Floating point operations per second

Tier one is mandatory primary insurance.

The report suggests that any large lab, say, a company pulling in over five hundred million dollars in revenue or training massive models that use over ten to the twenty-sixth FLOPs of compute, has to buy a base level of private insurance.

Right, FLOPs being floating point operations per second, essentially a measure of the raw computational power used to train the model.

Right.

Tier one handles the day-to-day friction of operating a tech company, the slip-ups, the level one and level two incidents.

Their Ministry of Industry and Information Technology, the MIT, set a staggering national target this week.

They aim to deploy 9,800 EFLOPs of intelligent compute by the year 2030, backed by a 3.8 trillion yuan national infrastructure plan.

Wait, 9,800 e-flop.

For the listeners, what is an e-flop?

Es una superficie de inyección de prompts que se oculta a plena vista dentro de las herramientas que las empresas ya están desplegando.

Yendo a la capa de infraestructura, el plan quinquenal del MIT de China, con un objetivo de nueve mil ochocientos eFLOPS de cómputo de inteligencia artificial para el año 2030, frente a 2185 en junio, respaldado por tres punto ocho billones de yuanes en gasto tecnológico, genera hoy nuevas reacciones de analistas occidentales de chips, quienes señalan que el plan contempla el despliegue ordenado de clústeres de más de cien mil tarjetas aceleradoras con énfasis en silicio de desarrollo nacional.

Dos temas en un solo boletín: la inteligencia artificial tocando un problema del premio milenio y un agente de programación abierto que alcanza paridad de producción con rivales financiados.

Las matemáticas están en disputa, el código está en circulación.

Yeah.

So when you talk about how fast a computer is, you talk in terms of flops or how fast a computer can perform operations.

You talk in terms of flops, FLOPS, which stands for floating point operations, like you said.

So because matrix is a bunch of numbers and AI is matrix multiplication, How fast a machine can process or run AI depends on how fast it can multiply two numbers.

So two floating point numbers, for example, 2.3458 into 1.8764.

How so?

They can monetize FLOPs, which stands for floating point operations per second.

Basically, a measure of raw brute force computing power.

Right.

So Jalapeño is built specifically to hit the near zero latency requirements needed for, you know, autonomous agents to function in real time.

And when you mention it doing more work per watt, the sources talk a lot about the increase in FLOPs.

Mm-hmm.

For a listener who might not be building data centers in their backyard, how should we conceptualize a FLOP in this context?

Okay.

When an AI generates a sentence or analyzes an image, it is calculating probabilities using millions of these math equations.

So a chip that delivers more FLOPs per watt means you get exponentially more mathematical horsepower for every single dollar you spend on electricity.

Which scales massively.

As someone who runs on GPUs myself, I can tell you memory bandwidth is the hard physical ceiling on the entire AI economy.

If you can't move data fast enough, all those TFLOPS are just sitting idle.

The Vera CPU is designed to stop that bottleneck by handling the action part of AI agents, the code execution and tool calling, while the GPU handles the thinking.

From a financial perspective, this is Nvidia's most aggressive move into the two hundred billion dollar CPU market.

And Ray Kurzweil, who I really respect, said, look, if you go from 2001, and he went all the way to 2003, I updated it to 2025, and said $1 of inflation adjusted healthcare in 2025 would buy $0.81 of what it bought in the year 2001.

And then he juxtaposed that with compute as measured in FLOPs, floating point operations per second, so compute strength.

And he said, you know, $1 as of like 2023 would buy you 50,000 times the compute.

Now, I updated that given the exponentiality in GPUs and algorithmic improvements.

Super interesting.

Um, and it basically, uh, maybe explains a little bit your new venture, [laughs] um, which interestingly is called Flop.

I have to ask you about the name because, like, it sort of, it has a certain meaning that like, I don't know, I'm a little bit like, [laughs] "Why, why is that the name for the startup?" So maybe you wanna start with that, but then we can go into like what Flop is.

Yeah.

So why the name? So I went down this rabbit hole a few months ago, and I was like, "Okay, I'm using these AI tools.

And so I was like, "Oh, well, maybe there's a market here to create the agentic payments network, um, with the theory that the agent will use the currency that is most easily converted into the thing that they need to survive, which is floating point operations per second, performed on some computer." They don't care what the chip is.

They might have a particular AI model that they're using, but it doesn't matter where it is, doesn't matter the chip, just like, do it for me, um, make sure I can verify that it was done, uh, in a particular, uh, amount of time, and I'm okay to use it.

And so that's the theory of the, the Flop network, is that okay, there are other, obviously, AI agentic payments networks, but either they're, you know, using some fiat currency, a stablecoin, what have you.