Friends of NFTBC.

As seasoned NFTBC readers will know, I am particularly drawn to stories of unconventional business success. Naturally, these sorts of stories are rarely obvious in the beginning - so any satisfaction in the pursuit comes from following them over time and living inside the story while it writes itself. All the better, if I can make some useful investment returns along the way. Cerebras Systems is an example where I have been following the story for about five years and an investor for just over two years. With Cerebras having recently filed an S-1 and an IPO looking imminent, now seems like the right time for a post. But I have been of two minds as to whether I should write one. AI, IPOs, hype - these are certainly not the traditional NFTBC territory you’re used to, and to some extent they did put me off from wanting to involve myself. But on the other hand Cerebras is, in my view, an extraordinary business - and a fascinating one. It’s worth writing about. I’m also well aware I’m posting this at a time of especially high enthusiasm for semiconductor stocks - and this unsettles me to some degree. But equally, if ever there was a propitious moment for Cerebras to raise needed capital on the stock mark, this is it.

This post is intended for curious readers with an interest in unconventional businesses doing interesting things. It’s probably less relevant for those who want to go deep into the weeds on engineering questions. But all are welcome. Additionally, the post will be much more concerned with the business itself rather than IPO dynamics.

Let’s begin.

[NFTBC does not give investment advice. Please do your own research.]

Introduction

Since its founding a decade ago, Cerebras has been in the business of inventing, producing and selling a highly unconventional computer system with specific design intent for AI applications. What makes Cerebras’ system so unusual is the way they’ve gone about it - using an entire monolithic 300mm silicon wafer as the basis of their processor, rather than dicing their wafers up into small chips like everyone else does. Doing it this way also necessitates rethinking the entire system as well as the manufacturing process and supply chain: from packaging to power to cooling to software. And there are good reasons why no one else does this - for decades it was supposed to be a dead end. Cerebras’ founders decided to do it anyway, succeeding against all the odds - they’ve now built something unique and very hard to replicate. As to why you might even want a ‘wafer-scale’ processor, I’ll cover some of that later. But suffice to say for now that the founders came to realise in late 2015 that the future computational requirements of AI, would look very different to those that the existing architectures were designed for (e.g. the graphics processing unit). And so, the Cerebras Wafer-Scale Engine (WSE) is what you get when one of the most determined, experienced and innovative systems design teams comes together at the right time, with the right backers, a bit of luck - and with a blank sheet of paper.

Cerebras’ current system, the CS-3, was released two years ago. The CS-3 was primarily intended for model training and provided an elegant solution for simple near-linear scaling out to large clusters. I thought this was the way things were going to go. But despite impressive demonstrated performance the CS-3 has remained very niche in training - training being where Nvidia’s moat is strongest. Within months of the CS-3 launch, in mid-2024 Cerebras quickly pivoted to inference in order to capitalise on the emerging shift to inference-time compute - or smarter ‘thinking’ models. In 2026, we’re all familiar with thinking models - they’ve become mainstream. And we’re all familiar with waiting while the model thinks. To calibrate your expectations here, the CS-3 typically runs models around 10-20 times faster than GPUs - in practical terms this means that chat responses come back more or less instantly without having to wait. Seeing is believing of course, so you can try out Cerebras Chat for free at this link. Or you can just watch Cerebras code an interactive pool game in 12 seconds:

Of all places, I came to be aware of Cerebras some years ago via GSK - yes, the stodgy old big pharma! (see my GSK post here). One of the lesser-known things about GSK is that they have perhaps the leading in-house AI and machine learning capabilities in the pharma industry. They were also Cerebras’ first commercial customer. Intriguingly, it was actually GSK who sought out and approached Cerebras - a complete unknown at the time and only recently out of stealth mode. The GSK AI/ML group had some big problems they were struggling to solve with GPUs. They looked into rightsizing the hardware to their problem, rather than the other way round - Cerebras had the hardware for the job. The two companies remain partners years later.

It took Cerebras around four years to design and launch their first system - a process defined by formidable challenges that they ultimately overcame. It took another six years to get traction with hyperscale customers. But they stuck at it, kept making their case and made a couple of important pivots. And so, after years of having to promote an offbeat roster of reference customers such as GSK, TotalEnergies, Mayo Clinic and G42, Cerebras can finally point to major megawatt-scale contracts with OpenAI and AWS. The foot is now firmly in the door. It’s no coincidence that the list of chip startups with major cloud deployments is particularly short. The barriers to entry are mighty. Turning up with something just a bit better simply will not do - you have to turn up with something an order of magnitude better, and that was the plan from the start. In many respects the WSE was ahead of its time but it’s now coming into its own.

In my case, it took a good couple of years for the potential of wafer-scale computing to dawn on me. By early 2024, I had become convinced enough to buy Cerebras shares in the private market. Any early excitement around Cerebras had now been squeezed out by Nvidia’s tightening grip on the AI narrative. Private market appetite for wafer-scale was minimal - which, of course, is exactly what made it interesting.

In the domain of frontier models and the underlying infrastructure that makes it all possible, we are, of course, talking about a space that is moving with extraordinary velocity. The number of people (or more accurately, teams of people) who are able to synthesise all of the associated technologies at the leading edge and hold a useful idea of where things might be heading, is incredibly limited - and almost all such people are active participants in shaping this future rather than investors passively trying to make sense of things from the sidelines. So we should be honest with ourselves: practically all of us are mere passengers. As I have said a few times before, investing for me is sometimes about striking the right balance between knowing too much and too little. It seems to be the case that some of Cerebras’ first venture backers felt the same way1 - they knew enough to grasp the significance of what the founders were proposing, but perhaps not enough to be scared off from trying. But most importantly, the reason Cerebras got funding at all was because the VCs were backing a team - in this case, a team who had worked together for years and had founded a successful systems venture previously (SeaMicro). They’ve done it twice now, and achieved something even more challenging the second time around. So in its most simplified form, the Cerebras thesis might be stated something like this: whoever had the requisite foresight, ingenuity, experience and persistence, against all the odds, to first create the Wafer-Scale Engine and then drive its adoption into the mainstream - it’s worth betting on those people and whatever they’re going to do next in the field of AI computing.

The AI market is moving extraordinarily quickly, and we are well-positioned to anticipate its twists and turns
- Cerebras Founders Letter

It’s worth stating at the outset that the range of outcomes for Cerebras is particularly wide and, in the context of an IPO and a possible AI infrastructure bubble, this will result in highly polarised views and speculation every which way. Cerebras produced around $500m in revenue for 2025, and per the S-1/A the IPO market cap looks like it could be in the region of $26 to $27bn. As the sceptics will rightly point out, these sorts of starting valuations rarely produce attractive prospective returns. Sceptics might also point to extreme customer concentration, not to mention the constant uncertainty of when the AI capex bubble will eventually pop. But on the other hand, if Cerebras executes to the full extent of the contracts disclosed in its prospectus, it could conceivably ramp to well over $10bn in revenue over the next five years - for context, this might just sneak Cerebras into the top 10 fastest ever growing businesses from product launch (roughly similar to Facebook’s 2004-2014 run). And, let’s say, if Cerebras can make it to just 5% of Nvidia’s current revenue run-rate by the early 2030s (that’s ~$15bn), you could still be looking at a $100bn market cap business even with an extreme de-rating from 50x revenue to 7x. And I want to be clear about this: Cerebras is not an “Nvidia killer” such as you see in some of the sensationalist headlines - the thesis is more about winning a small slice of a very large pie. We’ll look at scenarios later on. For my part, I think Cerebras has a genuine shot at getting big quite quickly. But it’s by no means a dead cert and there are important risks and considerations that we’ll look at.

I have deliberately decided to try to make this post as non-technical as possible - although it does assume some base level knowledge around certain terms and concepts. I’m going to link to some of Cerebras’ technical materials/presentations and I suggest diving in for those who want to go there. From following Cerebras over time, I have encountered a good deal of punditry, typically concerning why the company will remain limited to only a highly specialised niche or why its systems will surely prove economically unviable. Some of this commentary is simply a result of insufficient familiarity with the company, and some of it seems to assume that Cerebras will not continue to adapt or make advances in future system generations. But some observations are fair and worth making note of - which we’ll touch on later. I’ll try to avoid making too many technology-related claims of my own, especially where they might be contentious - to do otherwise, I risk getting in over my head. So I’ll try to stick to claims that seem fairly robust and highlight where I’m choosing to put my faith in others. It’s inevitable I’ll get some things wrong, so I shall endeavour to clear up mistakes later on.

But if there’s only one thing you need to understand about the Cerebras thesis it’s this: speed. It’s generally uncontentious to say that the principal bottleneck in inference output speed is having to constantly move massive amounts of data from memory to compute - move the data too slowly and the processor sits idling while waiting to be fed. Cerebras’ answer is to pack an entire wafer with the very fastest type of memory (SRAM) interspersed among the compute - data travels only tiny distances and through incredibly fat pipes. If you have to move data between individual chips (like everyone else does), the distances get longer and the pipes much thinner. This is why there’s nothing else remotely as fast as the CS-3 at inference:

Llama 3.3 70B API provider benchmarking on Artificial Analysis. Data from February 2026 (Cerebras has been pulling some model APIs this year to free up capacity so February is the most recent I could see)

Moreover, it has become clear that much higher output speed is able to command a price premium. In a similar vein, Anthropic have shown they can charge 6x more for their ‘fast mode’ at 2.5x the speed.

Cerebras is the only player who can produce a wafer-scale system such is required for extreme memory bandwidth. Extreme memory bandwidth enables extreme speed. Extreme speed potentially enables a breakout from the commodity token space through premium pricing and new use cases. Competitors are, of course, lining up to enter this gigantic market. And it’s also true that barriers to entry in chip design may have come down in some respects - there’s far more venture money washing around in semiconductors compared to a decade ago and the hyperscalers have deep pockets and motivation. Thus far, there’s not much evidence of anyone else going wafer-scale and, moreover, Cerebras owns important IP in that domain, having pioneered it. And, as we shall learn, it’s by no means easy. I therefore believe that barriers to entry in wafer-scale remain intact. Most alternative approaches to wafer-scale appear to retain the ‘thin pipe’ problem, go after cost rather than speed or else sacrifice vital programmability (which has obvious limitations if the models keep changing). And perhaps most importantly it’s important to weigh up the following considerations: 1) the CS-3 wasn’t even designed principally with inference in mind and 2) no frontier lab has yet designed a model specifically to take advantage of wafer-scale benefits and to optimise around the constraints. Both of these things are about to change. It’s clear that Cerebras and its partners have big ambitions for major system improvements from here - more on that later.

But who knows? Maybe the wafer-scale critics will have the last laugh in the end. Maybe the economics won’t play out competitively or maybe Cerebras will struggle to keep up with Nvidia’s relentless pace. Maybe they’ll struggle to manufacture these systems in the necessary volumes. Or perhaps they (or their largest customer) will run out of cash. We’ll review some of the risks later on. But what excites me most is that I don’t think we’ve even seen Cerebras’ A game yet. Here’s Cerebras co-founder and Chief Architect, Michael James, speaking at the end of last year:

If you look at the next designs that we’ll come out with, they might make [the WSE-3] seem quaint.

This? Quaint?:

I’m curious to see what comes next.

The Meaning of Speed

I was reading about George Stephenson, the great steam train pioneer, earlier this year. In Great Britain the idea of travelling twice as fast as the fastest horse-drawn mail coach was once considered absurd, unfeasible and dangerous, not to mention unnecessary - a dead end. And the country had previously invested heavily in a comprehensive network of canals for the efficient horse-drawn conveyance of commercial goods - the canal network was itself a paradigm shift and a key accelerant of the industrial revolution in Britain. So when in the 1820s Stephenson proposed the world’s first intercity steam locomotive railway between Liverpool and Manchester, he faced a good deal of scepticism:

George Stephenson’s idea was at that time regarded as but the dream of a chimerical projector. It stood before the public friendless, struggling hard to gain a footing, scarcely daring to lift itself into notice for fear of ridicule. The civil engineers generally rejected the notion of a Locomotive Railway; and when no leading man of the day could be found to stand forward in support of the Killingworth mechanic, its chances of success must indeed have been pronounced but small. But, like all great truths, the time was surely to come when it was to prevail.2

I am, of course, just having a bit of fun here. But I do enjoy these stories where some maverick succeeds with an audacious groundbreaking idea against the odds - and the analogy of moving from the canal age (i.e. GPU) to the steam locomotive age (i.e. Wafer-Scale Engine) is a tempting one. Stephenson got his railway in the end and it changed the meaning of speed for the whole world:

Notions which we have received from our ancestors, and verified by our own experience, are overthrown in a day, and a new standard erected, by which to form our ideas for the future. Speed – despatch – distance – are still relative terms, but their meaning has been totally changed within a few months: what was quick is now slow; what was distant is now near; and this change in our ideas will not be limited to the environs of Liverpool and Manchester – it will pervade society at large.3

The rest is history, as they say. Back in Stephenson’s day, there were those who pointed to the efficiency of the existing canal system to demonstrate why railways were unnecessary. They weren’t wrong. But it showed a lack of imagination. Even today’s canal baron, Jensen Huang, is finally hedging his bets and ponying up for some steam engines of his own - as evidenced by Nvidia’s recent acquisition of Groq (the closest Cerebras has to a fast inference peer). The first sections of intercity track are now being laid.

Staying on the topic of speed, Cerebras co-founder and CEO Andrew Feldman likes to draw the more modern analogy between dial-up/broadband internet and slow/fast inference. I’m old enough to have had my formative internet experiences with dial-up modems. And Feldman has a point - waiting for a model to respond in 2026 conjures up old memories of loading webpages 30 years ago. For any patient youngsters or nostalgic oldies, you can see what this was like here:

What did it mean when broadband arrived and we no longer had to wait? The internet changed: websites could become far more data rich, streaming, cloud computing, you name it - entirely new business models emerged and their effects continue to ripple through our world today. And as the world moves forward, expectations change - Feldman and others like to point to research carried out by Google and Amazon some years ago showing how mere milliseconds of latency can cause a reduction in traffic or even revenue.4

So what of fast inference? Here’s how Cerebras co-founder Sean Lie put it when they announced an experimental ultra-fast coding tool with OpenAI a couple of months ago (Codex-Spark):

What excites us most about GPT-5.3-Codex-Spark is partnering with OpenAI and the developer community to discover what fast inference makes possible—new interaction patterns, new use cases, and a fundamentally different model experience. This preview is just the beginning.
-Sean Lie, CTO and Co-Founder of Cerebras

Readers may have their own views on this, but in the prospectus (p121) Cerebras cites coding as a strong example, as well as real-time insights in deep research and voice applications. We’ll just have to see what emerges in the months and years ahead. But the trend is clear: more thinking time » more inference » more waiting (unless we do something about it):

Here ends the complementary section of this post - thank you for reading! In the remainder, we will cover the following topics:

The Road to Wafer-Scale
Brief Thoughts & Speculations on Tokenomics
The Roadmap: what we know + some possibilities
Business Model & Sizing The Opportunity
Risks & Other Considerations
Final Thoughts: Moat

[One other thing before we get going - you can find all of my post-IPO updates and thoughts on CBRS at this link]

Subscribe now

The Road to Wafer-Scale

The Cerebras origin story gets underway around 20 years ago.

SeaMicro

At the time, Gary Lauterbach was a highly experienced architect at Sun Microsystems. He had designed Sun’s latest generations of powerful server chips. Andrew Feldman was a former network equipment executive turned entrepreneur-in-residence at a VC firm. Both had, via different paths, arrived at the same conclusion: a new kind of efficient server hardware was needed to address the massively escalating challenge of datacenter power consumption. Feldman decided to start a company, but he needed a technical founder to join him. He had the right network and VC connections, while Lauterbach had the engineering expertise. A former colleague of Feldman (who

later joined) put them in touch and they founded SeaMicro in 2008. They raised venture capital and started recruiting. Early engineering recruits included JP Fricker (another former colleague), Sean Lie and Michael James - together, the five Cerebras founders. SeaMicro’s innovation was to replace power-hungry server chips with lots of low-power chips integrated together with a proprietary low-latency fabric. [Note: Many interconnected dies, low-latency fabric, software abstraction layer - these are some of the important ideas that eventually went into Cerebras.]

In 2012 they decided to sell the business to AMD for around $300m. They all went over to AMD (along with numerous other colleagues). But by 2015, they had time on their hands and were looking for something to do - indeed AMD decided to kill off the SeaMicro business in 2015, despite ultimately adopting some of the foundational IP into its own critical technology roadmap.

Getting the Band Back Together

Sometime in 2015 the five started to meet up once a week in a Palo Alto coffee shop. They knew they wanted to work together again but they needed to pick a computing problem to work on. As their meetings became more serious they started to take a room in the public library. And they picked a problem to work on: a new computing hardware system for artificial intelligence. Bear in mind in AI terms, this was a long time ago - AlexNet was only 3 year old, Google hadn’t even announced the first TPU and the famous Transformer paper was still a couple of years away. Eventually, via Feldman, they went to have an informal discussion with a VC - to their surprise they were talking term sheets by the end of the meeting. So they went away to formalise a pitch to take around the VCs. And they pitched the team first - they had done it before, they worked well together and had complementary skills in leadership, chip design, software, networking etc. Importantly, at this stage they weren’t even pitching a wafer-scale system - that only came later. Incredibly they got offers from everyone they pitched, raising around $30m from Foundation Capital, Benchmark and Eclipse in the Series A. Also keep in mind that a decade ago the VC purse strings were pretty tight when it came to semiconductors - Cerebras has raised considerably more venture money in the last 12 months than the entire industry did back in 2016:

Cerebras itself was incorporated in early 2016 and the deliberations continued. They contemplated all manner of approaches whether GPU-like or even massive multi-die arrays of the sort that Tesla tried (and discontinued). Eventually:

One day, Gary leaned back and he said, “But why would a graphics processing unit that for 20 years had been tuned for graphics, why would that be good at AI? Wouldn’t that be serendipitous if 20 years of tuning and improvements for one type of work left you well positioned for another type of work?” And what that did is, it opened up a problem space for us. We said, well, that would be weird. And let us look at what artificial intelligence software asks of the underlying hardware. What’s hard about it? And let’s look at the architecture of the GPU and see if it’s well designed for that problem. What we found, of course, was that the graphics processing unit was a perfect machine for graphics processing. Not a surprise there, but it was less bad than a CPU. Those were the only two options over there.
As we turned that idea over in our head, we realized something very unusual about AI work. There are two things a computer does. It does calculations and it moves data. And AI has an unusual combination in that it does a huge amount of very simple calculations. Almost trivial, but an unbelievable number of them. But it has a very challenging problem in moving data. Moving data to memory and from memory, and breaking up problems and moving intermediate results around tens, hundreds, thousands of cores. We thought to ourselves, if we could make innovations on the dimension of communication, on the movement of data, we would be able to accelerate this workload and do so at a tiny fraction of the power draw. And so we set ourselves the task of building the largest chip in the history of the computer industry. And by building a big chip, we keep more data on chip, we have to move it less, we keep it in this extremely high performance domain on silicon where it uses very little power to be moved.5

No alternative text description for this image — From left to right: Sean Lie, Gary Lauterbach, Michael James, JP Fricker, Andrew Feldman. Credit: Cerebras

Others had tried and failed with wafer-scale integration in the past - most notably famous chip architect Gene Amdahl in the 1980s. Cerebras sought out and spoke to people who had prior experience with big chips:

I tell people, it’s a little bit like the first guys who tried to climb Everest, they got halfway up and said, “Well, this is really hard.” They came back down and they’re in a bar, and they’re talking to some new guys who want to climb Everest. They said, “it gets really hard at this spot halfway up.” The new guys say, “alright, we’ll worry about that.” These new guys, they get all the way up to the top and come back down and they’re back in the same bar. They meet the guys who only got up halfway and they say “the first half wasn’t the hard part. It was really hard near the top.”
All the problems that had broken previous efforts, we solved quickly. But, there are always new ones.

Here are some of the problems they encountered along the way, briefly:

Cross-die connectivity. Building a wafer-scale chip begins with the same lithography process as regular chips - you just leave the wafer whole rather than cut it up into chips. But you have to figure out how build wires across the blank scribe lines where the wafers are usually cut, in order to connect the individual dies. Additionally, they had to figure it out within the confines of the existing manufacturing process. And then, they had to persuade the chip manufacturer, TSMC, to actually do this for them. They own the IP on this.
Yield. All wafers have defects and this is one of the reasons why wafer-scale integration wasn’t supposed to work as not many of these giant chips would be usable. Cerebras figured out how to achieve a higher yield than GPUs through a fault-tolerant process. Took them a year and a half to solve.
Thermal expansion: A giant wafer gets hot, expanding and contracting with different coefficients in surrounding materials. They had to invent new flexible components that would prevent this from becoming a problem.
Package assembly: The necessary equipment to assemble their package didn’t exist, and the existing specialists wouldn’t build it for them. They had to build it themselves
Power and cooling: Chips are normally powered at the edges - but they discovered this doesn’t work with massive chips. They had to invent new ways to bring power directly into the surface of the wafer. Then they had to figure out how to uniformly extract heat - which involved designing a new water cooling system. All in all, packaging cost them $200m to figure out and 2.5 years. Listen to any of the Cerebras founders and you’ll hear a lot about how many (expensive) wafers they cracked.
Software: Depending who you listen to, it’s either the power/cooling problem that was the hardest or it was the software. The WSE has hundreds of thousands of tiny processors, each with local memory and connected by an on-wafer network - abstracting away the complexity to make the system run as a single accelerator was an exceptionally difficult problem (as I’m told)

Building the WSE was not just a regular chip design problem. It required vertical integration and the invention of an entirely new system around it. Not to mention the best part of $500m in venture capital to bring the first system to market.

The Wafer-Scale Engine & CS System

In August 2019 Cerebras came out of stealth mode and Sean Lie presented the CS-1 at Hot Chips. The CS-2 followed in April 2021. To repeat, this is more of an orientation than a technical deep dive - so here’s a helpful schematic of the WSE to get us started:

You’ll see it comprises the largest square that can be cut from a 300mm wafer. Within the square, you can see the the columns and rows of dies - essentially chips that are left in situ rather than diced. The dies, in turn, are connected by wires that cross the grid lines you can see (scribe lines). And each die is composed of thousands of uniform cores containing the logic and memory.

The CS-3 arrived in March 2024. Each system is an incremental improvement on the next in terms of transistor density, core count, SRAM capacity, memory bandwidth and cluster size. The CS-3 was, and is, a beast (datasheet here):

4 trillion transistors
900,000 cores
44GB SRAM
21 PB/s memory bandwidth
214 Pb/s fabric bandwidth

Each system draws a maximum 27kW, with two systems per rack and scalable up to more than 2,000 nodes in total.

I highly recommend watching this 4 minute video that Cerebras used to have on their Youtube page (link now via Wayback Machine). It’s a teardown of the CS-2 with commentary on the purposes of, and engineering behind, some of the core assemblies. You can see for yourself how radically different the CS system has to be in order to accommodate the WSE - to say it’s not like anything else that exists is a major understatement. They had to invent the entire system without recourse to the typical off-the-shelf parts. There are two notable absentees from Cerebras’ bill of materials however: high bandwidth memory and advanced networking equipment. And I’ll briefly note Cerebras’ ancillary memory and interconnect systems, that are principally intended for training.

Brief Thoughts & Speculations on Tokenomics

What follows will be semi-speculative and in numerous instances, wrong. The intention is to draw attention to some of the vectors that will inform relative WSE economics in future - and the roadmap.

Let’s start with a little context around the industry benchmark: Nvidia’s AI factory. The GB300 NVL72 rack based system is a throughput monster for popular models. If you want to produce massive numbers of tokens per second (TPS) per unit of power, it’s probably unrivalled in most current everyday cases - including against the CS-3.

But there’s a tradeoff - increasing the per-user TPS causes a drop-off in total throughput. Beyond 400 TPS or so, Nvidia starts to become particularly inefficient, even with compact economical models:

Nvidia is most efficient therefore when massive numbers of users are batched together, with the compromise of a low TPS user experience. Nvidia simply cannot supply you with a 1,000+ TPS experience (let’s ignore for now the Groq acquisition, but I note that Groq is yet to show >1,000 TPS either). Ultra fast inference is the domain of Cerebras, and it’s likely to command different economics.

Now turning to Cerebras. There tend to be two main critical claims that get levelled at the CS-3 and its economic viability:

low throughput: high speeds can only be achieved with low user concurrency
limited space for model storage: with only 44GB of on-system memory all but the most compact models have to be split among systems if you want to retain high speeds, meaning higher capital cost and power draw - Cerebras can’t practically inference large models

Both are reasonable claims, but also made within the context of incomplete information. Initially, some seemed to be assuming that Cerebras’ high speeds were limited to single user concurrency - which we now know to be incorrect. Here are some of the few public details we do have. Cerebras confirmed about a year and a half ago that, at the time, they were operating with between 8 and 20 concurrent users depending on configuration - all at maximum FP16 model density. This was right at the start of Cerebras’ inference launch and I assume they’ve continued to make advances on the software side since then. These configurations, from what we can tell, were presumably for Llama3.1 8B, Llama3.1 70B and Llama3.1 405B at the time. 8B ran on a single system, 70B on four systems and 405B on some higher number.

Much later, in August 2025, an analysis by Artificial Analysis showed that single-user speed did not drop when Cerebras scaled GPT-OSS-120B from 1 user to 10 concurrent users (that chart I showed you above). GPT-OSS-120B is natively quantized and fits within an 80GB of memory or presumably two CS-3s.

We’re working with scant details here and those we do have are likely very stale by now. My inklings are as follows:

On a pure TPS/watt basis the CS-3 probably looks significantly less economical than a GB300 NVL72 with high batching. But in return a Cerebras deployment can inference 10-20x faster.6 These are therefore different economic propositions and should be treated differently. It’s the equivalent of comparing the economics of a bus service to a taxi service - the latter is much more expensive to run per passenger and it moves fewer passengers, but the individual experience is much faster and more convenient, thereby commanding a premium
Cerebras currently works best with small to medium size models that can fit onto 12 or fewer systems. They have offered some relatively large models such as 405B and Qwen 3 Coder 480B in the past but not in continuous deployment.
Cerebras’ architecture seems well-suited to efficient mixture-of-experts thinking models like GPT-OSS-120 and GLM 4.7 - very high speeds combined with premium pricing still delivers competitive price-to-performance.

But is there anything they can do to achieve higher efficiency and throughout? Will they be able to handle larger models? It’s clear that Cerebras believes all the answers to be yes.

The Roadmap: what we know + some possibilities

Cerebras is yet to communicate a formalised technology roadmap, but it’s been clear for a while that something is cooking. The plan here is to show you what they have communicated and hinted at, and to highlight some of the possible areas targeted for improvement. But I repeat: I’m no expert. I’ll start with something that’s been formally announced and we know for sure is coming soon: disaggregated inference. Per the prospectus:

Our roadmap includes development of a disaggregated inference-serving solution.
Inference disaggregation is a technique that separates AI inference into two stages: prompt processing, or “prefill,” and output generation, or “decode.” These two stages have different computational characteristics. Prefill is natively parallel and requires very little memory bandwidth. Decode, on the other hand, is inherently serial and memory bandwidth intensive. Decode is typically the bottleneck. It dominates total inference time, and defines the speed of the user experience.
Cerebras’s wafer-scale engine would be the fastest at both prefill and decode, but in relative terms, it is much faster at decode. Our wafer-scale architecture and ultra-high memory bandwidth delivers faster output token generation where speed matters most. Disaggregated inference would allow Cerebras to operate alongside other architectures, serving as the high-performance engine for decode while other systems handle prefill.

In March, Cerebras and AWS announced such a partnership for disaggregated inference running on AWS Bedrock to serve leading open source models and Amazon Nova models. Pre-fill is to be handled by AWS Trainium and decode by the CS-3. According to Cerebras, they’ll be able to 5X the throughput for high-speed inference in the same hardware footprint. This could change the nature of the above highlighted tokenomics quite considerably for the better. Both companies appear to be taking the relationship quite seriously (per the S-1):

if future product purchases or leases by AWS under this agreement increase as contemplated by our minimum capacity commitments to AWS, this agreement may account for a material percentage of our revenue in the future.

So Cerebras has developed a disaggregated inference-serving solution and it’s going live with AWS in the months ahead. I would be amazed if they weren’t planning to use this solution in their OpenAI partnership too, at massive scale, as it would be likely to significantly improve the economics. One might further imagine that OpenAI would intend to use its own ASIC for the pre-fill part eventually.

Next up: what about future system generations? Per the prospectus:

Our roadmap builds on the advantages of wafer-scale integration. We intend to invest heavily in research and development to continue to expand on-chip memory and memory bandwidth, improve interconnect density, and leverage advancements in process technology to increase transistor counts and reduce power in future WSE generations. As a result, we expect that future generations of WSEs will have faster compute, and more and faster memory and communication onto and off of the wafer

I’ll share a few notes on several topics:

Expanding on-chip memory (SRAM)

This is an obvious way to improve a future CS system and it’s what they’ve done before - the CS-1 had 18GB of SRAM, CS-2 40GB and CS-3 44GB. The natural way to expand memory previously has been through shrinking the transistors (moving to a smaller process node). But the problem is that SRAM scaling has more or less stopped in recent nodes:

sram-density-tsmc-n3b-n3e.pngFred Chen · Dec 15, 2022 at 7:04 AM

Cerebras have previously confirmed their next wafer-scale engine will be on TSMC’s 3nm node which would imply 60-70% improved transistor density compared to the CS-3 on 5nm, however SRAM might only scale a few percent. So what are the options? An obvious one could be simply to dedicate more wafer surface area to SRAM while accepting a more moderate increase in raw computing speed. Others have suggested 3D stacked SRAM to add density. A less likely route could be simply to stop cutting the edges off the wafer - they’ve said the principal reason they do it is for convenience (presumably so the system can sit within a standardised rack space).7

Perhaps one of the most potent levers here could be to architecturally introduce native support for quantization - storing and executing the model weights at lower precision. This is what Nvidia does. Recall that the CS-3 was not designed specifically with inference in mind or for storing more than 44GB on-chip - storing models weights at full precision (FP16) might not have been envisaged as a bottleneck at the time. Adding quantization could double or quadruple on-chip capacity. There are presumably numerous other architectural innovations and efficiencies that could eke out more memory density too - I don’t know enough to tell you about those.

[Update, October 2026: CBRS finally gave us more detail about their roadmap in August 2026. I wrote a couple of posts about this here and here, and you can find some of the official materials here and here. The short version is that CS-4 onwards comes with a radically new rack-scale modular architecture, ‘Nexus’. Each Nexus rack includes 3 WSEs, among numerous other design, networking and manufacturability upgrades. CS-4 (ramping now) contains essentially the same silicon as CS-3 but with double the power. CS-5 (in late 2027) brings an all-new WSE (likely the 3nm part I referred to above). CS-4 and CS-5 will each roughly double the speed and throughput relative to the generation it succeeds. And then CS-6 (in 2028) introduces wafer-scale stacked DRAM for much higher memory density and greatly reduced system footprint for running large models at low cost (an design direction I certainly hadn’t anticipated back in May. While CS-4 and CS-5 certainly sound promising and extremely competitive (perhaps especially in combination with disaggregation) but CS-6 sounds like it could be the next major innovative disruption that potentially opens up an even larger market opportunity where speed and lower cost start to converge. Co-packaged optics is coming (we are told) but with no confirmed timeline yet - see more on this in the next section]

Interconnect density

I think this is one of the most exciting things that Cerebras has teased in public - only a couple of times. The first signs showed up a couple of years ago at an industry conference attended by JP Fricker:

Today the I/O is on two edges of the chip but it would be better if the I/Os were distributed across the chip. When you reduce the channel length you can reduce the size of the SERDES, saving space and power.
We’d like to have a very large number of optical engines. At the moment they are externals but eventually we will get those lasers into the chip. These would be used for multiple communications lanes at reasonable data rates of 100 to 200Gbit/s rather than fat pipes.
We have our wafer-scale engine and we take a third party wafer-scale programmable optical interconnect and combine them, using the entire surface of the wafer to connect to a wafer.

Wafer-scale co-packaged optics. About a year later, we learnt that work was well underway via a partnership with optics specialist Ranovus, who teased a photo of their own wafer-scale optical interconnect solution:

Cerebras Integrates Wafer Scale Technology and Ranovus Co-Packaged Optics Technology — WSE-3 on the left, Ranovus wafer-scale solution on the right. Credit: Ranovus

Ranovus’ CEO seems to be saying that once fully implemented the solution will provide at least 160 Tb/s in communication bandwidth - this compares to just 1.2 Tb/s today on the CS-3, so more than 100x greater. Cerebras would then be very competitive with Nvidia on scale-out. Whether they’ll be able to implement this on the CS-4 or whether it will come in later generations remains to be seen.

While Cerebras is being somewhat tight-lipped about its roadmap, I think we should expect to see some exciting developments in due course and I would assume we’ll see a CS-4 in the coming months.

Software

Nvidia has been wringing out performance improvements from software for years. Cerebras say they’re just getting started with software optimisations e.g. speculative decoding, predicted outputs, kernel and graph level optimisations and kv cache and memory layout. As an example, when they launched inference with Llama3.1 70B in August 2024 they were delivering 450 TPS - they can do more than 2,000 TPS now.

OpenAI Co-design

The prospectus also disclosed some notable details about the OpenAI partnership:

Our partnership with OpenAI also allows for collaboration and co-design across both frontier model development and hardware architecture. This hardware-software co-development enables OpenAI to design models built for our hardware architecture and Cerebras to evolve hardware design in response to the needs of upcoming frontier model architectures. This creates a continuous feedback loop that can help our systems prepare for the next generation of AI, establishing a structural advantage for Cerebras.

As noted, there remains a good deal of scepticism that Cerebras will be able to serve frontier models. For their part Cerebras have already confirmed they intend to serve the “largest frontier models” with OpenAI before the end of 2026 - details remain sparse and there’s likely some wiggle room on what that means. Either way, it’s clear that OpenAI now plays an important role in the technology roadmap. Until there are major hardware advances (as detailed above) I’m guessing we’ll see OpenAI deploy highly optimised small- to medium-size thinking models on Cerebras, potentially with some longer thinking time compensating for lower model intelligence:

Business Model & Sizing The Opportunity

Other have done S-1 tear-downs - I’m not going to do that here. My focus will be a quick review of how the business model has evolved and what progress has looked like so far. We’ll then consider the new contracts Cerebras has announced with OpenAI and AWS and the nature of how things might evolve from here.

Business Model

Cerebras briefly flirted with an IPO back in 2024 (old S-1 here) and we can see they booked $25m in revenue in 2022, principally through hardware sales. At the time of the 2024 S-1, G42 accounted for well over 80% of revenue - which raised eyebrows at the time. Fast forward to 2025 and revenue has grown to over $500m - a 20X increase in three years. Gross margins are reasonable if unimpressive at 40%, and the R&D expense line is massive, at you might expect - that’s about all we can say at this stage, and it would be dangerous to make any strong inferences about the future from these figures.

In 2025 the largest customer was Mohamed bin Zayed University of Artificial Intelligence (like G42, another UAE sovereign entity) with around 60% of revenue - a further example of Cerebras’ non-typical route to gaining acceptance and scale. The business continues to pivot and evolve, most notably and increasingly in the direction of cloud computing. Today they operate a flexible deployment model across inference, training and HPC workloads:

Cerebras Cloud: open-source or proprietary models through an API
Partner Clouds: access through AWS Marketplace, OpenRouter, Hugging Face and others
On-premise
Hybrid

Cloud customers access compute through one of two models: 1) dedicated take-or-pay capacity contracts or 2) self-serve on-demand / PAYG. Through all channels, it’s clear that Cerebras has been growing as fast as it is able, as evidenced by tight capacity. Shortly after Cerebras began serving Codex-Spark for OpenAI in February they pulled access to all self-serve models other than OpenAI’s GPT-OSS in some channels and then pulled that too for a full month - presumably to free up capacity from (less profitable) PAYG customers:

While they continued to invest in datacenters throughout 2024-2025, it became clear they were going to need a lot more capital to fund growth in their cloud business - through two more venture rounds, a prepayment from OpenAI and the IPO, Cerebras will have raised in the region of $6-7bn since September.

The base of enterprise customers is already starting to broaden quite nicely into areas like financial services (e.g. Alphasense) and software coding (e.g. Cognition/Windsurf). But let’s be honest, even with booming growth this part of the business is likely to remain a small percentage in the near- to medium-term. Here’s what really matters:

OpenAI details

750MW of capacity to deploy in tranches from 2026-2028, each tranche with a term of 3-4 years and extendable to 5 years
Contracted value more than $20bn
OpenAI has an option for 1.25GW in incremental capacity in tranches to be deployed by the end of 2030
Certain exclusivity provisions regarding named competitors (presumably Anthropic and Google) [Note: if I was Cerebras I’d only have agreed to this continuing past 2027 if OpenAI opts into the additional capacity, giving an extra incentive for OpenAI to take it]
OpenAI has warrants to acquire >33m shares at a negligible exercise price - potentially reflecting up to 15% of the contemplated post-IPO share count. The warrants vest in tranches depending on various milestones such as a working capital loan (already vested), capacity delivery, exercise of additional capacity tranches, market cap etc - note that these will show up as contra-revenue
RPO - the prospectus notes a remaining performance obligation of $24.6bn of which 15% will be recognised in 2026-2027, 43% in 2028-2029 and the rest thereafter. A ‘significant’ amount of the balance is noted to be due to OpenAI (with no other customers mentioned)

AWS details

AWS will become the first hyperscaler to deploy Cerebras in its own data centers
a multi-year strategic collaboration to deploy AI inference compute infrastructure and related services, and includes a binding initial multi-year lease
“if future product purchases or leases by AWS under this agreement increase as contemplated by our minimum capacity commitments to AWS, this agreement may account for a material percentage of our revenue in the future”
Warrant for 2.7m shares with $100 exercise price, vesting tied to purchase volumes ‘substantially beyond’ the initial lease
Binding term sheet with definitive agreement yet to be negotiated and executed
Term sheet includes exclusivity provisions

Oracle?: I mention Oracle because in their last earnings call they causally mentioned they were offering the latest accelerators including “emerging designs from companies like Cerebras”. I suspect what’s going on here is a ‘Hardware Purchase’ under the terms of the Cerebras-OpenAI Master Agreement with OpenAI buying CS-3s to go into Oracle’s datacenters - with such purchases counting as committed capacity (and Oracle being one of OpenAI’s core infrastructure providers).

Sizing The Opportunity

It’s clear that Cerebras will increasingly become a cloud infrastructure business - a vertically integrated one with a unique high-value offering. With enough scale, I would assume Cerebras’ financial profile could loosely resemble well established highly-profitable cloud infrastructure players over time - but it will take time.

I think the best we can do is to look at the range of possibilities for revenue, according to publicly disclosed commitments. Here’s what we get if we assume the following: a) OpenAI committed capacity fulfilled with zero extensions/additions, b) $20bn of the RPO is OpenAI c) a ramp trajectory for OpenAI capacity (i.e. the bulk of the 2026-2027 RPO recognition is 2027-weighted) and d) ‘material’ in the case of AWS means 10% of revenue by the end of the decade:

It’s clear that things could start getting very interesting from 2027 onwards. Note the reason for the drop after 2029 is the assumption that OpenAI does not exercise the option to extend capacity tranches beyond the initial terms - obviously, in the real world Cerebras would not leave this capacity sitting idle.

Building on this we might further assume a) maximum OpenAI extensions/additions and b) similar contract terms to the initial tranches and c) AWS materiality is 20% by 2031:

It’s worth noting that in either case we’re simply talking about previously disclosed commitments - no growth in RPO, no short cycle or on-demand business, no new customers. I’m also very aware that I’ve barely touched on the model training opportunity in this post - I have assumed that Cerebras will be primarily an inference business in future. But it’s quite possible that Cerebras will eventually be able to make some inroad into Nvidia’s moat here especially with a foot in the OpenAI door now - a discussion for another time. We’re also assuming that Cerebras remains highly niche within the broader context of a massive market - and less than 5% the size of Nvidia at the end of the decade. What if Cerebras can breakthrough into the mainstream with future ‘less quaint’ systems? Recall that 2030 is further away from today than is the launch of ChatGPT in late 2022 - expect many more twists and turns.

[Update, October 2026: my 2027 revenue forecast actually turned out to be very close to where consensus has been since IPO. But my 2028 number is significantly lower than consensus which is at ~$7.5bn. Both consensus numbers, it turns out, were influenced by colour given to analysts around the time of the IPO roadshow (which I did not take part in). Notably, very few analysts even participate currently in the forecast beyond 2028, presumably due to the lack of visibility. In the intervening months, I’ve gradually come to the view that I was probably too conservative, even in my ‘stretch’ scenario]

Risks & Other Considerations

In the right conditions therefore its very conceivable we could see an extraordinary increase in market cap over the next five years - perhaps to well over $100bn relative to $26-27bn at IPO. But there are important risks and considerations:

Dilution. It’s been an exceptional challenge getting Cerebras into its current position and I have no issue conceptually with the relevant risk-takers getting paid. But there’s no escaping that Cerberas will be a dilution monster (I refer you to page 15 of the prospectus). Additionally senior management are incentivised on aggressive market cap targets ($75bn, $150bn, $250bn) which don’t factor in share count. It’s conceivable basic share count could increase by 50%+ in the next 5 years - but higher ranges are also likely to be associated with important positive milestones such as the additional OpenAI capacity.
Capital. It’s unclear how much more share capital they may need to raise post-IPO.
Execution. Very rapidly growing businesses almost always run into operational growing pains of some kind. They’ll also need to execute on their roadmap in a timely fashion - not easy.
OpenAI. Cerebras’ fate is tied to OpenAI to a large extent and the shares will most certainly trade around this dynamic.
Contra-revenue. Reported revenue will be negatively impacted by non-cash amortization of the share warrants.
AI capex bubble bursting. To state the obvious. However, given the nascency of the market for ultra-fast tokens and its less commoditised nature, I do wonder if it could be robust to cuts in capex related to commodity tokens. We’ll see…
Takeover. Cerebras could be a target (AMD seems to me like the most probable acquirer).

Final Thoughts: Moat

Once Cerebras is up and running with a vertically integrated hyperscale one-of-kind ultra-high-speed inference cloud I think it will be incredibly hard to compete with in its high-speed niche for users who need the fastest tokens:

structural speed advantages to wafer-scale
wafer-scale is exceptionally hard to replicate (know-how, timescale + IP)
frontier lab / wafer-scale co-design increases switching costs and performance advantage
barriers to gaining customer acceptance
capital barriers
essential adaptability: hardware and management that can pivot/evolve

Thanks for reading and best of luck to anyone participating in the journey. Speaking for myself, it’s possible I could end up selling my holding should the share price go truly bananas in the months ahead (lock-up depending), but equally if the shares collapse I might buy more (context depending). Whatever happens, it’ll certainly be exciting to watch how the business evolves in the coming years. Stay tuned for updates.

Notes From The Beauty Contest

Cerebras Systems: A Deep Dive