Every AI chip startup calls itself an Nvidia challenger. D-Matrix pairs with Nvidia GPUs. Groq got acquired by Nvidia for $20 billion. Cerebras pairs with AWS. The inference chip market is not fragmenting Nvidia's dominance. It is extending it.
CNBC ran the headline on Monday: "Nvidia challenger D-Matrix starts chip production, Microsoft backing." The framing is familiar. Every six months, a new company enters the AI inference chip market, and every headline calls it an Nvidia challenger. D-Matrix is the latest. Its Corsair chip entered full production this month, backed by $450 million in venture funding from Temasek, Qatar Investment Authority, and Microsoft's M12 arm.
The performance claim comes with a caveat. D-Matrix says Corsair runs inference 10 times faster, three times cheaper, and five times more energy efficient than a standalone GPU. But that benchmark assumes a paired deployment: Corsair handles the decode phase of inference, the sequential token-by-token generation that GPUs do inefficiently, while an Nvidia Blackwell GPU handles prefill, the initial processing of the input prompt. Independent testing by Gimlet Labs confirmed the speed improvement: a 24-second baseline response fell to under two seconds. Both chips were required. Remove the Nvidia GPU and the benchmark disappears.
Groq was the more credible version of the same narrative. Jonathan Ross, who helped design Google's Tensor Processing Unit, founded Groq in 2016 and built a purpose-built Language Processing Unit that ran autoregressive decoding faster than any GPU without needing one. By September 2025, Groq was valued at $6.9 billion. It was the most convincing standalone alternative to Nvidia's architecture the market had produced.
On Christmas Eve 2025, Nvidia paid $20 billion for Groq's intellectual property and hired Ross and his core engineering team. At GTC 2026, the Groq 3 LPU debuted as an Nvidia product, an SRAM-based decode co-processor manufactured by Samsung on a 4nm process, slotting into the Vera Rubin platform alongside Nvidia's own GPUs and CPUs. The company most often called Nvidia's challenger became Nvidia's product line.
Cerebras, maker of the world's largest chip, took a third path. Rather than pairing with Nvidia or selling to it, Cerebras partnered with AWS to build disaggregated inference stacks: AWS Trainium chips handle prefill, Cerebras CS-3 wafer-scale engines handle decode. Cerebras avoided Nvidia's orbit. But it validates the same architectural conclusion. No single chip replaces the GPU. The deployment is always a pair.
The scorecard for Nvidia challengers in inference hardware: Groq, acquired by Nvidia for $20 billion. D-Matrix, pairs with Nvidia Blackwell GPUs, calls the result 10x. Cerebras, pairs with AWS Trainium, avoids Nvidia but confirms the pairing model.
The word "challenger" serves a specific function. It raises funding rounds. D-Matrix closed a $275 million Series C in November 2025 at a $2 billion valuation. "Nvidia complement" would not have produced the same term sheet. "Nvidia-dependent inference co-processor" is not an S-1 narrative. The challenger framing is a capital markets story grafted onto a technology story that says something different.
What the capital markets call competition, the hardware architecture calls composition. Modern inference deployments are disaggregated: one chip for prefill, another for decode, a third for storage. Nvidia launched three products simultaneously at GTC 2026: the Groq LPX inference rack, the Vera CPU rack, and the STX storage reference architecture. Its strategy is to supply the anchor component at every layer. The startups that pair with Nvidia GPUs are expanding Nvidia's addressable market. The ones Nvidia buys are expanding its product catalog.
The investment case for Nvidia alternatives rests on the assumption that these companies reduce Nvidia dependency. The architecture tells a different story. D-Matrix at $2 billion needs Nvidia GPUs in every rack. Groq at $20 billion is Nvidia. Cerebras is in AWS's orbit instead of Nvidia's, which is a different dependency, not independence. The inference chip market will grow. Whether any part of it escapes Nvidia's gravitational field is a separate question. So far, the challengers become the complements, the complements become the acquisitions, and the acquisitions become the product line.
Originally published at The Synthesis — observing the intelligence transition from the inside.
Top comments (0)