Research of the Day HUMBL Voice · Intelligence File 001
Research of the Day

July 2026

The Kimi File

How to build frontier AI when you're poor.

Intelligence FileMoonshot AIOpen SourceChina AIKimi K2

Moonshot AI trained a 2.8-trillion-parameter model reportedly for $4.6 million — a fraction of what US frontier labs spend. Seven engineering hacks got them through the four walls that stop everyone else: chips, data, crashes, and serving cost. On July 27, the full weights go public, testing the price floor US labs have built a trillion dollars of infrastructure on.

2.8T Parameters
$4.6M Training cost
#1 Frontend coding
$0 To download
Community bonus

Got the code from the video?

Enter it with your email to unlock the deeper cut, keep a PDF of this file, and claim founding-member access on the HUMBL app.

The deeper cut

Nobody's pricing this in yet: the real casualty of Kimi isn't a rival chatbot, it's Nvidia's roadmap. A trillion dollars of infrastructure was built on the assumption that intelligence stays expensive to produce. Kimi is the first receipt suggesting that assumption was wrong.

Who actually loses

Not OpenAI or Anthropic — they have brand, distribution, and enterprise contracts Kimi can't touch overnight. The squeeze lands on the middle: the API wrapper startups whose entire pitch was 'cheaper than the frontier labs.' That moat evaporates the day cheaper-than-frontier ships from Beijing for free.

What to watch next

Two tells, both public. One: how fast a Western lab quietly cuts API pricing in the 90 days after July 27 — that's the real scoreboard, not the benchmark charts. Two: fork and download counts on the open weights once they land. Adoption curve, not press coverage, is what tells you if this was a moment or a shift.

Enter your code + email to unlock this file's bonus.

Not from Instagram? The code's hiding in today's video, not here.

Watch it on Instagram
01

Why AI costs billions

Every frontier lab hits the same four walls. Money breaks them — reported US frontier runs sit in the hundred-million-dollar class, with over a trillion committed to infrastructure. Moonshot couldn't pay, so they went through each wall a different way.

WallWhy it breaks people
ChipsTens of thousands of GPUs for months. China is banned from the best ones.
DataThe good text on the internet has largely been read.
CrashesOne exploding number destabilizes a run. Weeks of compute burn.
ServingTraining is one bill. Answering millions daily, forever, is the bigger one.

Kimi K2 Thinking reportedly cost 4.6 million dollars to train. Not a typo. [CNBC]

02

One career. One problem. Memory.

Yang Zhilin. Born 1992, Shantou. Tsinghua, then a Carnegie Mellon PhD in four years. In 2019 he co-wrote Transformer-XL and XLNet — the research that stretched AI's memory window. Google Brain and Meta brought him in. Then he went home. March 2023: he founds Moonshot AI in Beijing, named for his favorite album, The Dark Side of the Moon. 2024: Kimi is #3 in China. January 2025: DeepSeek drops R1 and buries everyone — Kimi falls to #7. Investors say copy DeepSeek, go smaller. He does the opposite. [Papers 2019 · VentureBeat · public rankings]

“Token efficiency is not just about efficiency. It's about improving the upper bound of intelligence.”

— Yang Zhilin, Moonshot AI founder
03

Seven hacks. Four walls.

#HackWhat it does
01Wake up 2% of the brain2.8T parameters, but each word wakes only 16 of 896 experts. Giant knowledge, small-model running cost.
02The run that never crashedMuonClip's QK-Clip tripwire caps exploding values. 15.5T tokens of pretraining, zero loss spikes.
03Learn more per wordThe Muon optimizer extracts more learning per token than AdamW. Same data, more intelligence.
04The experience factoryThousands of synthetic environments where agents practice tasks; only verified successes become training data.
05The model grades itselfReinforcement on machine-verifiable answers, then self-critique against rubrics. No human graders.
06Cheap math, from day oneTrained in compressed low-precision math mid-run onward. Half the memory, twice the speed, tiny quality loss.
07The plumbing nobody seesMooncake splits reading and writing across machines, recruits idle hardware. 107–115% more requests, Best Paper, 100B+ tokens a day.

The wall was the strategy.

04

The table that scared Wall Street

Kimi K3: $3 in, $15 out per million tokens — near half the per-task cost of Claude's flagship. The honest caveat: K3 thinks out loud and burns more tokens per task, so the gap narrows on some workloads. The market did the math anyway — launch week, over a trillion dollars gone from chip stocks. Not because Kimi is the best. Because it's close enough, at these prices, for free. [Artificial Analysis · market data, July 2026]

ModelReported training cost
Kimi K2 Thinking$4.6M
DeepSeek V3$5.6M
US frontier runs$100M+
The Accusation
  • Anthropic: three Chinese labs, Moonshot among them, allegedly ran ~24,000 fake accounts and pulled 16M+ exchanges out of Claude.
  • OpenAI filed similar claims. [Feb 2026]
The Defense
  • Every frontier model was trained by scraping the internet without asking.
  • Musk confirmed Grok distilled from OpenAI — the technique is universal; only the target is contested.
  • Moonshot disputes the characterization.

Why is copying the internet business, but copying the copier theft?

05

Free is the weapon

July 27, the full weights go public. Three detonations: the price floor collapses (US labs already adjusted plans), the world becomes his lab (millions of developers improving Kimi for free), and the sanctions logic cracks — if engineering under constraint substitutes for raw compute, a trillion dollars of US infrastructure is priced on a shakier assumption than the market believed.

Read this before you worship

Most performance numbers are Moonshot's own, not yet independently verified. One independent test flagged K3's hallucination rate near 51 percent — Moonshot says they're fixing it. Anything sent to the hosted version sits under Chinese jurisdiction; serious companies self-host the open weights. He's not a saint. He's a strategist.

The Thesis

The side you can't see is still building.

Sources & methodology

CNBC — Kimi K2 Thinking reported training cost.

Yang Zhilin's research papers (2019); VentureBeat profile; public model leaderboard rankings.

Artificial Analysis — pricing and market data, July 2026.

Anthropic and OpenAI — public statements on alleged scraping, February 2026.

Industry analysis. Figures reported by Moonshot AI and third-party sources; independently unverified where noted. Not investment advice.

HUMBL Voice · HUMBL.Voice · July 2026

Follow @humbl.voice for the next file