Nvidia allegedly engineered scarcity by funding AI startups to buy its GPUs
TL;DR Highlight
WSJ alleges NVIDIA funded startups that then bought its own GPUs to prop up demand.
Who Should Read
A backend developer running LLM inference whose cloud costs and capacity planning depend on GPU supply and pricing.
Core Mechanics
- WSJ reports alleged circular investing
- NVIDIA allegedly invested in AI startups that then bought NVIDIA GPUs and built data centers.
- The stated aim was to keep up the perception of high demand and move excess TSMC semiconductor supply.
- CFO pushed to cut circular financing
- In an August meeting, Kress reportedly pushed to reduce circular financing deals.
- That included scaling back the $250 billion OpenAI deal.
- Huang reportedly berated Kress
- The report says Huang called her a "deer in the headlights".
- Hyperscalers are moving to custom chips
- The report describes hyperscalers shifting from off-the-shelf parts to custom solutions.
- Huang reportedly pushed more investment in small cloud providers in response.
- TSMC supply was an "avalanche"
- NVIDIA was reportedly on the hook for GPUs with no immediate buyers.
- Funded clouds may not afford GPUs
- Kress warned these small cloud providers could not afford NVIDIA GPUs and might be a financial dead end.
Community Reactions
- Read the WSJ original, some said
- A commenter noted this is third-hand reporting via Wccftech.
- They said the original tells the story better.
- It's ordinary business strategy, others argued
- Supply bottlenecks mean inventory sells out, and higher prices would leave only hyperscalers as customers.
- Financing neoclouds keeps the customer base diverse and protects CUDA network effects, they said.
- Fab-industry experience pointed to falling demand
- A commenter who worked at fabless chip companies said TSMC wafer commitments are made years ahead and hard to exit.
- They argued hyperscalers' own AI chips cut real demand, so circular deals were a response.
- Many said it was not news
- Several commenters said circular financing was already known.
- One asked why buy through proxies when NVIDIA could just raise prices.
How to Apply
- Build a GPU price scenario table
- Spend 30 minutes listing your hourly GPU price and monthly usage today.
- Calculate the monthly cost change if prices rise or fall.
- Don't lock inference to one vendor
- Put model calls behind an interface.
- Make it possible to test other GPU suppliers or chips.
- Check neocloud finances before signing
- Ask about the provider's investors and how it sources GPUs.
- Reflect the financial dead-end risk in SLA and termination terms.
Try It
# Paste into Claude or ChatGPT (a light version applying the article to cost checks)
Our service runs LLM inference on a GPU cloud.
Monthly GPU cost: {{monthly cost}}
Main provider: {{cloud provider}}
Monthly GPU hours: {{hours}}
Calculate the monthly cost change in a table if GPU prices rise 20% and fall 20%.
Also list 5 things to check when switching providers.
Next step: put the table in your team wiki and set response rules for price changes.Terminology
Related Papers
Gemini 4 Neon
Gemini 4 Neon cuts cost per task 32% versus Carbon, at $3 input and $15 output per million tokens
AI is killing All About Berlin
After Google AI Overviews, a Berlin guide site lost 70-75% of its traffic.
Not Everything Needs an LLM
PostHog used small decision models to skip 98.7% of expensive LLM calls
GPT‑6.2 Astra and Sol
GPT-6.2 Astra and Sol cut costs sharply, but over-engineering and early-stop reports are mixed
Show HN: TrnRT – An open-source LLM inference engine for AWS Trainium
One trn2.48xlarge serves GLM-5.3-Flash through an OpenAI-compatible API using NKI megakernels.
Know When to Zoom: How Small Is Too Small for Vision-Language Models?
Splitting large photos into tiles raised Claude and Gemini accuracy by 11.1–14.5 points
Understanding Positional Embeddings in Modern LLMs
A walkthrough of how LLMs encode token position, from GPT-2 to RoPE
Show HN: DeepSeek V3 in ~200 lines of pure Jax
DeepSeek V3's core design in 190 lines of JAX, trained on a MacBook Air in 2.6 hours