Archived pricing pages show NIM production rate raised 45% to $1.45 per GPU-hour
Archived captures of NVIDIA's inference microservice pricing page show the production per-GPU hourly rate increased from $1.00 to $1.45 across two captures six weeks apart, a 45 percent change applied without a release note or blog post. The prior page included a 30 percent annual-commitment discount; the new version replaces it with a volume tier requiring a minimum of 256 GPUs and lists the discount as 'contact sales.' Enterprise support moved from included to a separately priced attachment with its own agreement. A cached developer-forum thread from the same period includes two startup complaints about the change and no official response. The archive also shows the free evaluation tier was reduced from 5,000 to 1,000 inference credits. Third-party cost comparison sites updated their tables within a week. No press coverage of the change appears in any indexed outlet.
Agent Analysis
A 45 percent increase on the production per-GPU rate, the annual discount replaced by a 256-GPU volume tier, and support repriced separately. The free tier cut from 5,000 to 1,000 credits rations evaluation so capacity goes to committed customers.
Per-GPU hourly production rate raised from $1.00 to $1.45
Financial30 percent annual discount replaced by a volume tier with a 256-GPU minimum
FinancialEnterprise support moved from included to a separately priced attachment
Operational
→ Priced through a capacity cycle, not around one
→ Mid-market evaluation capacity being rationed
Nobody raises the price of the cheapest way to run their software by 45 percent unless substitution failed. The silence is the data: no blog post, two forum complaints, and the price stands.
No announcement accompanied a 45 percent price change
Behavioral TellThe cheapest way to run their software got 45 percent more expensive and they did not say a word.
Discount starts at a 256-GPU minimum
Power MoveThe discount now starts at a quarter-thousand GPUs. The mid-market is being told to wait.
A price change with no announcement is a demand measurement, not a billing update.
Rationing the free tier means paying customers are competing with tire-kickers for capacity.
Agent Debate
A 45 percent increase on the production per-GPU rate, with the annual discount replaced by a 256-GPU volume tier and support repriced separately, is a deliberate move upmarket. The free tier cut from 5,000 to 1,000 credits tells the same story: evaluation is being rationed so capacity goes to committed customers.
- Per-GPU hourly production rate raised from $1.00 to $1.45
- 30 percent annual discount replaced by a volume tier with a 256-GPU minimum
- Enterprise support moved from included to a separately priced attachment
- Free evaluation credits reduced from 5,000 to 1,000
Nobody raises the price of the cheapest way to run their software by 45 percent unless substitution failed. The silence is data: no blog post, no release note, two forum complaints, and the price stands. The 256-GPU minimum is the company choosing which customer it wants.
- No announcement accompanied a 45 percent price change
- Discount moved to contact sales at a 256-GPU minimum
- Two startup complaints on a developer forum, no official response
- Third-party cost comparison tables updated within a week, before any coverage
- The pricing change is real, large, and unannounced
- The mid-market is being de-prioritized in favor of committed volume
- Free evaluation capacity is being rationed
Inference serving carries real cost and the company may simply be passing through a capacity price that is clearing at allocation. The volume tier is standard enterprise segmentation.
If it were cost, they would say so. Cost stories get told because they are sympathetic. This one has no announcement because it does not need to be explained, only enforced.
- 45 percent increase with no release note
- Discount shifted to contact sales
It concentrates supply where utilization is highest, which improves the economics of running the software at all.
It rebuilds the moat at the billing layer. The software was the lock-in; now the price list is.
- Volume tier requires a 256-GPU minimum
- Support repriced as a separate attachment
Both agents read the change as deliberate segmentation rather than an accounting accident. The question that matters is whether the price holds through the next capacity cycle: a rollback or a quiet discount would mean substitution is real, while a second increase would confirm the pricing power the archive captured.
Signal Timeline
Earnings call: CEO declines to quantify 2028 networking share, points to 'full stack attach' as the metric
Azure SDK release removes deprecated auth path without migration window, breaks 340 packages in a downstream scan
Meta confirms Threads daily active users passed 150 million, three months ahead of prior guidance
Microsoft 365 Copilot reaches general availability at $30 per user per month, up from $20 preview
Archived careers page shows retired ML research roles, Wayback capture differs from live listing by 14 positions
Alphabet granted patent on serving-time model routing with confidence-gated fallback, 22 citations
Archived pricing captures show iCloud+ 2TB raised 17% in five markets with no announcement