[Michael Eisenberg — 0:00]

What did Jensen playing at Nvidia miss, and that you figured out that is so novel?

‍

[Elad Raz — 0:05]

No one thought about it. We are now running workloads 10 times faster at half or even quarter of the power consumption of a GPU. NextSilicon is not a feature. At the beginning, most people don't understand the uniqueness around what we have built. What does it mean to program a chip for the future?

‍

[Michael Eisenberg — 0:24]

If you're engaged in AI, this is just about the hottest topic of our time right now.

‍

[Elad Raz — 0:29]

Absolutely.

‍

[Michael Eisenberg — 0:30]

So the big national lab supercomputer, one of them is built on NextSilicon now?

‍

[Elad Raz — 0:35]

Yes.

‍

[Michael Eisenberg — 0:35]

And how powerful is it?

‍

[Elad Raz — 0:37]

We have our own tricks over there and we are going to reveal them soon.

‍

[Michael Eisenberg — 0:00]

Ooh, want to reveal them now?

‍

[Michael Eisenberg — 0:00]

Welcome back to another episode of Invested. I'm thrilled to have with me Elad Raz, founder and CEO of NextSilicon. Welcome, Elad.

‍

[Elad Raz — 0:00]

Thank you. Thank you for having me.

‍

[Michael Eisenberg — 0:59]

Okay, we've done a few of these shows which I would call at the kind of more hard science, geekier end of the spectrum. This will definitely be one of those. So we should warn our audience in advance. Get ready to like, dig deep into the world of semiconductor, semiconductor architecture, data centers and the general need for compute in the world. Is that fair?

‍

[Elad Raz — 1:21]

Absolutely. And tell you more. You know, when I first had investor meetings, you know, I started with the deep tech. I mean, whatever episode you had about science and, you know, getting down, probably to understand NextSilicon, you need to get deeper. But it's not a warning. I think that we can, you know, convey the message about what NextSilicon is in a more generalized way that everyone will understand. So don't be scared.

‍

[Michael Eisenberg — 1:48]

Okay, so just to start with something a little lighter, tell everybody what was engraved on your first chip that you, that you produced or taped out.

‍

[Elad Raz — 1:58]

There is a habit, you know, when you produce a silicon chip and that's like for the past 40 years that you can play around with the transistors and the etching on the chip. You know, the way that the photoresist mask is, is laid out and generate symbols. And many companies have done it. Pat Gelsinger for example, wrote his initials on the 486 and 386 — PJ and others put cartoons and train. And when we started NextSilicon and we came to tape out, tape out is the term that you take your design and send it to TSMC, you had to decide what you want to etch over the chip. And I thought about it really hard and I went to the shul [synagogue] and go to the rabbi and ask him, "Hey, what do you think?" And he said, "Why wouldn't you choose the first word of the Bible? Bereshit." Bereshit, in the beginning, Genesis. That's in the holy scripture, in the Zohar, it says that this world encapsulates everything that was and everything that will be. Sounds appropriate. So we have done it. And this is the first 7 nanometer smallest word of bereshit in, I think in the entire world.

‍

[Michael Eisenberg — 3:11]

So given that it's everything that was and will be, that's quite a statement about compute. Let's just start in a reasonably simple place like why do we need, you know, why don't you give the, you know, three sentence pitch on what NextSilicon does. And then I want to talk about why does the world need so much more compute?

‍

[Elad Raz — 3:27]

Sure, we are all about compute. Compute entails many, many different things. But in order to win that race, I mean, think about it for a second. You start a company that focus on compute. Accelerate compute. Accelerate compute for science, for AI machine learning, for all of those different aspects. You need to build the company or design the company and start the company to be novel, original and very different than other multinationals that spends billions of dollars on their design. You can't start a company that trying to copy and just do incremental improvement over the leading companies, these days is obviously Nvidia, spending tens of billions of dollars on every chip design. You need something else. So that's our goal, you know, reinvent compute.

‍

[Michael Eisenberg — 4:24]

Reinventing compute. But what is that? Tell us in three sentences.

‍

[Elad Raz — 4:27]

Sure, in three sentences. Computers until now work in terms of processor that read instruction, execute instruction, right? So you do it like serial way. And there is a way to increase throughput by replicating many of those processor calls. At NextSilicon, we work in a different way. What we are saying is we build a chip that has the arithmetic unit on the chip itself in some kind of a grid. Very hard to program that. But we find a way to use algorithm, software algorithm that understand what matters now and literally reconfigure the silicon. Etch whatever your workload, you are running, scientifical code, AI machine learning, into the different transistor and then stream the data in. That's like the scary part. But we are going to...

‍

[Michael Eisenberg — 5:21]

We're going to unpack it soon.

‍

[Michael Eisenberg — 5:23]

Now, why does the world need much more compute today?

‍

[Michael Eisenberg — 5:27]

But you should explain it.

‍

[Elad Raz — 5:28]

This question was asked every time in the past eight years from first computers ever invented. Why do we need that? Why we need the big machine that cracked Nvidia? Suddenly you got the Enigma, a machine cracked by computers. You can't do it handcrafted. So there was a lot of money driving into compute in the early days from Department of Defense and Energy. Even here in Israel, 1953 the Israeli state was 5 years old. Weizemann Institute – let's build the first computer based on the Enigma. The name of the computer was WEIZAC. It cost 25% of the Weizmann Institute that year. You know how many discussions there was with the first president, Chaim Weizmann, about why do we need that machine? Who is going to use it? Now we want to get like this genius Pekeris bomb to work on that machine. But who else is going to use it? And guess what? Two years after, when the machine was powered on, everyone just queued in line to use those machines. Even on weekends, on Shabbat, they had to bring like a goy [gentile] of Shabbat, someone to operate that machine. So for around the history there was a question why do we need more compute? Even when I started NextSilicon, I spoke to one of my relatives and he asked me hey "Elad, my computer works just fine. I mean, why do you need a stronger computer?" But the answer is, and especially now in AI, the answer is kind of obvious. You want to send an agentic task, let the GPUs and you know, hear the fan rolling out and then token is spinning on your monitor. You need that faster. The more compute you can generate, the better result that you're going to get from your queries. Whether it's AI or even science, fidelity on oil and gas and more.

‍

[Michael Eisenberg — 7:21]

Okay, so if we look at the world. So this was eight years ago when you decided to do this. Today it's obvious, you know, eight years ago was around the founding moment OpenAI but nobody knew it existed. So at the time you made a decision, I think, and this is all above my pay grade, right, in terms of technological complexity, you made a decision to focus on HPC, high-performance computing. Try to tell our audience what was unique about the chip, your approach and what is unique about your desire to go after high performance computing. And we'll distinguish that in a second when you're done to what's going on in AI today.

‍

[Elad Raz — 8:01]

So let's start with the underlying core technology. We have produced two generations of chip that accelerate computationally, period. The way that we are doing that is using software algorithm that change the wiring on the chip, rather than have a general purpose chip that can get instruction and execute them.

‍

[Michael Eisenberg — 8:23]

Okay, stop for a second. Explain that to everybody. The difference between software that changes the quote unquote wiring on the chip versus a general purpose computing architecture, whether it's a GPU or a CPU.

‍

[Elad Raz — 8:35]

So let's start with the basic. Okay. And I'm going to use a name that I never use when I talk to investors, which is John von Neumann. He's a big scientist that explains the fundamental of a processor core. So you have a processor core, it has input, some memory state and an output that get instruction. Those are processors of today. You get the instruction of what to do, that's the program of the user, and the device itself, just run the instruction one after the other. Processor got more complicated, more vectorized, meaning you can do more operation, multicore everything. You know all of that and we won't get into the details. So don't worry. Those are processor core, traditional architectures. With NextSilicon we just threw it away. We said, hey, let's look at the processor code. There are so many logic that decode the instruction and try to understand hey, what those instructions means. How can I fetch the memory in order to execute those instructions. And until I get to the actual unit that compute the, you know, the schoolbook, multiply multiplication or the addition, like 2% of the transistor of the chip is, in CPU, allocated for compute. GPU is way better. We said, let's just stick with those transistors that doing the computational. That's it. Now the problem is, okay, so how you don't have the instruction how can you tell those different units what to do? So the answer we, and this is the IPs that we have developed, is saying, okay, let's connect the different ALUs in some kind of a mesh and we're just going to reconfigure the mesh to move the wires from A to B. Still, how can I program this thing? And there are like two aspects. One of them is, okay, let's go to the sophisticated user and ask him to draw on a circuit board exactly what to do. By the way, this is what literally when Los Alamos developed the A bomb, that's exactly what they have done. They put like bunch of calculator women. They were women, that took like two numbers, did an operation and passed it back to the different O and then you get the entire computation. We decided to say let's understand the computational in autonomous way. I mean the user wrote the program in a C code, let's run it on a CPU, let's understand the different computational kernel and automatically, autonomously just wire them on the chip, move the memory and accelerate, accelerate those computation. That's what we have done. So instead of working in 2%, we are working at close to 60%.

‍

[Michael Eisenberg — 11:25]

So 30 times the throughput on the chip.

‍

[Elad Raz — 11:27]

Exactly.

‍

[Michael Eisenberg — 11:29]

30 times. Three zero?

‍

[Elad Raz — 11:32]

Even, we have some workload that you get to 100. The problem is Amdahl's Law, meaning you have overhead moving the data in and out that you need to think about it like pipes. Okay, you can, you can put like a wide, a wide pipe of water, but if you have like thin pipes entering, you can't drive the machine as much as you want. So Amdahl's Law is saying that you're probably going to get the bottleneck not in the computation, but in the I/O and the data, everything.

‍

[Elad Raz — 12:01]

But yes, we are now running workloads 10 times faster at half or even quarter of the power consumption of a GPU. End to end.

‍

[Michael Eisenberg — 12:10]

Right, compute per watt is up 40x?

‍

[Elad Raz — 12:18]

Yes, 40x if it's 10x and a quarter of the power consumption.

‍

[Michael Eisenberg — 12:22]

Gavin Baker, who by the way, is one of our most popular episodes ever. And for anyone listening now, if you haven't listened to our Gavin Baker episode, you need to listen to Gavin Baker episode. Gavin Baker's an investor in NextSilicon. And one of the things he said on our show was that what you're doing is incredibly novel. It's very, very different. Explain to our audience what is so novel about it. What did Jensen playing at Nvidia miss, and that you figured out that is so novel?

‍

[Elad Raz — 12:47]

No one thought about it.

‍

[Michael Eisenberg — 12:48]

What does that mean?

‍

[Elad Raz — 12:49]

No one in academia, no one in industry. I mean, the concept of dataflow itself existed in literature. People tried to do that, but everyone failed.

‍

[Michael Eisenberg — 12:59]

What is dataflow?

‍

[Elad Raz — 13:00]

Dataflow is the way that you took the arithmetic unit and connect them together like delta that flowing through the device. That's the origin of a data flow. And every other architecture in the world today is based on... For example, GPU. There is something called SM. That's a processor code that run instruction and tensor cores doing matrix operations. Okay? We have just data flow. Just data flow. And some company tried 20 years ago to make this technology work. All of them failed.

‍

[Michael Eisenberg — 13:35]

Who was that?

‍

[Elad Raz — 13:36]

There were companies, I mean even HPE had the concept of the machine that there was like a memory in the center and then dataflow trying to fit it. Intel tried to do it with their spatial processor. I mean, many companies tried and the failure was because of the lack of programmability, the lack of software, it was very hard to program those workloads. Now the reason, let's take you back to the original decision. I mean why HPC? What is HPC? Why did you take the company into HPC. If I gave you like 10 minutes explanation about technology and I hope that you can have a skip button and just summarize a different.

‍

[Michael Eisenberg — 14:18]

Okay, no people are interested. You know, this is this. You know, if you're investing in the stock market or you're engaged in AI, this is just about the hottest topic of our time right now.

‍

[Elad Raz]

Absolutely, yeah.

‍

[Michael Eisenberg]

Go ahead.

‍

[Elad Raz — 14:27]

Nothing in the technology is oriented into a specific vertical. Not oil and gas and how you can find new pocket of gas and oil underneath like geoscience application beneath planet Earth. Nothing is related to space and have a prediction about how many dark matter there are and how galaxy was formed. Nothing about fluid dynamic that explains how the SpaceX rocket is being launched. I mean nothing is tailored into a specific market or even AI machine learning. Because at the end of the day AI and HPC uses linear algebra, you know, matrix multiplication, matrix vector operation, point-wise kernel. I mean the compute is the same. Nothing is aimed in, nothing in the  technology is blocking you from running any, any type of science or neural network. But eight years ago the world was different. I mean people forget how fast we have moved. Today architecture is transformers. That's the name of the big LLM. Whether it's Claude Opus or the new model or Groq, OpenAI with the codex in the agentic one or the GPT 5.4. So it's all about science. Eight years ago we were in a different space where the first AI accelerator were there, that's the one that started and the model were convolution, neural network or taking an image and classify it. Whether it's a dog or Chihuahua, cat or a dog. The reason why you start a company is because you geek out, like you want to build the technology. This is why you start a company. But at the end of the day every company needs to generate cash. So you can go today on machine learning eight years ago. But the market was undefined and Google had their own TPU back then. So it was very, it wasn't clear how you can generate money. Even now. But back then it was way more serious. And then we said okay, so we have novel compute architecture. Let's go to the basics. Every compute company, big one, started with to sell to the HPC market. Intel, AMD, Texas Instrument, even Nvidia. Yeah, you start with graphics but you know the big computation started with supercomputers. Build the supercomputers and we, Baruch Hashem [thank God], build a big supercomputer with the Department of Energy.

‍

[Michael Eisenberg]

In the United States. Yeah. It's a joint venture that primarily built at Sandia National Laboratory but it's a collaboration of Lawrence Livermore and Los Alamos and Sandia. Yeah.

‍

[Elad Raz]

In the United States. It's a joint venture that primarily built at Sandia National Laboratory but it's a collaboration of Lawrence Livermore and Los Alamos and Sandia.

‍

[Michael Eisenberg — 17:18]

Yeah. So the big national lab supercomputer or one of them is built on NextSilicon now.

‍

[Elad Raz — 17:23]

Yes.

‍

[Michael Eisenberg — 17:23]

Okay. Yeah. And how powerful is it?

‍

[Elad Raz — 17:27]

So I, I don't know when this episode is going to air but you know big news are coming soon. It's big. It's like in the top 500.

‍

[Michael Eisenberg — 17:37]

Top 500 supercomputers in the world?

‍

[Elad Raz — 17:39]

Yeah. Will be published but I'm not trying to reveal when what position

‍

[Michael Eisenberg — 17:44]

And in terms of cost of compute, when you compare it to the other 500 in the world.

‍

[Elad Raz — 17:48]

So you said 40x faster.

‍

[Michael Eisenberg — 17:50]

40x faster per watt.

‍

[Elad Raz — 17:52]

Yeah. And dollar.

‍

[Michael Eisenberg — 17:55]

And dollar. Well that's a step function change. But we'll come back to this in a second. So you decided to go into HPC because everybody starts in HPC but Jensen and Nvidia have gotten out of the HPC business because the AI business, or let's call it the LLM business or or transformers, is much more lucrative right now. Why don't you do that?

‍

[Elad Raz — 18:17]

We are.

‍

[Michael Eisenberg — 18:18]

Oh, you are doing that.

‍

[Elad Raz — 18:19]

Yeah.

‍

[Michael Eisenberg — 18:19]

Is that easy? Just like you say. Okay, I'll turn off HPC and turn on transformers.

‍

[Elad Raz — 18:25]

Yeah. Way more easy than we thought.

‍

[Michael Eisenberg — 18:27]

Why is that? Let's back up one second. Okay, so let's make some order for our listeners here. There are GPUs which is the game Nvidia is playing. There's TPUs which is the game Google and to some extent Amazon is playing.

‍

[Elad Raz — 18:43]

Let's take it in more of high level and understand what the user wants to do. And then on that we'll name the different accelerators because you have different workload. Some user want like edge computation saying here's an image, I want to classify it or I want to do voice to text. Okay. Kind of a wispr. Or diffusion models. Those models, transformer models are typically have hundred to billion parameters. Tiny models compared to the big LLM of a trillion parameter. And then the question is whether should I should I buy $10 million rack to accelerate those tiny models or I can go to hundreds of startup doing small cheap, low power, low cost to accelerate them. So this is one class of computing. And then you have, I think that the biggest range there are models that between 200 billion parameter to 2 trillion parameters. Because think about it, most users who adopt AI now are the early adopters of AI technologies, are software programmers with the Claude Code, and Codex, and even Gemini have a good kind of agentic solution for that. And they send a request, they want you to read the entire code base to get to millions of tokens. They want that their GPUs will crunch those models and speed up lots of token. That's the main and the prime use that we have today. And even agentic task of read all my contract and try to summarize that, you want a trillion thinking model. You want a big model that think that generate a lot of tokens, token is the little word, and read a lot of tokens. Those are the classical model. Between them, you have most of the hyperscaler and Google chasing after. TPU, the Nvidia Rubin, the new chip, or rubinultra.com, the Amazon Inferentia and Trainium. They're all around that section. And it used to be hey, there is like a classification between inference and decode. Today the nuance is more is it a pre-filled training chip or decode inference chip. And there is with the Groq acquisition there is another model of Cerebras.

‍

[Michael Eisenberg]

One second because I think that's important. I think when Gavin was here, it was right after or right before Nvidia bought Groq, right? Explain why did Nvidia buy Groq, why that's important, number one. And number two, explain Cerebras has raised a lot of money recently. I mean, it's an incredible story, Cerebras. Right? It almost went public but there was customer concentration, maybe some other issues. They pulled back from the IPO. It's now a deca-billion dollar company – or 20 or 30 billion dollars, I think. And raising lots of money in the private market. So let's just try to take our listeners through what the differences are.

‍

[Elad Raz — 21:53]

Sure. So let's classify them into these three categories. Training and inference that everyone spoke about it few years ago. People understand that there are few companies that train model and pre-tune a model. Meaning taking an existing model, getting your custom data and just tweak it. It's not something that costs a lot of money. You can do it anywhere.

‍

[Michael Eisenberg — 22:14]

Who does that?

‍

[Elad Raz]

The pre-tune?

‍

[Michael Eisenberg — 22:17]

Yeah.

‍

[Elad Raz — 22:17]

There are many companies that take open source model.

‍

[Michael Eisenberg]

Right.

‍

[Elad Raz — 22:20]

Whether it's the Chinese model, the Alibaba model or the Llama model and others that says okay, here is a model that is a base model. Base model is not the one that you ask him a question.

‍

[Michael Eisenberg — 22:34]

What do I do the pre training on?

‍

[Elad Raz]

Their proprietary data. That's open source.

‍

[Michael Eisenberg — 22:38]

So that's the data side. What kind of compute do I do it on?

‍

[Elad Raz]

So it depends on the volume but like single rack.

‍

[Michael Eisenberg]

Basic compute.

‍

[Elad Raz — 22:42]

Basic compute. Training model from scratch. It's super expensive and super big and there are few, few companies that are doing that. And then inference is a phase that, I mean GPUs is not one size fit or even within GPUs you have differences mostly around the memory and memory cost and capacity and bandwidth. When a user give you a query like, give you a paper in the LLM in the transformer, there are two phases. The first phase is called prefill. Let's take everything that the user give. I'm going to simplify the hell out of it. Let's do a big matrix multiplication. And here is the context. That's the prefill. It digesting the data into some kind of prompt. And decode, this is the phase that everyone loves to talk now especially to understand what Groq and Cerebras is, is the one that once you have the context generate a lot of small tokens. So the prefill phase is very similar to training because you are doing a big matrix multiplication. You don't need a lot of memory bandwidth. So you don't need expensive memory. That's like the digest of that. That's why you have chips that are dedicated for training or prefix. The workload is attributed by a big matrix and a vector. The problem with a big matrix and a vector, you are memory bound. You just stream the data from the metrics. You're doing the FMA, you're doing some point-wise kernel and reduction. That's it.

‍

[Michael Eisenberg — 24:21]

Okay, I'll explain what you just said to everybody.

‍

[Elad Raz — 24:24]

Memory bound means how fast. I mean everyone understand what token per second is how fast, you know, the screen is going to be. Okay, memory, different type of memory. You have SRAM, HBM. You know, SRAM is the one that Groq and Cerebras are all about, is how fast can I get the data from the memory or the context. The KV, the pre-fill phase, KV and I'm going to do the computation. You're not compute bound, you are memory bound. If you're using a cheap DDR, you're at around 1 terabyte of data movement. If you're using HBM, you're 30 terabytes, like 30x faster. So you can do more compute. If you are inside the SRAM, you can get to 150 terabytes way faster. Meaning you can speed the decode token faster.

‍

[Michael Eisenberg — 25:11]

And that's why Cerebras is using those giant wafers also.

‍

[Elad Raz — 25:14]

Yes, exactly. But the dumbest thing is, you know, no one calculate the cost.

‍

[Michael Eisenberg — 25:20]

What does that mean?

‍

[Elad Raz — 25:22]

So different memory has different type of cost. Okay. No one invent a new memory. The Cerebras didn't invent SRAM and neither Groq. I mean it's just existing memory that everyone use. We use SRAM. The question is how much memory you are using. HBM is expensive. HBM required advanced packaging. Meaning that in supply chain, when you generate the chip, when you generate the Maverick chip, our chip, your bottle neck...

‍

[Michael Eisenberg — 25:49]

Your chip is called Maverick.

‍

[Elad Raz — 25:50]

Exactly.

‍

[Michael Eisenberg — 25:50]

Not Genesis, it's called Maverick.

‍

[Elad Raz — 25:52]

Maverick. You have different type of memory and every memory has a different cost related to it. HBM are very expensive, but they have a medium capacity. DDR have low memory bandwidth, 1 terabyte relatively cheap, but you can get 2 terabyte of data. SRAM is hundred times less dense than that. Meaning they cost hundred times more because you need more chip to build them. Those are the three categories. Now you want, the question as you said is how do you measure your performance? Is it performance per watt? And you forget about the CapEx. You forget about the dollar amount that you put inside the chip. Just performance or performance per watt per dollar. Okay. CapEx is not free. You need to invest billions of dollars to build those data centers. That's why you see data centers that are still running Hopper. It's not that...

‍

[Michael Eisenberg]

Hopper is a previous generation of chip.

‍

[Elad Raz]

Yeah, exactly. I mean when Blackwell launched Jensen said, "Throw away the Hopper, buy Blackwell." CapEx is not zero cost.

‍

[Michael Eisenberg — 27:05]

Well, it also turns out that these chips, because of that, it turns out that these chips have a longer life than anyone thought before. You know, they'll depreciate over I don't know, seven or eight years and not three or four years.

‍

[Elad Raz — 27:15]

True. Yeah, but it's also related that in the decode phase it's all about the memory and memory technology didn't progress as

‍

[Michael Eisenberg — 27:22]

fast as the chip.

‍

[Elad Raz — 27:23]

So you can put way, way, way more compute. Jensen's Law is like 40x, you know, more compute. Right. Can you utilize the memory as well? Because that's the bottleneck. And then you need to build a big rack and connect them and doing collective to take the memory into one place. So the question is how you generate token per second. Token per second is the measurement that everyone loved. This is what started the LPU and Jonathan was the founder.

‍

[Michael Eisenberg]

What is LPU?

‍

[Elad Raz]

LPU is the name of Groq architecture, meaning there is a SRAM. You put the weights of the transformer there and you just trim the data and you can generate, you know, lots of token when you are SRAM. The problem is you need for a trillion parameter model. You need 2,200 of those chips. Super expensive. You need to normalize that.

‍

[Michael Eisenberg — 28:17]

What does it mean to normalize that?

‍

[Elad Raz — 28:19]

Normalize that is in the last GTC. I mean we are now in doing this recording after Passover in April. Eight months ago there was the GTC when Jensen spoke about the Groq acquisition and

he had like a nice chart saying we Nvidia Rubin, that's the name of the Nvidia GPU, are in the mid range, you know, until 2 trillion parameter model with half a half a million context window. But if you want faster token per second, here is a joint effort between rec of Rubin and rack of Groq LPU that combine them together and get 15 times more token per second. Great. How much it cost? 50 times more. Instead of $6 per 1 million token, 9 times more expensive, it's going to cost you $45 per million token. Normalize mean will user are going to pay the price to run those models faster, you get two times more token per second at a cost of nine times more dollar. That's the big question around the Groq, Cerebras. I mean will the price make sense for user to pay this extra price for slightly faster models? That's the big question around the SRAM based.

‍

[Michael Eisenberg — 29:43]

Okay, so said simply, and then we'll come back to NextSilicon. Said simply, everything is measured right now in AI land, not an HPC land, but AI land. It's measured in cost per token. The new architectures which combined the Nvidia GPUs alongside either Groq or Cerebras to increase memory footprint, we'll call it, memory capacity has increased the cost. It's doubled the speed but increased the cost of producing a token significantly. What are the applications that people are willing to pay that for?

‍

[Elad Raz]

Yeah.

‍

[Michael Eisenberg]

Okay. Enter NextSilicon now.

‍

[Elad Raz — 30:22]

Enter NextSilicon. Yeah, so in NextSilicon we obviously can play around the entire range. We have decided to focus.

‍

[Michael Eisenberg — 30:33]

What does that mean you can play around the entire range? What does that even mean? Nobody can do everything.

‍

[Elad Raz — 30:35]

Yeah, absolutely. You have, you need different products to different market. You need the SRAM base, you need HBM based, you need LPDDR for the pre-fill or training. That's like different type of product. The core technology itself is how to run compute faster. And then the question that you need to ask me. Well Elad, who cares? I mean if you said, you just told the entire audience that AI is memory bound problem. So take a memory, maximize it. That's the result that you can get. How can you run it faster?

‍

[Michael Eisenberg]

Right. There goes that Amdahl problem again.

‍

[Elad Raz — 31:18]

Exactly. Yeah but this is like short memories of let's design a chip that looks on past performance and not future. What does it mean to program a chip to the future? So a, understand that or the audience – you need to. You are an investor at NextSilicon. You know how painful it is to build a semiconductor company. Just a perspective, $150 million for every ASIC design. Meaning that the company burns $150 million and then you get a first sample from TSMC, from the foundry. There can be mistakes and mistakes are expensive and then you need to go to the drawing board and do another revision. It's like super equity capital expensive. It's not easy. I'm never going to build a semiconductor company again.

‍

[Michael Eisenberg — 32:10]

Because Jensen's still doing what he's doing 30 years. I expect that you'll be in a rocking chair before you get another chance.

‍

[Elad Raz — 32:17]

Oh probably. Okay. So that, that's expensive. So when you design the chip, you need to design a chip for three years into the future rather than three years in the past. Today architecture are transformers. Transformers are bottle neck by HBM memory or SRAM memory. But algorithm change quickly. Do you remember in the beginning we said well first AI model came it was convolution, neural network or image classification until the chatgpt moment that people say this thing works and we can, you know, unlock the imagination. And then a lot of people spoke about AGI and other... at the end of the day, whether it's Yann LeCun AGI, this doesn't matter. Algorithm change. With algorithmic change you can get way, way better performance than the actual chip. At NextSilicon we don't care what workload is running, whether it's AI decode or HPC code. Why this thing is so powerful, because future models 28, 29 probably not going to be the transformers we know today. Even now there are papers at Google and even Nvidia that says let's compress the memory and, because we are memory capped, so let's compress the memory, get to the computational core, decompress that KV cache and then we can do more work.

‍

[Michael Eisenberg — 33:49]

Okay. Said simply – what you're betting on is the following. Because I am the fastest, bestest, cheapest compute processor on planet earth. Are you by the way?

‍

[Elad Raz — 34:02]

Yes.

‍

[Michael Eisenberg — 34:04]

Okay. And everybody else is memory balance. Why don't we you know, tighten the pipes a little bit or you know, compress that so we can jam more memory into the system, push it through NextSilicon, do some magic on it and basically take advantage of the increased compute and not be memory bound anymore.

‍

[Elad Raz]

That's an example that we are not going to compress it. The users are doing those research now and show 8x better performance just by memory compression.

‍

[Michael Eisenberg]

Even without that?

‍

[Elad Raz — 34:22]

No, no. even without NextSilicon, on a GPU because it's easier to program. I mean building an AI accelerator just to run transformers is the dumbest idea ever because algorithm change. I'm betting that three or four years from now you will see more complex algorithm that accelerate the computation or demand 1000 times less computation and memory bandwidth that we can run those models 1000 times faster. Now this 1000x number is important. Why? Because your brain operate at 20 watt. Those big racks are, or GPUs, are in 22 kilowatt per operation and same operation cost 1000 times when you run them on GPU. That's not going to be solved by getting to 0.01 nanometer on the transistor.

‍

[Michael Eisenberg — 35:25]

Your brain operates at 20 watts. Mine probably operates at less than that.

‍

[Elad Raz — 35:28]

I'm not sure about that.

‍

[Michael Eisenberg — 35:29]

Let's just be very clear about the bet. The bet is the world is going to need dramatically faster compute. The bottleneck problems, let's call it, which is memory.

‍

[Elad Raz]

For now.

‍

[Michael Eisenberg]

For now, is going to get solved or get solved to be able to take advantage of the compute. But right now you don't provide that much more value because of the memory problem.

‍

[Elad Raz — 35:54]

Yeah, but we have our own tricks over there and we are going to reveal them soon.

‍

[Michael Eisenberg — 36:02]

Ooh, want to reveal them now?

‍

[Elad Raz]

No.

‍

[Michael Eisenberg — 36:05]

When's soon?

‍

[Elad Raz]

Your partner is going to shoot me.

‍

[Michael Eisenberg — 36:07]

Okay, now when Nvidia or Amazon or Google take a look at you, right? These are people, these are companies producing GPUs and TPUs that a priori are less fast, less accelerated than your computer. What's their answer? What do they say? By the way, I should point out, I pointed this out in a tweet recently, that if you look at kind of the unlock and I asked Grok, by the way, Grok with a K, not with a Q. Grok with a K, which is the X, you know, kind of bot there. SuperGrok. The question I said, okay, what has enabled Amazon's, what is the original technology that enabled Amazon's Trainium and Nvidia's scale up of their compute? And the answer by the way, is Israel in both cases because I think it was Annapurna, I can't remember Savannah or Annapurna sold to Amazon, became the basis for Trainium and Mellanox of course, which is the interconnect which has enabled Nvidia. What?

‍

[Elad Raz — 37:12]

That's the NV link. I mean today you talk about system.

‍

[Michael Eisenberg — 37:17]

Yeah.

‍

[Elad Raz — 37:17]

Because you can take a model and shard it across many GPUs. The bottleneck is how fast you can connect those different GPUs.

‍

[Michael Eisenberg — 37:24]

Yeah, that comes from Mellanox, of course, by Eyal Waldman, and he's been on this podcast before. So if I'm, if I'm Google and Amazon and Nvidia and I'm looking at NextSilicon, I go, nah. Why? What do they, what would they say to say, okay, this is not real or not important?

‍

[Elad Raz — 37:42]

Right now. The demos that we have made was around the HPC.

‍

[Michael Eisenberg]

Not the demos, you have actual customers.

‍

[Elad Raz — 37:49]

The customers. I mean...

‍

[Michael Eisenberg]

The use cases are around HPC.

‍

[Elad Raz]

Exactly.

‍

[Michael Eisenberg]

HPC, in case I didn't say it earlier, is high performance computing. But go ahead.

‍

[Elad Raz — 37:56]

Right now we are taking scientifical code in the HPC. That's what we are running and winning. The core technology is so much different. It's not a feature. NextSilicon is not a feature. It's not, okay, let's take the GPUs and connect NextSilicon into that. It's different architecture, it's a replacement. This is where we are. And as every novel technology, GPU included, at the beginning, most people don't understand the uniqueness around what we have built. The way for us to compete with that is just execute.

‍

[Michael Eisenberg — 38:36]

If you had to use historical analogies, you're more CPU than GPU, is that fair to say?

‍

[Elad Raz — 38:42]

No, it's not CPU and GPU. It's not an FPGA. It's something different. But let me take an example. When Intel first heard about the GPUs, they said, who cares about the GPU? Okay, we have a CPU. This is the way the people see us. They say, okay, they have an interesting technology, they are a technology and not, obviously we are not a market leader in any category, but this is where we are now.

‍

[Michael Eisenberg — 39:09]

And how big is your penetration into the HPC world?

‍

[Elad Raz — 39:13]

We are in every Department of Energy, Three Letter Agency. I mean we are there.

‍

[Michael Eisenberg — 39:21]

You're there. And now that Jensen left the HPC market, has that opened up more opportunity for you?

‍

[Elad Raz — 39:27]

Absolutely.

‍

[Michael Eisenberg — 39:28]

Why did he leave it?

‍

[Elad Raz]

Good question. I think that you have limited resources inside your silicon, and there's fixed amount of transistors. It used to be 100 billion transistors. Now it's 200 billion transistors on the Blackwell chip. Should I allocate more resources for the HPC or should I allocate more resources for the AI? In HPC computers are built around 10 to 100 million dollar. The big leadership computers AI gigawatt factories start with a billion dollar 10 of billions of dollar per gigawatt.

‍

[Michael Eisenberg — 40:13]

So you're still playing in the, in the little leagues there.

‍

[Elad Raz — 40:16]

Yes. Okay, I'm in the little leagues.

‍

[Michael Eisenberg — 40:17]

Like one of the big advantages that Nvidia has is CUDA, right? Which is the software layer that's created a lot of lock in. How are you going to break into that?

‍

[Elad Raz — 40:30]

You just run CUDA.

‍

[Michael Eisenberg — 40:31]

You're going to run CUDA?

‍

[Elad Raz — 40:32]

We are running every program language that the user wants us to run. We can run it.

‍

[Michael Eisenberg]

Why?

‍

[Elad Raz]

The way that NextSilicon operate is not by programmable language. We don't really care how the user wrote the code. We get the code. We understand the different computational and at runtime, once we understand how it works, we reconfigure the chip, we place the different transistor...

‍

[Michael Eisenberg]

You actually reconfigure the chip?

‍

[Elad Raz]

Exactly. And whether the input is a C code or C++ for the HPC or CUDA or Triton to write a new envelope or, we just execute that.

‍

[Michael Eisenberg — 41:13]

One of the things that Gavin said when he was on the podcast, he said that chip companies take three generations to kind of get your act together and find the chip that works in the market. You're Gen 2 now, right?

‍

[Elad Raz — 41:26]

Correct.

‍

[Michael Eisenberg]

Do you agree with him?

‍

[Michael Eisenberg]

In general terms? Yes. Because even now, as you said, we are in the little league. It just take time for companies to generate market fit.

‍

[Michael Eisenberg — 41:39]

It's three generations?

‍

[Elad Raz — 41:41]

It takes time. Why? Because it take you like three to four years until you get the first chip from TSMC. So that's already four years. First generation, at almost every company I know, it's not exactly the same as the customer meets. You need another generation. By the time you start selling your...  we're selling the Maverick 2, you're already in your third generation in terms of lifetime. So yes, I agree with that.

‍

[Michael Eisenberg — 42:12]

When does your third generation tape out? Okay, let me ask you another question. I Want to get to your personal background one second. So you served in the 8200 unit?

‍

[Elad Raz — 42:25]

Correct.

‍

[Michael Eisenberg — 42:25]

Okay, that's well known for cyber, not for chip or chip design. What happened to you?

‍

[Elad Raz]

Personally, I think that I mean it's incredible that the State of Israel educate young entrepreneur, young young talented pre-school to go after cyber. Okay. So, I was in the first generation of the cyber but I love to create. I love to write code. I love compute. This is what I've done and after my military service I decided that well I've done it for too much time. I was really good at it but it's time let's build a company. I didn't know what I want to do I started a service company around high performance computing, high performance networking. I got a lot of interesting project I think that during that lifetime of the company integrity project, you and I met for the first time. This company was later got acquired by Eyal Waldman in Mellanox and then I got to see the big supercomputers and I just fell in love with that and I say okay this is what I want to do and I started NextSilicon.

‍

[Michael Eisenberg — 43:46]

Just like that?

‍

[Elad Raz — 43:47]

Exactly like that.

‍

[Michael Eisenberg — 43:48]

Where did you recruit your team from?

‍

[Elad Raz — 43:51]

All over.

‍

[Michael Eisenberg — 43:52]

How many people work at NextSilicon now?

‍

[Elad Raz — 43:54]

At the moment, around 400.

‍

[Michael Eisenberg — 43:56]

400. And where did you recruit the technical talent from?

‍

[Elad Raz — 43:59]

Yeah, so the majority of the company is R&D, engineering. Divided a bit more on the software side but let's call it half and half between hardware and software. So I was, I'm the software guy, this is what I've done. So the team of the software is mostly coming from intelligence company or other companies I had relationship with. I was able to get Eyal Nagar, my co founder, he was in a lot of startups, did all the tape out in his life. I mean I never built a chip, and he was able to get the best talent from all over Israel, from Apple and Google and a bit from Mellanox. I mean all over.

‍

[Michael Eisenberg — 44:51]

I want to talk about for a second this semiconductor industry. So you mentioned TSMC and you mentioned it takes three years or four years to get a chip out. Does it need to be this slow or is it just the structure of the industry where TSMC is essentially a monopoly in chip tape out any manufacturing basically. Couldn't we accelerate this industry in any way or not really?

‍

[Elad Raz — 45:13]

Oh yeah, of course, of course, but not in our lifetime. And what do I mean by that? You know, when I came and started NextSilicon as a software engineer, I said what's the problem? I mean hardware is like a software. Let's like include the libraries and accelerate things and you know, the engineers in the hardware told me Elad, you don't get it, it's different. And there are many discipline around hardware design. And for the audience, I'm going to divide my talk. There is fabless companies like NextSilicon that design a chip and the end result of them is a graphic file that you send to TSMC for manufacturers. This is like how to print the chip. And there is the actual chip that has own their own fabs and so many, many different vendors around two aspects. So let's talk around the design phase. So you have the design, you have verification, you have dark magic of analog, how to lay down a transistor that send data in and out and then you have the place and route – how you take the design of the chip, press a button and then you get the layout of the chip. Those are like different aspects. And I said what's the problem? Let's write a Python script that does everything with Claude Code. But the amount of IP and complexity in that industry which is dominated by few companies, you have Synopsys, Cadence, and Mentor doing EDA tools. I mean how to take a code and get it into the end design for manufacturing, Ansys that got acquired and others. And the problem there is the cost of a mistake. Cost of a mistake means that your chip will not work. Chip will not work is another $150 million or $50 million if you find a quick fix. No one wants to take that risk. So everyone wants to sign off their tool that everything works and there is simply not enough money there to innovate.

‍

[Michael Eisenberg — 47:31]

Like this whole AI race and the chips race in general. The United States introduced the Chips Act. It's no secret that the tension between the US and China around semiconductors is very large. Every decision at the end of the day at TSMC to supply you with a taped out chip is like a geopolitical decision. So...

‍

[Elad Raz]

Absolutely.

‍

[Michael Eisenberg]

If you were writing policy for the State of Israel right now or had $10 billion and a magic wand, what would you create to ensure that we're able to design, develop, tape out, and produce semiconductors going forward?

‍

[Elad Raz — 48:13]

So let's divide it into manufacturing and design. Manufacturing, in the State of Israel,  how much that I'm Zionist and I love my state, going to be impossible to build. We are not talking about 10 billion, we are talking about $30 billion if you want to be in the leading edge, and this is if you know exactly what you are doing. It's a know how that Intel, TSMC, and Samsung have mastered over the past two decades and China are trying to beat them and they are like just two generation in between.

‍

[Michael Eisenberg — 48:52]

We have Intel fabs here.

‍

[Elad Raz — 48:54]

You have good people, you have Intel fab. You don't have Intel technology.

‍

[Michael Eisenberg — 48:58]

Right, that's true.

‍

[Elad Raz — 49:00]

So in order to rebuild it, you still need this decade of development. By that time TSMC is going to raise. So you talk about more about $100 billion to have to build your own technology from scratch. In the design phase, absolutely. I mean we are not doing anything around accelerating or de-risking there. Most of the multinational that design a chip are US based. Many Israelis going to the military service and then going to build their own startup. Because that's the ecosystem, that's where the investors are. Even for us NextSilicon, fundraising at the beginning was super challenging because you come to investors that understand cyber and ARR and get like a fast money. A billionaire a chip is, you need patience, you need time, different type of stories. And we need. We need to build it, we have to build it. It's important. I think that the design phase is more important than manufacturing because that's your power.

‍

[Michael Eisenberg]

Can you disrupt the Synopsis? Is it disruptable?

‍

[Elad Raz — 50:12]

I think that for the EDA tool it is. But the market is not as big as everyone hoped for that worth the investment.

‍

[Michael Eisenberg — 50:21]

So going back to something about Nvidia. You said in 2024 that Blackwell, the Blackwell announcement was the best thing that ever happened to NextSilicon. Why did you say that? What does that mean to you?

‍

[Elad Raz — 50:34]

I said that on Blackwell and I said that also around Rubin.

‍

[Michael Eisenberg — 50:41]

Yeah. You said that on Ruben too. Yes, that's true.

‍

[Elad Raz — 50:42]

Yeah. We're scared shitless of every company that develop AI chip and you never know how they are going to progress. And I think that the best thing that happened to us with Hopper and Blackwell. Both of them were manufactured and designed with 4 nanometer and we look at the Blackwell specification, remove the sparsity, remove, you know, the other nuance that pretty cheap. But you take the core compute of Blackwell, core compute of Hopper, you compare them one by one. You see Blackwell is 2x than Hopper. Great. But it also consumed 2x more power. Nothing in the architecture had been improved between Hopper and Blackwell. NVLink for sure. You know other nuances, but that's like their core. Hopper to Rubin almost 2x more power, 2x more performance. This is...

‍

[Michael Eisenberg]

Linear scaling.

‍

[Elad Raz]

Linear scaling. At the end of the day...

‍

[Michael Eisenberg — 51:42]

You want to say that you're exponential in the power to compute.

‍

[Elad Raz — 51:44]

Between generation, way more than that. And we are just in the beginning of our design. We have so many places to improve efficiency that we have because we spend $100 million on a chip and not $1 billion on a chip.

‍

[Michael Eisenberg]

What does that mean?

‍

[Michael Eisenberg]

In terms of engineering.

‍

[Michael Eisenberg — 52:03]

You spend 100 million, they spend a billion. So that's 10x. But how does that impact exponentially?

‍

[Elad Raz — 52:06]

If you work on the architecture better on the transistor layout, you can squeeze more performance by those changes. Not 10x but 2x.

‍

[Michael Eisenberg — 52:15]

Are you arguing that your capital efficiency has actually caused you to be more innovative?

‍

[Elad Raz — 52:19]

Of course. Have to be. To start with, you know every company has its own risk. It's a roller coaster. Ups and down. In semiconductor is way worse, these ups and downs you need a lot of help from Hashem [God]. You know that miracle. Hashem is God Almighty. That's what you need in order to succeed in the semiconductor. And yeah, you have limited resources. You have to innovate.

‍

[Michael Eisenberg — 52:50]

Interesting. So a couple of closing questions. Since the advent of AI, Claude Code, what changed at NextSilicon?

‍

[Elad Raz — 52:57]

Everything in every company, across the board. I mean people don't understand what is coming and it's going to be a fun roller coaster. I think that in NextSilicon we have 400 talented people. With those AI tools even I mean the hardware, you remember that I said it's the most risk-averse kind of nuance. Even there, they are open for those type of innovation. And I'm literally see people getting 10x more productivity using those tools.

‍

[Michael Eisenberg — 53:32]

Have you mandated that everyone needs to use...

‍

[Elad Raz]

Absolutely.

‍

[Michael Eisenberg]

AI code?

‍

[Elad Raz — 53:37]

Of course. I mean a company that don't have a leaderboard scores of how many tokens is being used is...

‍

[Michael Eisenberg]

The more the better.

‍

[Elad Raz]

Of course.

‍

[Michael Eisenberg — 53:46]

How big is your token budget now?

‍

[Elad Raz — 53:50]

Depends. Like we are using everything. Me personally, I'm using Gemini, Codex, and Claude. We have people that generate $300 a day, $100 a day per developer. But I think it's mostly, we are now in April 26th, just in timeline. You know, those type of questions get a tendency not to age well. We are now in a phase that training for employees is the most important thing. Like personal example. So I find myself walking around and I am seeing a developer and like he run an agentic task and I ask him okay, how many agents have you opened? He answered only one. I said let's open another one and work on a different problem in parallel.

‍

[Michael Eisenberg — 55:19]

And how does it really help even? Because it takes two years to tape out a chip anyway, two three years to tape out a chip anyway. So you're running faster on the software side of it but you can't move the main you can't move the atoms.

‍

[Elad Raz]

So software is a no-brainer because everyone understand. Let's talk about the design. In design and verification, specifically in verification, you write a lot of tests in order to find hardware bugs. So you manually need to skip them. To start with all the basic test, you're doing that in a second. But an agentic test on all your hardware design until you find some bugs.  All of those those are applicable to hardware as well.

‍

[Michael Eisenberg]

And how much does that, if you have a call it a two-year cycle from beginning to end, how much would that take out?

‍

[Elad Raz — 55:31]

Not much, because you have nine months in manufacture. But within the one year of design you can just do more.

‍

[Michael Eisenberg — 55:39]

Got it. You can do more?

‍

[Elad Raz — 55:41]

More features, better quality, good frequency and power consumption.

‍

[Michael Eisenberg — 55:46]

So the jump from generation to generation should be bigger.

‍

[Elad Raz]

Yes.

‍

[Michael Eisenberg]

But you can't dramatically shorten the time from, call it beginning of design to tape out.

‍

[Elad Raz — 55:55]

I don't think so.

‍

[Michael Eisenberg — 55:56]

Maybe a month here to max. Nothing dramatic. That thing is interesting by the way, to think about the world of atoms has its own pace. The world of code has a entirely different pace these days. And the hardware can bear what the hardware can bear over time. So you mentioned that you go around and you're kind of setting an example and asking, you know, you did one agent. Let me help you do another agent. Apparently, you still code in assembly on a Commodore 64. I think that might have been my first computer. Commodore 64. It had the cartridges in the back. They still have that? The one you have?

‍

[Elad Raz — 56:34]

Yes, yes, yes.

‍

[Michael Eisenberg — 56:35]

You still have a Commodore 64?

‍

[Elad Raz]

I have Commodore. I don't have the cartridges, I have for the Atari one. For the Commodore I have the big disc, you know, the floppy disc.

‍

[Michael Eisenberg]

Oh okay.

‍

[Elad Raz]

Yeah, that's the Commodore disk that I have. But yeah, I love all computers.

‍

[Michael Eisenberg — 56:52]

You actually code on that thing?

‍

[Elad Raz — 56:54]

Yeah, yeah.

‍

[Michael Eisenberg — 56:56]

You don't have a VGA monitor still?

‍

[Elad Raz — 56:58]

Yeah, I'm a bit lazy so I'm doing everything in emulator before moving it to the actual hardware. But, yes.

‍

[Michael Eisenberg — 57:04]

Two last questions for you. We're living like in this moment of history right now where a chip company, Nvidia in this case, was like the most valuable business in human history. What do you think that means? Not just financially, but philosophically about the world where this notion of silicon can become the most valuable thing in human civilization.

‍

[Elad Raz — 57:30]

No surprise there. Think about it for a second. When 1000 years ago we all were fighting around food. Okay. That's the basic element that everyone needs. Food, water, resources. I think that around 300, 200 years ago it was around energy resources. That's most of the worlds. Now we are in the age of compute. Those are the phases better, more energy you can produce will be utilized by compute. Compute is the wave.

‍

[Michael Eisenberg — 58:03]

So when they call it a wafer, it's not just silicon, it's like a food reference that everyone.

‍

[Elad Raz — 58:09]

I never thought about that.

‍

[Michael Eisenberg — 58:11]

Could be a religious reference too. Christianity. And to finish, I started by asking you what was etched on the first NextSilicon chip and you said it was Bereshit, or Genesis, and on your most recent chip?

‍

[Elad Raz — 58:27]

So Maverick 2, we haven't found a better word than Bereshit to place on the Maverick 2 for a long time. Like I ask everyone and you know, people gave us some kind of different references but nothing that was more powerful. And I told the engineer, just keep the word Bereshit because it's like it's Bracha, just keep it.

‍

[Michael Eisenberg — 58:52]

Bracha means blessing.

‍

[Elad Raz — 58:53]

Yeah. And then Maverick 2, sorry. And then unfortunately October 7th came and that was like few weeks before the tape out. I just told everyone a day after it was like Saturday and Sunday I sent the email to TSMC, said well I want to put Am Yisrael Chai, people of Israel live. And I had to explain them why it's non politicized and you know why. And I told them I'm not going to tape out unless you put it. That's like Israeli resilience and let's etch it. And that's, that's what exists on the Maverick 2. I'm

‍

[Michael Eisenberg — 59:36]

Am Yisrael Chai on the Maverick 2.

‍

[Elad Raz]

Yeah.

‍

[Michael Eisenberg]

It's inspiring.

‍

[Elad Raz — 59:39]

Thank you.

‍

[Michael Eisenberg]

And thoughtful. And that's what's running the computing and high performance computing all over the place.

‍

[Elad Raz — 59:45]

Exactly.

‍

[Michael Eisenberg]

Much more than resilient, it's powerful.

‍

[Michael Eisenberg — 59:49]

Elad, Thank you for joining us.

‍

[Elad Raz — 59:51]

Thank you.

‍

[Michael Eisenberg — 59:52]

Thank you for joining us on Invested. Thank you to Elad. If you enjoyed this podcast, please rate us five five stars on Spotify, Apple Podcasts. Please please please subscribe to the YouTube channel and if you're enjoying the content, be sure to share with your friends as well.

‍