LLM Savings: Your Weekly Cost-Cutting Update!

The LLM Cost‑Cutter

Weekly briefing on advances that shrink the cost of large language models (Dec 23–Dec 30, 2025)

Pricing announcements from AI labs this week

  • No new per‑token price cuts announced by OpenAI, Anthropic, or Google in the past 7 days.
  • DeepSeek’s aggressive pricing remains in effect: $0.28 per million input tokens (cache miss), $0.028 per million (cache hit), and $0.42 per million output. Google’s Vertex AI also lists current partner/lab prices, with DeepSeek V3.2 shown at $0.56/$1.68 (Vertex hosted), reflecting platform markups versus DeepSeek’s direct API.

1) Energy and power: On‑site generation ramps up

  • What happened (Dec 30): Two power‑supply moves underline a shift toward “self‑powering” AI campuses.
    • FTAI Aviation launched “FTAI Power,” repurposing abundant CFM56 jet engines into fast‑deploy gas turbines for AI data centers—a shortcut around long lead times for new utility power. Plain English: they’re turning jet engines into mini power plants to get AI centers running sooner.
    • Caterpillar’s generator business is spiking on AI demand; a Utah AI campus plans >700 natural‑gas generators toward 1.5 GW by 2028, highlighting the scale of off‑grid or grid‑light strategies.
  • Why it matters: Time‑to‑power is now a dominant driver of AI cost. Reliable, local generation avoids years of grid delays and peak tariffs, stabilizing electricity prices that flow straight into $/token. Over time, more predictable, cheaper electrons mean cheaper inference and training.
  • Sources:

2) Grid policy: FERC pushes PJM to speed co‑located data‑center hookups


3) Infrastructure finance: SoftBank buys DigitalBridge to scale data‑center capital

  • What happened (Dec 30): SoftBank agreed to acquire DigitalBridge for ~$4B, bolstering an AI‑first strategy and adding a major digital‑infrastructure manager (data centers, fiber) to its toolkit. Plain English: a cash‑rich tech investor is buying a company that funds and builds the places where AI runs.
  • Why it matters: More capital and specialized developers entering the AI infra stack usually means faster build‑outs and stronger competition for capacity—key ingredients for lowering compute rental prices that feed into model API costs.
  • Source: https://www.investors.com/news/softbank-acquire-digital-bridge-artificial-intelligence-digital-infrastructure-sftby-stock-dbrg-stock/

4) Hardware & serving software: Blackwell efficiency + vLLM optimizations


5) Real‑time market signal: Public cloud GPU prices (CoreWeave)

  • What’s new (pages refreshed within the last 3 days): Public instance cards show B200 8‑GPU nodes at ~$68.80/hr, H200 at ~$50.44/hr, H100 at ~$49.24/hr, and Blackwell RTX Pro 6000 Server Edition 8‑GPU at $20/hr. Plain English: the going rate for top‑tier inference/training boxes is visible and trending competitive.
  • Why it matters: Transparent hourly rates make it easier to negotiate reserved capacity and forecast $/million tokens. As Blackwell supply rises and software efficiency improves, these numbers typically drift down—pressuring API prices across the board.
  • Sources:

6) Cooling and data‑center efficiency: Liquid cooling ecosystem expands


7) Research watch: New quantization theory to push 4‑bit mainstream

  • What’s new (Dec 24): “HIGGS” (NAACL ’25, posted this month) introduces a linearity theorem linking layer‑wise quantization error to perplexity, and proposes a data‑free scheme that beats popular formats like NF4 in their tests. Plain English: a clearer recipe for how low you can quantize without getting worse answers—and how to set per‑layer bit budgets.
  • Why it matters: If labs can safely run 4‑bit weights/activations and compress KV caches with minimal quality loss, memory traffic and energy per token drop sharply. Expect faster, cheaper inference on the same GPUs and better fit on cheaper ones—another lever on $/token across the stack.
  • Source: https://aclanthology.org/2025.naacl-long.543/

8) Platform pricing snapshot: Google Vertex AI’s current schedules

  • What’s new to note: Vertex AI’s live pricing page shows current rates for Claude 4.5 tiers and several Chinese‑lab models (including DeepSeek) when hosted on Google Cloud—useful for cross‑checking markups versus direct API. It also details runtime charges for Agent Engine and related services that can become hidden costs in “agentic” apps.
  • Why it matters: Even if token prices fall, added runtime or grounding fees can raise total cost per task. Comparing direct API vs. platform‑hosted models (and enabling prompt caching) is a practical way to keep overall spend down.
  • Sources:

Quick takes and how they compound


Actionable checklist for the week

  1. Benchmark NVFP4 and latest vLLM builds on your current GPUs; record tokens/sec and watts. Sources: https://docs.nvidia.com/deeplearning/frameworks/vllm-release-notes/index.html
  2. Price‑shop reserved B200/H200 blocks against public cards; consider RTX Blackwell 6000 clusters for inference‑only workloads. Sources: https://www.coreweave.com/pricing
  3. Model mix: route high‑volume/simple tasks to DeepSeek or platform “flash/haiku” tiers; keep complex reasoning on premium models. Sources: https://api.deepseek.com/quick_start/pricing/ | https://cloud.google.com/vertex-ai/generative-ai/pricing
  4. Watch grid‑policy windows in PJM for co‑location; they can materially cut interconnection timelines and carrying costs. Sources: https://www.datacenterdynamics.com/en/news/ferc-directs-pjm-to-formulate-new-rules-for-the-colocation-of-large-electric-loads/ | https://www.bakerbotts.com/thought-leadership/publications/2025/december/ferc-issues-order-providing-guidance-for-co-locating-power-plants-with-data-centers-within-pjm

Notes on dates: All items above reflect developments posted or updated between Dec 23 and Dec 30, 2025, or enduring price pages crawled within the last few days that affect current costs.