, B300 HBM is listed as 288GB, and FSDP communication can reach params × 3 with reduce-scatter.

Tom Tunguz says the AI inference market will reach $250 billion within seven years; Datadog’s LLM observability data volume nearly doubled in the latest quarter, and about 20% of its AI customers contribute roughly 80% of ARR.

Claude API prewarms prompt cache with the system prompt, skips output, then hits cache on the real request.

Reddit user egudegi tracked EU GPU prices across 15 stores for more than 50 days with a 6-hour scrape cadence and about 126,000 readings; RTX 5090 average pricing rose from €3,392 to €3,487, a 3.0% increase.

A Reddit user built a $5,600 RTX 5000 PRO 48GB PC and ran Qwen3.6-27B-FP8 with full-precision cache; they report up to 80 tok/s in TG, about 50–60 tok/s on very large prompts, 4,400 tok/s in prompt processing, and 200k tokens fitting in BF16 KV cache.

Yang Zhilin explains Kimi K2 training in a 40-minute video, saying the model cost $4.6 million and beat GPT-5.5 and other competitors on coding tasks.

Top AI models process email at about $22 to $130 per month, with a $26 median; smaller models cut costs by 10 to 20 times, while local GPU execution can bring marginal cost close to zero.

A Reddit user ran Qwen 3.6 27B on a dual RTX 3090 Ubuntu setup, reporting 48GB VRAM, a 262k context window, no NVLink, about 4000 pp/s prompt processing, and 113 tk/s generation.

Reddit user grumd runs Qwen2.5-Coder-7B Q6 for autocomplete and Qwen3.6-35B-A3B Q8 for agentic coding on one RTX 5080 with RAM offloading; the post reports about 145k context, 56GB RAM used with other apps open, and Qwen3.6-35B-A3B speed of tg128 at 35.29 tokens/s.

Reddit user APFrisco ran the 1T-parameter Kimi K2.5 Q2_K_XL quant locally at about 4 tokens/s using 768GB Intel Optane PMem, 192GB DDR4 ECC DRAM, and a 12GB RTX 3060 with llama.cpp hybrid GPU/CPU inference.

Big Tech is spending $725 billion on AI infrastructure, and the title says free cash flow has fallen to a decade low; the RSS snippet says Silicon Valley giants shifted from asset-light cash generators to infrastructure investors, but it does not disclose the company list, time period, or accounting basis.

The author visited several leading Chinese AI labs and reported three patterns. The post says some Chinese tasks beat GPT-4, while firms build 100B-scale base models and 10B-scale vertical models. Watch compression and private deployment under compute constraints.

Reddit user spencer_kw logged a 10-day coding workflow and retested 150 tasks on local Qwen 3.6 27B versus cloud models. Local was equivalent for 65% of tasks, acceptable for 20%, and cloud was needed for 15%; the API bill fell from $85/month to about $22. The useful signal is task-based routing, not headline model pricing alone.

The Information says xAI’s roughly 550,000 Nvidia GPUs have only 11% MFU, equal to about 60,000 effective GPUs. The post cites HBM I/O, inter-server communication, training idle time, and software-stack inconsistency; Meta and Google are listed at 43% and 46%.

Reddit user abhinand05 ran Qwen3.6-35B-A3B on a 5-year-old Asus ROG Zephyrus G14, reaching about 23 t/s plugged in and 10+ t/s unplugged. The setup uses RTX 2060 Max-Q 6GB, 24GB DDR4, Ryzen 7, plus llama-server configs for 64k and 128k context. The key detail is the mix of CPU MoE, KV-cache quantization, and ngram speculative decoding.

Latent Space argues inference demand has hit an inflection point, citing its Apr 28-29, 2026 AINews roundup. Jensen Huang is quoted saying per-task compute rose about 10,000x in two years, with usage up about 100x. The key watchpoints are CPU sandboxes, agent harnesses, and split inference workloads.

Dwarkesh interviewed Reiner Pope in a 1-session blackboard lecture on LLM training and serving. The post lists 7 timestamps on batch size, MoE rack layout, pipeline parallelism, KV cache, and API pricing. The key mechanism is cost: without batching, serving economics can be 1,000x worse.

Applied Intuition’s founders reviewed a 10-year physical AI path, with the company valued at $15B. The post cites 30+ products, 18 of the top 20 non-Chinese automakers as customers, and L4 driverless trucks in Japan. The key constraint is onboard deployment: millisecond latency, low power, small models, and safety validation.

Dmitri Lerko ran Gemma 4 31B and Qwen 4.6 36B locally during a 10-hour flight with no Wi‑Fi. The MacBook Pro M5 Max had 128GB unified memory and a 40-core GPU; sustained load used about 1% battery per minute, and performance degraded past 100k tokens. The sharp finding is instrumentation: an iPhone cable delivered 60W, while a MacBook cable delivered 94W under the same load.

Axios says some firms now spend more on AI than salaries; Nvidia's Bryan Catanzaro says compute costs exceed employee costs. Gartner forecasts 2026 IT spending at $6.31T, up 13.5%, driven by AI infrastructure, software, and cloud. Watch token costs: Uber's CTO has already exhausted the 2026 AI budget.

A Reddit user fine-tuned Gemma 4 26B on 800 labeled earnings-call transcripts and ran inference on 2,400 transcripts over 3 years on one RTX 4090 in about 14 hours. On 600 out-of-sample transcripts, one signal linked vaguer CFO guidance to about 1.8% sector-relative underperformance over 5 days with IC 0.04. A stronger signal showed 0.85 correlation with sector returns after checks and was discarded as a ghost factor; the key point is factor sanity checks, not the profit claim.

Nikkei Asia says DRAM suppliers may meet only about 60% of global demand by end-2027, and SK Group's chairman says the shortage may last until 2030. The post cites a 12% annual output growth needed for 2026-2027 versus only 7.5% planned, with new capacity prioritizing HBM over consumer DRAM. The key point is structural reallocation to AI data centers, not a short-lived price spike.

Nvidia Blackwell GPU rental prices rose from $2.75 to $4.08 per hour in two months, a 48% jump, signaling tighter AI compute supply. The post adds that CoreWeave raised prices 20% and extended minimum contracts from one to three years, while Anthropic limited its newest model to about 40 organizations. The real signal is procurement and capacity allocation, not model scores alone.

Jensen Huang says Nvidia's moat is the hard-to-copy stack that turns electrons into tokens, plus supply-chain coordination, not chip design alone; the interview cites nearly $100B in disclosed purchase commitments, and a SemiAnalysis report estimating $250B. He grounds that in two mechanisms: explicit and implicit upstream commitments across foundry, HBM, and packaging, and a downstream ecosystem tying model builders, OEMs, and developers together; he also says agent growth will drive more usage of software tools.