The short version
Cerebras Systems thinks the entire architecture of AI hardware is wrong. Not slightly wrong. Wrong at the foundational level. Their argument is pretty direct: the way the industry has always built chips, by dicing a silicon wafer into hundreds of small dies and stitching them together with interconnects and memory hierarchies, hits a wall when you try to scale to the models the world wants to run today. The fix, they say, is to skip the dicing entirely and build a chip the size of a whole wafer.
That idea came from a founding team that had previously sold a semiconductor startup to AMD for $334 million. Ten years later, they've raised $715 million, built relationships with national labs and pharma companies, closed a $1 billion-plus deal with a UAE sovereign AI fund, and grew revenue 10x in a single year. They've also spent most of that time operating in Nvidia's shadow, watched hyperscaler customers start building their own chips, and navigated a national security review that pushed the IPO back by more than a year.
The memory bandwidth wall
To get why Cerebras exists, you need to understand a constraint most AI chip coverage glosses over. The bottleneck in training and running large neural networks isn't compute. It's the time it takes to move data between memory and the cores doing the math.
Modern GPUs are organized around a memory hierarchy. You have fast on-chip SRAM (expensive, small, a few hundred megabytes), then HBM stacked on the package (tens of gigabytes, fast but physically separate from the cores), then system memory below that. Every matrix multiplication in a neural network means hauling weights up from wherever they live in that hierarchy, doing the math, and writing results back down.
For small models, this is fine. For GPT-scale models and bigger, the weights don't fit on-chip. Every forward pass means shuttling hundreds of gigabytes across a bandwidth-limited bus. Nvidia has made real improvements here with NVLink and faster HBM generations, but the gap between compute throughput and memory bandwidth has only widened as models have grown. An H100 can do roughly 3,900 teraflops of bfloat16 math per second. Its HBM bandwidth is 3.35 terabytes per second. For the models researchers actually care about today, the chip spends a lot of its time waiting for data.
Things get messier when you scale up. Running a 70-billion-parameter model across a cluster of H100s means sharding the model across cards, synchronizing gradients at every step, and dealing with communication overhead that compounds as you add nodes. NVLink handles this okay at small scale but it becomes a serious software and infrastructure problem at large scale. Cerebras's argument is that the whole distributed training stack is basically a workaround for the wrong underlying hardware. Solve the architecture problem and the software problem mostly goes away.
The wafer-scale architecture
The Wafer-Scale Engine does something simple but unusual: instead of slicing a 12-inch silicon wafer into hundreds of individual chips, Cerebras just keeps the wafer intact. The result is a single die roughly 56 times larger than an H100. Not 56 percent larger. Fifty-six times. At 46,225 mm², the WSE-3 is bigger than any chip produced at volume by any company in the industry.
Wafer-scale integration isn't a new idea. Back in 1979, a company called Trilogy Systems raised $230 million to build a wafer-scale supercomputer. It failed because of the yield problem: any silicon wafer will have defective transistors scattered randomly across its surface, and at wafer scale, avoiding all of them is basically impossible. Cerebras's manufacturing answer is a technique called redundant routing. The WSE is built with enough spare pathways and cores that the chip can route around defective sections at the factory and still ship working. Getting this to work commercially took several years of close collaboration with TSMC.
The end result is a chip with 900,000 cores and 44 gigabytes of on-chip SRAM. That SRAM sits physically next to every compute core, so the bandwidth between memory and compute is orders of magnitude higher than anything you'd get with off-chip HBM. For models that fit on-chip (or close to it), the difference is real. Cerebras says a CS-3 system, which is one WSE-3 plus the power and cooling infrastructure around it, delivers the performance of a room full of GPU servers with a fraction of the code complexity.
| Generation | Year | Transistors | Cores | On-chip SRAM | Notable |
|---|---|---|---|---|---|
| WSE-1 | 2019 | 1.2T | 400,000 | 18 GB | First wafer-scale chip in history |
| WSE-2 | 2021 | 2.6T | 850,000 | 40 GB | Powers Andromeda supercomputer |
| WSE-3 | 2024 | 4.0T | 900,000 | 44 GB | CS-3 system ships; SwarmX + MemoryX |
Beyond the chip itself, Cerebras built two system-level products to extend what the WSE can do. MemoryX adds up to 1.2 terabytes of external memory per system (enough for 24 trillion model parameters) and streams weights into the WSE on demand, for when a model is too big to live on-chip. SwarmX connects up to 2,048 CS-3 systems into a single cluster and, according to Cerebras, scales near-linearly as you add nodes. That last part matters more than it sounds. GPU clusters scale sublinearly because of communication overhead, which means you end up needing more hardware than you'd expect to hit a given performance target. If the SwarmX numbers hold up, it's a real differentiator.
What Cerebras actually sells
Cerebras has three distinct revenue lines. They matter because the margins and growth dynamics are pretty different across each.
| Product | Price point | Customer type | What it offers |
|---|---|---|---|
| CS-3 hardware | ~$2M per unit | Sovereign governments, national labs, defense agencies | On-premise deployment, air-gapped where needed, full data control |
| Supercomputing packages | $50M–$100M+ | Sovereign AI funds, research consortia | Condor Galaxy-class multi-system clusters, tuned for frontier model training |
| Cloud / inference API | Pay-per-hour or flat fee | Developers, enterprises, research teams | Hosted access via Andromeda; fastest publicly available inference API |
The hardware business is lumpy by nature. A $100 million supercomputing deal moves the revenue needle dramatically in the quarter it closes, then contributes nothing the next quarter unless there's a follow-on order. That lumpiness shows up in Cerebras's financials and complicates the IPO story a bit. Investors pricing the company on a revenue multiple really need a view on the backlog and pipeline, not just what's been recognized so far.
The cloud and inference API business is steadier and should carry higher long-run margins. Cerebras launched its inference API in mid-2024 and almost immediately got attention for delivering the fastest publicly available inference on open models, actually outpacing Groq, which had built its whole brand around inference speed. The API comes in three tiers: free, developer, and enterprise. The real money is in enterprise, which charges flat monthly fees for dedicated inference capacity.
The software story is worth mentioning too, and it's easy to underestimate. Cerebras claims running a GPT-3-scale model on their hardware takes 565 lines of Python versus 20,000 lines across multiple languages on Nvidia clusters. That's a marketing number, but the underlying point is real. For research teams that would otherwise spend months building a distributed training stack, eliminating that complexity is genuinely valuable. And for national labs and pharma companies running specialized models on proprietary data, getting fast results with simple code often matters more than maximizing raw FLOP efficiency on cheaper hardware.
Who's buying and why
Cerebras's customer base falls into three groups, and understanding the logic behind each purchase explains how the company has real traction despite competing against an incumbent with a 10-year ecosystem head start.
National laboratories and government research
Argonne National Laboratory, Lawrence Livermore, Sandia, and Los Alamos are all Cerebras customers. These aren't buying hardware on a bet. Their workloads look nothing like the transformer training runs that dominate AI coverage. They're simulating fluid dynamics, modeling nuclear reactions, folding proteins at scale, running molecular dynamics at resolutions no commercial GPU cluster has achieved. Cerebras has published benchmarks showing 300x speedups over prior Argonne infrastructure on cancer research workloads, and a 179x speedup versus the Frontier supercomputer on molecular dynamics. Frontier was the world's top-ranked supercomputer at the time. These are extraordinary numbers, and they reflect a real advantage in workloads with irregular memory access patterns that GPU architecture handles poorly. For the national labs, Cerebras isn't a speculative buy. It's a tool that has demonstrably solved problems they couldn't solve otherwise in reasonable timeframes.
Pharma and life sciences
GlaxoSmithKline, AstraZeneca, Mayo Clinic, Genentech, and Bayer are all on the customer list. The use cases are similar to the national labs: large simulation and model training workloads, often on proprietary biological data that can't be sent to a public cloud. AstraZeneca has described Cerebras as enabling work that previously took weeks to complete in days. That pattern shows up again and again with life sciences customers. When computation is the rate-limiter on expensive wet lab experiments, the cost of faster hardware starts looking small relative to what a compressed cycle time is worth.
Sovereign AI
The most financially significant customer is G42, the Abu Dhabi-based AI company backed by the UAE government. The two Condor Galaxy supercomputers deployed with G42 in 2023 were each valued at around $100 million. A follow-on contract for seven more systems was then signed for $900 million total. That makes the G42 relationship worth over $1 billion, which is the backbone of Cerebras's revenue story. It's also the source of the biggest near-term risk, which I'll get to.
The sovereign AI pattern extends well beyond G42. Governments that want AI capability without routing data through US hyperscaler clouds are natural Cerebras customers. On-premise deployment at sovereign data centers, air-gapped from outside networks, on hardware that isn't subject to the same export scrutiny as Nvidia's highest-end chips. That's a real value proposition, especially for countries that have watched US government scrutiny freeze Nvidia access for others. DARPA being on the customer list gives a different kind of strategic credibility on top of that.
The numbers that matter
Cerebras's S-1, filed in September 2024, shows a real business. 2023 revenue came in above $250 million, up 10x from 2022. Manufacturing capacity was also scaled 10x the same year just to keep up with backlog. For a hardware startup, those are striking numbers.
The numbers come with real caveats though. A large fraction of 2023 revenue came from the G42 contract, which means the business is highly concentrated. That's standard S-1 risk factor language, but it's more significant here because the G42 relationship itself attracted a national security review. The US government looked at whether advanced AI hardware shipped to a UAE entity with ties to Chinese technology companies posed export control risks. That review was a big reason the IPO got pushed back, and even after Cerebras navigated it, the G42 exposure remains a point of investor scrutiny.
Hardware margins are also structurally lower than software or cloud margins, and Cerebras hasn't disclosed detailed numbers. The CS-3 is a $2 million system that needs liquid cooling and substantial power infrastructure. The economics look more like selling enterprise servers than selling software licenses. The path to the kind of margins that justify a high revenue multiple runs through the cloud and inference API side of the business, which the company has been pushing hard since mid-2024.
The competitive landscape
Cerebras isn't operating in isolation. The AI hardware market has attracted more venture capital than almost any other category in deep tech, and Nvidia hasn't been standing still. Here's how the landscape actually breaks down:
| Company | Approach | Funding / Status | Primary threat to Cerebras |
|---|---|---|---|
| Nvidia | GPU + CUDA ecosystem; H100/B200/GB200 series | $2.6T market cap, dominant | Ecosystem lock-in; every new developer starts on CUDA |
| Groq | LPU (Language Processing Unit) for inference; 350+ tokens/sec on Llama 3 | $1B+ raised; $2.8B valuation | Direct inference API competitor, strong developer traction |
| SambaNova | Reconfigurable dataflow architecture; full-stack offering | $1.1B raised; $5B valuation | Overlapping enterprise and government customers |
| EtchedAI | Sohu: transformer-only ASIC, claims 20× B200 on inference | $125M raised; pre-revenue | Higher inference efficiency if claims hold at scale |
| AWS Trainium / Google TPU | In-house chips; hyperscaler cloud, not sold externally | Billions of internal capex | Reduces TAM by capturing hyperscaler internal workloads |
| AMD MI300X | GPU with higher HBM (192 GB); drops into existing CUDA-adjacent toolchains | $231B market cap | Better memory capacity on familiar programming model |
The most important competitive dynamic isn't between Cerebras and Groq or Cerebras and SambaNova. It's between Cerebras and the collective decision by Amazon, Google, Microsoft, and Meta to build their own silicon. The hyperscalers together spent around $200 billion on capex in 2024. A meaningful chunk of that went toward chips that reduce their Nvidia dependence and, incidentally, remove them from Cerebras's addressable market entirely. Google's TPU v5 is reportedly competitive with the H100 on training runs. AWS Trainium2 is being deployed at scale inside Amazon. None of these chips are sold externally, but they represent the biggest AI training workloads in the world being absorbed internally.
Where Cerebras has a real, defensible advantage is in the workloads I described earlier: large-scale scientific computing, life science simulation, and on-premise sovereign deployments. Those markets are growing and they're not being touched by in-house hyperscaler chips. The open question is whether they're large enough to justify the IPO valuation.
What the IPO is pricing in
Cerebras filed its S-1 in September 2024. The listing was delayed first by the national security review of the G42 relationship, then by general market conditions for deep-tech hardware IPOs. By the time it actually hits the market in 2026, the company has been building this story for a decade and has real revenue to show for it.
The challenge in valuing Cerebras is that it sits at the intersection of multiple comp sets that analysts genuinely disagree about. Is it a semiconductor company? Nvidia trades at around 30x forward revenue; AMD is closer to 6 to 8x. Is it an AI infrastructure company? Closer to the Nvidia multiple. Is it a government contractor with lumpy, concentrated revenue? Much lower multiples. Honestly it's a mix of all three depending on which part of the business you're looking at.
| Company | Revenue (last 12M) | Market cap | Revenue multiple | Business similarity |
|---|---|---|---|---|
| Nvidia | $130B+ | $2.6T | ~20× | AI chips; dominant market position |
| AMD | $25B | $230B | ~9× | AI chips; challenger position |
| Marvell | $6B | $60B | ~10× | Custom silicon, data center |
| Palantir | $2.8B | $170B | ~60× | Gov't / enterprise AI; high growth |
| Cerebras (IPO) | ~$500M–$700M est. | TBD | Depends on narrative | AI hardware; concentrated customers; sovereign focus |
The range of plausible outcomes is wide. At a Palantir-style software growth multiple, Cerebras would trade well above $10 billion. At an AMD-style hardware multiple on disclosed revenue, it's more like $2 to $4 billion. Where it actually clears will depend on the story the company and its bankers manage to tell, and whether institutional investors see the G42 concentration as a massive pipeline or a single-customer liability.
The bull case
The bull case comes down to three things: the architecture, the customers, and the timing.
| Bull case | Bear case |
|---|---|
| Memory bandwidth is a real bottleneck; WSE-3 solves it in a way no competitor does | Nvidia keeps improving HBM bandwidth and NVLink; gap closes over time |
| National lab and pharma customer base is sticky, growing, and not Nvidia-dependent | Revenue highly concentrated in G42; single customer risk is substantial |
| Sovereign AI is a structural tailwind; every government wants on-premise AI | Hyperscalers capturing training workloads with in-house chips shrinks the TAM |
| Inference API is the fastest available; developer adoption is growing | Groq and EtchedAI are credible inference competitors; it's a race |
| 10x revenue growth in 2023; backlog large enough to double manufacturing | Hardware margins are structurally low; path to software-like margins is unproven |
| 100+ patents across founders; wafer-scale manufacturing is hard to replicate | TSMC dependency for manufacturing; competitors getting CHIPS Act subsidies |
On architecture: the wafer-scale bet has been validated by customers who actually know what they're doing. Argonne and Lawrence Livermore don't buy hardware because of press releases. They buy because benchmark results show their workloads run meaningfully faster, and they've demonstrated 100x-plus speedups on problems that actually matter. That's a real moat in scientific computing, and it's not one Nvidia can close quickly. GPU architecture is optimized for parallelizable matrix operations, not for the irregular memory access patterns that dominate physical simulation.
On timing: the inference API launch in mid-2024 hit the market right as demand for fast, cheap inference on open models was starting to explode. Mistral, Llama 3, Qwen, DeepSeek-style models are increasingly competitive with closed-source alternatives, and the ecosystem around them is growing fast. Cerebras's API delivers measurably faster token generation than competitors at scale. If that position holds as model sizes grow, it's an entry into a market worth billions a year, with software-level margins.
The bear case
The bear case isn't that the technology doesn't work. It does. The bear case is about market structure.
The AI hardware market is getting bifurcated in a way that's not great for Cerebras. At the high end, hyperscalers are vertically integrating. Amazon, Google, Microsoft, and Meta collectively drive the majority of AI training compute demand in the world, and they're all building chips to service that demand internally. That was never really Cerebras's market. But it does mean the total addressable market for third-party AI training hardware is smaller than the headlines imply. The buyers are concentrated in customers who genuinely can't use hyperscaler chips: sovereign governments, regulated industries, organizations with unusually demanding performance requirements.
At the lower end, the inference market is getting commoditized quickly. Groq has been building inference infrastructure for years and has strong developer mindshare. EtchedAI's Sohu chip is designed specifically for transformer inference and claims metrics that, if they hold up at scale, would make it more efficient than the WSE-3 for that use case. Tenstorrent is selling inference cards for $599 to $799, targeting developers who want Cerebras-like simplicity at a fraction of the cost. Inference will probably support multiple winners, but it's not a segment where Cerebras has a clear structural advantage long-term.
The G42 concentration is the most immediate risk. If that relationship hits any more political complications, whether through tighter US export controls, UAE policy shifts, or G42 funding pressure, the revenue impact would be immediate and severe. And even setting aside the geopolitical angle, a business where one customer accounts for most of the revenue just trades at a discount to a diversified one. That's not unfair.
My take
The strongest version of the Cerebras story isn't about being a better training chip than Nvidia. That battle is nearly impossible to win. Nvidia's CUDA ecosystem has a decade of developer inertia behind it and the switching costs are real. The better story is about being the right hardware for workloads that look nothing like standard AI training: physical simulation, drug discovery, sovereign deployment, and inference at the bleeding edge of latency.
In those markets, Cerebras has real, demonstrated advantages. The national lab benchmarks aren't marketing copy. They're published research results from institutions that have no reason to oversell. The pharma customer base reflects a pattern that makes intuitive sense: when computation is the bottleneck on expensive biological experiments, organizations will pay meaningful premiums for anything that compresses the cycle. The sovereign AI market is real and growing, even if the G42 concentration creates near-term risk.
Where I'm less convinced is the inference API story. Cerebras can claim the fastest tokens-per-second numbers right now, but inference is where competition is most intense and hardware improvements are coming fastest. Speed leads in inference tend to last months, not years, before someone catches up. If the cloud business depends on holding that lead indefinitely, that's a fragile foundation to build on.
The question worth asking about any hardware company going public is: what's the enduring advantage that holds as technology improves? For Cerebras, the most honest answer is the architecture itself, wafer-scale integration that places on-chip memory next to every compute core, and the decade of manufacturing expertise with TSMC that makes that architecture commercially viable. That advantage is real and defensible. Whether it's worth $5 billion or $15 billion depends on how large the market for non-hyperscaler AI hardware turns out to be, and that market is still being defined.
I wouldn't bet against the technology. I'd be watchful about the customer concentration, the pace of inference competition, and how quickly Cerebras can expand beyond the sovereign AI and national lab anchors. The IPO isn't a conclusion. It's a checkpoint in a decade-long bet on a genuinely different architecture. The next few years of results will tell us whether that bet paid off.