The release of reasoning models has unlocked a new scaling law for AI: the reasoning computing power scaling law. Recent mathematical proofs by experts like Sloan Award winner Ma Tengyu and Google Brain reasoning team founder Denny Zhou have demonstrated a profound reality: as long as the “Chain of Thoughts” (CoT) is long enough, a Transformer model can theoretically solve any problem.
In the quest for Artificial General Intelligence (AGI), it turns out that compute and test-time compute is all you need. But this revelation brings a massive, hidden problem to the surface.
The $15 Million-Token Problem While Chain of Thoughts allows Transformers to run much more effectively, this extended reasoning process is incredibly expensive. For example, the price to generate 1 million tokens using the o1-preview model is $15. In comparison, a less compute-intensive model like GPT-4o mini costs just $0.15 per million tokens.
This 100-fold price difference exposes the true bottleneck holding AI back: compute efficiency. Brute-forcing complex reasoning through traditional digital hardware requires moving massive amounts of data back and forth, consuming unsustainable amounts of power and driving up the cost of AI access. If we want models to truly reason at scale, we must completely solve the compute efficiency crisis.
The Inevitability of Analog Fifty years ago, when a young Steve Jobs saw a prototype of a graphical user interface at the Xerox Research Center, he instantly realized the future: “It was obvious to me that all computers would work like this someday… You couldn’t argue about the inevitability, it was so obvious”.
Today, the AI hardware industry is facing a similar moment of inevitability. It is increasingly obvious that traditional digital chips cannot sustainably fuel the reasoning capabilities of tomorrow’s AI. To push forward, computing must fundamentally change and take inspiration from the human brain.
The Vision of “Mortal Computing” Nobel laureate and Turing Award winner Geoffrey Hinton, widely regarded as one of the fathers of AI, believes that the inevitable future of intelligent systems lies in a concept he calls “Mortal Computing”.
This new computing paradigm completely shifts away from traditional digital architectures—which separate memory and processing—and moves toward analog computing that closely mirrors the biological operations of the human brain. The foundational building block of this future is the memristor (multi-level RRAM), a breakthrough device capable of mimicking biological synaptic functions.
Unlock the Era of Mortal Computing with TetraMem To make the advanced reasoning of models scalable and affordable, the industry needs a hardware revolution. TetraMem is pioneering this inevitable future with its breakthrough multi-level RRAM analog in-memory computing (IMC) solution.
By monolithically integrating analog computing memristor arrays directly onto CMOS driving circuits, TetraMem’s architecture entirely bypasses the energy-draining data movement of traditional digital chips. Operating much like a biological brain, these integrated memristor arrays can “grow like a tree in a self-similar fashion,” directly enabling the Mortal Computing paradigm envisioned by Geoffrey Hinton.
Don’t let legacy digital hardware and massive compute costs limit your AI’s reasoning potential. Discover how TetraMem’s analog in-memory computing solutions can deliver unprecedented efficiency and unlock the next era of intelligent computing today.