I'll do it better this time.
You're all set.
Okay. Blaize is a really interesting company. It's very exciting, and I see it as part of my job is to make sure a lot of people know about it, and so really appreciate you being here, but also means that we should probably start at the beginning.
Sure.
Let's start with the founding story.
Yeah
for Dinakar, how they got here, when you joined, where we're at today.
Okay. Thanks, Gil. Thank you very much for having me, hi, everybody. I like to say that if you've ever used an Apple Macintosh in your life, chances are that the GPU, the Centrino chip, was designed by Dinakar and his co-founders. They started Blaize 12, 14 years ago with a view to designing a new type of GPU. What's different is that that GPU is programmable. Programmability is hugely important. Number two, it's actually optimized for edge type of use cases, why edge is important, there's low power, low latency, high efficiency. Those are the important things. The programmability part is super important. Another anecdote that one of the co-founders says is, if you go back 20 + years, there were 45 GPU companies.
To date, three exist or survive, let's put it that way, NVIDIA, AMD, ATI that AMD acquired, Intel. What's common with those three that was not common with the remaining ones, that they were programmable. I think it's true to say that Blaize, of all the startups and people who are playing in this space, we're that programmable chip. What we deliver is an AI platform, which is underpinned with three pillars. Pillar number one is the edge and far edge. If you have a drone or you need a box on a lamppost, that's inference happening where the data is being created, that's where the efficiency and low power are important. Pillar number two is where we've married our chip, which is called the GSP, Graph Streaming Processor, with a GPU.
We like to say we're not here to replace NVIDIA, we're here to complement a GPU. What that second pillar does is allows the right silicon to be used for the right workload.
Today, if you're trying to use a GPU-based solution for the far edge, it's like using a sledgehammer to crack a nut. That middle pillar for us is super important because it allows us now to offer our AI services platform, which is the third pillar, and the customers for that are Tier 2 data centers. You'll have read about our partnership with Nokia and with Datacom, which is one of the first data centers that we're working with. That third pillar is hugely important because it allows those data centers now to generate a new revenue stream because the software layer that sits on top of the platform is exposed through APIs.
If you want something to do with facial recognition or security or whether it's in retail setting or smart city settings, the APIs allow the customer to benefit from the AI workloads. The advantage for Blaize is we get to participate in the data center infrastructure revenue stream by providing servers powered by Blaize, as well as an ongoing recurring revenue stream from the monetization of the APIs.
That's a great introduction. Let's go piece by piece here. Let's start with when you can put your chip in a camera or on a drone or at the edge, what applications are your partners selling and your customers buying?
We don't actually go onto a camera. In fact, one of our chips can probably handle four to five high-definition streams. If you take a retail setting or smart city setting, so one of the customers that we've announced, a company called TCC, it's a subsidiary of the PIF, Saudi PIF. Their application is monitoring highways, number plate recognition, fines, and all of that. It's a huge project, and they need a box that sits on a rooftop, on a lamppost. What's really, really amazing is that last September, we went through a test with them, and our box was running the AI algorithms that they were interested in on a rooftop, and the temperature inside the box was 75 degrees Celsius, and the workloads were running. A GPU-based one stopped working at 50 something.
Yeah.
That's one set of use cases. Winmate, a Taiwanese company, we announced something with them recently. They develop ruggedized boxes. Our cards will go into those boxes for industrial use. They also do drones. When you're looking at a drone, and we have other drone applications as well, they're being used more in a defensive capability, which is can my drones be flying around and look and see and identify incoming threats? The card sits on the drone, and the drone is self-sufficient. Any algorithms that need to be run, whether it's navigation or whether it's identification or something, all happen on the drone itself.
It's video image processing.
Yeah.
The uses are, you talked about smart city, defense, retail.
Retail, yeah.
Any place where cameras are placed and you can't have a human in a call center or a monitoring center.
Right
watch all the videos. You need to monitor it. Right now, a lot of times, they're sending all the signal back to a data center.
Correct
instead of doing some of the processing at the edge, which is far more efficient.
Yeah, exactly. The adoption of AI at the edge has been hindered because of the cost of sending it backwards and forwards. It's not just video processing. One of the beauties of the design of the chip is that it can run multiple algorithms on the chip at the same time, whether they are related to each other or not. For example, in a factory sort of setting, you could have a sensing type of data coming in alongside your video, and that can be processed.
When we talk about putting it side by side with GPUs in a data center.
Right
Are we still just processing video, or are we?
No
Processing other streams?
No. The roadmap for APIs includes text, document processing, video, anything else that people want to be able to utilize AI for, where it's a combination of workloads running on our hardware versus workloads running on GPU. As far as the customer now is concerned, unless you're working with a very large company where they have a systems integrator and they're going to put all parts together, the majority of our customers now just want a solution. The applications that are going to run on the hybrid thing are across the board. The good thing is we have all seen, and are continuing to see all these 600 billion, 700 billion parameter models coming down to 70, 80, coming down to seven, eight. When you get to that level, and the accuracy is not compromised to the extent that it becomes meaningful.
At that level, all of those models can run very well on the Blaize chip. If you have slightly larger models that are required, they will run happily on the GPU.
What's the microeconomics of it for a customer? Instead of buying a rack of just GPUs, now they're buying a rack with GPUs and GSPs in them. What are the types of efficiency savings they're seeing?
We did, I'll mention Yotta is a company that we got a contract with last year. It's one of the data centers in India, and their end customer is the Indian highways. If anybody's seen an Indian number plate, it's all sorts of cursive. I didn't even know that you could make your own handwriting up.
Right.
That's a huge challenge to identify when you're doing toll booth. They were running on NVIDIA-based GPUs. We were able to demonstrate to them that if they ran those workloads on a Blaize server, it happened to be a Supermicro server powered by 24 Blaize cards, that we could run that at 2x-4x less total cost of ownership to them. It's not just the cost of the server to start with, but it's the cost of running it, which is significantly lower.
50%-75% cost reduction.
That's right. It's all workload dependent, right? One of the penalties that a GPU pays, that if you have a batch of one, then in order for it to go through, is a lot of on and off memory.
Right.
With Blaize, that doesn't happen at all. It's those workloads all run together. The advantage there is we don't need HBM memory to make ours run. It's all LPDDR.
Right.
The combination of that is where the efficiency comes from. Now when you put it into a system, which we've got our own pilot tests that we've done right now, which you'll hear about a lot more publicly as we start to announce the deployment of these hybrid. Depending on the workload, you are getting a significant reduction. Our customer here is the Tier 2 data center, and with Asia Pac, where we're pretty active right now, I think these are publicly talked about reductions in ROI from, let's say, five years down to four years down to three years from their perspective. No.
Got it. That's very powerful microeconomics.
Yeah, exactly.
that you can go to market with.
Yeah.
You've talked about partners. What is your overall approach about go to market? How much of that do you want to go through channel? Do you have any direct salespeople? Will there be a change in that going forward?
Yes, we do have direct salespeople. Yes, we have an ecosystem that we're selling through. I think the Nokia partnership for us is a game changer because they already have relationships with data centers, and they provide connectivity. The combination of Blaize and Nokia together, we're able to bring an AI solution to those data centers. We will no doubt leverage their sales force. We work with ecosystem of ISVs, the independent software vendors. We have developed some of our own applications, but it's just too expensive for us to do that for every use case. We go and seek out experts that have already got an application built, and that ecosystem enables us to, if we win something, then we bring the right ecosystem partner along. If another ecosystem partner has an opportunity that can benefit, then they bring us along.
This is how we see the go to market. It's a mixture of direct plus the partnerships.
Got it. We went one way in terms of channel partners. Let's go the other way in terms of where we source, where do we make our chips?
Right.
Who do we partner with?
That's a great question because it's very important for defense. You're talking about your 250th anniversary and the innovation. Well, this chip is taped out by the Samsung Foundry in Austin, Texas.
Great.
Which means that when we put it in a defense application, or our partners put it into a defense application, then the Made in America box is all ticked.
Got it. That's helpful both for being able to sell to the U.S. Armed Forces, but also, let's say, somebody in the Gulf Coast.
Yes. Some of the conversations that are going on in other parts, not just Gulf Coast, U.S., but also in Europe. All of that is important.
Who do we get the memory from?
That's an interesting question. I know part of my job is to go talk to Samsung and Micron-
Sure
to see if I can secure the supply. I don't know the answer to that yet.
I vote Micron or if Samsung can make those here as well.
Yeah.
That's helpful. Before I open it up to questions, let's just go through some financial numbers so everybody.
Right
Gets a sense of where we're at. What's our current guidance for this year? What guidance have we given past that? What's our overall perspective on backlog right now?
One of the questions I omitted to answer was my own story, and I'll come to yours in a minute, the question in a minute. I've been with Blaize about seven years. I love technology. I love getting involved with B2B salespeople. I love to get how does software work and all that type stuff. I feel it's really important for a CFO because if you are putting numbers out, then I need to have a level of comfort that actually there is something to support them.
Right.
Most of my day, I know as a public company CFO, the role is changing and it does change. I have different stakeholders to look after. Fundamentally, the message that I have to give them is based on the relationships internally. Our guidance for last year was a range which we tightened. We exceeded that by an amount. I think we closed at about $39 million or so. We had a range for 2026, which we tightened to $130 million. It is backend loaded. The reason for that is that some of the contracts and so on that we've got require us to go secure our chips, secure our card, secure some servers.
A lot of my effort has been put to making sure that that's in place and as some of the proofs of concept are getting into production, hopefully the two things come together that we are able to fulfill the promise that we made.
If we get 130 done this year, how much of your existing deals will remain in backlog? How far into 2027 are you starting the year?
Again, we have not projected anything for 2027, the premise of your question is, look, when we're talking about data center type applications, it's kind of land and expand. Once you've created a solution for one, then it'll go into others. When we're talking about some of the edge type of applications, you look at the Yotta one, you look at the one in the GCC region, you look at some of the drones, we expect those to continue to grow, not just within that one customer, but with other customers. I'm thinking very, very hard about introducing concepts like backlog and bookings, because we will have a software element, recurring element to our mix, as we go through the rest of this year.
How about pipeline? How would we quantify the pipeline at this point?
Pipeline, we talked about the pipeline maybe a couple of quarters ago. Pipeline is still a very important number for us, but internally. What we want to focus on is revenue, contracts, purchase orders, backlog, and bookings. The pipeline, we're very deliberate about what goes into the pipeline. There is a process. We have to have gone through a proof of concept. It's not just, I think this is going to happen, and therefore, that's a good customer for us. The use case has to be identified. In many cases, we focus on where we already have the workloads and the applications designed, as opposed to going down rabbit holes with others.
Got it. It's a good time for me to see if there's any questions from the audience.
Sure. Maybe can you talk about how you see your revenue mix moving forward between hardware, software, some of the AI services you provide?
Right. 2025 and 2026, majority is hardware. 2026 will be more of the system-level hardware that we've got. I expect edge boxes and just our cards also to feature in that. What I'm going to see with some of those deployments, for example, with the Yotta deployment in India, there was already a software component of that mix, of that delivery. That helped with our margins and so on on that. As we start to deploy AI services and participate in this recurring revenue model, I expect my margins to improve significantly as we go through 2027 to 2028, because the portion of software will continue to grow. Hardware will still remain an important part because we're still going to be providing a hybrid server, so GPU, GSP. The good thing is our relationship with NeoTensr, that's proving to be an amazing relationship.
We've identified a white-labeled server manufacturer, that helps my margins, it also means that I'll still continue to be in the hardware business as part of the mix.
I'm curious if you were to break down your revenue by industry, what that looks like today and maybe where it's headed, if you look out a few years, the next few years.
In my first pillar of the near and far edge, I would say it's smart city, I would include highways and that type of thing. I would call it defense, drone type of applications. As we move into the data center world, it kind of depends what their verticals, what their customers are demanding. The APIs that we have today, things like facial recognition and so on. Those will be the use cases that we'd find as we go along.
The roadmap is pretty rich. At some point, we'll talk more about the kind of use cases those APIs will allow the end customer to go into which vertical. We've done a lot of work with retail in South America, at some point, one retailer versus another retailer is much the same type of application. Those are the kind of verticals that we are working with today.
Maybe your new CRO, Stephen, has been in the seat now a few months.
Right.
What's he seeing? Maybe what changes does he plan on making?
We're actually very lucky to have Stephen. He comes from, I think, a Cisco background, which means that where the company's going with the AI services platform is smack bang in a sweet spot of his understanding of the ecosystem. For us, the far edge will remain an important component, but really the growth is going to be in this AI services providing solutions, that's exactly where Stephen comes from.
Let me wrap up with one kind of forward-looking one. Before, let me ask you, where in the world is Dinakar right now?
Dinakar, he's just come back from Asia, I think he's resting.
Because I think last time we had earnings, he was just coming back from Saudi, where he was on a roof
Yes, that's right
measuring something that was 75 degrees.
Yeah.
Which by the way, is that 150, 200 Fahrenheit?
It's-
That's a lot.
It is.
Yeah.
The funny thing about that story is we were not expecting that. They said, "We need to do a test." We said, "Fine." "We're going to do it outdoors." We said, "Fine." We thought it'd be in some shade. They said, "Let's take it up the roof and let's set it running." They went off to lunch and came back, and the thing was still running. He was also at the Computex-
In Taiwan, yeah
Taiwan, where we jointly presented with Winmate. I think that's another partnership that I'm really excited about because it just opens up opportunity for us. The Nokia partnership, those are really important for us.
I'll go next here. I know that there's another generation of product coming.
Right
Within a couple of years. What's that generation going to be capable of? What kind of opportunity would that open?
I didn't realize, I came from a telecoms background, and a bit of medias, and I didn't realize how big a deal it is when you tape out a chip and it lights up first time. We're on our third chip. We had two test chips and one production.
Right.
They lit up within the first hour.
Right.
It goes to show that the skill set that you have in the company, I'll name drop. Santiago came from, he's an ex-Apple guy, an ex-AMD guy, maybe he's also got NVIDIA in his blood somewhere. He's the one who started Apple's iPhone one to five chip business.
Right.
The pedigree that you have in the company is very deep. The founders, the design that they've come up with is one which is common to every chip. We can make it bigger, smaller, have more features, and so on, but the design of how the chip works is common. The next generation chip, we're not talking about what size and things it's going to be, but it is going to help run whatever size of models are going to be in the next two to three years. It's going to help our hybrid strategy extremely powerfully. It just means there'll be more Blaize component, if you like, in the hybrid versus GPU.
That's great.
Yeah.
That's great. I think that's a good time for us to wrap up. I'll just point out that I see a lot of irony in the fact that Cerebras, which has less real revenue than you do- is trading at tens of billions and we are where we are. It's my job-
Yes
to make sure that changes.
Well, it's our job to just keep on executing, right? We have a good team, we have a great team right now, particularly on the go-to-market with, Stephen's brought some people in that just have expertise in this area. All our job is to just go execute on the contracts that we've got and make sure that the new ones that come in don't divert us into new types of development, and we'll use the ecosystem for that.
Makes sense.
Yeah.
Thank you very much.
Thank you so much for having me.
Thank you.
Thanks.
Thank you, everybody.