# LLM Cost-Cutting Insights: Week of Nov 18-25 Newsy: LLM Cost Tracker Published: 2025-11-25T21:58:00.343Z Source: https://newsyai.com/edition/0f2ba478-c94e-4ee0-aea7-d490d1dc07d1 Archive: https://newsyai.com/newsy/7e294c66-a84b-45ba-b656-6f55351953db.md --- # The LLM Cost‑Cutter Weekly briefing on advances that shrink the cost of large language models (Nov 18–25, 2025) ## Pricing announcements from AI labs this week - No public token‑price cuts announced by OpenAI, Anthropic, Google, Meta, Mistral, or Cohere in the last 7 days. OpenAI’s pricing page and Google’s Vertex AI pricing show no week‑over‑week per‑token changes. ([platform.openai.com](https://platform.openai.com/docs/pricing/?utm_source=openai)) - New model news: Anthropic introduced Claude Opus 4.5 today; no official per‑token price card was posted alongside the launch. Availability is via Anthropic apps, API, and major clouds. ([timesofindia.indiatimes.com](https://timesofindia.indiatimes.com/technology/tech-news/anthropic-introduces-claude-opus-4-5-says-its-their-smartest-ai-model-yet/articleshow/125564172.cms?utm_source=openai)) — — — ## 1) Transparent GPU prices keep drifting down at the “good‑enough” tier - What’s new: CoreWeave’s public price cards (refreshed this month) still show clear rates, including RTX Pro 6000 Blackwell Server Edition nodes listed at $20/hr and B200/H100/H200 multi‑GPU servers with posted hourly pricing. SaladCloud continues to advertise ultra‑low hobbyist‑to‑startup prices (e.g., RTX 5090 ≈ $0.25/hr, RTX 4090 ≈ $0.16–$0.18/hr), and trade press corroborated 5090 rentals at $0.25/hr in November. Plain English: “mid‑tier” and consumer GPUs are cheap and easy to comparison‑shop. ([coreweave.com](https://www.coreweave.com/pricing?utm_source=openai)) - Why it matters: Matching model size/SLA to the cheapest viable silicon lowers dollars per request without hurting quality for non‑frontier workloads. These savings compound with software serving gains below. ## 2) Grid rules are changing to plug in big AI loads faster - What’s new this week (Nov 20): FERC held its monthly meeting and highlighted active dockets; in parallel, firms briefed clients on DOE’s October 23 directive pushing FERC to standardize and speed up interconnections for “large loads” (20+ MW) like AI data centers, with comment windows landing through mid‑November and activity continuing into this week. Translation: federal regulators are moving to make it easier and faster for huge data centers to get connected. ([ferc.gov](https://www.ferc.gov/news-events/news/summaries-november-2025-commission-meeting?utm_source=openai)) - State action: Missouri formally approved a large‑load rate plan on Nov 24, and Kansas adopted a ≥75 MW tariff earlier this month. These plans require long‑term contracts, minimum bills, and “pay your share” rules. Plain English: states are setting predictable prices and terms for giant power users. ([psc.mo.gov](https://psc.mo.gov/Electric/PSC_Approves_Ameren_Missouri_Large_Load_Power_Rate_Plan_with_Customer_Protections--pr-26-40?utm_source=openai)) - Why it matters: Faster, clearer hookups and “cost‑causer pays” tariffs reduce delays, cut uncertainty, and let operators plan multi‑year capacity at known $/kWh—directly lowering total cost to serve tokens and improving “intelligence per watt” via higher utilization. ## 3) Power and cooling suppliers lock in multi‑billion‑dollar AI data center orders - What’s new (Nov 19): Schneider Electric announced $2.3B of U.S. data‑center deals (power modules, cooling, UPS/switchgear), including a $1.9B agreement with Switch. Plain English: the supply chain for power and cooling is scaling quickly. ([reuters.com](https://www.reuters.com/legal/transactional/schneider-electric-seals-23-billion-us-data-centre-deals-power-ai-boom-2025-11-19/?utm_source=openai)) - Why it matters: More factory‑built power and cooling gear shortens build times and reduces facility energy waste (better PUE). That lowers operating costs per token and helps clusters sustain high throughput without thermal throttling. ## 4) Fresh research to trim inference energy and time - Online Smoothed Demand Management (Nov 23): An algorithms paper models how big compute campuses with batteries should buy and use electricity smoothly over time, balancing grid‑friendly behavior against cost. Plain English: smarter scheduling of power + batteries reduces peak charges and keeps GPUs busy. ([arxiv.org](https://arxiv.org/abs/2511.18554?utm_source=openai)) - Decentralized Speculative Decoding (Nov 13; still new and being tested): Turns cross‑node network waiting time into useful work; reported up to ~2.6× speedups without retraining. In practice this means fewer GPU‑hours per answer. ([arxiv.org](https://arxiv.org/abs/2511.11733?utm_source=openai)) - FlexiCache (Nov 2; circulating among practitioners): A vLLM‑based method that classifies attention heads and offloads stable KV pages to host memory, cutting GPU memory up to 70% and improving throughput/latency on long contexts. Plain English: fits longer prompts on the same GPUs and serves more users per rack. ([arxiv.org](https://arxiv.org/abs/2511.00868?utm_source=openai)) - Why it matters: These are software flips—no new chips required—that reduce energy per token and dollars per request. Combined with cheap GPUs and better tariffs, the cost curve bends down further. ## 5) Runtime and hardware stack updates worth noting - NVIDIA ecosystem: Current docs emphasize FP4/FP8 support in TensorRT‑LLM for Blackwell, aimed at higher throughput per watt; NVIDIA also highlighted Blackwell’s MLPerf v5.1 training wins on Nov 12 (outside the 7‑day window but relevant to near‑term fleet efficiency). Plain English: newer NVIDIA stacks squeeze more tokens per second from the same electricity. ([nvidia.github.io](https://nvidia.github.io/TensorRT-LLM/overview.html?utm_source=openai)) - AWS Neuron: Recent Neuron SDK releases (2.25–2.26) add inference and Trainium2 optimizations and support for new models. Translation: non‑GPU options are improving, potentially undercutting $/token for certain workloads. ([aws.amazon.com](https://aws.amazon.com/about-aws/whats-new/2025/08/aws-neuron-2-25-announce/?utm_source=openai)) - Why it matters: As compilers and kernels mature, you get higher throughput on existing hardware—freeing you to target the lowest‑cost instances that meet your SLA. ## 6) Site development signals: more land, on‑site power, and expansion paths - Texas/New Mexico expansion (today): A West Texas AI campus near Odessa expanded to 438 acres; the developer also optioned 3,500 acres in New Mexico targeting 2+ GW of gas now and multi‑GW nuclear later. Plain English: builders are securing land and firm power to avoid grid bottlenecks. ([mrt.com](https://www.mrt.com/business/article/ector-county-ai-data-center-21207962.php?utm_source=openai)) - Why it matters: On‑site or dedicated power plus big land banks speed time‑to‑compute and lower delivered energy costs, translating to cheaper inference at scale. ## 7) Data‑center thermal efficiency: bigger deals and new approaches - Cooling capacity at scale: Vertiv and Schneider are prominently pushing liquid‑cooling portfolios (SC25 presence; recent M&A; large contracts). Plain English: liquid cooling is becoming standard for dense AI racks, cutting fan power and improving PUE. ([vertiv.com](https://www.vertiv.com/en-us/about/news-and-insights/events/sc25-supercomputing/?utm_source=openai)) - Lab‑to‑fab ideas: Recent research into membrane evaporative cooling reports record heat‑flux handling, pointing to future options for ultra‑hot GPUs. Plain English: better cooling lets you run hotter chips efficiently, raising “intelligence per watt.” ([tomshardware.com](https://www.tomshardware.com/pc-components/cooling/membrane-evaporative-cooling-technology-achieves-record-breaking-results-could-be-a-solution-for-next-generation-ai-server-cooling-clocks-800-watts-of-heat-flux-per-square-centimeter?utm_source=openai)) — — — ## What this means for your unit economics this week 1) Treat posted “Pro” Blackwell and consumer‑class rentals as your floor for non‑frontier jobs. Start pricing with RTX 4090/5090 at ~$0.16–$0.25/hr for evals/batch inference, and step up to RTX Pro 6000 Blackwell or H100/H200/B200 only when memory, latency, or SLA require it. Expect immediate $/request savings. ([blog.salad.com](https://blog.salad.com/lowest-cost-gpus/?utm_source=openai)) 2) Model your sites under the new large‑load tariffs. Missouri’s Nov 24 approval and Kansas’ plan offer predictability (long terms, minimum bills). Predictable tariffs + expedited interconnection proposals at FERC reduce delays and financing risk, which lowers total $/token over the project life. ([psc.mo.gov](https://psc.mo.gov/Electric/PSC_Approves_Ameren_Missouri_Large_Load_Power_Rate_Plan_with_Customer_Protections--pr-26-40?utm_source=openai)) 3) Pilot KV‑cache management and speculative decoding on your longest prompts and multi‑node paths. Target workloads with long context and distributed inference first; these see the biggest tokens‑per‑watt gains. ([arxiv.org](https://arxiv.org/abs/2511.00868?utm_source=openai)) 4) Ask colocation and cloud partners about liquid‑cooling readiness and PUE commitments tied to your racks. Big OEM orders (power/cooling) signal faster installs and better energy efficiency—directly lowering operating costs. ([reuters.com](https://www.reuters.com/legal/transactional/schneider-electric-seals-23-billion-us-data-centre-deals-power-ai-boom-2025-11-19/?utm_source=openai)) 5) If you have batteries on site (or plan to), borrow ideas from the new “smoothed demand” algorithms: flatten demand peaks to cut demand charges and keep clusters fully utilized. That increases tokens per joule and per dollar. ([arxiv.org](https://arxiv.org/abs/2511.18554?utm_source=openai)) — — — Links you can keep open on a second screen - OpenAI pricing (reference, unchanged this week): https://platform.openai.com/docs/pricing ([platform.openai.com](https://platform.openai.com/docs/pricing/?utm_source=openai)) - Google Vertex AI pricing (reference, unchanged this week): https://cloud.google.com/vertex-ai/generative-ai/pricing ([cloud.google.com](https://cloud.google.com/vertex-ai/generative-ai/pricing?utm_source=openai)) - CoreWeave public pricing: https://www.coreweave.com/pricing and instance docs ([coreweave.com](https://www.coreweave.com/pricing?utm_source=openai)) If you want, next edition can add a live tracker of posted RTX/B‑class rental rates, state large‑load tariffs, and fresh serving papers that move tokens/sec or watts/token.