01 — Nvidia: the tollbooth on the entire AI economy
Nvidia doesn't just sell chips. It sells CUDA, the software ecosystem that every AI model, every research lab, and every cloud provider has spent years building on. Switching away means rewriting years of code. That software lock-in, combined with supply scarcity, is how a chip company maintains 70–80% gross margins.
The hardware moat
Every AI company must pay Nvidia. There is no alternative — yet.
The H100 GPU costs approximately $3,320 to manufacture and sells for $28,000 — an 88% gross margin at the unit level. SA The entire global AI buildout flows through Jensen Huang's production lines. Nvidia holds 88% of the AI accelerator market for cutting-edge training chips. SA
$3,320
H100 manufacturing cost
$28,000
H100 market price
73.4%
Nvidia gross margin Q3 FY2026
NVDA
$57B
Revenue, single quarter (Q3 FY2026)
Big Tech AI capital expenditure 2025
RBC
| Company |
Primary AI focus |
2025 capex scale |
2025 capex est. |
| Amazon (AWS) |
Cloud infrastructure, Trainium chips |
|
~$100B+ |
| Microsoft |
Azure AI, OpenAI partnership |
|
~$80B+ |
| Alphabet (Google) |
TPUs, Gemini, DeepMind |
|
~$75B+ |
| Meta |
Llama models, recommendation AI |
|
~$65B+ |
| Nvidia (supply side) |
H100/B200 GPU manufacturing |
|
$57B/qtr |
| TOTAL (Big 8) |
Combined AI infrastructure spend |
|
$427B |
Context: Microsoft, Google, Amazon, and Meta generated $350 billion in revenue in Q1 2025 alone — funding this buildout largely from internally generated cash, not debt.
MACK The four firms held $490 billion in cash as of Q3 2025. These are not bets — they are the most profitable companies in history doubling down.
So what does this mean?
One company sits at the centre of the entire AI economy, collecting a toll on every model trained and every inference served. That's unprecedented concentration — but it's also why the infrastructure exists for you to use.
The $427 billion that Big Tech is pouring into AI infrastructure means compute is becoming abundant and accessible. The tollbooth is real, but the road it built runs in both directions. As an individual, you don't need to own the infrastructure — you just need to use it before it reprices.
02 — The inequality argument: who captures AI's value?
AI doesn't just create wealth — it concentrates it. The IMF, the Center for Global Development, and independent researchers all point to the same pattern: the productivity gains flow to capital owners, not workers. The question is whether individuals can position themselves on the right side of this divide.
2×
AI growth benefit in advanced economies vs. low-income countries — the gap widens despite improvements in AI access.
IMF
Gini +7pp
IMF projects wealth Gini could rise by up to 6.89 percentage points under high AI adoption with firm-level decisions factored in.
IMF1
$67B
US AI private investment in 2023 alone — 8.7× more than China's $7.7B. Capital concentration is geographic too.
CGD
IMF dual finding: AI and inequality (April 2025)
IMF1
AI could reduce wage inequality
By displacing high-income workers whose tasks AI can perform, wage inequality could actually narrow. High-paying cognitive jobs are more exposed than manual ones — the traditional pattern reverses.
But wealth inequality gets worse
Capital owners capture the productivity gains. When firms choose how much AI to adopt, cost savings from automating high-wage tasks drive higher adoption rates — and the returns flow to shareholders, not workers.
High-income workers are complementary
In practice, AI augments many high-income workers' productivity rather than replacing them — lawyers, doctors, analysts use AI to do more. Those workers also own more capital, compounding their advantage.
Low-income countries fall further behind
Advanced economies' AI productivity gains could be 2× or more those of developing nations. Manufacturing automation in rich countries reduces the low-wage labour cost advantage developing economies rely on.
So what does this mean?
The wealth divide isn't just between companies — it's between people who use AI and people who don't. The IMF is clear: productivity gains flow to capital owners and AI-skilled workers. Everyone else falls behind.
The actionable insight here isn't despair — it's urgency. The 56% wage premium for AI skills (from our jobs piece) means individuals can capture real value without owning infrastructure. Learn the tools now, while the skill premium is highest and the barrier to entry is lowest.
03 — The honest counter-argument: concentration isn't new, and it doesn't always hold
Every technology revolution starts with concentration. Railroads, oil, telecoms, the internet — monopolistic infrastructure phases that eventually gave way to competition, regulation, or commoditisation. The question isn't whether this pattern will repeat. It's when.
The concentration case
—Five companies control the infrastructure. OpenAI, Anthropic, DeepMind, Meta AI and Google depend on either Nvidia chips or each other's cloud platforms
—The AI "tollbooth" has no historical precedent at this speed — railroads took decades to build the moat
—$427B capex in 2025 creates barriers so high that new entrants cannot self-fund at the frontier
—AI-driven productivity gains flow primarily to capital owners; workers get wage premiums only if they hold AI skills — a minority do
—Data centre AI facilities coming online in 2025 face ~$40B in annual depreciation costs but generate only $15–20B in revenue at current utilisation rates
GWK
The diffusion case
+Every previous infrastructure monopoly eventually diffused: electricity, telephone, internet. Concentration peaks before it spreads
+The solo founder economy from piece 03 — $500M companies with 40 people — is only possible because concentrated cloud infrastructure is accessible at commodity prices
+The 56% wage premium for AI skills (piece 02) means individuals can capture value without owning the infrastructure
+Open-source models (Llama, Mistral, DeepSeek) are breaking the closed-model moat — advanced AI is now accessible for free to anyone
+AI spending contributed meaningfully to US GDP in 2025 — the investment boom itself creates jobs in construction, power, and manufacturing before the software benefits materialise
MACK
The honest read: Concentration is real and accelerating right now, but the historical pattern is that infrastructure monopolies eventually become utilities. The question is timing — and who gets hurt in the transition. The IMF's warning is about the transition period, not the endpoint.
So what does this mean?
Both sides are right — and that's the point. Concentration is real today, but the forces of diffusion are already at work. Open-source models, falling compute costs, and sovereign AI investment are all chipping away at the moat.
For individuals, the takeaway is clear: you don't need to wait for the monopoly to break. The infrastructure is concentrated, but access to it is not. A ChatGPT or Claude subscription costs less than a streaming service. The commodity phase is already beginning — and the people who built skills during the concentration phase will have the biggest advantage when it arrives.
04 — The geopolitics: who controls the models controls the infrastructure
DeepSeek's R1 model was released the same day as Trump's inauguration. Within a week it was the #1 free app on the US App Store. Nvidia lost $590 billion in market value in a single day — the largest single-day market cap loss in US history. CSIS
The DeepSeek earthquake: what January 20, 2025 changed
DeepSeek proved that algorithmic innovation can substitute for brute-force compute. A Chinese lab, working under US export restrictions, produced a frontier reasoning model at a fraction of the cost. The implications ripple through every assumption about AI concentration.
US strategy
Export controls
Restrict China's access to advanced AI chips (Nvidia H100/A100 export ban 2022, extended April 2025). Strategy: delay China's AI progress by cutting off compute. DeepSeek's success suggests this approach has limits — "you can't export-control ideas."
CSIS
China strategy
Open-source efficiency
State-backed AI with open-source collaboration. DeepSeek achieved frontier performance at a fraction of the compute cost. Stanford: China is now #1 for AI patent applications. DeepSeek R1 shows 91.2% pro-China bias — AI as geopolitical instrument.
ORF
Sovereign AI — the third path
$200B+ sovereign spend
UAE, India ($1.25B IndiaAI Mission), France, Singapore, Saudi Arabia building domestic AI infrastructure. The insight: depending on US or Chinese models means depending on their values and their data.
MSFT
The open-source disruption
$450 → $50
Training a reasoning model comparable to OpenAI's o1 cost UC Berkeley researchers $450 before DeepSeek. After DeepSeek's open-source release: $50. The cost of frontier AI capability is collapsing.
ORF
So what does this mean?
The AI monopoly is already cracking from the outside. DeepSeek proved that you don't need a $100 billion budget to produce frontier AI. Open-source models are collapsing costs at a rate that no export control or corporate moat can contain.
For you, this is the most important signal in the entire piece: the cost of using frontier AI is falling faster than any technology in history. What cost $450 six months ago now costs $50. What costs $20/month today may be free next year. The window to build AI skills at bargain prices is open right now.
$427B
Big Tech AI capital expenditure in 2025 — up from $256B in 2024. More than many nations' GDP.
88%
Nvidia's share of the AI accelerator market. H100 costs $3,320 to make, sells for $28,000.
$450→$50
Cost collapse for training a frontier reasoning model, thanks to DeepSeek's open-source release.
5 things you can do this week
to position yourself on the right side of the value capture.
1.
Subscribe to Veltrix Collective. We track the AI tools, rankings, and economic shifts weekly so you don't have to. Understanding who controls the infrastructure is the first step to navigating it. One email per week, every Tuesday — no hype, just signal.
2.
Try an open-source model this week. Download Llama or DeepSeek via Ollama (free, runs locally on your laptop), or use HuggingChat in browser. Compare the output to ChatGPT or Claude. Open-source AI is the antidote to the concentration problem — and it's already frontier-quality.
3.
Audit your AI cost stack. If you're paying for AI tools, check whether open-source alternatives could reduce your dependency on Big Tech infrastructure. n8n and Make can automate workflows without vendor lock-in. Ollama runs models locally for zero marginal cost. Every dollar you save is a dollar the tollbooth doesn't collect.
4.
Use AI to analyse one investment or business decision. Feed a quarterly earnings report into Claude or ChatGPT. Ask it to identify risks, compare margins, and flag concentration risks. The wealth divide from the IMF data isn't abstract — it shows up in every portfolio and every P&L. Use AI to see it.
5.
Read the IMF papers cited in this piece. "AI Adoption and Inequality" and "The Global Impact of AI: Mind the Gap" are both free to read. Understanding the wealth concentration dynamics — Gini +7pp, the 2× growth gap — helps you position on the right side of the value capture equation.
The data is clear: $427 billion in infrastructure spend is creating the most concentrated technology moat in history — but open-source AI is collapsing costs from $450 to $50 and falling. The window to build skills while the tools are cheap and the premium is high is open right now. Don't wait for the commoditisation phase to start learning.
Source references
NVDA
Nvidia SEC Filing — Q3 FY2026 Earnings (Oct 2025)Q3 revenue $57B, gross margin 73.4%, operating income $36B. 62% YoY growth. Data centre $51.2B.
sec.gov → SA
Silicon Analysts — Nvidia GPU Market Share 2024–202687% peak share, 75% by 2026. H100: $3,320 manufacture, $28,000 market price. 88% unit margin. 90%+ training market through 2030.
siliconanalysts.com → RBC
RBC Wealth Management — Big Tech AI Expansion (Feb 2026)$427B Big Tech capex 2025 (up from $256B 2024). $490B cash held. $400B trailing free cash flow. $562B projected 2026.
rbcwealthmanagement.com → MACK
Mackenzie Investments — The AI Buildout: Boom or Bust? (Dec 2025)MS/Google/Amazon/Meta: $350B revenue Q1 2025 alone. AI contribution to US GDP growth 2025. Historical capex comparison.
mackenzieinvestments.com → IMF1
IMF Working Paper 2025/068 — AI Adoption and Inequality (April 2025)Wealth Gini +6.89pp under high AI adoption. AI could reduce wage inequality but increase wealth inequality. Capital income channel.
imf.org → IMF2
IMF Working Paper 2025/076 — The Global Impact of AI: Mind the Gap (April 2025)Advanced economy AI benefit 2× or more vs. low-income countries. Cross-country inequality widening. Cannot be fully offset.
imf.org → CGD
Center for Global Development — Three Reasons AI May Widen Global Inequality (2025)US $67.2B AI private investment 2023 (8.7× China). 61 notable AI models in US. Geographic concentration of capital and talent.
cgdev.org → CSIS
CSIS — DeepSeek, Huawei, Export Controls and the US-China AI Race (Mar 2025)Nvidia lost $590B in a day. DeepSeek R1 overtook ChatGPT on App Store. Export controls can slow but not stop China. Huawei/SMIC as alternative stack.
csis.org → ORF
Observer Research Foundation — DeepSeek and Global AI Innovation (Nov 2025)91.2% pro-China bias in R1. Training cost collapsed from $450 to $50 post-DeepSeek open-source release. Jevons Paradox dynamic.
orfonline.org → GWK
GWK Investment Management — When Will AI Investments Start Paying Off? (Oct 2025)$40B annual depreciation on 2025 AI facilities vs. $15–20B revenue at current utilisation. "Defensive insurance" theory. Brad DeLong analysis.
gwkinvest.com → MSFT
Microsoft AI Economy Institute — Global AI Adoption H2 2025 (Jan 2026)UAE $200B+ sovereign AI. India $1.25B IndiaAI Mission. Sovereign AI as national security. Global North/South adoption gap widening.
microsoft.com → MED
Medium / Truthbit — Will Nvidia's AI Chip Monopoly Be Broken? (Nov 2025)Training market: Nvidia consolidates to 90%+ by 2030. Inference market fragments to 45% custom ASICs by 2028. Bifurcation thesis.
medium.com →
05 — Don't watch from the wrong side of the tollbooth
Veltrix Collective
The rails are being laid.
Are you on the right side of the tollbooth?
$427 billion in AI infrastructure spend. Open-source models collapsing costs from $450 to $50. The wealth divide is widening — but the tools to stay on the right side are cheaper than ever. We track them weekly so you don't have to.
Weekly, every Tuesday · No spam ·
Privacy policy · Unsubscribe anytime
