SambaNova: What Killed the AI Chip Dream?

SambaNova: What Killed the AI Chip Dream?

The right technology. The right backers. The wrong timing — and the wrong competitor.

In April 2021, SambaNova Systems raised $676 million at a $5.1 billion valuation. SoftBank led the round. Temasek and GIC joined. It was one of the largest AI chip raises on record, and it made SambaNova the undisputed poster child of the AI hardware wave — a Stanford-born startup that was going to do to Nvidia what Nvidia had done to everyone else.

Four years later, it was exploring a sale. Intel acquisition talks collapsed. It raised a Series E at $2.2 billion — less than half its 2021 peak. On secondary markets, the implied valuation is closer to $800 million.

That's an 84% de-rating in five years.

What went wrong? The honest answer is: not one thing. SambaNova is a story about the collision between genuinely interesting technology and an almost impossible competitive reality.

What SambaNova Actually Built

SambaNova was founded in 2017 by two Stanford professors — Kunle Olukotun and Christopher Ré — and Rodrigo Liang, a former Oracle executive. The founding thesis was sharp: the GPU-centric architecture Nvidia had built was fundamentally mismatched to AI workloads. Too much compute, not enough memory bandwidth, too expensive to run at scale.

Their answer was the Reconfigurable Dataflow Unit (RDU) — a chip designed from scratch for the matrix operations that underpin large language models, trading flexibility for throughput. Unlike a GPU, which is a general-purpose parallel processor that AI workloads have been adapted to run on, the RDU was purpose-built for AI from day one.

The pitch was compelling. As models scaled into the hundreds of billions of parameters, memory became the bottleneck, not compute. SambaNova's architecture was designed precisely for this. Its latest chip, the SN50, claims:

The technology is real. That's the frustrating part of this story.

The 2021 Raise and What It Bought

The $676 million Series D came at peak exuberance. 2021 was the year of the SPAC, the $10 billion seed round, and the general conviction that anyone building AI infrastructure was worth a lot of money. SambaNova fit the narrative perfectly: deep tech, credible founders, obvious market, believable story about displacing Nvidia.

The valuation implied SambaNova would need to become a very large business very quickly. The maths:

At a $5.1 billion post-money, assuming a 10x revenue multiple at exit, you need $500 million in revenue to justify the price paid. SambaNova hit $100 million in ARR in June 2025 — eight years after founding, and $1.1 billion in total capital later.

This is not unusual in deep tech. Hardware businesses are brutally capital-intensive. A single advanced chip design run (tape-out) can cost $50–100 million. Then you need to build the software stack, hire world-class engineers, and persuade enterprises to rip out existing infrastructure and replace it with yours. The sales cycle alone can run 18 months. Revenue at scale takes time that 2021-era valuations simply don't allow for.

The Nvidia Problem

Every AI chip startup faces the same structural challenge: Nvidia's moat is not the chip. It's CUDA.

CUDA is the software ecosystem Nvidia has spent 15 years building — the programming model, the libraries, the tools, the frameworks that every AI researcher and engineer uses. PyTorch runs on CUDA. TensorFlow runs on CUDA. The entire ML research community writes in CUDA.

When a startup builds a better chip, it doesn't just need to prove the hardware is faster — it needs to convince customers to rewrite or re-optimise their entire software stack to run on different silicon. That is an enormous switching cost, and it is Nvidia's real competitive moat. The chips are almost secondary.

SambaNova has worked hard to reduce this friction — its platform supports PyTorch and integrates with HuggingFace pipelines — but "compatible" is not the same as "native." Every new model architecture, every new training technique, every framework update that CUDA supports natively requires SambaNova to catch up.

It's a race run on a treadmill.

The Intel Saga

The Intel acquisition story is its own case study in misaligned incentives.

Intel's CEO Lip-Bu Tan had been serving as executive chairman of SambaNova's board since May 2024 — before acquisition talks began. When Intel signed a non-binding term sheet to acquire SambaNova for approximately $1.6 billion in December 2025, Tan was effectively negotiating with himself. An obvious conflict of interest, even if structurally managed.

The deal made strategic sense for Intel. The company had been trying to rebuild its AI chip credibility since its Gaudi accelerators failed to gain traction against Nvidia's H100. Acquiring SambaNova's technology and team would have accelerated Intel's AI roadmap by years. The alternative — building from scratch — was slower and no less expensive.

But the deal collapsed in late 2025, reportedly over price and integration concerns. Intel then participated in SambaNova's Series E as a strategic investor at roughly 9% ownership, alongside a multi-year collaboration agreement. It got technology access without the acquisition risk. SambaNova got a cheque and a credible name on its cap table.

For investors who came in at the $5.1 billion 2021 valuation, this was cold comfort.

What Happens Next

SambaNova is reportedly exploring a sale again, having hired an investment bank to run a process. The SN50 launch is its best near-term asset — if the performance claims hold up in real enterprise deployments, it could attract genuine interest from hyperscalers or defence contractors looking for Nvidia alternatives.

The most plausible acquirers:

• Intel — still in the frame despite the failed talks
• A hyperscaler — Amazon, Microsoft, or Google, each building out proprietary silicon
• A defence or enterprise strategic — where Nvidia lock-in is a geopolitical concern as much as a commercial one

What makes SambaNova's situation instructive is that it isn't a story of bad technology or bad founders. It's a story about three compounding problems: (1) the structural difficulty of competing with a platform monopoly, (2) the danger of 2021-era valuations for capital-intensive hardware businesses, and (3) the way a failed M&A process can crystallise a company's predicament in very public fashion.

The AI chip market is not going away. The inference wave — running models in production at scale, cheaply and quickly — creates real demand for Nvidia alternatives. SambaNova has a legitimate product in a legitimate market.

The question is whether it can survive long enough, and at the right ownership structure, to capitalise on it.