NVIDIA and Over 20 Tech Giants Co-Sign Letter Supporting Open-Weight AI
Alina Collins
Nvidia, Microsoft, Meta and over 20 tech giants urged the U.S. government to back open-weight AI, just as the White House debates banning Chinese open models — the open-vs-closed fight is now reshaping how AI profits get divided.
Who signed — and who didn't?
Nvidia, Microsoft, Meta, Alphabet, Andreessen Horowitz and more than 20 companies signed an open letter urging the White House to support open-weight AI.
OpenAI and Anthropic did not sign. This means → the lineup itself is the signal: chip sellers, cloud providers and infrastructure builders on one side; closed-source labs charging per token on the other.
Jensen Huang's repost drew over 60 million views on X. Satya Nadella, Mark Zuckerberg, Sundar Pichai and Demis Hassabis all voiced public support.
What is "open weight" — and why does it matter?
Open weight means releasing a trained model's parameters so anyone can download and run it on their own servers — no per-token fees to a closed-source vendor.
In plain terms = a closed model is a taxi — you pay by the mile. An open-weight model is buying a car — you get it once, then drive it yourself at no per-trip cost.
The letter states: "Open weights let every organization match the right model to the right task at the right cost." This reflects a collective industry pushback against closed-source pricing power.
How far have Chinese open models already spread?
Moonshot AI's Kimi K3 has 2.8 trillion parameters and a one-million-token context window, priced on OpenRouter at $3 / $15 per million input/output tokens.
Compare Anthropic's Claude Opus 5 at $5 / $25 — Kimi K3 costs roughly 40% less. This means → Chinese open models are using aggressive pricing to embed themselves in the global developer toolchain.
Kimi K2.5 has already logged over 12.3 billion tokens of usage on OpenRouter. OpenAI and Anthropic accuse Chinese startups of copying models through "distillation" — training a smaller model on a larger model's outputs. Treasury Secretary Scott Bessent warned of sanctions if Chinese firms "cross the IP-theft red line."
How does "model routing" erode the closed-model toll booth?
Model routing — automatically assigning tasks by difficulty, sending easy jobs to cheap open models and reserving expensive frontier models for hard reasoning — is becoming the standard enterprise AI architecture.
Microsoft's router can redirect 60%–80% of traffic to cheaper models with no measurable quality loss; Amazon Web Services internal tests show smart routing saves roughly 16%–56% in cost, hitting 63.6% in some retrieval-augmented-generation benchmarks.
This means → OpenAI's and Anthropic's per-token billing model faces structural erosion. Unless they prove frontier capability is worth the premium, traffic will keep leaking.
Could falling unit costs actually boost total demand?
The Jevons paradox — when a resource gets cheaper per unit, usage expands so much that total consumption rises — may replay in AI: cheap intelligence unlocks workloads that were previously uneconomical — always-on enterprise agents, parallel sub-agents, million-token document analysis, real-time audio-video comprehension.
Morgan Stanley's team raised 2027/2028 capex forecasts for Meta, Amazon, Microsoft, Google and SpaceX to roughly $1.2 trillion and $1.4 trillion; the 2026 U.S. Big Tech capex estimate jumped from $433 billion a year ago to $805 billion.
Citi notes the AI bottleneck is shifting from raw compute to HBM capacity and bandwidth, GPU interconnects and cluster scheduling; Goldman Sachs sees excess returns spreading to the full data-center infrastructure stack — CPUs, storage, liquid cooling, optical interconnects.
What does this battle ultimately come down to?
The open camp's logic (Nvidia, Microsoft, Meta): more open models → more developers → more compute and cloud demand → the shovel sellers earn more.
The closed camp's logic (OpenAI, Anthropic): frontier intelligence must be controlled → safety and compliance cost money → per-token billing is a fair value-capture model.
In plain terms = this is not a technology debate — it is a profit-distribution fight. Whoever becomes the AI era's "toll booth" takes the largest share. Whether open-weight models truly drive total compute consumption higher through the Jevons effect will decide the winner.
Content is for reference only, not financial advice.