The machinery on the assembly line is world-class. On paper, it can produce an enormous volume of goods per hour. And it does.

Yet, the business still misses its targets. Why? Because outcomes are rarely limited by the assembly line itself.

The supply of raw materials to the machinery and the delivery of finished products to customers play equally vital roles. Let me be a bit poetic here:

Raw materials arrive late, the line waits. Finished goods pile high.

In the end, delivery is routes and time. 

And when movement is the game, the bill runs high. 

Each trip costs energy, in money and time.

This is the central lesson: world-class machinery does not guarantee a high-performing, high-throughput factory.

• Speed is not just how fast the machinery can produce.
• Latency is the total time from order to delivery.
• Energy efficiency is about the total cost of keeping the whole operation moving.

They are related, but not the same problem, and they all contribute to overall performance.

The Compute Analogy

This is a perfect analogy for modern computing. The CPU and GPU are the machinery on the assembly line. They are very good at the arithmetic that turns data into answers.

However, many modern workloads are limited by the supply chain and delivery equivalents in chips and semiconductors. Data has to travel from storage to memory, from memory to the processor, and back again. The raw materials—data—spend a surprising amount of time in transfer before they become finished products in customers’ hands: answers.

That transfer time creates a triple threat to performance:

• It hurts speed because the processor stalls while waiting to be fed.
• It hurts latency because the system spends time moving data before it can produce an answer, and then spends time delivering that answer to where it is needed.
• It hurts energy efficiency because moving bits costs power, dissipates waste heat, and repeated transfers compound the cost.

From Graphics to AI: Changing the Bottleneck

Remember, GPU stands for Graphics Processing Unit. It was originally designed and optimized for graphics—first for gaming in the 1990s—and later for other math-heavy tasks. The bottleneck back then was arithmetic, and GPUs were the solution that gave us the most bang for the buck.

But modern workloads—AI inference in particular—have different characteristics, which makes the bottlenecks show up differently. The computational characteristics of inference put immense pressure on memory behaviour and data transfer. In many cases, the limiting factor is not just the math anymore. It is the movement and the waiting. And the delivery problem is getting bigger too.

The Rise of the “Edge” Factory

Sensors are everywhere now. They generate raw data where the action is. If every sensor stream has to be shipped to a distant “factory” (a data center) before anything useful happens, latency and bandwidth also become part of the product.

That is why edge computing is increasingly important. It is the computing version of building smaller factories closer to customers and shipping less raw material—or sometimes no raw material—across the network.

Investing in New Architectures

Of course, this does not mean CPUs or GPUs are obsolete. It means there are many other bottlenecks now. We need:

• Less distance between memory and compute.
• Less shuttling of data inside the system.
• Less distance between sensing and decision.

The Von Neumann architecture used by many modern computers today is about 80 years old. The first CPU is more than half a century old. The first GPU is almost 30 years old. It is time for new architectures.

This is a core part of our investment thesis in the next frontier of computing and its applications, specifically in Advanced Computing Hardware, one of the five areas we invest in. For many years, we have made investments in semiconductor companies, including Zinite, Hepzibah, ABR, and Blumind, that, through architectural innovation, address performance bottlenecks across speed, latency, and energy efficiency in ways faster GPUs alone will not solve.

We are super excited about this massive opportunity and are looking for new investments. If you are a deep tech researcher or founder in this area, please reach out to us at pitch@twosmallfish.vc.