Moonshot AI Wants 30% Revenue Share from Microsoft, Amazon & Google for Kimi K3

TL;DR

  • What's happening: Moonshot AI is in early-stage talks with Microsoft, Amazon, and Google to claim up to 30% of revenue generated from Kimi K3 services on their cloud platforms, Reuters reported on August 26, 2026.
  • Why it matters: Kimi K3 is already live on Azure and AWS, it's frontier-class on coding benchmarks, and a 30% cut from three of the world's largest cloud providers would be a genuinely new revenue model for open-weight AI.
  • What's unresolved: No deal is signed. The exact percentage, data access terms, and token-auditing mechanism are all still being negotiated - and every party declined to comment.
Moonshot AI seeks 30% revenue share from Microsoft, Amazon, and Google for hosting Kimi K3

Reporting note: This story is compiled by the AI Tech Safar editorial team directly from Reuters' original report, Moonshot's official model card and license terms, and Artificial Analysis benchmark data - cross-checked against AWS, Microsoft, and Bloomberg's own coverage. Every number here traces back to a primary source linked at the bottom.

Imran Khan Pathan, Editor: As someone who runs Claude daily for coding work, K3's benchmark numbers are close enough that I'd genuinely consider testing it myself if access outside China-hosted infrastructure gets easier.

What's Actually Being Negotiated?

Reuters broke the story on August 26, 2026: Moonshot AI is in early-stage talks with Microsoft Azure, Amazon Web Services, and Google Cloud about revenue-sharing agreements tied to Kimi K3 hosting on their platforms.

The ask is up to 30% of revenue from K3-related services - meaning API calls, token consumption, and managed deployment fees generated when these cloud giants serve the model to their enterprise customers.

Three issues remain unresolved, per Reuters:

  • The exact revenue split (30% is the ceiling Moonshot is seeking, not a confirmed number)
  • Data access terms - what usage data Moonshot can see
  • Token auditing - how consumption gets tracked and verified

All four parties - Microsoft, Amazon, Google, and Moonshot - declined to comment. Reuters was explicit: there is no certainty these talks lead to agreements.

This isn't just a licensing negotiation. Moonshot's public license already bakes in a commercial trigger: any company running Kimi K3 as a Model-as-a-Service business and exceeding $20M in aggregate revenue over any 12-month period must enter a separate commercial agreement with Moonshot. The cloud deal being discussed is the formalization of that trigger at hyperscaler scale.

What Is Kimi K3? (For Those Just Tuning In)

Kimi K3 is Moonshot AI's flagship model - and the world's first open-weight model to hit the 3-trillion-parameter class.

The specs, fast:

  • 2.8 trillion total parameters, 104 billion activated per token (Mixture-of-Experts architecture with 896 routed experts, 16 active at a time)
  • 1 million token context window (1,048,576 tokens exactly)
  • Native multimodal - vision and text in the same model, not a bolt-on adapter
  • Weights released July 27, 2026, roughly 11 days after the initial announcement

Moonshot's own model card is honest about where it stands: "overall performance still trails the most powerful proprietary models, Claude Fable 5 and GPT 5.6 Sol." That's a rare thing for a lab to say at launch. It also means the claim is credible when they say it's frontier-class - because they're not overclaiming.

How Does Kimi K3 Actually Perform?

Here's the benchmark picture across the four metrics that matter most for agentic coding work:

Benchmark Kimi K3 Claude Fable 5 GPT-5.6 Sol Winner
DeepSWE 67.5% 70% 73% GPT-5.6 Sol
Terminal-Bench 2.1 88.3% ~88% 88.8% Effectively tied
Frontend Code Arena 1,679 Elo 1,631 Elo Below Fable 5 Kimi K3
Reliability 63.4 66.9 50.9 Claude Fable 5

Scores from Moonshot's official model card (July 2026) and Artificial Analysis evaluations. Kimi K3 results use the Kimi Code harness at max reasoning effort.

The honest read: K3 doesn't beat the best proprietary models overall, but it's close enough to matter. On Terminal-Bench it's essentially a three-way tie. On frontend development it's the clear leader. The reliability gap vs. Claude Fable 5 is real but not disqualifying - and GPT-5.6 Sol's reliability score of 50.9 is notably weak, which is its own story.

For a model you can download and self-host, these numbers are genuinely unprecedented among Chinese AI models 2026.

Why Microsoft, Amazon, and Google Are Even Interested

The short answer: open-weight models are cheap to deploy and Kimi K3 is already there.

K3 went live on Azure via Fireworks AI on Microsoft Foundry within days of the weights dropping. AWS published a full SageMaker HyperPod and EKS deployment guide on July 30. Google Cloud announced Day 0 support in its developer forum. These weren't revenue-sharing deals - they were infrastructure moves, cloud providers getting ahead of enterprise demand.

The kimi ai model is already generating real usage on these platforms. That's exactly why Moonshot has leverage: they didn't need to ask permission, and the model is already running.

From the cloud providers' perspective, offering a frontier-class open-weight model is a competitive differentiator. Azure Copilot has reportedly tested K3 for model routing. If you're Microsoft and you can offer a 2.8T-parameter coding model that rivals GPT-5.6 Sol on some benchmarks - at a fraction of the inference cost - that's a genuine product advantage.

The 30% ask is steep. But Moonshot's position is that they built the model; the cloud providers are just the pipes.

The 30% Demand - Is It Reasonable?

Apple's App Store charges 30%. That's the number everyone reaches for, and it's not a coincidence Moonshot landed there too.

But the comparison only goes so far. App Store revenue is from consumer apps with massive margins. Cloud AI inference is a thin-margin infrastructure business. A 30% cut on K3 API revenue would meaningfully compress margins for Azure, AWS, and GCP on every K3 token served.

That said, Moonshot's negotiating position is real:

  • $300M ARR as of June 2026 - up from $200M in April. That's 50% growth in two months.
  • $35B valuation after a $3.5B funding round closed in July 2026
  • Reportedly seeking a $50B pre-money valuation ahead of a Hong Kong IPO
  • The moonshot ai funding trajectory suggests investors believe the revenue model works

More importantly: there is no other open-weight model at this performance level. DeepSeek is competitive but different. Meta's Llama series doesn't touch K3 on agentic coding. If cloud providers want frontier open-weight inference in their catalogs, Moonshot is currently the only supplier.

Typical API reseller margins run 20–40% gross on AI model services. A 30% royalty on top of that leaves cloud providers with thin or negative margins on K3 specifically - which is probably why the talks are still early.

What This Means for the US-China AI Race

K3 doesn't prove China won. It proves the gap is closing faster than most Western analysts expected.

Six months ago, the working assumption in most US policy circles was that export controls on advanced chips were meaningfully slowing Chinese AI development. K3 - trained despite those restrictions - complicates that narrative. The exact training setup isn't public, but the results are.

Three things K3 actually changes:

  1. Pricing pressure is real. Kimi API pricing starts at $0.30/MTok for cache-hit input and $15/MTok for output. That's competitive with US providers, not dramatically cheaper - but the open-weight release means enterprises can self-host and pay nothing per token beyond infrastructure. That's a structural threat to OpenAI and Anthropic's API revenue.

  2. The open-weight strategy is working. By releasing weights, Moonshot gets global distribution without a sales team. Every developer who deploys K3 on their own infrastructure is a user Moonshot didn't have to acquire. The moonshot ai kimi brand spreads with the model.

  3. Export controls are under scrutiny. If a Chinese lab can train a 2.8T-parameter frontier model despite chip restrictions, the policy question isn't whether controls exist - it's whether they're calibrated correctly. That debate is now unavoidable in Washington.

The US-China AI competition isn't a single race with a finish line. It's a multi-front contest over scale, distribution, and monetization. K3 is a strong move in all three.

What's Still Unresolved

A lot, actually. Here's what we don't know as of August 27, 2026:

  • No deal is signed. Reuters was explicit that talks are preliminary and may not result in agreements.
  • The exact percentage is unknown. 30% is what Moonshot is seeking. What the cloud providers will accept - if anything - isn't public.
  • Token auditing is unsolved. How Moonshot verifies how many K3 tokens Azure or AWS serves is a genuinely hard technical and contractual problem. Cloud providers don't typically give third parties that kind of visibility into their infrastructure.
  • Data access terms are unclear. Whether Moonshot gets any usage data from cloud deployments is unresolved - and likely contentious.
  • Regulatory and export status is murky. K3 weights are publicly available, but whether US cloud providers can formally commercialize a Chinese AI model at this scale - and under what regulatory conditions - hasn't been tested.
  • Distillation allegations are unresolved. US officials and Anthropic have accused Moonshot of distilling outputs from US models to train K3. Moonshot hasn't confirmed or denied this. If true, it adds a significant legal dimension to any commercial deal with US cloud providers.

FAQ

Why would Microsoft, Amazon, and Google even agree to pay a Chinese AI company?

Because Kimi K3 is already running on their platforms and customers are already using it - the cloud providers deployed it within days of the weights dropping, before any revenue deal existed. Paying Moonshot formalizes an arrangement that's effectively already happening, and it's cheaper than trying to build a comparable open-weight frontier model in-house.

Could this deal fall apart over the distillation allegations?

It's a real risk factor, not a formality. US officials and Anthropic have accused Moonshot of distilling outputs from American models to train K3, which Moonshot denies. If that allegation gains traction - especially with Treasury Secretary Bessent already floating a blacklist - it could make a formal commercial partnership with a Chinese lab politically radioactive for three major US companies, regardless of the technical merits.

What happens to smaller AI companies if this revenue-sharing model catches on?

It could reshape how open-weight models get monetized industry-wide. If Moonshot successfully gets 20-30% of cloud revenue from a model it gives away for free, that becomes a template other labs - including smaller ones - may try to replicate, shifting monetization from direct API sales toward revenue-sharing with the infrastructure layer.

Is Kimi K3 actually cheaper to run than GPT-5.6 Sol or Claude Fable 5?

On paper, yes for self-hosting - the open-weight release means enterprises can deploy it on their own infrastructure and pay nothing per token beyond compute costs, unlike closed API-only models. Whether it's cheaper in practice depends on whether you already have the infrastructure to self-host a 2.8-trillion-parameter model, which most companies don't.

Does this change anything for developers using Kimi K3 today?

Not immediately - the negotiations are about cloud provider economics, not end-user pricing or access. If a deal is reached, the more likely long-term effect is broader, more official support for K3 inside Azure, AWS, and Google Cloud's managed AI tooling, rather than a change to what developers pay today.

Related Reading on AI Tech Safar

Useful Sources

-

Comments

Post a Comment

Popular Post

Agentic AI Explained: What It Is, How It Works, and Why 2026 Is the Tipping Point

Image
Written by Imran Khan Pathan I'm a Graphics Designer who crossed over into AI tools building and tech blogging after a decade in print media - working with Vistaprint India, Ace Infoway, and Caterpillar Signs. I've been hands-on with AI tools since 2021, and I've built real, working AI-powered apps: a workflow tool for Chartered Accountants, a Society Maintenance management app, and an SEO tool - all built using AI. I run AI Tech Safar, where I cover AI news, tools, and industry trends for people who actually build and use this stuff. Connect on LinkedIn → Ai Tech Safar TL;DR Agentic AI is AI that can plan, act, and self-correct across multi-step tasks - without a human prompting every move. It's fundamentally different from chatbots like ChatGPT: it does things, not just says things. Real-world deployment is accelerating fast: 49% of mid-market firms had at least one production agentic workflow by Q2 2026. 2026 is the year it moves from pilot to produc...
Read more

Is Claude Down? Status, Outages & Fix Guide [2026]

Image
TL;DR: Claude went down for roughly 3 hours on August 24, 2026 (04:50–07:36 UTC), hitting Claude Opus 5, Fable 5, Mythos 5, and Opus 4.8 - and then had a second login incident the same afternoon. August 13 was the previous 3-hour outage, just 11 days earlier. TechTimes counted 28 outages in 30 days for commercial Claude - while Claude for Government logged zero outages in the same window. This article breaks down every major 2026 incident, what breaks during a claude outage, and exactly what to do when Claude AI is not working. Reporting note: This timeline is compiled by Imran Khan Pathan at AI Tech Safar directly from Anthropic's own status page (status.claude.com), cross-checked against Downdetector report counts and TechTimes' outage tracking. Every timestamp and error count in this article traces back to a primary source linked at the bottom - updated as new incidents are confirmed. What Happened on August 24? (The Latest 3-Hour Outage) Monday morning. 9:50 PM ...
Read more

Which Jobs Is AI Actually Replacing in 2026? (The Real Data)

Image
TL;DR AI is displacing ~16,000 US jobs per month in 2026, with AI cited as the leading reason for US job cuts for four consecutive months. 37% of companies plan to have replaced workers with AI by end of 2026 (HR Dive / IBM survey). 300 million jobs globally are exposed to AI automation (Goldman Sachs) - but "exposed" ≠ "eliminated." The roles disappearing fastest are repetitive, text-heavy, rules-based jobs. Clinical, skilled-trade, and relationship-dependent roles are holding. The Scale of the Problem - How Many Jobs Are Actually at Risk? Let's be precise here, because the numbers get misquoted constantly. WEF Future of Jobs Report 2025 projects 92 million jobs displaced globally by 2030 - offset by 170 million new roles created, for a net gain of 78 million. That's not a jobs apocalypse. It's a massive, painful reshuffling. Goldman Sachs puts 300 million jobs globally exposed to AI automation over a roughly 10-year adoption windo...
Read more