Step-by-Step Teaching

STEP 1: Core Vocabulary — Know Your Terms

Before reading any data, you need these definitions:

🔑 Key insight before proceeding: Token volume and spending share are two different measurements. A model can dominate tokens but barely register in spending — or vice versa.

---

STEP 2: The Central Finding — Open Weight Is Rising Fast

What happened?

Open-weight token share:    71.5% of all tokens processed
Open-weight spend share:    ~23% of all dollars spent

Why is there such a big gap between 71.5% and 23%?

This is the most important concept in the report. Here's the logic:

High token volume + Low price per token = Low spend share
Low token volume  + High price per token = High spend share

Concrete example from the data:

🧠 Understand this: Open-weight models are winning the volume game because they are cheap. But cheap × huge volume still doesn't equal the dollars that expensive models generate from fewer requests.

---

STEP 3: The Market Is Growing AND Getting Cheaper Simultaneously

This seems contradictory. Here's how to understand it:

The Numbers

• September spend = 13x December spend (+1,180%)
• Average price per token = 47% lower than December
• Median customer more than doubled their own spend

The Logic

More teams using AI
× More tokens per team
× More models per team
= Much higher total spend

Even though each token costs less

Think of it like streaming music:

Songs got cheaper (or free), but people listen to far more music, so total industry revenue grew.

Visual Mental Model

December:  [Small market] [High price/token]  → Baseline spend
September: [13x larger market] [Half price/token] → 13x spend

🔑 Key insight: Falling prices did not shrink the market. Lower prices enabled more usage, which grew total spend dramatically.

---

STEP 4: Model Switching — The Market Is Becoming Fluid

The Data Points

• ~1 in 2 teams changed their most-used model in September
• ~1 in 3 teams changed their most-used lab — an all-time record
• Teams processing 1B+ tokens used an average of 42 different models

What This Means Structurally

Old model of AI adoption:
  Pick one lab → Commit → Build deep dependency

New model (emerging):
  Pick a gateway → Route to best model per task → Switch freely

Why Does This Matter?

A lab can receive more money from a customer while that customer depends on it less. The customer:

• Spends more overall on AI
• Spreads that spend across more models
• No single lab captures the relationship

🧠 Understand this: The AI model is becoming a commodity component inside a larger system, not a platform you lock into.

---

STEP 5: Chinese Models + American Providers — A Geopolitical-Commercial Pattern

The Baseline Fact

December: <10% of Chinese model spend went through US providers
September: ~70% of Chinese model spend went through US providers

How This Works

Chinese lab (e.g., Moonshot AI) → Creates the model weights
American provider (e.g., AWS, Azure, etc.) → Hosts and sells access
Enterprise customer → Pays American provider → Uses Chinese model

By Lab Breakdown

🔑 Key insight: The same model can simultaneously compete with US labs and generate revenue for US providers. Geopolitical origin of the model ≠ who profits from its deployment.

---

STEP 6: How to Measure a Model Launch — Three Different Lenses

The report introduces a framework for evaluating launches. This is a transferable analytical skill.

The Three Metrics

1. REACH        → What % of teams tried it in first 14 days?
2. SPEND SHARE  → What % of total dollars did it capture?
3. RETENTION    → Of teams that tried week 1, how many came back week 2?

Why You Need All Three

The Core Lesson

A model can be the most adopted (Jev) without being the most profitable (Opus 5.5). A model can have the best retention (Luna) while capturing little spend. These are not failures — they reflect different price points and use cases.

---

STEP 7: Synthesis — The Three Structural Shifts

Now connect everything into a coherent picture:

Shift 1: Volume vs. Value Divergence

Open-weight models dominate usage but not revenue. The market has bifurcated:

• High-volume, low-cost tier (DeepSeek Flash, Jev)
• Low-volume, high-value tier (Kimi K3, Opus 5.5, GPT-6 Astra)

Shift 2: From Lock-in to Composability

Teams use 15–42 models. No single lab owns the customer relationship. The gateway layer (not the model) becomes the strategic asset.

Shift 3: Geography of AI Is Decoupling from Geography of Revenue

Chinese model origin ≠ Chinese revenue capture. US providers are monetizing Chinese model weights. The model creator and the model seller are increasingly different entities.

---

Quick Self-Check Questions

Test your understanding:

1. Why can open-weight models have 71.5% of tokens but only 23% of spend? (Answer: They are priced much lower per token)

2. How can total AI spend grow 13x while price per token falls 47%? (Answer: Volume grew far faster than prices fell)

3. A model reaches 35% of teams but captures 0.27% of spend. Is this a failed launch? (Answer: No — it's a cheap, high-volume decision model. Success depends on what you measure)

4. Why would an enterprise use 42 different models? (Answer: Different models for different tasks; easy switching via gateway; no lock-in penalty)

5. How can a Chinese lab's model generate revenue for a US company? (Answer: US providers host and sell access to open-weight Chinese models)

---

Bottom Line: This report teaches that in a maturing AI market, volume, spend, adoption, and retention are four separate signals — and understanding the gap between them reveals the true structure of how AI is being bought, sold, and used at scale.