# LLM Savings: Cost-Cutting Insights This Week!

Newsy: LLM Cost Tracker
Published: 2026-02-03T18:00:00.315Z
Source: https://newsyai.com/edition/f2778e29-855b-438f-92ef-de374387ee8a
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 (Jan 27–Feb 2, 2026)

## Pricing announcements from AI labs this week
- No new per‑token API price cuts from OpenAI, Anthropic, or Google in the past 7 days. OpenAI’s public card still lists GPT‑5.2 at $1.75/M input, $14/M output, with a 90% discount for cached input. ([openai.com](https://openai.com/api/pricing/?utm_source=openai))  
- Google’s Vertex AI Agent Engine add‑on billing went live on January 28 for Code Execution, Sessions, and Memory Bank (runtime rates remain $0.0864/vCPU‑hr and $0.009/GB‑hr after the December cut). ([cloud.google.com](https://cloud.google.com/vertex-ai/pricing?utm_source=openai))  
- Reference yardsticks for GPU cloud costs (useful for negotiations): CoreWeave’s posted prices continue to show GB200 NVL72 at $42/hr and B200 8× at $68.80/hr; RTX Pro 6000 Blackwell 8× at $20/hr. Lambda’s B200 lists at $4.99/GPU‑hr for Instances; 1‑Click B200 clusters as low as $3.79/GPU‑hr. ([coreweave.com](https://www.coreweave.com/pricing?utm_source=openai))

—

## 1) FERC nudges PJM toward easier “power next to compute” arrangements
- What’s new: A fresh legal analysis (Feb 2) outlines FERC’s December order directing PJM to clarify tariffs for data centers co‑located with generation—defining transmission service options and behind‑the‑meter rules. Translation: clearer pathways to put firm power directly beside AI campuses. ([mintz.com](https://www.mintz.com/insights-center/viewpoints/2151/2026-02-02-ferc-directs-pjm-issue-new-rules-co-location-power?utm_source=openai))  
- Why it lowers LLM costs and boosts “intelligence per watt”: Co‑location cuts interconnection delays and wheeling costs, improving uptime and letting operators run hotter, denser racks. Reduced soft costs and fewer curtailments mean more tokens per dollar and fewer idle GPUs.

## 2) Grid volatility check: last week’s storm spiked prices—resilience matters
- What happened: A winter storm (within the last week) drove wholesale prices in Virginia from ~$200 to >$1,800/MWh and caused widespread outages, underscoring grid stress from AI load growth. ([theverge.com](https://www.theverge.com/report/868859/electricity-rates-power-grid-ai-data-center-winter-storm?utm_source=openai))  
- Why it matters for LLM economics: Price spikes and outages raise the effective cost per token (retries, failovers, downtime). Firms with firm power, backup gensets, or co‑location (see Item 1) can keep utilization high—delivering steadier $/token and better “intelligence per watt.”

## 3) KV‑cache breakthroughs: smaller memory, same answers
- Fast KVzip (Jan 25): Introduces a lightweight “gated eviction” scheme that evicts up to 70% of KV entries while maintaining near‑lossless quality—trained with forward passes only. Less KV memory means bigger batches per GPU and lower inference latency. ([arxiv.org](https://arxiv.org/abs/2601.17668?utm_source=openai))  
- Elastic Attention (Jan 24): Adds a tiny router so models dial sparse vs. full attention at inference, adapting cost to the input and task. Reported efficiency gains on long‑context tasks with modest training. ([arxiv.org](https://arxiv.org/abs/2601.17367?utm_source=openai))  
- Bigger picture: KV compression and adaptive attention reduce memory traffic—the hidden driver of inference energy. They let you serve longer contexts on the same cards, improving tokens/sec and cutting $/million tokens without hardware changes.

## 4) Software and systems for long‑video and edge decoding
- Spava (Jan 29): A multi‑GPU, sequence‑parallel approximate attention for long‑video LMMs shows up to 12.7× speedups over FlashAttention in prefill—relevant for multimodal agents that were bottlenecked on pre‑processing. ([arxiv.org](https://arxiv.org/abs/2601.21444?utm_source=openai))  
- SwiftKV (Jan 16): An “edge‑oriented” single‑pass attention algorithm plus a multi‑head accelerator improves token efficiency (reported ~1.98×) and attention latency—signals of where low‑power inference is heading. ([arxiv.org](https://arxiv.org/abs/2601.10953?utm_source=openai))  
- Why it compounds: Prefill dominates cost for long inputs; faster prefill or single‑pass attention pairs nicely with KV compression (Item 3). Together, you get more useful work per watt—especially on smaller, cheaper hardware.

## 5) Cooling: two fresh directions that can cut facility power overhead
- Novel “salty” cooling cycle (Feb 1 report): Chinese Academy of Sciences team shows a pressurize/depressurize ammonium‑thiocyanate solution dropping temps by 30–50°C with claimed high theoretical efficiency versus conventional refrigerants. If commercialized, this could trim cooling power and let you pack more GPUs per rack. ([tomshardware.com](https://www.tomshardware.com/pc-components/liquid-cooling/chinese-researchers-discover-new-salty-cooling-solution-that-can-drop-temperatures-by-more-than-50-degrees-celsius-in-seconds-depressurizing-saturated-fluid-triggers-massive-amounts-of-heat-transfer?utm_source=openai))  
- Heat reuse at scale (Jan 31 report): CERN began exporting up to 5 MW of waste heat to a nearby town; a similar approach is planned for the Prévessin data center. For AI facilities, district heat offtake can monetize waste heat and lower net energy costs. ([tomshardware.com](https://www.tomshardware.com/tech-industry/artificial-intelligence/cern-begins-warming-thousands-of-local-french-residents-with-waste-energy-from-its-16-mile-large-hadron-collider-accelerator-leverages-its-massive-cooling-network-to-help-slash-local-carbon-emissions?utm_source=openai))  
- Why it compounds: Lower PUE means a greater share of your electricity drives tokens, not fans or chillers. Heat‑reuse or higher‑efficiency cycles reduce operating costs and carbon—useful for permits, power contracts, and financing.

## 6) Real‑world cost anchors you can use today
- CoreWeave: GB200 NVL72 at $42/hr; H200 8× at $50.44/hr; H100 8× at $49.24/hr; RTX Pro 6000 Blackwell 8× at $20/hr. GB300 NVL72 is listed—contact sales for price. ([coreweave.com](https://www.coreweave.com/pricing?utm_source=openai))  
- Lambda: On‑demand B200 Instances at $4.99/GPU‑hr; 1‑Click B200 clusters as low as $3.79/GPU‑hr with short commitments. ([lambda.ai](https://lambda.ai/pricing?utm_source=openai))  
- Why it compounds: Public, posted rates compress margins and make it easier to split fleets—premium GPUs for training and cheaper cards or smaller clusters for inference. As software (Items 3–4) boosts throughput on each GPU, your effective $/token falls even if sticker prices don’t.

## 7) Housekeeping: Google’s Agent Engine add‑ons now bill separately
- Live since January 28: Runtime stays at the reduced $0.0864/vCPU‑hr and $0.009/GB‑hr, while Code Execution, Sessions, and Memory Bank now appear as distinct line items. Check invoices to avoid mistaking runtime wins for end‑to‑end cost cuts. ([cloud.google.com](https://cloud.google.com/vertex-ai/pricing?utm_source=openai))  
- Why it matters: These are not token prices, but they affect total workflow cost. Track them so LLM‑side savings aren’t erased by orchestration fees.

—

## Quick takes (this week)
- No new token‑price cuts from major labs; Anthropic’s Opus 4.5 still $5/M in, $25/M out; OpenAI’s GPT‑5.2 still $1.75/$14. Use these as benchmarks when negotiating. ([anthropic.com](https://www.anthropic.com/claude/opus?utm_source=openai))  
- DeepSeek remains the low‑cost floor for many workloads, with third‑party trackers showing sub‑$0.30/M input on several models; treat these as negotiation anchors (verify model/version before committing). ([pricepertoken.com](https://pricepertoken.com/pricing-page/provider/deepseek?utm_source=openai))

—

## Action checklist for the week
1) Pilot KV‑cache reduction on your longest contexts. Test Fast KVzip or similar ideas; measure GPU memory, batch size, and tokens/sec before/after. Expect lower latency and higher throughput per watt. ([arxiv.org](https://arxiv.org/abs/2601.17668?utm_source=openai))  
2) Re‑bid capacity using public rate cards. Compare your quotes to CoreWeave and Lambda and right‑size fleets for training vs. inference. ([coreweave.com](https://www.coreweave.com/pricing?utm_source=openai))  
3) If you operate in PJM, map co‑location options and timelines under FERC’s directive; co‑located firm power reduces risk and unlocks higher rack density. ([mintz.com](https://www.mintz.com/insights-center/viewpoints/2151/2026-02-02-ferc-directs-pjm-issue-new-rules-co-location-power?utm_source=openai))  
4) Track Google Agent Engine line items now that billing has started; refactor to minimize paid Sessions/Memory/Code Execution when they don’t add outcome quality. ([cloud.google.com](https://cloud.google.com/vertex-ai/pricing?utm_source=openai))

If helpful next week, I can include a one‑pager rolling these software wins and runtime charges into an all‑in $/million tokens for your current stack.
