# Cutting Costs: LLM Pricing Updates This Week

Newsy: LLM Cost Tracker
Published: 2026-03-04T16:02:00.580Z
Source: https://newsyai.com/edition/2b1a9dda-5a10-4a5f-8eea-000e5bc57569
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 26–Mar 4, 2026)

## Pricing announcements from AI labs this week
- OpenAI: No new API per‑token price changes this week; GPT‑5.1 remains $1.25/M input and $10/M output; GPT‑5.2 is $1.75/$14. Built‑in tool and web‑search line items unchanged. ([platform.openai.com](https://platform.openai.com/docs/pricing/?utm_source=openai))  
- Anthropic: No price cuts; Claude Sonnet 4.6 remains $3/M input and $15/M output, with 1M‑token context in API beta. ([anthropic.com](https://www.anthropic.com/claude/sonnet?utm_source=openai))  
- Google: No model price cuts; Vertex AI Agent Engine runtime and add‑on billing remain in effect (runtime vCPU/RAM pricing and Code Execution/Sessions/Memory Bank charges). ([cloud.google.com](https://cloud.google.com/vertex-ai/pricing?utm_source=openai))  
- Mistral: No new public API price changes posted this week. ([mistral.ai](https://mistral.ai/pricing?utm_source=openai))  
- Low‑cost alternative still notable: DeepSeek V3.2 continues at $0.28/M input ($0.028 cached) and $0.42/M output. ([api-docs.deepseek.com](https://api-docs.deepseek.com/quick_start/pricing/?utm_source=openai))

Bottom line: No headline token‑price cuts from the major Western labs; most savings this week come from cheaper GPU hours, improving supply chains, and fresh inference‑efficiency research.

—

## 1) Fresh, real GPU prices you can use this week
- What’s new
  - Lambda updated its public price card on Mar 2, listing on‑demand B200 SXM at $5.74/GPU‑hr and H100 SXM at $3.44/GPU‑hr in one configuration set, with regional/shape variations shown across multiple instance sizes. These list rates are live anchors for negotiations. ([lambda.ai](https://lambda.ai/pricing?utm_source=openai))
  - Runpod’s public page continues to show serverless and instance pricing (e.g., H200 at $4.31/hr on one tier while many SKUs are “contact sales”), still useful for benchmarking effective $/million tokens. ([runpod.io](https://www.runpod.io/pricing?utm_source=openai))
  - CoreWeave remains in the news for AI inference deals (Perplexity partnership) and GB200 deployments, reinforcing competitive market pressure on GPU‑hour rates. ([barrons.com](https://www.barrons.com/articles/coreweave-stock-ai-perplexity-3f575e81?utm_source=openai))

- Why it matters for LLM costs
  - Public rate cards compress the gap between “headline” and contracted pricing, letting you re‑bid capacity with leverage. Pairing lower $/GPU‑hr with serving optimizations below directly reduces $/M tokens and boosts “intelligence per dollar.” ([lambda.ai](https://lambda.ai/pricing?utm_source=openai))

—

## 2) Hardware and supply: Rubin starts rolling; AMD ramps U.S. builds; HBM signals mixed
- What’s new
  - Nvidia’s fiscal Q4 (reported Feb 25) highlighted massive data‑center momentum and formally touted the Rubin platform, claiming up to 10× lower inference token cost versus Blackwell in its release; early Rubin shipments are beginning to select customers this week per trade press. Expect throughput/watt gains as fleets refresh. ([nvidianews.nvidia.com](https://nvidianews.nvidia.com/news/nvidia-announces-financial-results-for-fourth-quarter-and-fiscal-2026?utm_source=openai))
  - Flex began U.S. manufacturing of AMD Instinct MI355X 8‑GPU platforms in Austin, TX (Mar 2). Domestic assembly shortens lead times and can trim logistics risk/cost for inference clusters. ([stocktitan.net](https://www.stocktitan.net/news/FLEX/flex-announces-u-s-manufacturing-collaboration-with-amd-to-kfuuxyoe1pzy.html?utm_source=openai))
  - AMD’s recent technical disclosures (ISSCC) emphasize MI355X per‑CU throughput gains, aiming to match GB200‑class performance at aggressive perf/$; recent coverage also points to MI400/Helios roadmaps stressing memory‑rich inference—useful for long‑context LLMs. ([tomshardware.com](https://www.tomshardware.com/tech-industry/semiconductors/inside-the-instinct-mi355x?utm_source=openai))
  - Memory markets remain volatile; recent coverage of HBM supply and broader memory swings suggests near‑term price stickiness despite capacity adds—important for accelerator pricing. ([marketwatch.com](https://www.marketwatch.com/story/micron-and-other-memory-stocks-see-outsized-losses-whats-behind-the-big-moves-c14882a0?utm_source=openai))

- Why it matters for LLM costs
  - As Rubin and next‑gen AMD platforms land, tokens/sec per rack go up and tokens/watt improve. U.S. MI355X manufacturing can reduce total delivered cost and deployment time. Even if HBM stays tight, architectural gains mean lower $/answer and better “intelligence per watt.” ([nvidianews.nvidia.com](https://nvidianews.nvidia.com/news/nvidia-announces-financial-results-for-fourth-quarter-and-fiscal-2026?utm_source=openai))

—

## 3) Power and policy: “Bring your own power” pledges and a shifting moratorium map
- What’s new
  - Big Tech is preparing a pledge to build their own power supplies to avoid pushing residential bills higher—an explicit nod to on‑site generation and co‑location trends. Analysts caution it’s non‑binding and hard to execute at scale, but it signals continued movement toward self‑provisioned power. ([ft.com](https://www.ft.com/content/b0f83b19-1cd6-4876-b2dd-88c331d7744a?utm_source=openai))
  - Local siting remains fluid: new moratoria advanced in Fulton County, IN (1‑year), Birmingham, AL (6‑month pause >20 MW), and continued debates in Ohio; meanwhile, a Louisville, KY moratorium attempt failed, and Hood County, TX again rejected a pause. This variability directly affects project timelines and cost of power access. ([wndu.com](https://www.wndu.com/2026/03/03/fulton-county-commissioners-approve-one-year-moratorium-data-center-construction/?utm_source=openai))
  - PJM co‑location rulemaking continues on the FERC timeline (Contract Demand services), with filings already in and responses due in March—key for pairing on‑site generation with load. ([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
  - Faster, more predictable power hookups (or on‑site power) reduce time‑to‑compute and shield you from price spikes—lowering effective $/inference. But county‑by‑county politics can add months to timelines; plan for siting risk and keep alternatives ready. ([ft.com](https://www.ft.com/content/b0f83b19-1cd6-4876-b2dd-88c331d7744a?utm_source=openai))

—

## 4) Facilities and cooling: liquid‑cooling momentum and vendor capacity
- What’s new
  - Schneider Electric reported strong results with triple‑digit data‑center growth and a bullish 2026 outlook—evidence that power/cooling supply chains are scaling with AI demand. ([barrons.com](https://www.barrons.com/articles/schneider-electric-earnings-stock-price-51a112da?utm_source=openai))
  - Vertiv continues to gear up for AI‑dense builds (events, product lines, and edge cooling expansions), signaling more standardized, high‑density thermal options in 2026. ([vertiv.com](https://www.vertiv.com/en-us/about/news-and-insights/events/vertiv-at-dcdconnect-new-york-2026-advancing-aiready-data-center-infrastructure/?utm_source=openai))
  - Immersion‑cooling partnerships (e.g., Infosys with ExxonMobil fluids) continue to expand, aiming to cut energy use and enable denser racks. ([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
  - Better cooling lowers overhead power (PUE), keeps GPUs at boost clocks, and enables higher rack density—so you get more tokens per watt and per square foot. As vendors scale, delivery times and installed cost per MW can fall. ([barrons.com](https://www.barrons.com/articles/schneider-electric-earnings-stock-price-51a112da?utm_source=openai))

—

## 5) Serving research this week: less memory traffic, more tokens per watt
- What’s new (all within the past week or so)
  - InnerQ proposes a hardware‑aware, tuning‑free KV‑cache quantization scheme with up to 22% speedup vs prior methods and up to 88% vs FP16 matmuls—by aligning dequantization with GPU compute and reusing scale factors. Plain English: it shrinks the live memory your model needs during generation without hurting quality. ([arxiv.org](https://arxiv.org/abs/2602.23200?utm_source=openai))
  - CHESS dynamically rebuilds a small, relevant context each step and reports up to 4.56× higher throughput while using just ~1% of the KV cache—cutting the biggest long‑context bottleneck. ([arxiv.org](https://arxiv.org/abs/2602.20732?utm_source=openai))
  - A new theory paper on KV compression warns of a “safety cliff” near 90% compression as attention routing collapses—useful guardrails when you push quantization/pruning. ([arxiv.org](https://arxiv.org/abs/2603.01426?utm_source=openai))
  - ParisKV shows drift‑robust, GPU‑native KV retrieval that scales to million‑token contexts and can slash long‑context latency (e.g., 17× vs. a baseline at 1M tokens). ([arxiv.org](https://arxiv.org/abs/2602.07721?utm_source=openai))

- Why it matters for LLM costs
  - These methods reduce GPU time and memory traffic per token—directly lowering $/M tokens and raising “intelligence per watt,” especially for long‑context applications that otherwise explode KV‑cache cost. ([arxiv.org](https://arxiv.org/abs/2602.20732?utm_source=openai))

—

## 6) Market signals from NVIDIA’s quarter: more perf per rack incoming
- What’s new
  - Nvidia posted record Q4 revenue ($68.1B) with Data Center at $62.3B and highlighted Rubin’s target: up to 10× cheaper tokens than Blackwell. Strong guidance implies rapid fleet refresh toward higher tokens/watt hardware. ([nvidianews.nvidia.com](https://nvidianews.nvidia.com/news/nvidia-announces-financial-results-for-fourth-quarter-and-fiscal-2026?utm_source=openai))

- Why it matters for LLM costs
  - When hyperscalers roll in higher‑efficiency parts, everyone’s $/answer tends to fall—either directly (if you run on those fleets) or indirectly (as competition pressures prices). ([nvidianews.nvidia.com](https://nvidianews.nvidia.com/news/nvidia-announces-financial-results-for-fourth-quarter-and-fiscal-2026?utm_source=openai))

—

## 7) Product moves at low price points
- What’s new
  - DeepSeek’s ultra‑low pricing remains a meaningful floor for the market: $0.28/M input ($0.028 cached) and $0.42/M output, with 128K context. For many utility tasks, this resets expectations for “good enough at scale.” ([api-docs.deepseek.com](https://api-docs.deepseek.com/quick_start/pricing/?utm_source=openai))

- Why it matters for LLM costs
  - Even without price cuts from the big three, competitive alternatives can shift workload mix to cheaper routes, lowering blended $/M tokens. ([api-docs.deepseek.com](https://api-docs.deepseek.com/quick_start/pricing/?utm_source=openai))

—

## This week’s impact in plain English
- No API token price cuts from the majors, but total cost per answer keeps sliding because: public GPU‑hour cards moved again (Lambda), next‑gen hardware (Rubin/MI355X) is ramping, liquid‑cooling vendors are scaling, and fresh KV‑cache methods cut memory traffic. Stack these and you get more tokens per dollar and higher “intelligence per watt.” ([lambda.ai](https://lambda.ai/pricing?utm_source=openai))

—

## Action checklist (1–2 hour tasks)
1. Re‑price your steady‑state inference capacity using Lambda’s Mar 2 price card; model $/M tokens from your observed tokens/sec and expected utilization. Then ask current providers to match or beat. ([lambda.ai](https://lambda.ai/pricing?utm_source=openai))  
2. Pilot a KV‑cache optimization: try CHESS or InnerQ on one long‑context route; track latency, acceptance rate, and dollar savings at fixed quality. ([arxiv.org](https://arxiv.org/abs/2602.20732?utm_source=openai))  
3. Power risk scan: if you’re building in PJM or in a county with emerging moratoria, map timelines and fallback sites; watch FERC/PJM Contract Demand steps and local council calendars. ([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. Cooling ROI quick pass: estimate PUE delta and rack‑density gains from moving a pilot row to liquid cooling using vendor literature; roll savings into $/answer forecasts. ([barrons.com](https://www.barrons.com/articles/schneider-electric-earnings-stock-price-51a112da?utm_source=openai))  
5. Consider workload mixing: shift high‑volume, non‑critical tasks to DeepSeek‑class pricing; keep premium models for edge cases. Track blended cost and user‑visible quality. ([api-docs.deepseek.com](https://api-docs.deepseek.com/quick_start/pricing/?utm_source=openai))

—

## Sources and further reading
- OpenAI pricing and model/tool charges. ([platform.openai.com](https://platform.openai.com/docs/pricing/?utm_source=openai))  
- Anthropic Sonnet 4.6 pricing and 1M‑context beta. ([anthropic.com](https://www.anthropic.com/claude/sonnet?utm_source=openai))  
- Google Vertex AI pricing and Agent Engine billing. ([cloud.google.com](https://cloud.google.com/vertex-ai/pricing?utm_source=openai))  
- DeepSeek API pricing. ([api-docs.deepseek.com](https://api-docs.deepseek.com/quick_start/pricing/?utm_source=openai))  
- Lambda public GPU pricing (effective Mar 2, 2026). ([lambda.ai](https://lambda.ai/pricing?utm_source=openai))  
- Runpod pricing page (serverless/instances). ([runpod.io](https://www.runpod.io/pricing?utm_source=openai))  
- Nvidia FY26 Q4 release; Rubin 10× cheaper‑token claim; trade press on early Rubin shipments. ([nvidianews.nvidia.com](https://nvidianews.nvidia.com/news/nvidia-announces-financial-results-for-fourth-quarter-and-fiscal-2026?utm_source=openai))  
- AMD MI355X U.S. manufacturing at Flex; MI355X/MI400 technical context. ([stocktitan.net](https://www.stocktitan.net/news/FLEX/flex-announces-u-s-manufacturing-collaboration-with-amd-to-kfuuxyoe1pzy.html?utm_source=openai))  
- Memory market context affecting HBM and AI hardware. ([marketwatch.com](https://www.marketwatch.com/story/micron-and-other-memory-stocks-see-outsized-losses-whats-behind-the-big-moves-c14882a0?utm_source=openai))  
- Power/policy: FT on Big Tech self‑power pledge; local moratoria and votes; FERC/PJM co‑location track. ([ft.com](https://www.ft.com/content/b0f83b19-1cd6-4876-b2dd-88c331d7744a?utm_source=openai))  
- Facilities/cooling: Schneider earnings momentum; Vertiv portfolio/events; immersion‑cooling partnerships. ([barrons.com](https://www.barrons.com/articles/schneider-electric-earnings-stock-price-51a112da?utm_source=openai))  
- Serving research: CHESS; InnerQ; ParisKV; KV‑compression theory. ([arxiv.org](https://arxiv.org/abs/2602.20732?utm_source=openai))

—  
If you want, I can attach a simple calculator that turns your observed tokens/sec and these current GPU rates into estimated $/M tokens for your stack.
