The AI chip industry is no longer limited to just building faster GPUs. A new generation of companies is trying to overhaul the entire system architecture to reduce data movement, energy consumption, and reliance on high-bandwidth memory. Cerebras, Groq, Etched, d-Matrix, Tenstorrent, and others start from different ideas but share the same goal: to find a viable alternative to the model that has made Nvidia the market leader.
The key points of new AI chip startups in 30 seconds
- Cerebras, Groq, Etched, and other firms are reaching valuations in the billions of dollars.
- Each architecture seeks to reduce a specific bottleneck, from memory to inter-chip communication.
- Cerebras uses a processor the size of a wafer; Groq prioritizes SRAM and predictable execution.
- Etched and Taalas favor highly specialized chips, while d-Matrix brings computation closer to memory.
- Valuations reflect high expectations, but many technologies still need to demonstrate scale, cost-efficiency, and commercial adoption.
As of July 2026, the market picture shows an industry in active search. Several companies are valued over $1 billion, with some reaching much higher figures—though it’s important to distinguish between market capitalization, private funding rounds, market rumors, and deals that don’t constitute full acquisitions.
Groq’s case is a good example. The announced agreement with Nvidia in December 2025 was described in some analyses as a $20 billion acquisition. However, both companies clarified it was a non-exclusive inference technology license accompanied by hiring part of Groq’s team. Nvidia later confirmed it did not buy Groq’s products, contracts, or equity stakes. Groq continued operating independently, raising an additional $650 million in June 2026.
Cerebras did go public. The company started trading on Nasdaq in May 2026, with shares initially set at $185—above the initially proposed range. Its valuation has varied since; thus, figures like $48 billion seen in some comparisons are snapshots of the market, not a permanent valuation.
The common problem: moving data costs too much
Most of an AI accelerator’s workload involves performing mathematical operations on massive matrices. However, compute capacity isn’t the only limit. Data must continuously move between memory, the processor, and other chips within the system.
This movement consumes time and energy. It also requires high-bandwidth memories, ultra-fast internal networks, and advanced packaging systems. When a model is distributed across many GPUs, some potential performance is lost waiting for weights, activations, and results to arrive at the right place.
Most startups attempt to address this challenge from a specific approach. They don’t all compete directly with Nvidia in the same markets nor offer interchangeable products, but nearly all question some element of the conventional architecture.
Cerebras aims to reduce inter-chip communication by building a processor that almost entirely occupies a wafer. Groq replaces some dependence on external memory with large amounts of integrated SRAM and static scheduling. d-Matrix moves certain operations inside or near the memory. Etched restricts processor flexibility to focus on transformer-based models.
The shared idea is straightforward: if data movement has become one of AI’s main costs, an architecture that reduces these movements can gain in latency, efficiency, or cost.
Cerebras removes many boundaries between chips
Cerebras built its approach around the Wafer-Scale Engine, a processor fabricated on a complete silicon wafer. Its WSE-3 covers 46,225 mm², integrates four trillion transistors, and contains 900,000 AI-oriented cores, according to the company’s specifications.
A traditional processor is obtained by cutting a wafer into many smaller chips. Cerebras keeps most of the surface intact and uses proprietary mechanisms to handle defects, power, cooling, and internal communication.
This solution reduces the need to split workloads across dozens or hundreds of packages. The company claims it also simplifies training large models and decreases data movement between devices.
There are trade-offs. Manufacturing, powering, and cooling such a large processor requires complex engineering. It also creates a platform very different from conventional GPU clusters, meaning customers must work with Cerebras’ software and systems.
Cerebras has successfully turned its architecture into commercial computing products and services—an achievement that many semiconductor startups don’t reach. Its IPO strengthened the perception of a market for alternative accelerators, although its future will depend on revenues, utilization, and its ability to attract clients beyond large, one-off deals.
Groq bets on SRAM and predictable execution
Groq designed its language processing unit, known as LPU, with a different approach. Its architecture uses hundreds of megabytes of SRAM inside the chip as primary storage for certain data, rather than just as cache memory.
This SRAM is fast and offers very low latency, but it takes up more space and is more expensive than other memories. This limits how much can be integrated into each processor, so Groq distributes models across multiple interconnected chips.
Another difference is in execution. Its compiler programs operations statically, making behavior predictable. This aims to avoid some of the complexity associated with GPUs, which are designed to handle a wide variety of workloads and make numerous runtime decisions.
Groq has focused especially on inference—running trained models to generate responses. Low latency is particularly valuable in assistant, agent, and interactive applications where users expect rapid token generation.
The license granted to Nvidia validates some of Groq’s intellectual property but doesn’t mean Groq has disappeared. Its continued operation as a provider of GroqCloud shows that such arrangements can allow a large company to access technology and talent without fully integrating all of Groq.
Etched sacrifices flexibility to focus on transformers
Etched represents a riskier approach: designing an application-specific integrated circuit (ASIC) tailored almost exclusively for transformer models.
GPUs are flexible because they can handle many types of calculations. This versatility demands hardware, instructions, and software that aren’t always fully utilized for any one specific model.
Etched argues that removing much of this generality allows the chip’s surface area to be dedicated mainly to operations that transformers need most. The potential benefits include higher performance-to-cost and performance-to-watt ratios. But the risk is clear: if dominant model architectures shift, a highly specialized processor could become obsolete.
In July 2026, the company announced Series C funding that valued it at $10.3 billion and noted raising about $800 million through various deals. These figures are provided by the company and don’t yet reflect revenue, delivered orders, or market share.
Their rise demonstrates investor willingness to fund specialized designs and highlights the gap between private valuation and the ability to manufacture chips at volume, develop software, and compete with established platforms.
d-Matrix brings computation closer to memory
d-Matrix starts with a different diagnosis: the processor spends too much time waiting for data. Its in-memory digital compute architecture integrates mathematical operations near storage cells.
The Corsair platform uses SRAM-based chiplets for generative model inference. The company claims this approach can improve latency, energy efficiency, and cost-performance, though comparisons are based on specific configurations and models.
In-memory computing doesn’t eliminate data transfer altogether but reduces some trips between processor and memory. Its challenge is maintaining accuracy, capacity, programmability, and scalability.
In June 2026, d-Matrix announced that Corsair had entered full production. This step is important because it transitions a prototype architecture into a product ready for customer deployment, though widespread success depends on actual deployments and model support in live systems.
Tenstorrent aims to compete via intellectual property
Tenstorrent combines AI accelerators, RISC-V processors, and an IP licensing business. Its strategy isn’t solely to sell complete cards or systems.
The company offers CPU cores, chiplets, and acceleration blocks that others can incorporate into their own designs. This approach could open markets in data centers, automotive, consumer electronics, and specialized systems.
RISC-V is an open instruction set architecture that allows chip design without depending on ARM or x86 licenses. Tenstorrent integrates RISC-V with its Tensix cores to build heterogeneous platforms.
The benefit is greater commercial flexibility. A company can adopt Tenstorrent’s technology without necessarily buying a full system. The challenge lies in developing mature software environments and convincing manufacturers to adopt a new architecture.
Not all chip startups produce the same products
Comparisons of chip startups often mix different kinds of businesses. Cerebras, Groq, Etched, and d-Matrix develop AI accelerators, but Substrate specializes in semiconductor manufacturing.
Substrate aims to develop an alternative to advanced lithography tools with a new form of X-ray lithography. Their goal is to build U.S.-based fabs capable of producing advanced chips with potentially less costly equipment than current top-tier machines.
This project doesn’t directly compete with Nvidia’s GPUs. It seeks to modify another part of the supply chain: how transistors and interconnects are printed onto wafers.
Similarly, companies focused on photonics, networking, cooling, superconductors, or packaging address different infrastructure needs for AI, each solving different problems.
Therefore, there’s no definitive list of “all” chip startups. The market includes accelerators, memories, interconnects, IP, manufacturing services, optical solutions, and complete systems.
Valuations grow faster than products
Sector figures reveal massive investor interest. Cerebras has surpassed tens of billions in public valuation; Etched is at $10.3 billion after a private round; SambaNova reportedly reached $11 billion after new funding.
These numbers can’t be directly compared. Public market caps fluctuate daily. Private valuations come from negotiated deal prices and might include conditions. An IP license is not equivalent to buying a company outright.
Moreover, semiconductor development requires vast capital. Designing advanced chips, producing initial wafers, building systems, developing compilers, and supporting products can cost hundreds of millions before meaningful sales emerge.
Nvidia’s dominance isn’t just due to its GPU. It benefits from CUDA, software libraries, networking, partnerships, and an extensive developer community built over years. Even a better-performing accelerator might struggle to replace that entire ecosystem at once.
An academic study published in 2026 compared several accelerators, finding that the best choice varies with model size, context length, and request volume. It also noted differences in idle power and software maturity. The takeaway is that no single architecture dominates all workloads.
Real competition probably won’t produce a sole Nvidia replacement. Instead, it may lead to a fragmented market, with general-purpose GPUs for broad tasks and specialized accelerators for low-latency inference, niche models, edge devices, or energy-constrained facilities.
Frequently Asked Questions
Which AI chip startup has the highest valuation?
Cerebras is publicly traded, with a valuation that has exceeded $40 billion. Values fluctuate with the market and shouldn’t be directly compared to private valuations like Etched’s.
Did Nvidia buy Groq for $20 billion?
No, Nvidia didn’t acquire the entire company. It signed a non-exclusive inference technology license and hired part of Groq’s team. Groq continues to operate independently.
Why are so many companies trying to reduce data movement?
Because transferring data between memory, processors, and chips consumes energy and increases latency. In large models, this movement can limit performance more than the raw computational capacity.
Can a startup replace Nvidia?
It might compete effectively on specific tasks but displacing Nvidia across the entire market is more challenging. Besides hardware, it would need to match Nvidia’s software, tools, networking, support, and developer ecosystem.
