From our other sites
The story
It has come to light that NVIDIA is considering a plan to separately charge corporate customers who buy its GPUs a CUDA software usage fee of roughly $2,000 (about ¥320,000) per chip, per year. CEO Jensen Huang mentioned the figure during a Q&A at GTC Berlin on October 22. It amounts to a kind of subscription charge layered on top of hardware purchases — an extra license fee that kicks in after the sale — and appears aimed at monetizing CUDA’s grip on the market as the de facto standard for AI development. Reactions on 5ch were split between outrage over “being billed after the fact for something you already bought” and pointed observations that the shift toward AI-specific ASICs could erode CUDA dependence anyway.
Nvidia Weighing Annual GPU Usage Subscription of 2.7 Million Won (About ¥320,000) — Expanding the AI “Toll Road” via CUDA
Nvidia is considering a plan to separately charge corporate customers who purchase its graphics processing units (GPUs) a software license fee of roughly $2,000 (about ¥320,000) per chip, per year. This is a kind of subscription-style billing structure that keeps charging for the software driving the chip even after the hardware itself has been sold — seen as an attempt to convert the dominance of CUDA, the development platform that has become the de facto standard of the AI ecosystem, into revenue.
The figure was mentioned directly by CEO Jensen Huang during the Q&A session at the “GTC Berlin” conference held in Berlin, Germany, on October 22 (local time) this year.
Source: finance.biggo.jp / Original article here
What people said
The shift to AI-specific, power-efficient chips is already happening.
Otherwise you can't build large-scale data centers — the power draw and heat would be unmanageable.
Those AI-specific ASICs are being developed and are already replacing GPUs.
Apparently using them for inference cuts power consumption to about 1/10.
And that's probably why CoreWeave — the company that's been renting out GPU servers — had to issue corporate bonds to raise cash, with lenders demanding extra interest because they're worried about repayment risk.
(Nvidia has guaranteed that company's debt.) Doesn't this suggest Nvidia itself needs cash?
Won't we see a GPU oversupply crash sometime soon?
Prices will stay elevated a while longer yet.
I'm never signing up for a subscription.
It's always the same story — service quality drops or the price goes up later anyway.
CUDA is licensed by Nvidia, so they can just collect it as a license fee.
In other words, totally legal.
Having the ability to develop AI in-house is basically equivalent to possessing nuclear weapons.
Don't you get why the people at the top keep screaming 'regulate AI'?
I'm never buying anything like the GTX Spark again.
Nvidia and Rambus have always been the industry's black sheep.
Speaking of which, I wonder what Rambus is up to these days…
AMD's a solid company that sells at fair prices instead of gouging people, and it still never gets rewarded for it lol
Quite the opposite —
this will accelerate open source, so we're actually getting closer to an era where everyday users can use AI casually.
To soak up the mountain of dollars printed during COVID, the US jacked up Treasury yields massively. The side effect was an inverted yield curve that put banks in crisis mode.
So they built an AI bubble to generate earnings and ride it out. The inversion is expected to resolve by the end of next year.
The Nasdaq hit its all-time high back in June and hasn't climbed since.
You can get a risk-free 4.8% yield on 10-year US Treasuries.
That means less money flowing into AI investment. They say it's still around 6 trillion yen (roughly $40 billion) a week going in, and business is apparently booming regardless.
It's not technically a monopoly —
it's just that no other company has managed to build something at the same level yet.
Oracle's adopting TPUs, so the move away from Nvidia might actually be picking up steam.
Even so, if it amounts to market monopolization, antitrust law would still apply.
the part getting turned into a subscription is the ★ one! We've seen this road before lol
To put it in Microsoft terms:
Windows → CUDA (as the underlying platform)
★Excel/Office → Nvidia's enterprise AI software
The PC itself → the GPU
Re: #147. That one's actually mine, by the way
CUDA matters less than it used to, so the move away from it might actually gain real momentum.
The internet is open —
it's not some monopolized technology.
GAFAM: '( ・∀・)つ〃∩ heeey~ heeey~ heeey~' (mocking taunt emoticon)
A usage fee just to use a GPU you already bought → WHAT?! Don't mess with us, boycott time.
It was bound to end up like this.
The latter. Apparently it's like Excel, according to Chappy (ChatGPT).
It says right at the top this is for corporate customers.
Though if you're running as a sole proprietor, you might end up on the billing list too.
Even if you build one, you can't put up competitive numbers unless you're cozy with the OS makers and software vendors.
That's exactly why no newcomer can break in.
A GPU specialized for Japanese-language processing might actually have a shot.
Background and Key Points of This Topic
CUDA is the de facto standard platform for AI development that Nvidia has spent nearly 20 years building, and the company’s edge has rested less on raw GPU performance than on developers’ familiarity with — and lock-in to — this software ecosystem. The plan reported here would layer a separate license fee on top of hardware sales, effectively a ‘subscription,’ aimed at growing recurring revenue rather than per-unit GPU pricing. On the thread, reactions split between outrage at ‘being billed after the fact for something you already bought’ and a more resigned take that it’s ‘just a legitimate license fee bundled with CUDA anyway.’ Another point raised is that data-center power constraints are already pushing companies toward AI-specific ASICs — moves like Oracle adopting TPUs to cut CUDA dependence are genuinely underway, not just thread speculation. On the other hand, it’s worth noting the original article explicitly specifies this applies to ‘corporate customers,’ a detail that’s easy to lose track of when discussing whether individual users would be billed.
※This article is excerpted and summarized from the 5ch (News Speed+) thread ‘Nvidia Reportedly Considering AI Usage Subscription of ¥320,000 a Year Per GPU — CEO Jensen Huang‘.
Leave a Reply