Cornelis Raises 205M To Challenge Nvidia Dominance

Fexingo (0:01)

The headline out of Silicon Valley this evening is actually two distinct stories that are converging in a way I think most investors are missing. First, OpenAI has acquired smartphone camera maker Glass Imaging for $300 million.

Second, a smaller player called Cornelis Technologies just raised $205 million specifically to build chips that challenge Nvidia's stranglehold on AI inference.

SPEAKER_2 (0:27)

$205 million is a serious war chest for a start-up, especially when you're trying to take on a company with a valuation that fluctuates around $210 per share. Why now?

Fexingo (0:40)

Because the economics of running large language models are breaking down. Nvidia's data center GPUs are incredible for training, but they are notoriously power hungry and expensive for inference, which is where 90% of the actual usage happens. Cornelis is betting that specialized silicon can cut energy costs by half while delivering comparable throughput. It's not about beating Nvidia at training.

It's about winning the long tail of everyday AI requests.

SPEAKER_2 (1:09)

So they're targeting the margin squeeze that every cloud provider is feeling right now. If you can lower the cost per query, you unlock use cases that were previously too expensive to run profitably.

Fexingo (1:21)

Exactly, and look at the market reaction today. The SOXX index is down nearly 6% and even ASML is off more than 10%.

Investors are nervous about KPEC spending because they don't know if the ROI on these massive data centers is there yet. But here's the thing. If inference becomes cheap enough, the volume explodes. That's where the money is made, not in selling chips to the top 10 hyperscalers, but to everyone else.

SPEAKER_2 (1:50)

It feels like we've been hearing this narrative before. Every few years, someone says they have the chip that will unseat Nvidia.

What makes Cornelis different from the last wave of failed startups?

Fexingo (2:02)

Software compatibility is usually the killer. But Cornelis isn't trying to rewrite the CUDA ecosystem from scratch. They're using a compiler layer that translates existing PyTorch and TensorFlow workloads directly onto their hardware.

So developers don't have to retrain or recode their models. It's an interoperability play rather than a pure architecture play that lowers the friction significantly.

SPEAKER_2 (2:29)

I see.

So they're essentially offering a side door into the Nvidia Fortress rather than building a new house next door. That's a much smarter go to market strategy.

Fexingo (2:40)

Precisely. And let's tie this back to the other headline, OpenAI Buying Glass Imaging. On the surface, that looks like a camera company purchase. But think about what it means for edge AI. If OpenAI wants to deploy models on smartphones, they need optimized neural processing units that aren't tied to the cloud. This suggests a bifurcation in the industry.

Massive cloud-based training clusters versus lightweight, specialized inference engines at the edge.

SPEAKER_2 (3:10)

That makes sense. The cloud can't handle all the latency-sensitive tasks.

You need intelligence closer to the user. But does that help Cornelis? Their pitch seems focused on data centers, not phones.

Fexingo (3:24)

Good point. But the technology transfer is real. The same principles that make inference efficient on a phone, low voltage, high parallelism, are the exact same principles that reduce costs in a data center. If OpenAI is optimizing for mobile, the pressure on data centers to match that efficiency increases. Everyone is racing toward the same physics problem.

Right.

SPEAKER_2 (3:48)

Efficiency is the only metric that matters when you're paying for electricity and rack space. It's no longer just about raw speed.

Fexingo (3:57)

It never was, really. Speed is vanity. Efficiency is sanity. And looking at the semiconductor stocks today, the market is realizing that the era of unlimited growth at any cost is over. AMD is down 2.5%, Intel is down 7%.

Even Broadcom is slipping. There's a collective sigh of relief that maybe the hype cycle is cooling, but also fear that demand might be softening faster than expected.

SPEAKER_2 (4:26)

Is it demand softening, or just a correction after the insane run up? Because the actual adoption of AI tools hasn't slowed down. People are still using them daily.

Fexingo (4:38)

That's the disconnect.

Adoption is up, but profitability is lagging. Companies are spending billions on compute, but the revenue streams aren't keeping pace. That's why investors are punishing the suppliers. They want to see clear paths to margin expansion. Cornelis represents that path, if they can deliver on their promises. $205 million buys them 18 months to prove it works at scale.

SPEAKER_2 (5:06)

18 months is a long time in tech.

One prototype delay, and the funding dries up. Do you think they'll actually capture meaningful market share, or are they just nibbling at the edges?

Fexingo (5:18)