Cerebras files a $25B IPO challenging NVIDIA's 80% AI chip monopoly, wielding a WSE-3 chip 57 times larger than NVIDIA's and backed by an OpenAI $20B inference deal.

Cerebras Files $25B IPO to Break NVIDIA’s AI Chip Monopoly

by
💡 Key Takeaways

  • Cerebras Systems filed for a Nasdaq IPO on April 17 at a $22–25B valuation, pitching its WSE-3 — an AI chip 57 times larger than NVIDIA’s B200 — as the future of inference computing.
  • An OpenAI contract worth $20B+ for 750MW of inference infrastructure and a revenue swing from –$485M to +$87.9M anchor the bull case.
  • Global AI chip startup funding hit $8.3B in 2026, with challengers like Fractile (25x faster inference claim) and SambaNova exploiting the structural shift from training to inference — the very domain where GPU architecture is weakest.

46,225 square millimeters. That is roughly the size of a dinner plate. It is also the die area of Cerebras Systems’ WSE-3 (Wafer-Scale Engine 3), a single AI chip 57 times larger than NVIDIA’s flagship B200 GPU.

On April 17, the company behind that dinner-plate-sized chip filed for a Nasdaq listing under ticker CBRS, targeting a $22–25 billion valuation and roughly $2 billion in proceeds. This is not just another AI chip IPO — it is a formal declaration of war against NVIDIA’s dominance in data-center AI silicon.

The $25 Billion IPO — Cerebras Throws Down the Gauntlet

Cerebras’ Nasdaq filing comes after 18 months of regulatory turbulence. The company first submitted an S-1 in September 2024, but a CFIUS (Committee on Foreign Investment in the United States) review triggered by investor G42 forced a withdrawal.

Morgan Stanley is leading the offering, which targets a $22–25 billion valuation. That range tracks closely with the $23 billion post-money mark set during the company’s February 2026 Series H, when Tiger Global led a $1 billion round with participation from AMD and Fidelity.

In total, Cerebras has raised approximately $2.8 billion across eight funding rounds. The investor roster reads like a who’s who of institutional AI bets, and the IPO now gives public-market investors their first direct access to a pure-play inference chip maker.

Key point: After an 18-month CFIUS delay, Cerebras is making a second run at a public listing. The $22–25B valuation makes it the largest pure AI chip IPO attempt in history.

A Chip the Size of a Dinner Plate — WSE-3’s Disruptive Architecture

What makes the WSE-3 fundamentally different from conventional GPUs is not just its size. The chip packs 4 trillion transistors and 900,000 AI-optimized cores onto a single silicon wafer — numbers that make standard GPU specifications look quaint by comparison.

The real differentiator, however, is memory bandwidth. The WSE-3 delivers 27 PB/s (petabytes per second) of on-chip memory bandwidth — roughly 200 times the bandwidth of NVIDIA’s NVLink interconnect. Tom’s Hardware estimates that a single WSE-3 provides theoretical performance equivalent to approximately 62 H100 GPUs.

Futuristic AI chip wafer-scale processor in a data center environment (AI-generated)
The WSE-3 integrates 4 trillion transistors on a single wafer, targeting inference performance equivalent to 62 NVIDIA H100 GPUs — Generated by OpenAI DALL-E 3

In inference workloads — particularly the token-by-token decode phase of large language models — data must be fetched sequentially from memory. GPUs, designed for massively parallel matrix operations, hit a structural bottleneck here. The WSE-3’s SRAM-based architecture eliminates this bottleneck by keeping all model weights on-chip, avoiding the latency penalties of external memory access.

In simpler terms: If a standard GPU is a sports car on a highway, the WSE-3 is the highway itself built into the chip. Instead of shuttling data back and forth, it keeps everything in one place — eliminating the travel time entirely.

From a $485 Million Loss to $87.9 Million Profit — The Financial Turnaround

Cerebras’ financial trajectory is among the most dramatic in the AI chip sector. Revenue climbed from $78.7 million in 2023 to $510 million in 2025 — a 6.5x increase in just two years, with 76% year-over-year growth in the most recent period.

More striking than the top-line expansion is the bottom-line swing. In 2024, Cerebras posted a net loss of $485 million. By 2025, that had flipped to a net income of $87.9 million — a nearly $573 million turnaround in a single fiscal year.

The catalyst behind this transformation was a fundamental restructuring of the customer base. Cerebras moved away from its heavy dependence on G42 and secured OpenAI — one of the world’s largest AI infrastructure spenders — as its anchor client. That strategic pivot underpins the entire IPO narrative.

Key point: A 6.5x revenue surge and a $573 million swing from loss to profit in one year give Cerebras a financial story that few pre-IPO AI companies can match.

The OpenAI Contract — A $20 Billion Anchor for the IPO Narrative

The OpenAI-Cerebras deal stands as one of the largest single contracts in AI infrastructure history. The agreement covers $20 billion or more in inference infrastructure, starting with a 750MW deployment and including options to expand to 2GW by 2030.

The contractual structure binds the two companies at the equity level. OpenAI can earn up to 33.4 million shares in Cerebras through purchase-milestone warrants — a mechanism that aligns the customer’s growth directly with the supplier’s stock performance.

Close-up of an AI chip semiconductor wafer
Surging demand for AI inference is driving billions in investment toward GPU alternatives — Photo by Brian Kostiuk on Unsplash

In March 2026, Cerebras also signed a distributed inference architecture partnership with AWS. Under this arrangement, WSE-3 chips handle the decode phase — the sequential, memory-intensive step — while AWS Trainium chips manage the prefill computation. This hybrid approach lets each chip play to its architectural strengths.

💡 Related Posts
Nvidia Bets $2B on Silicon Photonics to Break AI’s Data Wall

Training to Inference — Why the GPU Monopoly Is Cracking

The simultaneous emergence of Cerebras, Fractile, SambaNova, and a dozen other AI chip challengers is not coincidental. It reflects a structural shift in where the AI industry’s computational demand is concentrated. The center of gravity is moving from model training to inference.

Patrick Schneider-Sikorsky of the NATO Innovation Fund put it bluntly in a CNBC interview: “Inference is dominant. Existing GPU architectures are not designed for it in the ways that matter most.”

The technical logic is straightforward. Training requires dense matrix multiplication across thousands of parallel cores — exactly what GPUs excel at.

Inference, by contrast, is bottlenecked by memory bandwidth, latency, and energy efficiency. LLM token decoding, in particular, involves sequential memory access patterns that expose a fundamental architectural weakness in GPU design.

As global AI inference demand overtakes training demand, the market for non-GPU alternatives is expanding rapidly. Cerebras’ SRAM-based on-chip memory and Fractile’s inference-optimized architecture are purpose-built to exploit this gap.

In simpler terms: Training is like building a library — you need raw construction power. Inference is like finding and reading the right book — speed depends on how fast you can navigate the shelves, not how many builders you have.

The $8.3 Billion AI Chip Rebellion — Fractile, SambaNova, and Positron

Cerebras is far from alone. Total global funding for AI chip startups in 2026 has reached $8.3 billion. Cerebras’ $1 billion round leads the pack, followed by MatX ($500M), Ayar Labs ($500M), and Etched ($500M) in what amounts to a coordinated capital assault on NVIDIA’s market position.

Behind the mega-rounds, a second tier of well-funded challengers is taking shape. SambaNova raised $350 million and unveiled its SN50 chip.

Positron closed a $230 million Series B. Axelera and Olix each secured more than $200 million.

The most provocative claim comes from Fractile, a UK-based startup that raised $200 million at a $1 billion valuation. The company asserts its architecture delivers 25 times faster inference than NVIDIA at one-tenth the system cost. The NATO Innovation Fund is among its investors.

AI data center server infrastructure
The $8.3 billion funding wave reflects growing market conviction that GPU alternatives can win the inference war — Photo by Joshua Sortino on Unsplash

The trend extends beyond the United States. European AI chip startups attracted $800 million in 2026, compared to $4.7 billion in the US. In the Netherlands, Euclyd — backed by a former ASML CEO — is in the process of raising €100 million, signaling that the inference chip race has gone transatlantic.

Key point: The $8.3 billion in AI chip startup funding is not just venture enthusiasm — it is the market’s capital-denominated verdict that NVIDIA’s GPU monopoly has structural cracks in the inference layer.

NVIDIA’s $20 Billion Defense — The Groq Acquisition and the CUDA Moat

NVIDIA has not been passive. In December 2025, the company acquired Groq’s inference chip assets for $20 billion and invested $4 billion in photonics companies. Annual R&D spending exceeds $18 billion — more than double the combined funding of every AI chip startup on the planet.

Yet NVIDIA’s most formidable defense has never been hardware. It is the CUDA software ecosystem — the programming framework, libraries, and enterprise integrations that the vast majority of AI developers rely on. Switching costs remain enormous, and the largest AI training operations are deeply embedded in NVIDIA’s stack.

Still, the 80% market share figure faces pressure from multiple directions simultaneously. Cerebras’ OpenAI contract alone represents a significant inference workload migrating away from GPUs.

Fractile’s NATO-backed inference architecture targets defense and sovereignty-sensitive deployments. AWS’ Trainium expansion signals that even NVIDIA’s largest cloud partners are hedging.

💡 Related Posts
$5.6M, 1 Trillion Parameters: DeepSeek V4 Drops Nvidia

The inference war is not about whether NVIDIA will lose its crown entirely. The company’s training dominance is structurally secure for the foreseeable future. The real question is whether the inference market — which is rapidly outgrowing training — will remain a GPU monopoly or fragment into a multi-architecture landscape.

The $8.3 billion bet from investors worldwide suggests the fragmentation has already begun. Two metrics will tell the story over the next 12 months: the Cerebras CBRS ticker’s opening trade and the percentage of OpenAI’s inference workload that shifts to non-GPU silicon.

References

Stay Ahead of the Curve

Get the latest insights on AI, tech, and markets — delivered in both Korean and English.

Explore More Posts