The AI ASIC Market, Part 3: The Independent AI Chip Companies Taking on NVIDIA

TLDR

• Groq's $20 billion acquisition in December 2025 is the defining event of the independent AI chip era, validating that novel inference architectures can command large valuations and reshape comparables.
• Cerebras went public on May 14, 2026, closing day one at a ~$56 billion fully diluted valuation — the largest U.S. tech IPO since Snowflake in 2020. Groq's sale to NVIDIA reset the comparables for independent AI silicon, and Cerebras priced into that re-rated market.
• Etched's transformer-only chip is the most extreme bet, contingent on the AI industry not migrating to non-transformer architectures during the multi-year build-out window.
• The funding arc across all three companies has compressed dramatically since 2024, with valuations now rising on architectural conviction rather than production deployments.

In Part 2 of this series, we mapped the design enabler half of the independent AI chip universe — Broadcom and Marvell, the two companies quietly powering more than 80% of hyperscaler custom AI silicon through ASIC design services and networking IP.

Direct NVIDIA Challengers in the AI ASIC Market

Groq, Cerebras, and Etched are direct competitors to NVIDIA — independent companies designing, manufacturing, and selling their own AI ASICs under their own brand. They sell finished chips into the same data center inference market that NVIDIA's GPUs currently dominate. Each company is also a concentrated bet on a specific architectural thesis — and the gating variable for all three is whether that bet survives contact with production reality.

Two more independent AI chip companies — Tenstorrent and Tensordyne — operate as architectural alternatives with strategic logic that diverges meaningfully from the three covered here. We'll cover them in Part 4.

Groq: The $20 Billion Exit and What It Proves

Groq's acquisition by NVIDIA is the defining story of the independent AI chip era. The $20 billion deal — structured as a perpetual IP license plus talent acquisition of approximately 90% of Groq's staff, including founder Jonathan Ross — closed in December 2025. It's the largest transaction in NVIDIA's history, and it validates everything the independent ASIC thesis claimed.

Groq was founded in 2016 by Jonathan Ross (the engineer who designed Google's original TPU) and a team of former Google engineers. After leaving Google, Ross built a completely different architecture: the Tensor Streaming Processor (TSP), later rebranded as the Language Processing Unit (LPU) following the LLM boom. The architectural premise was radical — eliminate all the "reactive" hardware components that make GPUs flexible but slow for inference. No branch predictors. No arbiters. No reorder buffers. No caches. Replace all of it with a compiler-controlled, fully deterministic execution model running on-chip SRAM instead of external HBM.

The result was a chip that couldn't do much besides inference, but did inference extraordinarily fast. Independently verified benchmarks from Artificial Analysis told the story: 241 tokens/second on Llama 2 70B in early 2024 — more than double every other provider at the time, forcing Artificial Analysis to extend its chart axes. Versus standard GPUs, Groq's LPU delivered approximately 10x the throughput for LLM inference at 90% lower power consumption per compute operation.

NVIDIA paid a 2.9x premium over Groq's September 2025 valuation. The deal structure — perpetual IP license plus talent acquisition — looks designed to avoid triggering standard M&A review processes, which is the subject of open regulatory scrutiny discussed below.

Why NVIDIA paid $20 billion. The clearest answer came from NVIDIA itself at GTC 2026. GPUs are optimized for the prefill phase of inference, which is compute-bound. LPUs are optimized for the decode phase, which is memory-bandwidth-bound. By pairing the two in rack-scale systems, NVIDIA creates a heterogeneous inference platform optimized for both phases simultaneously. NVIDIA couldn't close this architectural gap from its own GPU roadmap in any reasonable timeframe — so it bought the solution. The world's most valuable chip company concluded that a competing architecture was necessary, and paid $20 billion to own it.

The Groq 3 LPU shipped fast after the acquisition closed. At GTC 2026 in March — three months after deal close — NVIDIA unveiled the Samsung-manufactured 4nm chip, shipping in Q3 2026. The architectural choice that defines it: SRAM in place of HBM, which delivers an order of magnitude more memory bandwidth per die than a Rubin GPU and is what makes the LPU faster at decode-phase inference than any GPU NVIDIA could have built itself.

The LPX rack — 256 LPUs paired with a Vera Rubin NVL72 — is how NVIDIA is shipping the technology. Early Q3 2026 customers include Meta, OpenAI, and Anthropic, with AWS, Google Cloud, Azure, and Oracle Cloud deploying Vera Rubin instances in H2 2026. NVIDIA's deployment guidance is to build data centers with approximately 25% LPU capacity — a measured integration that protects GPU revenue while addressing the specific decode-phase bottleneck LPUs solve.

The antitrust risk is real. On March 20, 2026, Senators Elizabeth Warren and Richard Blumenthal sent a letter to Jensen Huang raising pointed questions about whether the Groq deal was deliberately structured to evade antitrust review. The FTC's broader crackdown on acqui-hire transactions — the Microsoft-Inflection deal is the reference point — combined with NVIDIA's ~86% GPU market share creates a regulatory environment where this deal faces genuine scrutiny. An adverse ruling could force NVIDIA to unwind the license, divest the technology, or accept penalties. The Samsung manufacturing arrangement and Q3 2026 shipment commitments suggest NVIDIA believes the structure holds.

GroqCloud continues independently. Before the acquisition, the platform had amassed over 2 million developers and accounts at 75% of Fortune 100 companies, and it continues operating post-deal — meaning Groq's LPU-based inference is available both through NVIDIA's integrated rack systems and as a standalone cloud service.

The broader legacy: Groq's $20 billion exit reshapes every valuation in the independent AI chip sector. For Etched, Cerebras, and the architectural alternatives covered in Part 4, the $20B benchmark is now the operative comparable — and NVIDIA itself views specialized inference silicon as a strategic necessity rather than an optional supplement. Cerebras' IPO at a $56 billion day-one valuation is the first real-world test of that thesis — and it priced well above the benchmark.

Cerebras Systems: Wafer-Scale and the Biggest AI IPO Since NVIDIA

Cerebras is now the largest publicly traded independent AI chip company. After a multi-year path that included a withdrawn 2024 S-1 over CFIUS scrutiny and an April 2026 resubmission, the company priced its IPO at $185 per share on May 14, 2026 — above the raised $150–$160 range — raising $5.55 billion and valuing the company at approximately $56.4 billion on a fully diluted basis. The pricing also reflected a re-rated comparable set: Groq's acquisition by NVIDIA established that independent AI inference silicon could command large valuations, and Cerebras priced into that environment rather than into the pre-Groq one its 2024 S-1 had targeted. Shares opened at $350, traded as high as $385, and closed near $311 on the first day for a gain of approximately 68%.

The architectural thesis is the most distinctive in the independent camp. Rather than packaging multiple chips on a wafer, Cerebras uses the entire 300mm silicon wafer as a single processor. The Wafer-Scale Engine 3 (WSE-3) is 57 times larger than NVIDIA's H100 GPU and packs 4 trillion transistors with 44GB of on-chip SRAM — large enough to eliminate HBM dependency entirely, which is a meaningful architectural advantage given the ongoing HBM supply constraints covered in Part 1. Cerebras claims up to 20x faster inference than NVIDIA equivalents for specific LLM workloads.

Cerebras reported $510 million in 2025 revenue (up 76% year-over-year) and $24.6 billion in remaining performance obligations as of December 2025, with 15% expected to recognize in each of 2026 and 2027 — the financial scaffolding that makes the $22–35B IPO target legible.

The customer concentration story is where Cerebras' IPO thesis lives or dies. G42 (the UAE-based conglomerate that forced the 2024 IPO withdrawal via CFIUS scrutiny) accounted for 87% of H1 2024 revenue. By 2025, G42 was down to 24% of revenue. Mohamed bin Zayed University of Artificial Intelligence (MBZUAI), a public UAE institution, picked up 62% of 2025 revenue. The concentration shifted — not fully diversified geographically — but it shifted, and G42 has restructured its investment to non-voting shares, clearing CFIUS.

The anchor commercial deals are what pushed the IPO past pre-listing expectations and into a $56 billion fully diluted valuation. The January 2026 OpenAI contract was announced as $10 billion for 750MW of compute through 2028; an April 2026 expansion could push the total relationship to over $20 billion, with OpenAI getting warrants for up to 33.4 million Class N shares and an option to buy additional 1.25GW of capacity through 2030. Cerebras also received a $1 billion loan from OpenAI at 6% interest. OpenAI gets an additional 17% stake if Cerebras maintains a $40 billion valuation average for a month — not far off from the $35 billion IPO target. Then in March 2026, Cerebras announced an AWS deal making AWS the first hyperscaler to deploy Cerebras chips in its own data centers — a structural shift from Cerebras selling cloud services to Cerebras hardware going into hyperscaler infrastructure directly.

Other Cerebras customers include IBM, the U.S. Department of Energy, the U.S. Department of Defense, Meta, Amazon, GSK, and the Mayo Clinic. The DoE and DoD presence is particularly important for regulatory optics given the historical UAE customer concentration.

Etched: The Transformer-Only Bet

Etched makes the most concentrated architectural bet in the independent AI chip landscape. Founded in 2022 by three Harvard dropouts — Gavin Uberti, Chris Zhu, and Robert Wachen — the company permanently burns the transformer architecture into silicon. The Sohu chip strips out all hardware needed for other neural network types to maximize throughput for the one architecture that now powers virtually every frontier AI system.

The performance claims, all company-reported, are staggering: an 8-chip Sohu server reportedly replaces 160 H100 GPUs on Llama-70B inference, with claims of 15x faster, 10x cheaper, and 9x more power-efficient per million tokens versus competing hardware. The bet is that hardcoding the transformer's matrix multiplication pattern into silicon — rather than running it on flexible GPUs — is worth the loss of architectural flexibility. The NVIDIA-Groq $20B acquisition is Etched's direct comparable valuation reference, and at $5 billion, analysts at Jon Peddie Research suggest the valuation may be "conservative."

The honest risk framing on Etched matters more than it does for most independent AI chip companies, because the concentration of the bet is so extreme.

1. Unproven performance claims. The 20x and 160-to-1 numbers are company-reported benchmarks, not widely reproduced at production scale. They will be tested hard once the chip is in customer hands.
2. Architecture concentration risk. If the AI industry moves meaningfully to state-space models, mixture-of-experts architectures that break transformer assumptions, or any other non-transformer design, Sohu becomes obsolete. Etched has stated it would design a new chip in that scenario — but that's a multi-year pivot, during which competitors with more flexible architectures would have a significant window. This is the most existential single risk of any independent AI chip company covered here.
3. HBM supply competition with NVIDIA. Every HBM allocation Etched gets is an allocation NVIDIA would prefer to have.
4. Software ecosystem build-out. CUDA is 15+ years old. JAX has Google's full weight behind it. Etched has to build developer mindshare from scratch while validating silicon and proving production performance simultaneously.

None of these risks invalidate the thesis. They just mean the $5 billion valuation assumes Etched executes through a specific and narrow corridor of outcomes.

How the Three Independent Challengers Compare

The three companies in this article share a category — direct competitors to NVIDIA in inference silicon — but the structural distinctions between them are sharp. Groq has been validated by the largest acquisition in NVIDIA's history. Cerebras completed a blockbuster IPO on the strength of $20+ billion in OpenAI commitments, an AWS hardware deployment deal, and the re-rated valuation environment Groq's $20B exit created five months earlier. Etched is making the most concentrated architectural bet of the three with the smallest current revenue base and the most extreme single-point-of-failure risk profile.

The valuation arcs tell the structural story:

Three things worth pulling out of the chart:

1. Groq's line stays flat through 2024 before accelerating sharply into the NVIDIA acquisition. The architectural validation came late and forcefully rather than as a gradual ramp.
2. Cerebras stretches highest, and the actual IPO outcome exceeded the pre-listing target range. The public market priced the wafer-scale thesis aggressively; whether that valuation holds is now the open question, particularly given customer concentration and a valuation sits at more than 130x trailing sales.
3. Etched's line is the most compressed in time, going from Series A to a $5B valuation in roughly 18 months — on the smallest commercial proof base of the three.

The market is currently willing to fund three different architectural answers to the same observation: that inference at scale rewards purpose-built silicon. The Groq exit confirmed that one of these architectures could command a $20B valuation. Whether Cerebras and Etched maintain or exceed that benchmark depends on execution outcomes that are still ahead of them.

What the Independent AI Chip Trade Means for Bitcoin Miners

For Bitcoin mining operators and infrastructure investors, the structural takeaway from this group is the same lesson the Bitmain and MicroBT story taught a decade ago: custom silicon for specific workloads always wins at scale. The Groq, Cerebras, and Etched valuations are the AI compute market arriving at the same conclusion mining has already lived through — hashprice and J/TH economics force the mining industry toward purpose-built ASICs because flexible silicon can't compete on watts per useful operation, and the same logic is now playing out in AI inference.

The capital availability signal matters too. Bitcoin mining infrastructure companies sit on overlapping power, cooling, and site economics that map directly onto AI compute infrastructure — and miners that can credibly articulate their position in the broader custom silicon thesis have access to significantly more capital than they did in 2022–2023. The AI infrastructure capital flywheel will not leave adjacent industrial compute opportunities untouched indefinitely.

For infrastructure operators considering AI/HPC pivots, the architectural diversity in this group also matters tactically. Different inference architectures carry different physical infrastructure requirements — interconnect density, cooling profiles, and power-delivery design all vary across SRAM-heavy, wafer-scale, and transformer-hardcoded silicon. Operators evaluating colocation or hosting opportunities for AI compute should assume the hardware mix in their data halls will be more heterogeneous than the GPU-dominated picture of 2023–2024.

What's Next in the Independent AI Chip Market

Groq, Cerebras, and Etched are the three companies most directly confronting NVIDIA on its core inference territory. Two more — Tenstorrent and Tensordyne — operate on architectural alternatives with different strategic logic. Tenstorrent is building an open RISC-V ecosystem and licensing IP rather than chasing hyperscaler deployments; Tensordyne is making a pre-silicon bet on a logarithmic number system architecture. Both get full coverage in Part 4.

*Updated May 15, 2026 with additional information about Cerebras' IPO.