Inferam

AI and Equities | Pricing the AI future | Making AI compute-light | Discovering paths to AGI | Rank 1, Institutional Investor and AsiaMoney

[Inferam @InferamAI]

Some opportunities test your ability. Others test your patience.

[Inferam @InferamAI]

Does hyperscaler data center capex end up like dark fiber in 2001?

Back then, internet traffic grew exactly as expected, but bandwidth prices fell even faster, from $1,200 per Mbps to $90 in six years, and the fiber owners never earned back what they spent.

The same setup is forming in AI compute. Models will need 80-90% less compute for the same output within the next 1-2 years, which strands the GPUs all this debt is buying.

To be clear, I am not calling defaults, hyperscalers have cash-rich core businesses. The problem is return on capital. Credit seems to be figuring this out before equity.

[zerohedge @zerohedge]

"Cover Ratios for Hyperscaler Bonds Declining: The cover ratio measures how many dollars of investor orders a bond deal receives for every dollar of bonds issued. For hyperscalers, it has fallen from nearly 5x in February 2026 to below 2x in July, suggesting investors may need wider spreads to absorb additional hyperscaler supply" - Torsten Slok

[Inferam @InferamAI]

300x cheaper in 3 years.

[Inferam @InferamAI]

Name one tool in history that made thinking cheaper and left humanity thinking less. Writing? No. Calculators? No. The internet? No. AI will not be the first.

[Inferam @InferamAI]

True, and that's why this gap is not priced. Labs can only announce what is verifiable, so public milestones systematically undercount progress in the taste-heavy layer: conjectures and definitions.

[Inferam @InferamAI]

The most repeated fear about AI is that it will make humans lazy thinkers. I think this is a weak argument. Full case in a few days.

[Inferam @InferamAI]

The most valuable skill is updating your thinking without defending your past.

[Inferam @InferamAI]

End to end workflow of an LLM agent:

User
|
▼
Application
(builds the context)
|
▼
LLM
(next-token prediction)
|
▼
Application
(parses the tool call, executes it)
|
▼
Tools
(PDF, Search, Python, Image)
|
▼
Tool Results
(appended to the same context)
|
|__ loops back to the LLM until no tool call is emitted
|
▼
LLM
(generates report text)
|
▼
Application
|
▼
Document Generator
(PDF / HTML / DOCX)
|
▼
Final File

The model never runs a tool. It only predicts tokens.
The application does the running.

[Inferam @InferamAI]

A clear objective removes more confusion than endless motivation.

[Inferam @InferamAI]

How Production AI Systems Actually Work.

[Inferam @InferamAI]

AI infrastructure discussions are increasingly about inference throughput, not training FLOPs.

[Inferam @InferamAI]

People who learn fastest aren't chasing answers. They're reducing blind spots.