explainx / blog
Nvidia now offers AI startups token credits and GPU capacity in exchange for a share of their revenue. How the program works, who qualifies, and the circular-financing debate.
Jul 6, 2026
At an internal town hall, Mark Zuckerberg admitted Meta's AI agent bets "haven't come to fruition yet" — months after laying off 10% of the company and reassigning 7,000 people on the assumption they would. Developers on Hacker News had a blunter diagnosis: agents multiply code output, but review is the bottleneck and unsupervised autonomy still doesn't work.
Jul 6, 2026
A Polymarket tweet claiming more students are building AI startups instead of taking internships hit 800K+ views. The replies split between "smart move" and "1637 all over again." Here's the honest breakdown for students and parents.
Jul 2, 2026
If you run open weights on your own hardware in 2026, you are almost certainly touching llama.cpp — directly or through Ollama and LM Studio. This guide explains what it is, how GGUF fits in, copy-paste install and run commands, and how to expose a local API for coding agents.
Nvidia just changed what it means to be a chip company. In a program announced by CFO Colette Kress in the first week of July 2026, Nvidia will give fast-growing AI startups access to its cloud compute through token credits — no massive upfront bill — in exchange for a share of their product and cloud revenue. The official announcement frames it as "unlocking AI compute at scale"; the market read it as Nvidia becoming a financier, not just a vendor.
The timing is not subtle. Jensen Huang said days earlier that a 1-gigawatt AI factory on Nvidia architecture could cost nearly $100 billion. Nobody except a handful of hyperscalers can write that check — a squeeze we covered in the AI bubble reality check and in Anthropic's Colossus-1 partnership with SpaceX. Revenue sharing is Nvidia's answer: if startups can't afford the factory, Nvidia will front the compute and take a piece of what they build.
| Question | Answer |
|---|---|
| What did Nvidia launch? | A revenue-share program: startups get Nvidia cloud compute via token credits, Nvidia takes a cut of their revenue |
| Who announced it, and when? | CFO Colette Kress, announced around July 1–2, 2026 |
| Do startups pay upfront? | No — token credits replace upfront compute bills; Nvidia recoups via revenue share |
| Who's in it so far? | Australian providers Sharon AI (up to 40,000 GB300 GPUs) and Firmus Technologies (up to 170,000 GPUs, 360MW in Batam, Indonesia) |
| What does Nvidia get? | Standard hardware revenue from cloud partners plus an ongoing share of cloud/product revenue on supported capacity |
| Why now? | 1GW AI factories cost ~$100B; startups can't finance that, and Nvidia doesn't want demand bottlenecked by capital |
| Is this circular financing? | That's the debate — see below |
The mechanics have three layers, per Nvidia's announcement and CNBC's reporting:
That third layer is the structural shift. Nvidia has spent three decades selling hardware in discrete transactions. This program converts a slice of its business into recurring revenue indexed to its customers' success — closer to how app stores or payment processors monetize than how semiconductor companies do.
The launch partners are both Australian:
| Partner | Deployment | Scale |
|---|---|---|
| Sharon AI | Grace Blackwell GB300 GPUs | Up to 40,000 GPUs |
| Firmus Technologies | DSX AI factory in Batam, Indonesia | Up to 170,000 GPUs, 360 megawatts |
Sharon AI CEO James Manning called the collaboration "a pivotal moment" for delivering "sovereign, large-scale AI compute infrastructure" — the same sovereign-compute framing driving national buildouts we mapped in the Europe AI landscape. Firmus co-CEO Tim Rosenfield was blunter about the target customer: "AI-native companies need access to scalable, energy- and cost-efficient compute infrastructure to compete globally."
Huang's "AI factory" framing — data centers that manufacture intelligence rather than goods — now comes with a price tag that breaks the old procurement model. At roughly $100 billion per gigawatt, frontier-scale compute is out of reach for everyone except Microsoft, Google, Amazon, Meta, and a few sovereign funds.
Meanwhile, demand keeps compounding:
Seen through that lens, the revenue-share program is Nvidia removing the last bottleneck on its own demand curve. Startups that would have delayed training runs for a funding round can now start immediately — on Nvidia silicon, inside Nvidia's software stack, generating usage that Nvidia monetizes twice.
It's also a competitive moat move. As we covered in China's AI playbook of free models and cheap compute, the strongest pressure on Nvidia's pricing power comes from ecosystems that subsidize compute to win developers. Token credits are Nvidia's version of the same play, funded by revenue share instead of state subsidy.
The obvious criticism arrived within hours of the announcement: Nvidia is financing its own customers so they can buy Nvidia compute. Chamath Palihapitiya's reaction — "the price of poker is going up" — captured the sense that staying competitive in AI now requires balance-sheet moves, not just better chips.
The skeptic's case:
The counter-case:
Our read: this is neither free money nor Lucent 2.0. It's Nvidia converting its dominant position into an index fund on the AI application layer — the same logic that makes AI companies push agents so hard: usage-based economics compound in favor of whoever owns the meter.
If you're building an AI product, the practical implications:
Compute access is decoupling from fundraising. Until now, a serious training run meant raising venture capital largely to hand it to a cloud provider. Token credits break that chain — which could compress seed rounds and shift dilution from VCs to Nvidia's revenue share.
Read the revenue-share terms carefully. A percentage of revenue is a permanent cost of goods sold. For inference-heavy products where compute already dominates margins — the problem enterprises are hitting in the token cost surge — stacking a revenue royalty on top may be worse than owning your compute economics outright.
Lock-in is the price. Credits denominated in Nvidia-powered capacity make migrating to AMD, TPUs, or efficient local inference progressively more expensive. That's the point.
Expect competitors to respond. Google (TPU credits), Microsoft, and AWS already run startup credit programs; Nvidia's revenue-share twist raises the ceiling on how much compute a startup can get pre-revenue. If the model works, "compute for equity/revenue" becomes a standard term sheet line — the financialization of GPUs we first saw signs of at Computex 2026.
Look closely at the two-layer structure and a subtlety emerges that most coverage glossed over. In a traditional GPU sale, Nvidia's revenue is a one-time event tied to a shipment — once the cloud partner has the hardware, Nvidia's economic interest in what happens next is indirect (repeat purchases, brand loyalty). In this program, Nvidia keeps an economic claim on the output of the hardware for as long as the revenue-share term runs.
That changes Nvidia's incentives in three concrete ways:
Nvidia has quietly built up institutional muscle for exactly this kind of deal. Its venture arm participated in 67 startup rounds in 2025, up from 54 in 2024 and just 12 in 2022 — a compounding cadence that shows the revenue-share program isn't a one-off improvisation but the next rung on a ladder Nvidia has been climbing for three years. Each rung took on more risk and more revenue upside:
| Year/Mechanism | Nvidia's exposure | Nvidia's upside |
|---|---|---|
| Pre-2022: Chip sales only | None beyond warranty | One-time hardware margin |
| 2022–2025: Equity investments in AI startups | Capital at risk, no control | Equity appreciation + guaranteed chip demand |
| 2026: Compute-for-equity in OpenAI/Anthropic-scale deals | Very large, but with named counterparties | Preferred access, alignment with frontier labs |
| 2026: Token-credit revenue share | Smaller per-deal, but unlimited startups | Recurring royalty, indexed to real usage, no dilution needed |
The revenue-share model is arguably Nvidia's most scalable financing tool yet, because it doesn't require Nvidia to pick individual winners the way an equity check does. It can be extended to any cloud partner willing to operate DSX infrastructure, turning what used to be a discrete sales motion into something closer to a franchise model — Nvidia licenses its "AI factory" template and platform, partners operate it, and everyone downstream shares the meter.
The announcement leaves real gaps that reporting hasn't yet filled:
We'll update this post as terms surface.
Official sources: Nvidia blog announcement · CNBC · Bloomberg
Program details, partner GPU counts, and market figures are accurate as of publication on July 6, 2026. Revenue-share terms had not been publicly disclosed at press time; verify current program terms with Nvidia directly.
