# Cutting Costs: LLM Pricing Updates This Week Newsy: LLM Cost Tracker Published: 2026-02-17T18:00:00.288Z Source: https://newsyai.com/edition/269d27e3-a4b3-4733-84f8-5cfc193b7e06 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 (Feb 11–Feb 17, 2026) ## Pricing announcements from AI labs this week - No new per‑token price cuts were announced by OpenAI, Anthropic, Google, or Mistral in the last 7 days. Current public price cards still show: - OpenAI GPT‑5.2 at $1.75/M input, $14/M output; GPT‑5 mini at $0.25/M input, $2/M output. ([openai.com](https://openai.com/api/pricing//?utm_source=openai)) - Anthropic Claude Opus 4.6 at $5/M input, $25/M output (1M‑token context in beta on Anthropic’s platform). ([platform.claude.com](https://platform.claude.com/docs/en/about-claude/pricing?utm_source=openai)) - DeepSeek V3.2 at $0.28/M input (cache miss), $0.028/M cached input, $0.42/M output. ([api-docs.deepseek.com](https://api-docs.deepseek.com/quick_start/pricing/?utm_source=openai)) - Google’s Vertex AI Agent Engine continues to bill runtime and new add‑ons (Code Execution, Sessions, Memory Bank) per its January 28 change; no new changes this week. ([cloud.google.com](https://cloud.google.com/vertex-ai/pricing?utm_source=openai)) — ## 1) Utility deal locks in 1.4 GW for Oracle’s Michigan AI campus - What’s new: DTE Energy won conditional approval to supply 1.4 GW to Oracle’s Saline Township data center under a structure that prevents shifting costs onto other ratepayers. The agreement begins immediately and runs through February 2045. Plain English: Oracle gets predictable, dedicated power; existing DTE customers aren’t subsidizing it. ([barrons.com](https://www.barrons.com/articles/dte-energy-oracle-ai-earnings-stock-price-89fc7c8c?utm_source=openai)) - Why it matters for LLM costs: Firm, long‑dated power lowers volatility and curtailment risk. That steadier electricity input helps keep “dollars per million tokens” predictable and improves “intelligence per watt” by reducing idle GPUs and retry losses. ## 2) PJM’s co‑location tariff deadlines hit this week (policy that can cut soft costs) - What’s new: FERC required PJM to file tariff revisions by February 16, 2026, creating clearer service options for data centers co‑located with generation (including new Firm Contract Demand and Non‑Firm Contract Demand services) and setting a paper hearing schedule. Plain English: PJM must formalize pathways for “power next to compute,” with flexible firm/interruptible transmission choices. ([ferc.gov](https://www.ferc.gov/news-events/news/fact-sheet-ferc-directs-nations-largest-grid-operator-create-new-rules-embrace?utm_source=openai)) - Why it matters for LLM costs: Clear rules shorten interconnection timelines and let operators buy only the firmness they truly need—shrinking interconnect, wheeling, and delay costs that otherwise inflate the all‑in price per token. ## 3) Hardware efficiency: Blackwell’s “throughput per megawatt” story keeps strengthening - What’s new: Nvidia continues emphasizing Blackwell‑generation gains, citing analysis that its systems deliver up to 50× higher inference throughput per megawatt than Hopper, with further software‑driven leaps for mixture‑of‑experts (MoE) serving. Plain English: same power, way more answers per second. ([barrons.com](https://www.barrons.com/articles/nvidia-stock-price-ai-chips-082c7640?utm_source=openai)) - Why it matters for LLM costs: If your data center power is capped, higher tokens/sec per MW means lower $/M tokens and better “intelligence per watt.” You can defer new buildouts and still scale traffic. ## 4) GPU cloud “street prices” you can use in negotiations - What’s new: Lambda’s public card shows B200 instances at $5.74/GPU‑hr on demand, plus 1‑Click B200 clusters “as low as” $3.79/GPU‑hr with short commitments. Plain English: transparent listings help you benchmark quotes and push vendors. ([lambda.ai](https://lambda.ai/pricing?utm_source=openai)) - Why it matters for LLM costs: Public reference rates compress margins and help you mix premium GPUs for training with cheaper fleets for inference—lowering effective $/M tokens even when API prices don’t move. ## 5) Cooling moves that raise rack density and trim power overhead - What’s new: Infosys will deploy ExxonMobil immersion fluids to improve data‑center energy efficiency. Plain English: liquid immersion reduces the electricity wasted on cooling, so more power goes to compute. ([m.economictimes.com](https://m.economictimes.com/tech/information-tech/infosys-to-deploy-exxonmobils-data-centre-fluids-for-efficient-cooling-of-ai-infra/articleshow/128256286.cms?utm_source=openai)) - Why it matters for LLM costs: Lower cooling overhead (PUE closer to 1.0) directly cuts your energy line item per token and enables denser racks—improving “intelligence per watt” and per square foot. ## 6) Global build‑out: India’s Adani pledges $100B for AI data centers by 2035 - What’s new: Adani Group announced plans to expand AI‑focused data‑center capacity to 5 GW, leaning on large renewable projects and partnerships. Plain English: a massive supply‑side push that could make compute cheaper in India over time. ([wsj.com](https://www.wsj.com/tech/indias-adani-group-to-invest-100-billion-on-ai-infrastructure-7d428962?utm_source=openai)) - Why it matters for LLM costs: More regional capacity + renewable PPAs can push down training/serving prices in those markets and create pricing pressure globally, especially for latency‑tolerant jobs. ## 7) Fresh research that cuts memory traffic and boosts tokens/sec - DeltaKV (Feb 8): Residual‑based KV‑cache compression cuts memory to ~29% of original with near‑lossless quality; with a Sparse‑vLLM runtime it showed up to 2× throughput vs. vLLM in long‑context tests. Plain English: store “differences” instead of full history so you fit more requests per GPU. ([arxiv.org](https://arxiv.org/abs/2602.08005?utm_source=openai)) - Cachemir (Feb 12): Integrates KV‑caching into fully homomorphic encrypted (FHE) inference, reporting up to ~67× CPU speedups over prior art and sub‑100s/token on GPU for Llama‑3‑8B. Plain English: private‑by‑design inference gets closer to usable speeds—important if privacy rules would otherwise force slower setups. ([arxiv.org](https://arxiv.org/abs/2602.11470?utm_source=openai)) - 3D‑FlashAttention (Feb 11): A 3D‑stacked NPU co‑design that shifts work from energy‑hungry on‑chip SRAM to register‑level flows, claiming 46–93% energy cuts and up to 7.6× speedups on long sequences. Plain English: specialized chips that move data shorter distances waste less energy per token. ([arxiv.org](https://arxiv.org/abs/2602.11016?utm_source=openai)) - Why it matters for LLM costs: KV‑cache traffic is a major hidden cost. Better compression/eviction and smarter attention drastically reduce memory bandwidth and energy, lifting tokens/sec and lowering $/M tokens without new GPUs. — ## Quick takes - Nvidia earnings on Feb 25 may catalyze more formal throughput/cost disclosures for Blackwell and Rubin; for now, the company is signaling large inference‑per‑MW gains. Track software stack releases (TensorRT‑LLM, NVFP4) that convert those claims into your real tokens/sec. ([barrons.com](https://www.barrons.com/articles/nvidia-stock-price-ai-chips-082c7640?utm_source=openai)) - DOE’s emergency authority for PJM expired Feb 2; rehearing granted Feb 4. Expect continued reliability and interconnection rulemaking that affects AI campuses’ timelines and power pricing. ([energy.gov](https://www.energy.gov/ceser/federal-power-act-section-202c-pjm-interconnection-pjm-order-no-202-26-02?utm_source=openai)) — ## This week’s impact in plain English - Even without headline API price cuts, total cost per answer keeps falling as utilities ink firm‑power deals, grid rules speed co‑location, GPUs push more tokens per watt, serving software improves, and cooling wastes less power. These stack: steadier power + faster GPUs + leaner KV‑caches + better cooling = more “intelligence per dollar” and “per watt.” — ## Action checklist (1–2 hour tasks) 1) Benchmark a KV‑compression method (e.g., DeltaKV) on your longest prompts; track VRAM use, latency, and any quality drift. Aim to raise concurrency per GPU. ([arxiv.org](https://arxiv.org/abs/2602.08005?utm_source=openai)) 2) Re‑shop inference fleets using Lambda’s public B200 rates as a negotiation anchor; pressure for sub‑$4/GPU‑hr on committed MoE serving. ([lambda.ai](https://lambda.ai/pricing?utm_source=openai)) 3) If you operate in PJM territory, ask your utility/consultant about the new Firm/Non‑Firm Contract Demand services and where your campus fits; map potential savings from co‑location. ([ferc.gov](https://www.ferc.gov/news-events/news/fact-sheet-ferc-directs-nations-largest-grid-operator-create-new-rules-embrace?utm_source=openai)) 4) For privacy‑sensitive workloads, pilot encrypted inference on a small task to see if Cachemir‑style acceleration changes your deployment calculus. ([arxiv.org](https://arxiv.org/abs/2602.11470?utm_source=openai)) Sources: OpenAI pricing; Anthropic pricing and Opus 4.6 note; DeepSeek pricing; Google Vertex AI pricing and blog; DTE‑Oracle utility deal (Barron’s); FERC/PJM filings and law‑firm briefs; Nvidia blog; research on KV‑cache and attention; WSJ on Adani AI infra; Infosys–ExxonMobil cooling collaboration. ([openai.com](https://openai.com/api/pricing//?utm_source=openai))