Could Cerebras looming IPO “Out-Chip” Nvidia

Cerebras Systems. Founded in 2015 in Sunnyvale, California, Cerebras was built by a team of semiconductor veterans – several of whom held senior engineering roles at AMD, Nvidia’s biggest GPU rival. From day one, they set out to solve a problem that had been quietly frustrating the AI world for years: standard chips, no matter how powerful, are small. AI models are enormous. And making thousands of small chips talk to each other quickly enough to train or run a cutting-edge model is messy, expensive, and slow.

Their answer wasn’t to make a better version of what already existed. It was to throw out the rulebook entirely.

What Cerebras does

Cerebras is an AI computing company. It designs and manufactures specialised processors for training and running AI models – the same fundamental task that Nvidia’s GPUs dominate today. But the way it goes about that task is radically different.

The company sells its hardware to organisations that need massive AI compute – governments, universities, research labs, and increasingly, commercial cloud providers.

More recently, it has evolved beyond just selling chips. Cerebras now operates its own AI cloud service, running its processors inside data centres and offering inference-as-a-service to developers who want blazing-fast AI responses without managing any hardware themselves.

It’s a shift from chipmaker to full-stack AI infrastructure provider – and that pivot matters a great deal to how investors should think about its long-term value.

One giant chip to rule them all?

Here’s where things get genuinely interesting.

Imagine taking an entire silicon wafer – the dinner-plate-sized disc from which thousands of ordinary computer chips are normally cut – and instead of slicing it up, turning the whole thing into a single, massive processor. That’s exactly what Cerebras has done with its Wafer Scale Engine, now in its third generation.

The WSE-3 is not incrementally bigger than a standard chip. It is incomprehensibly bigger. At 46,225 square millimetres, it is 57 times larger than Nvidia’s flagship H100 GPU.

It packs 4 trillion transistors, 900,000 AI-optimised cores, and memory bandwidth that Cerebras claims is over 2,600 times greater than what an Nvidia GPU cluster can achieve.

Why does size matter so much?

Because in AI computing, one of the biggest bottlenecks is moving data between chips. When you’re running a model across thousands of GPUs, enormous amounts of time are wasted just on the communication between those chips. Cerebras eliminates that problem by keeping everything on one die. No interconnects, no traffic jams, no wasted cycles waiting for data to travel across a network.

The result, at least for inference workloads – the real-time task of answering a user query or generating text – is speed that Cerebras claims is up to 15 times faster than comparable GPU-based setups.

That’s the claim that caught OpenAI’s attention. And it’s the claim that CEO Andrew Feldman made personal when he told the Wall Street Journal that Cerebras had taken the fast inference business at OpenAI directly from Nvidia.

The Market Opportunity: Why Timing Matters

The AI infrastructure market is growing at a pace that would have seemed fictional five years ago. Hyperscalers – the Amazons, Googles, and Microsofts of the world – are spending hundreds of billions of dollars building data centres to power AI services.

Inference is becoming the dominant workload. As AI moves from the lab to everyday products, the demand isn’t just for training massive models once – it’s for running those models billions of times a day, as fast as possible, as cheaply as possible.

That’s precisely the workload Cerebras has engineered its chip to win. The inference market is expected to dwarf the training market over the next decade, and Cerebras is positioning itself as the specialist in exactly that space.

The numbers tell an interesting story

Revenue hit $510 million in 2025, up 76% from the prior year. On the surface, the company also reported net income of $87.9 million – a dramatic swing from a $484.8 million net loss in 2024.

Look closer, though. The operating loss was actually $145.9 million, and the headline profit was almost entirely driven by a $391 million accounting gain from the remeasurement of a contract liability – a one-time item with nothing to do with the day-to-day business. Strip that out, and Cerebras was still losing money from operations.

The company is spending heavily on R&D, which consumed 48% of annual sales in 2025. Operating cash flow was negative, though only slightly, at -$10.1 million.

The more compelling number is the backlog. Cerebras reported $24.6 billion in remaining performance obligations as of December 31, 2025.

The incoming Cerebras IPO

Recently, the company filed its S-1 registration statement with the SEC, officially kicking off its path to a public listing on Nasdaq under the ticker CBRS. A mid-May debut is the target. For anyone watching the AI infrastructure race, this is one of the most anticipated IPOs in years.

The company is targeting a valuation of around $35 billion in the IPO, despite its most recent private round pricing it at $23 billion in February 2026. Pre-IPO shares on secondary platforms like Forge and EquityZen have been trading at implied valuations of $26–28 billion, suggesting sophisticated investors see the IPO price as potentially conservative, with room for a first-day pop.

At $35 billion, Cerebras would be trading at roughly 43–49x its 2025 revenue – aggressive compared to Nvidia’s ~25x forward multiple, and well above AMD’s ~10x. But context matters.

Cerebras isn’t being valued on what it earned last year. It’s being valued on what that $24.6 billion backlog implies about the next five years. If those performance obligations convert to revenue on schedule, the current multiple looks far less frightening.

The takeaway for investors

Cerebras is a genuine technological disruptor entering public markets at a fascinating inflection point. Its OpenAI and AWS partnerships are real validation. The backlog is enormous. Revenue is growing fast.

But this is not a low-risk investment. The company is still operationally unprofitable, the customer base is heavily concentrated, and the chip market it’s competing in – dominated by Nvidia, with aggressive moves from AMD and hyperscaler custom silicon – is brutal. The valuation being sought is ambitious.

For investors, the most sensible approach is to treat Cerebras as a high-conviction speculative position rather than a cornerstone holding. The technology is real, the partnerships are meaningful, and the market for AI inference compute is enormous. If Cerebras can diversify its customer base and execute on its backlog, the IPO price could look cheap in hindsight.