Why Google and Meta Can Undersell Everyone — The Cross-Subsidization Moat
Google sells Gemini API tokens cheaper than pure-play AI companies. Meta gives Llama weights away for free. SpaceX prices Starlink aggressively against traditional telecom. Submarine cable consortia find Google and Meta building their own cables.
These are not acts of charity. They are cross-subsidization — using profits from one business to underwrite another. When executed correctly, it’s an unassailable structural advantage over any pure-play competitor.
Here’s how it works.
The Core Concept
A pure-play company must make a profit on every product it sells. Its revenue is that product.
A vertically-integrated hyperscaler can sell a product at or below cost, because the product drives revenue in an adjacent, more profitable business.
Pure-Play AI Company:
Revenue = API fees
Profit = API fees − costs
→ Must charge above cost to survive
Vertically-Integrated Hyperscaler:
AI Product → User engagement → Core business revenue
Profit = (Core revenue) − (AI costs + everything else)
→ Can give AI away if engagement → core profitThis is not dumping or predatory pricing — it’s portfolio economics. The hyperscaler calculates ROI across the entire ecosystem, not per-product.
Case Study 1: Google — Gemini Protects the Search/Ads/Cloud Moat
How It Works
Google’s core business is Search advertising ($200B+/year). Every user query is a monetizable event. Gemini’s purpose is not to sell tokens — it’s to keep users searching on Google instead of opening ChatGPT or Perplexity.
The calculation:
| Item | Cost/Impact |
|---|---|
| Gemini API cost per query | ~$0.001-0.01 |
| Lost Search ad revenue per query if user goes to ChatGPT | ~$0.05-0.20 |
| Net benefit of keeping user on Google | $0.04-0.19 per query |
Google can subsidize Gemini because losing Search dominance would cost 10-100x more than running Gemini at a loss.
Same logic applies to Google Cloud:
- Pure-play AI inference providers (Together, Fireworks, Replicate) must cover GPU costs + margin
- Google Cloud offers GPU/TPU access as part of a broader cloud ecosystem
- If Gemini API brings a customer into Vertex AI, that customer then buys BigQuery, Spanner, Cloud Storage — high-margin services
- The AI is the loss leader; the platform is the product
Comparison Table
| Pure-Play AI (OpenAI) | Google (Gemini) | |
|---|---|---|
| Revenue sources | API fees only | Search ads, Cloud, Android, YouTube |
| AI pricing floor | Marginal cost + margin | Potentially zero (subsidized by Search) |
| Customer lock-in | None (you can switch APIs) | Deep (Cloud ecosystem, Android, Workspace) |
| Training cost recoup | Must amortize via tokens | Can amortize across entire business |
| Can they win a price war? | No — no other revenue | Yes — AI is a rounding error on Search revenue |
This is why Gemini Flash costs less than GPT-4o mini despite comparable capability — Google doesn’t need Gemini to be profitable.
Case Study 2: Meta — Llama Protects the Social Graph
How It Works
Meta’s core business is advertising on its social platforms ($130B+/year). The social graph — who you know, what you like, where you go — is the moat. Llama’s purpose is not to sell models — it’s to ensure no competitor controls the AI layer that sits between Meta and its users.
Why give Llama away for free:
Commoditize the model layer — If everyone has access to Llama-class models, no single company (OpenAI, Anthropic, Google) can become the “AI gatekeeper” that Meta must pay or partner with.
Drive adoption of Meta’s ecosystem — Llama powers AI features inside WhatsApp Business, Facebook Feed ranking, Instagram Reels recommendations. These keep users engaged, which drives ad revenue.
Crowdsource innovation — Open-source Llama means thousands of developers build tools, fine-tunes, and integrations. Meta gets the ecosystem benefit without paying for the R&D.
Defend against platform shift — If AI assistants become the primary interface (replacing browsers/apps), Meta needs its own model to insert into that interface. Giving Llama away ensures it becomes the default for third-party tools.
The calculation:
| Item | Cost/Impact |
|---|---|
| Llama R&D cost | ~$100M-1B per generation |
| Value of defending ad revenue | $130B+/year |
| Cost of being locked out of AI layer | Potentially existential |
| Net value of open-sourcing Llama | Massively positive if it prevents platform displacement |
Comparison Table
| Pure-Play AI (OpenAI) | Meta (Llama) | |
|---|---|---|
| Model pricing | API per-token | Free (open weights) |
| Revenue source | Token sales | Social platform ads |
| Open-source incentive | None (undermines API sales) | Strong (commoditizes competitors) |
| Customer relationship | API customer | Social platform user |
| Can they match $0? | No | Yes — and they do |
Meta can permanently offer Llama at $0 because the model itself is not the product — the social graph is. Every Llama download makes the AI layer less ownable by competitors.
Case Study 3: SpaceX — Starlink Protects the Mdirection: right
SpaceX: “SpaceX” { shape: diamond style.fill: “#f9ab00” }
Starlink: “Starlink(LEO Broadband)” { style.fill: “#005288” style.stroke: “#f9ab00” }
Launch: “Launch Services(Falcon 9 / Starship)” { style.fill: “#f9ab00” }
Mars: “Mars Colonization(Core Mission)” { style.fill: “#e85d26” }
Gov: “Government Contracts(DoD, NASA)” { style.fill: “#4b7bec” }
Moat: “Defense Moat:revenue fundsdevelopment,enables Mars” { style.font-size: 11 shape: square style.stroke-dash: 2 }
Starlink -> Launch: “Starlink revenue ($6B+/yr)Starship R&D” Starlink -> Mars: “Mars depends on+ in-space” Starlink -> Gov: “Starshield (military)from Starlink” Launch -> Mars: “Starship is the onlythat canMars payload”
PurePlay: “Pure-Play ISP(Singtel, Comcast, etc)” { style.stroke-dash: 3 style.stroke: “#666” }
Revenue: “Revenue:subscription(must covercost)” { style.font-size: 11 shape: square style.stroke-dash: 2 style.stroke: “#666” }
PurePlay -> Revenue: “Must maintainmargin onsubscriber” n every subscriber”
### How It Works
SpaceX's stated mission is **Mars colonization**. Everything else — Falcon 9, Starlink, Dragon, Starship — serves that goal. Starlink is the revenue engine that funds Starship development.
**The key insight:** Starlink is not a telecom company that happens to do rockets. It's a rocket company that built a telecom to fund its rockets.
**Cross-subsidization in action:**
| Revenue Stream | Use | Margin |
|---------------|-----|--------|
| Starlink consumer ($120/mo) | Funds Starship R&D | High (vertically integrated — SpaceX builds the satellites, launches them on their own rockets, and operates the network) |
| Starlink enterprise/government | Premium pricing for defense/commercial | Very high (Starshield contracts) |
| Falcon 9 launch services | Covers operational costs | Medium |
| NASA/DoD contracts | Funds advanced R&D | High margin, long-term |
**Why Starlink can undercut traditional ISPs:**
- Traditional ISPs (Singtel, Comcast, BT) must cover: fibre buildout, last-mile infrastructure, regulatory compliance, shareholder returns
- Starlink's marginal cost is: building a satellite ($500K), launching it ($15M for 60 satellites = $250K each), and operating the network
- Starlink doesn't need to maximize ISP profit — it needs enough cash flow to build Starship
- If Starship achieves full reusability, launch costs drop 10x, making Starlink economics even better
**The pure-play ISP cannot win this game:**
| | Pure-Play ISP (Singtel) | SpaceX (Starlink) |
|--|------------------------|-------------------|
| **Infrastructure cost** | Fibre trenching, landing stations, last-mile | Satellite manufacturing + launch |
| **Revenue mandate** | Profit on connectivity | Fund Mars / Starship development |
| **Can they price below cost?** | No (regulatory + shareholder) | Yes (mission-driven, private company) |
| **Vertical integration** | Low (buys equipment from vendors) | Extreme (builds satellites, rockets, ground stations, user terminals) |
| **Unit economics** | Must improve every quarter | Acceptable if Starship timeline advances |
This is why Starlink pricing has already dropped from $99/mo (beta) to $90/mo for standard, with promotional discounts — and why traditional ISPs can't match it without destroying their own margins.
---
## Case Study 4: Submarine Cables — Google and Meta Build Thdirection: right
Hyperscaler: "Google / Meta" {
shape: diamond
style.fill: "#4285f4"
}
Cable: "Private Submarine\nCable (Echo, Bifrost,\nApricot)" {
style.fill: "#34a853"
style.stroke: "#4285f4"
}
Cloud: "Cloud Revenue\n(Google Cloud / AWS)" {
style.fill: "#4285f4"
}
Ad: "Ad Revenue\n(Search / Social)" {
style.fill: "#fbbc04"
}
Users: "User Experience\n(Lower latency)" {
style.fill: "#ea4335"
}
Partner: "Partners get\ncapacity at cost" {
style.font-size: 11
shape: square
style.stroke-dash: 2
}
Cable -> Cloud: "Dedicated fibre pairs\nfor cloud traffic"
Cable -> Ad: "Faster page load →\nhigher ad revenue"
Cable -> Users: "Lower latency for\nall services"
Cable -> Partner: "Builds ecosystem\ngoodwill"
PurePlay: "Telecom Consortium\n(SEA-ME-WE-6)" {
style.stroke-dash: 3
style.stroke: "#666"
}
Revenue: "Revenue:\nSell capacity to carriers\n(must return\non investment)" {
style.font-size: 11
shape: square
style.stroke-dash: 2
style.stroke: "#666"
}
PurePlay -> Revenue: "30+ carriers share\ncost and revenue\n= higher prices"
higher prices"Submarine cable economics follow the same pattern:
| Consortium Cable (SEA-ME-WE-6) | Hyperscaler Cable (Echo/Bifrost) | |
|---|---|---|
| Cost | $300-700M shared across 30+ carriers | $300-700M paid by Google/Meta |
| Purpose | Sell capacity at a profit | Serve internal traffic |
| Capacity allocation | Pro-rata by investment | 100% owned |
| Pricing | Market rates (must return on investment) | Internal transfer price (zero if unused) |
| Competitive impact | Sets wholesale floor | Sets no floor — excess given to partners |
When Google builds Echo, they don’t need to make a profit on the cable. They need lower latency for Google services → better user experience → more ad revenue. The cable itself is a rounding error on Google’s balance shedirection: down
Title: “The Cross-Subsidization Moat” { shape: text style.font-size: 16 }
Layer1: “Company” { style.stroke: “#333” style.fill: transparent }
G: “Google” { style.fill: “#4285f4” } M: “Meta” { style.fill: “#1877f2” } X: “SpaceX” { style.fill: “#f9ab00” }
Layer2: “Loss Leader” { style.stroke: “#333” style.fill: transparent }
GL: “Gemini (cheap API)” { style.fill: “#ea4335” } ML: “Llama (free)” { style.fill: “#00c300” } XL: “Starlink (cheap ISP)” { style.fill: “#005288” }
Layer3: “Protected Revenue” { style.stroke: “#333” style.fill: transparent }
GR: “Search Ads ($200B+)” { style.fill: “#34a853” } MR: “Social Ads ($130B+)” { style.fill: “#f5a623” } XR: “Mars / Defense ($B+)” { style.fill: “#e85d26” }
G -> GL -> GR M -> ML -> MR X -> XL -> XR -> GL -> GR M -> ML -> MR X -> XL -> XR ```
Every example follows the same architecture:
- Core business generates massive, defensible revenue (Search ads, social graph, space access)
- Loss leader is priced at or below cost (Gemini, Llama, Starlink)
- Loss leader protects core by preventing competitors from disrupting the moat
- Pure-play competitors cannot match the price because they lack the adjacent revenue
What This Means for Consumers
| When you buy… | You’re paying | But the company actually makes money from… |
|---|---|---|
| Gemini API tokens | Below cost | Search ads and Google Cloud |
| Llama model weights | Nothing | Facebook/Instagram ad targeting |
| Starlink subscription | Near cost | Starship development and defense contracts |
| Google/Meta submarine cable capacity | Below wholesale | Cloud and ad revenue |
You are not the customer of the loss leader. You are the means to an end — more engagement, more data, more ad inventory, more Mars progress.
What This Means for Competitors
If you are a pure-play AI company (OpenAI, Anthropic, Mistral) competing against Google or Meta:
- You cannot win on price. Google can offer Gemini at zero and still profit via Search. Meta already offers Llama at zero.
- You must win on differentiation. Better reasoning, better safety, better at a specific vertical — something Google/Meta cannot replicate by subsidizing.
- You must build your own moat. For OpenAI, that might be enterprise trust (Microsoft partnership, Azure exclusivity). For Anthropic, it might be safety certification. For Mistral, it might be on-device efficiency.
The hyperscalers have already won the price war before it started. The only winning move is not to play it.
What This Means for You
When evaluating infrastructure or API pricing:
- Ask why it’s cheap. If Google sells tokens below OpenAI’s cost, the answer is not “Google is generous” — it’s that Google makes money elsewhere.
- Model switching costs. Free Llama weights today → Meta’s platform dependency tomorrow. No free lunch.
- Map the moat. Every company has a primary profit engine. Everything else is a loss leader or a side effect. Find the actual product.
- Don’t race to the bottom. If you compete against a cross-subsidized product, compete on something the subsidizer can’t easily replicate — trust, niche, vertical integration.
Bottom Line
Google, Meta, and SpaceX can sell below cost because the product is not the product. The product is Search ads, the social graph, and Mars. AI, connectivity, and open-source models are just means to those ends.
Pure-play competitors are structurally disadvantaged — they must profit on the product itself, while hyperscalers treat it as a marketing expense.
This is not unfair competition. It’s portfolio economics at scale. And it applies to everything from submarine cables to LLM tokens.
The Google vs SpaceX architecture comparison in a previous post (June 17, 2026) established the infrastructure value framework. This post extends the economics into cross-subsidization. Revenue figures are approximate based on public filings.