TL;DR
AI infrastructure, not the AI IPO wave, holds the bigger investment opportunity right now.
Anthropic, SpaceX, and OpenAI are racing toward public listings worth close to 3 trillion dollars combined. But only 5% of installed GPUs are actually running, because power and data center capacity have not caught up with demand.
This gap between digital growth and physical build-out is the real story behind the AI boom.
High revenue run-rates from AI labs can be misleading, since cloud deals let the same dollar count on both sides of the books. Construction delays and rising costs per gigawatt are pushing investors toward a new metric: compute per watt.
Key points
Only 5% of deployed GPUs are currently running.
Mistake to avoid: treating run-rate revenue as proven profit.
Practical takeaway: watch power and rack efficiency plays, not just model company valuations.
Critical insight
The real bottleneck in AI right now is concrete and electricity, not chips or code.
Table of Contents
Introduction
$ANTHZZX ( ▼ 0.02% ) just filed for an S-1 with a valuation of 965 billion dollars.
$SPCX ( ▼ 4.95% ) started trading on June 12 with a goal to raise 75 billion dollars.
$OPEAZZX ( ▼ 0.83% ) plans to go public in September.
This is the biggest wave of AI IPOs since the internet boom.
But there is a number that matters more than these headlines. Only 5% of installed GPUs are actually running. The other 95% sit idle because there isn’t enough power and data center capacity to run them yet.
This gap is the reason AI infrastructure deserves more attention than the AI IPO deals. The demand for AI is real, but the physical layer behind it, including power, data centers, and rack components, can’t grow as fast as software. This physical layer decides how fast every layer above it can grow.
I. The 5-Layer Model of the AI Industry
NVIDIA's Jensen Huang describes the AI economy using a 5-layer model, from bottom to top:
Energy powers the chips.
Chips sit inside the infrastructure.
Infrastructure runs the models.
Models power the applications.
Applications are where real economic value finally appears.
Each layer pulls demand from the layer above it. When an application takes off, it needs more models, more infrastructure, more chips, and more power.
This model makes one thing clear: money is flowing into the infrastructure layer faster than into the application layer.
II. AI IPOs Are Growing Fast But Profit Margins Are Still Unclear
AI model companies are no longer startups. Anthropic reported a run-rate revenue of about 47 billion dollars, up from around 10 billion dollars just one year ago. About 80% of this revenue comes from enterprise API contracts, not individual users.
This growth rate is even faster than Google's growth between 2003 and 2005, when Google's revenue grew from about 1.5 billion dollars to 6.1 billion dollars in two years.
1. Why High Run-Rate Doesn’t Tell the Whole Story
A high run-rate doesn’t always mean stable revenue. Anthropic has committed up to 100 billion dollars to AWS in exchange for an investment from $AMZN ( ▼ 3.46% ), and it has a similar deal with Google.
This means part of the revenue from AI model companies is actually money that cloud providers pay back so these same companies can buy compute. That makes the same dollar appear on both sides of the financial reports.
This is why the S-1 filing matters so much. The market needs to see real gross margins, customer concentration, compute commitments, and customer retention before it can value these AI IPO deals correctly.
2. Competition From China Is Squeezing Margins Too
Competition from China is also putting pressure on profit margins for Western AI model companies. Alibaba's Qwen has already passed Meta's Llama in total downloads on Hugging Face.
According to estimates from Andreessen Horowitz, about 80% of startups using open-source models now run on Chinese base models like DeepSeek and Kimi, which match the performance of US frontier models at a much lower cost.
III. Infrastructure Build-Out Is Falling Behind Demand
Demand for AI infrastructure is now far ahead of what can actually be built.
1. Construction Timelines and Costs Are Both Slipping
According to satellite analysis from JPMorgan, more than 60% of data center capacity planned for 2027 hasn’t even started construction. Among projects planned for that year, only 6.3 GW is actually under construction, compared to 21.5 GW that has been announced.
Construction costs are also rising fast every year. Jensen Huang said a 1 GW AI factory now costs between 50 and 60 billion dollars, and this number is heading toward 80 to 100 billion dollars. Goldman Sachs forecasts total AI infrastructure spending will reach 8 trillion dollars over the next six years.
2. Physical Barriers Are Slowing Down Construction
Several specific barriers are slowing down construction speed:
Grid connection queues in the US now take three to five years.
Large power transformers need 18 to 24 months to produce, because global manufacturing capacity is still limited.
Local communities near data center hubs are pushing back against noise, water use, and rising energy costs.
Software can scale in a matter of weeks, but the physical layer underneath it, from land to concrete to cooling systems, still has to follow normal construction timelines. The gap between these two speeds is the clearest investment opportunity in the AI industry right now.
IV. Compute Per Watt Becomes the Key Metric
When companies can’t build faster and can’t get more power, energy efficiency becomes the deciding factor.
1. 800V DC Architecture Cuts Power Loss in the Rack
The traditional power delivery chain loses about 17% of energy before it reaches the GPU. 800V DC architecture removes several conversion steps, raising efficiency from about 83% to over 92%, and cutting copper use by 45%.
NVIDIA's Vera Rubin, which just entered full production, runs on the Kyber rack architecture, which uses this 800V DC standard.
2. Power Density Per Rack Rises With Each Chip Generation
Power density per rack is rising fast with each chip generation: Hopper used about 40kW, Blackwell raised that to 120kW, and Vera Rubin is expected to reach 600kW to 1MW, a 25 times increase in three years.
Components that handle voltage conversion, data transfer, and heat insulation are now as important as the GPU itself. Jensen Huang has invested directly in Corning and Lumentum (optical fiber), Infineon (power semiconductors), Marvell (optical technology), and Qnity (materials for 800V DC).
Conclusion
The wave of AI IPOs from Anthropic, SpaceX, and OpenAI is grabbing all the market's attention, but the real value sits in the AI infrastructure layer underneath, from physical infrastructure to compute per watt.
With 95% of GPUs still sitting idle due to a lack of power and data centers, the clearest investment opportunity belongs to the companies solving this exact bottleneck, not to the most talked-about IPO deals.
Whether the stock already reflects that positioning is a separate question. The business case itself is specific, verifiable, and tied to trends that aren’t going away.
You remember our prediction that Bitcoin would return to $80K when the entire market believed BTC would hold $100K and continue moving up.
And we’ve shared high-potential tokens that are positioned for 200% growth in one month, while the broader market looks quiet and sluggish.
This series will be updated more frequently in the PRO edition moving forward.
Monthly Plan: Was
$29/mo→ Now $3.99/moAnnual Plan: Was
$199/yr→ Now $29/year 🤯
Rate us today!
Your feedback helps us improve and deliver better Crypto content!
Key Takeaways
Only 5% of installed GPUs are running, mainly due to lack of power and data center capacity.
Anthropic's run-rate revenue hit 47 billion dollars, up from about 10 billion a year ago, with around 80% from enterprise contracts.
High run-rate doesn’t guarantee real profit, since deals with AWS and Google let the same dollar show up on both sides of the books.
More than 60% of data center capacity planned for 2027 hasn’t started construction, while costs per gigawatt keep rising.
800V DC architecture and rising power density per rack are pushing compute per watt into the center of the AI infrastructure race.
⚠️ Disclaimer: This newsletter is for informational purposes only, just for fun and knowledge. This is not investment advice. Your money, your responsibility!
If you’re interested in other topics and want to stay ahead of how Crypto is reshaping the markets, from whale strategies to the next major altcoin narrative, you can explore more of our deep-dive articles here:
*indicates premium insights available to Pro readers only.
