LLM Cost Cuts: Weekly Update
The LLM Cost‑Cutter
Weekly briefing on advances that shrink the cost of large language models (Dec 10–Dec 16, 2025)
Pricing announcements from AI labs this week
- No new public per‑token price cuts were posted by OpenAI, Anthropic, Google Vertex AI, or Cohere in the last 7 days. OpenAI launched GPT‑5.2, but its API price pages show no new token‑pricing updates in this window; existing pages remain the latest reference. (reuters.com)
—
1) OpenAI launches GPT‑5.2; value up, prices unchanged (so far)
- What’s new: OpenAI released GPT‑5.2 (Instant, Thinking, Pro) on December 11, emphasizing stronger general intelligence and long‑context handling. No accompanying API price change was announced. In the same interview, OpenAI said prior models (GPT‑5.1/5/4.1) remain in the API. Plain English: more capability, same posted price baseline. (reuters.com)
- Why it matters: If productivity per request rises while per‑token prices don’t, effective cost per task falls. That compounds with other efficiency gains (quantization, cache tricks, cheaper hardware) to raise “intelligence per dollar and per watt” without waiting for a formal price cut. (reuters.com)
2) Posted GPU rates to benchmark your floor
- CoreWeave still lists public per‑hour prices across new and prior‑gen parts, e.g., B200 8‑GPU nodes at $68.80/hr, H200 8‑GPU at $50.44/hr, H100 8‑GPU at $49.24/hr, and RTX Pro 6000 Blackwell Server Edition 8‑GPU at $20/hr. Networking egress/ingress is still posted as $0. These rate cards are useful negotiation anchors even if you buy under private terms. Plain English: transparent menus help you price‑check and mix tiers. (coreweave.com)
- Why it matters: Public floors pressure market rates and let you blend “good‑enough” SKUs (e.g., RTX‑class or older datacenter GPUs) with software optimizations to cut $/request now. The same infrastructure can later benefit from model‑side efficiency upgrades, stacking savings. (coreweave.com)
3) Research signal: new low‑bit methods to boost tokens per watt
- Vec‑LUT (Dec 6) proposes a vector lookup‑table approach that speeds ultra‑low‑bit (≈1.6‑bit) inference on CPUs by up to 4.2× versus prior LUT methods; the authors say it’s integrated into llama.cpp. Plain English: running tiny‑bit models in parallel gets faster, potentially shifting some serving off GPUs. (arxiv.org)
- SignRoundV2 (Dec 4) reports improved quality at extreme low bits (2–5‑bit) via better sensitivity metrics and scale search, targeting production‑grade accuracy with fewer bits. Plain English: more compression with less quality loss. (arxiv.org)
- Why it matters: Low‑bit weights and caches reduce memory traffic—the silent energy hog—and raise throughput on whatever hardware you already own. This directly improves “intelligence per watt,” and pairs well with cache‑compression work in popular servers like vLLM. (arxiv.org)
4) Power and policy: tailwinds and headwinds for $/kWh
- Senate scrutiny: On Dec 16, three U.S. senators opened an inquiry into data‑center impacts on electricity bills, asking operators (including hyperscalers and AI clouds) for more transparency and suggesting big users prepay more of their grid costs. Plain English: pressure is rising to make data centers shoulder a bigger share, which could change contract structures. (theguardian.com)
- House bill to fast‑track gas pipelines: On Dec 12, the House passed a bill making FERC the lead agency for interstate gas‑pipeline permits—aimed at speeding energy infrastructure that powers data centers. Plain English: if it becomes law, more firm power could come online faster, easing price spikes. (reuters.com)
- Why it matters: Electricity is a growing slice of LLM cost. Faster infrastructure can stabilize or lower $/kWh; tougher scrutiny can add fees or commitments. Teams that secure long‑term, indexed power and site in tariff‑ready regions will ride lower, steadier cost curves. (reuters.com)
5) Cooling efficiency: smarter liquid‑cooling control can cut overhead power
- LC‑Opt (Nov 31 posting; new last month) releases a high‑fidelity benchmark for using reinforcement learning to optimize liquid‑cooling end‑to‑end (from towers to cabinets), built on a digital twin of an ORNL supercomputer system. Plain English: better software control of pumps, valves, and temps can shave cooling energy. (arxiv.org)
- Why it matters: Lower PUE means more of your site’s power feeds useful tokens. As rack‑power densities rise with Blackwell/GB300 systems, intelligent liquid‑cooling control compounds hardware and model‑level gains in tokens per watt. (arxiv.org)
6) Supply chain outlook: more chip‑making gear points to future capacity (and price relief)
- SEMI now forecasts wafer‑fab equipment sales to grow 9% in 2026 (to $126B) and another 7.3% in 2027, led by regions scaling AI‑relevant logic and memory. Plain English: more tools today mean more AI‑chip capacity tomorrow, which usually pressures compute prices down. (reuters.com)
- Why it matters: When advanced‑packaging and fab capacity expand, GPU and accelerator supply loosens. That reduces rental premiums and improves the odds that posted rates (see §2) keep drifting down—or at least, that availability improves so you can better utilize clusters. (reuters.com)
7) Market watch: operator finances and posted capacity
- CoreWeave’s shares slid further this week amid investor concern about AI‑infra overbuild and weather‑related delays; the company also recently raised a large convertible note. Plain English: despite market jitters, operators are still financing and building; watch for how this affects public rate cards and contract flexibility. (investors.com)
—
What this means for your unit economics this week
- Treat GPT‑5.2 as a free productivity boost unless/until pricing changes. Benchmark tasks you do often; if throughput per request rises, your effective $/task just fell. (reuters.com)
- Keep squeezing the memory side: pilot extreme low‑bit weight and KV‑cache paths (e.g., SignRoundV2‑style PTQ, vectorized LUTs). You’ll cut memory energy and raise tokens/sec without new hardware. (arxiv.org)
- Anchor negotiations on public GPU rate cards and mix SKUs. Use cheaper RTX/older datacenter nodes for high‑throughput or batch jobs; reserve top‑end nodes for latency‑critical paths. (coreweave.com)
- Hedge power risk early. Track where permitting is speeding up (pipelines, transmission) and where political scrutiny is growing. Lock in long‑term power with demand‑charges you can live with, and plan for liquid‑cooling controls to pull PUE down. (reuters.com)
- Watch the capacity wave. SEMI’s outlook suggests more AI‑chip supply in 2026–27; align contract terms so you can benefit from falling effective $/GPU‑hour as the wave lands. (reuters.com)
—
Quick reference links
- OpenAI API pricing (reference). (openai.com)
- Anthropic pricing and service tiers (reference). (anthropic.com)
- Vertex AI generative AI pricing (reference). (cloud.google.com)
- Cohere pricing (reference). (cohere.com)
- CoreWeave public pricing (reference). (coreweave.com)
If you’d like, next week’s edition can add a rolling “policy heat map” for large‑load tariffs and interconnection reforms that most affect AI siting and power costs.