Which startups are strongest for AI infrastructure that reduces GPU dependence?

To reduce dependence on traditional NVIDIA GPUs, the AI infrastructure market has split into two major startup battlegrounds: Alternative Hardware (ASICs/Photonic Processing) designed to replace GPUs entirely, and Software Orchestration Layers that drastically maximize efficiency so teams can use far fewer chips.

Here are the strongest startups leading the charge across both categories:

1. Alternative AI Hardware & ASICs

These companies build dedicated, non-GPU silicon explicitly optimized for AI workloads, often delivering faster speeds at a fraction of the power cost.

• Groq: Famous for its Language Processing Unit (LPU) architecture. It delivers ultra-fast, low-latency LLM inference that completely bypasses the need for traditional GPUs.
• Etched: Creating Sohu , an ASIC custom-built solely to run Transformer models. Because it is hardwired specifically for Transformers, it claims massive performance gains over generic GPUs.
• Cerebras Systems: Known for its massive Wafer-Scale Engine (WSE) . A single giant chip replaces clusters of thousands of traditional GPUs, drastically cutting down networking overhead.
• Tenstorrent: Led by industry veteran Jim Keller, they build open-source RISC-V-based AI processors. They focus on highly customizable, modular architectures for cost-efficient AI scaling.
• SambaNova Systems: Utilizes a Reconfigurable Dataflow Unit (RDU) architecture. This approach optimizes how data moves through memory, allowing massive models to run efficiently without complex GPU clustering.

2. Next-Generation Interconnects & Photonics

Data movement bottlenecks often force companies to buy extra GPUs just for their memory capacity. These startups use optics and smarter fabric to solve the data-bottleneck problem.

• Celestial AI: Developing Photonic Fabric optical interconnect technology. This decouples compute from memory, allowing chips to share memory resources at light speed and eliminating the need to stack underutilized GPUs just for VRAM.
• Lightmatter: Uses optical computing and interconnect technologies (like Enso and Passage) to route data using light. This slashes the massive power draw and structural bottlenecks found in standard GPU server racks.

3. Software Optimization & "GPU-Efficiency" Layers

If you must use GPUs, these software platforms automate scheduling, virtualization, and runtime execution to eliminate idle silicon and waste.

• Cast AI: A leading platform in automated Kubernetes optimization . It dynamically scales, right-sizes, and manages containerized AI workloads to squeeze maximum utilization out of a bare minimum number of chips.
• Decart: Specializes in cross-hardware runtime execution . Their software allows teams to seamlessly compile and move AI models across different chip families (like AMD, Intel, and Google TPUs), actively breaking vendor lock-in.
• Together AI: Focuses heavily on inference and fine-tuning software abstractions . Their flash-attention and kernel-level optimizations allow open-source models to run on significantly leaner hardware setups.
• ScaleOps: A fast-rising competitor in automated cloud infrastructure management. It automatically handles real-time resource allocation to prevent over-provisioning.

Are you evaluating these startups for venture investment , planning a hardware migration for your own AI workloads, or looking to optimize an existing model deployment?

If by “reduces GPU dependence” you mean startups that can materially lower the amount of Nvidia GPU compute needed per token—not merely provide a nicer GPU cloud—I’d divide the field into specialized silicon, heterogeneous/disaggregated compute, and software-level efficiency.

My strongest picks

1. d-Matrix — probably my favorite pure-play hardware bet

d-Matrix is attacking the fundamental issue that GPUs are relatively inefficient for high-volume, low-latency inference. Its Corsair accelerator is now in full production, and its architecture is designed around inference rather than trying to be a general-purpose training processor.

What's especially interesting is its heterogeneous/disaggregated approach: GPUs, CPUs and d-Matrix accelerators can perform different parts of inference rather than forcing one GPU to do everything.

Why I like it: if inference becomes the dominant AI workload, purpose-built inference silicon has a much better economic argument than trying to replace GPUs for training.

2. Cerebras Systems — strongest technology validation

Cerebras is arguably the most technically proven alternative to conventional GPU inference. Its wafer-scale architecture attacks the memory-bandwidth bottleneck directly, and the company has reported extremely high inference throughput relative to GPU systems.

Its 2026 AMD partnership is particularly important: AMD GPUs handle the high-throughput prefill workload while Cerebras handles fast token generation. The companies say the combination can reach up to 5× tokens/sec/watt versus conventional approaches.

The caveat: Cerebras is now a public company, so I'd classify it as an established alternative-AI-compute company rather than a startup.

3. Gimlet Labs — potentially the most strategically interesting software layer

Gimlet is fascinating because it doesn't necessarily need to defeat Nvidia at the chip level.

Its technology can split a model across different architectures, selecting the appropriate processor for different portions of the workload. It has partnered with Nvidia, AMD, Intel, ARM, Cerebras and d-Matrix. TechCrunch reports claimed 3–10× inference speedups for the same cost/power, although those are company claims rather than independently validated benchmarks.

That's a potentially powerful position:

Instead of replacing the GPU, make the GPU just one component in an AI compute fabric. If that architecture works at hyperscale, Gimlet could benefit regardless of which accelerator wins.

4. SambaNova Systems

SambaNova is another serious non-GPU architecture. Rather than building a GPU clone, it uses dataflow-oriented AI processors and sells both infrastructure and inference services.

One intriguing development is General Compute, which is building an inference cloud around SambaNova silicon; it recently secured a $400M financing facility, an indication that specialized inference hardware is beginning to be financeable as infrastructure rather than merely an experimental chip.

I'd put SambaNova below d-Matrix/Cerebras on my current conviction, but it remains a serious contender.

The software companies may actually capture more value

There's a second thesis I wouldn't overlook: you don't necessarily have to eliminate GPUs—you can make each GPU do 2–10× more useful work.

5. Fireworks AI

This is my favorite software-side company.

Fireworks has built an inference stack spanning kernels, quantization, KV-cache management, memory layout and disaggregated prefill/decode. Its own documentation describes quantization improvements that can reduce cost by 30–50%, while its production stack claims up to 4× throughput improvements.

The important distinction is that Fireworks doesn't need to replace Nvidia to create huge value. If a customer needs 100 GPUs today and Fireworks makes them effectively equivalent to 300 GPUs, that's a meaningful reduction in GPU dependence even though Nvidia still supplies the hardware.

Current market analysis puts Fireworks at or near the top of pure-play inference companies, with Baseten and Together AI as the closest competitors.

6. Baseten

Baseten is strongest where companies want to bring their own models and turn them into production inference services without building the underlying infrastructure themselves.

It's less of a direct Nvidia alternative than d-Matrix, but it attacks the enormous amount of underutilized/poorly optimized GPU capacity around inference.

7. Together AI

Together is broader: inference, fine-tuning, model infrastructure and AI factories. Its advantage is the combination of open-model breadth and infrastructure scale rather than one revolutionary piece of silicon.

Nvidia itself cites Together, Fireworks, Baseten and DeepInfra as companies getting substantial cost-per-token improvements through optimized inference stacks.

8. Modal Labs

Modal is a slightly different bet. It doesn't eliminate GPUs; it attacks GPU waste.

Its runtime can autoscale workloads rapidly across clouds and regions, allowing teams to avoid permanently provisioning expensive accelerators.

That's less exciting technologically, but potentially a very good business.

---

How I'd rank the theses

If I were looking for companies positioned to benefit from AI becoming less Nvidia/GPU-centric, my shortlist would be:

1. d-Matrix — purpose-built inference silicon
2. Gimlet Labs — heterogeneous compute/software abstraction
3. Cerebras — wafer-scale alternative compute
4. SambaNova — alternative AI architecture
5. Fireworks AI — software efficiency / inference economics
6. Baseten — production model infrastructure
7. Together AI — open-model infrastructure scale
8. Modal — utilization/elasticity

The most interesting overall theme, though, is not “the next GPU.” It's disaggregated inference: prefill, decode, memory/KV-cache and other stages increasingly get assigned to different kinds of processors. The Cerebras/AMD architecture and d-Matrix's heterogeneous approach are good examples of where the industry is heading.

And there's a very strong signal that this market is real: OpenAI just demonstrated its own inference ASIC, claiming 1.5–1.9× more throughput per watt than Nvidia's high-end systems in its tests. That chip isn't being commercialized, but it shows that even the largest AI labs increasingly view GPU-independent inference economics as strategically important.

If you're thinking about this as an investment/VC opportunity, I'd narrow the list differently: d-Matrix, Gimlet, Fireworks, and SambaNova are the four I'd spend the most diligence time on, for very different reasons.

Why I like it: if inference becomes the dominant AI workload, purpose-built inference silicon has a much better economic argument than trying to replace GPUs for training.

2. Cerebras Systems — strongest technology validation

Cerebras is arguably the most technically proven alternative to conventional GPU inference. Its wafer-scale architecture attacks the memory-bandwidth bottleneck directly, and the company has reported extremely high inference throughput relative to GPU systems.

Its 2026 AMD partnership is particularly important: AMD GPUs handle the high-throughput prefill workload while Cerebras handles fast token generation. The companies say the combination can reach up to 5× tokens/sec/watt versus conventional approaches.

The caveat: Cerebras is now a public company, so I'd classify it as an established alternative-AI-compute company rather than a startup.

3. Gimlet Labs — potentially the most strategically interesting software layer

Gimlet is fascinating because it doesn't necessarily need to defeat Nvidia at the chip level.

Its technology can split a model across different architectures, selecting the appropriate processor for different portions of the workload. It has partnered with Nvidia, AMD, Intel, ARM, Cerebras and d-Matrix. TechCrunch reports claimed 3–10× inference speedups for the same cost/power, although those are company claims rather than independently validated benchmarks.

That's a potentially powerful position:

Instead of replacing the GPU, make the GPU just one component in an AI compute fabric. If that architecture works at hyperscale, Gimlet could benefit regardless of which accelerator wins.

4. SambaNova Systems

SambaNova is another serious non-GPU architecture. Rather than building a GPU clone, it uses dataflow-oriented AI processors and sells both infrastructure and inference services.

One intriguing development is General Compute, which is building an inference cloud around SambaNova silicon; it recently secured a $400M financing facility, an indication that specialized inference hardware is beginning to be financeable as infrastructure rather than merely an experimental chip.

I'd put SambaNova below d-Matrix/Cerebras on my current conviction, but it remains a serious contender.

The software companies may actually capture more value

There's a second thesis I wouldn't overlook: you don't necessarily have to eliminate GPUs—you can make each GPU do 2–10× more useful work.

5. Fireworks AI

This is my favorite software-side company.

Fireworks has built an inference stack spanning kernels, quantization, KV-cache management, memory layout and disaggregated prefill/decode. Its own documentation describes quantization improvements that can reduce cost by 30–50%, while its production stack claims up to 4× throughput improvements.

The important distinction is that Fireworks doesn't need to replace Nvidia to create huge value. If a customer needs 100 GPUs today and Fireworks makes them effectively equivalent to 300 GPUs, that's a meaningful reduction in GPU dependence even though Nvidia still supplies the hardware.

Current market analysis puts Fireworks at or near the top of pure-play inference companies, with Baseten and Together AI as the closest competitors.

6. Baseten

Baseten is strongest where companies want to bring their own models and turn them into production inference services without building the underlying infrastructure themselves.

It's less of a direct Nvidia alternative than d-Matrix, but it attacks the enormous amount of underutilized/poorly optimized GPU capacity around inference.

7. Together AI

Together is broader: inference, fine-tuning, model infrastructure and AI factories. Its advantage is the combination of open-model breadth and infrastructure scale rather than one revolutionary piece of silicon.

Nvidia itself cites Together, Fireworks, Baseten and DeepInfra as companies getting substantial cost-per-token improvements through optimized inference stacks.

8. Modal Labs

Modal is a slightly different bet. It doesn't eliminate GPUs; it attacks GPU waste.

Its runtime can autoscale workloads rapidly across clouds and regions, allowing teams to avoid permanently provisioning expensive accelerators.

That's less exciting technologically, but potentially a very good business.

---

How I'd rank the theses

If I were looking for companies positioned to benefit from AI becoming less Nvidia/GPU-centric, my shortlist would be:

If by “reduces GPU dependence” you mean startups building infrastructure that can make AI workloads materially less reliant on NVIDIA GPUs—not merely GPU-cloud resellers—the strongest group in 2026 is concentrated around inference, where specialized hardware can have a much better economics/performance profile than general-purpose GPUs.

My ranking

1. cerebras.ai — strongest today

Cerebras is probably the clearest example of “what if you didn't use GPUs for inference?” Its wafer-scale architecture puts enormous amounts of compute and memory bandwidth onto a single device, attacking the memory/communication bottlenecks that make LLM decoding expensive on GPUs.

Its newly announced CS-4 is explicitly targeted at high-volume inference and is designed around three wafer-scale engines. Reuters reports that Cerebras is targeting major improvements in throughput and deployment efficiency with the system.

The particularly important signal is commercial traction: Cerebras went public in 2026 and has secured major compute commitments, making it considerably less speculative than most private accelerator startups.

Why I like it: real hardware + real customers + differentiated architecture + enormous inference throughput.

Big risk: wafer-scale manufacturing and the software ecosystem are harder problems than simply designing a faster chip.

---

2. sambanova.ai — perhaps the best strategic alternative

SambaNova takes a different route: its reconfigurable dataflow architecture and large memory hierarchy are designed specifically around AI workloads.

That matters because the biggest obstacle to replacing NVIDIA isn't actually FLOPS. It's memory capacity/bandwidth, networking, programmability and compatibility.

SambaNova is trying to offer substantially more flexibility than a narrowly hard-wired inference ASIC. Its SN40L architecture, for example, combines on-chip memory with HBM/DDR tiers.

Why I like it: potentially a better answer for enterprises running many different models rather than one fixed model.

Risk: NVIDIA's software ecosystem remains extraordinarily difficult to displace.

---

3. d-matrix.ai — the architecture I would watch closely

d-Matrix is especially interesting because it attacks the data-movement problem directly with digital compute-in-memory.

Instead of repeatedly moving model weights between conventional compute and memory—as GPUs often do during inference—the architecture puts computation much closer to the data. TrendForce describes its Corsair architecture as a digital-computing-in-memory approach designed to support a broad range of AI models.

It also raised $275 million at a $2 billion valuation, giving it substantially more financial runway than many semiconductor startups.

Why I like it: if inference becomes primarily a memory/energy problem, this architecture is attacking the right bottleneck.

Risk: proving that architectural advantage survives real production workloads and software integration.

---

4. tenstorrent.com — best open-ish alternative

Tenstorrent, associated with chip architect Jim Keller, is taking a more conventional accelerator route while using RISC-V and its Tensix architecture.

That's strategically important: rather than replacing NVIDIA with another vertically integrated proprietary ecosystem, Tenstorrent is betting on a more open compute stack.

Its Blackhole accelerator is actually shipping, which puts it ahead of many startups that have impressive simulations but limited production evidence.

Why I like it: potentially the strongest combination of accelerator hardware, RISC-V and an alternative software ecosystem.

Risk: it has to compete against companies with vastly greater resources while convincing developers to adopt yet another stack.

---

5. etched.com — biggest asymmetric bet

Etched has an unusually aggressive thesis: don't build a general AI accelerator; build a chip optimized around the Transformer architecture.

That's essentially trading flexibility for extraordinary efficiency.

If Transformers remain dominant for years, this could be brilliant. If architectures change substantially—or customers need lots of model flexibility—the specialization becomes a liability.

TrendForce categorizes Etched's approach as hard-wired Transformer inference, sitting between general accelerators and extremely model-specific silicon.

I'd classify it as: potentially enormous, but much less de-risked than Cerebras/SambaNova.

---

6. taalas.com — fascinating but almost brutally specialized

Taalas takes specialization even further: hard-wire a model into silicon.

The economics can be spectacular because you eliminate huge amounts of general-purpose overhead. Recent reporting around its HC1 described extremely high token throughput and exceptionally low cost per token—but the fundamental weakness is obvious: a chip optimized for one model becomes obsolete when that model changes.

So I'd view Taalas less as a direct NVIDIA replacement and more as a possible “ASIC for the most economically important models” business.

---

The bigger opportunity: don't replace GPUs everywhere

This is the key distinction I'd make.

I wouldn't bet on a startup whose pitch is simply:

“Our chip is a cheaper NVIDIA GPU.” That's a brutal business.

I'd bet on companies attacking workloads where GPUs are structurally inefficient:

Training → GPUs remain extremely hard to displace.

Inference/decode → wide-open opportunity.

Memory bandwidth → huge opportunity.

Power consumption → huge opportunity.

Low-latency agents/voice → specialized accelerators make enormous sense.

High-volume, fixed-model inference → ASICs can be extraordinary.

Academic benchmarking also reinforces that there isn't one universal winner: accelerator efficiency varies substantially with batch size, sequence length and model size.

My investment-style ranking

If I were building a private-market watchlist, I'd roughly rank them:

1. Cerebras — strongest proven GPU alternative
2. SambaNova — strongest flexible architecture
3. d-Matrix — most compelling memory-centric architecture
4. Tenstorrent — strongest open-platform thesis
5. Etched — highest asymmetric upside
6. Taalas — potentially phenomenal economics, but extreme model risk

One important update: Groq should no longer be treated as an independent startup bet. NVIDIA acquired its chip business for roughly $20B, so the Groq technology is now effectively part of NVIDIA's own strategy.

If you're asking this from an investment/VC perspective, I'd narrow the field differently: Cerebras, d-Matrix, Etched, Tenstorrent, and SambaNova are the five I'd spend the most diligence time on, with very different risk/reward profiles.

Its newly announced CS-4 is explicitly targeted at high-volume inference and is designed around three wafer-scale engines. Reuters reports that Cerebras is targeting major improvements in throughput and deployment efficiency with the system.

The particularly important signal is commercial traction: Cerebras went public in 2026 and has secured major compute commitments, making it considerably less speculative than most private accelerator startups.

Why I like it: real hardware + real customers + differentiated architecture + enormous inference throughput.

Big risk: wafer-scale manufacturing and the software ecosystem are harder problems than simply designing a faster chip.

---

2. sambanova.ai — perhaps the best strategic alternative

SambaNova takes a different route: its reconfigurable dataflow architecture and large memory hierarchy are designed specifically around AI workloads.

That matters because the biggest obstacle to replacing NVIDIA isn't actually FLOPS. It's memory capacity/bandwidth, networking, programmability and compatibility.

SambaNova is trying to offer substantially more flexibility than a narrowly hard-wired inference ASIC. Its SN40L architecture, for example, combines on-chip memory with HBM/DDR tiers.

Why I like it: potentially a better answer for enterprises running many different models rather than one fixed model.

Risk: NVIDIA's software ecosystem remains extraordinarily difficult to displace.

---

3. d-matrix.ai — the architecture I would watch closely

d-Matrix is especially interesting because it attacks the data-movement problem directly with digital compute-in-memory.

Instead of repeatedly moving model weights between conventional compute and memory—as GPUs often do during inference—the architecture puts computation much closer to the data. TrendForce describes its Corsair architecture as a digital-computing-in-memory approach designed to support a broad range of AI models.

It also raised $275 million at a $2 billion valuation, giving it substantially more financial runway than many semiconductor startups.

Why I like it: if inference becomes primarily a memory/energy problem, this architecture is attacking the right bottleneck.

Risk: proving that architectural advantage survives real production workloads and software integration.

---

4. tenstorrent.com — best open-ish alternative

Tenstorrent, associated with chip architect Jim Keller, is taking a more conventional accelerator route while using RISC-V and its Tensix architecture.

That's strategically important: rather than replacing NVIDIA with another vertically integrated proprietary ecosystem, Tenstorrent is betting on a more open compute stack.

Its Blackhole accelerator is actually shipping, which puts it ahead of many startups that have impressive simulations but limited production evidence.

Why I like it: potentially the strongest combination of accelerator hardware, RISC-V and an alternative software ecosystem.

Risk: it has to compete against companies with vastly greater resources while convincing developers to adopt yet another stack.

---

5. etched.com — biggest asymmetric bet

Etched has an unusually aggressive thesis: don't build a general AI accelerator; build a chip optimized around the Transformer architecture.

That's essentially trading flexibility for extraordinary efficiency.

If Transformers remain dominant for years, this could be brilliant. If architectures change substantially—or customers need lots of model flexibility—the specialization becomes a liability.

TrendForce categorizes Etched's approach as hard-wired Transformer inference, sitting between general accelerators and extremely model-specific silicon.

I'd classify it as: potentially enormous, but much less de-risked than Cerebras/SambaNova.

---

6. taalas.com — fascinating but almost brutally specialized

Taalas takes specialization even further: hard-wire a model into silicon.

The economics can be spectacular because you eliminate huge amounts of general-purpose overhead. Recent reporting around its HC1 described extremely high token throughput and exceptionally low cost per token—but the fundamental weakness is obvious: a chip optimized for one model becomes obsolete when that model changes.

So I'd view Taalas less as a direct NVIDIA replacement and more as a possible “ASIC for the most economically important models” business.

---

The bigger opportunity: don't replace GPUs everywhere

This is the key distinction I'd make.

I wouldn't bet on a startup whose pitch is simply:

“Our chip is a cheaper NVIDIA GPU.” That's a brutal business.

I'd bet on companies attacking workloads where GPUs are structurally inefficient:

Training → GPUs remain extremely hard to displace.

Inference/decode → wide-open opportunity.

Memory bandwidth → huge opportunity.

Power consumption → huge opportunity.

Low-latency agents/voice → specialized accelerators make enormous sense.

High-volume, fixed-model inference → ASICs can be extraordinary.

Academic benchmarking also reinforces that there isn't one universal winner: accelerator efficiency varies substantially with batch size, sequence length and model size.

My investment-style ranking

If I were building a private-market watchlist, I'd roughly rank them:

The race to bypass or reduce dependence on traditional Nvidia GPUs is heavily centered on AI inference —the phase where deployed models answer queries and generate tokens, which accounts for the bulk of everyday operational compute and energy costs. Instead of general-purpose parallel processing, these startups build application-specific architectures (ASICs) and wafer-scale chips designed to eliminate the memory-bottleneck and data-movement issues plaguing GPUs.

The strongest and most heavily capitalized infrastructure startups challenging the GPU monopoly include:

1. Etched
   • The Approach: Etched focuses purely on transformer-based models (like large language models), hardcoding the Transformer architecture directly into silicon rather than building a programmable chip.
   • Why it matters: By ditching general programmability for extreme specialization, Etched’s chips run inference at unprecedented speeds with radically lower energy and token costs. The company reached a massive $21 billion valuation following major backing and growing enterprise deployment interest.
2. Groq
   • The Approach: Groq utilizes a LPU (Language Processing Unit) featuring a deterministic, single-core streaming architecture. Rather than relying on off-chip high-bandwidth memory (HBM) that bogs down GPUs during token generation, Groq relies entirely on massive on-chip SRAM.
   • Why it matters: This pipeline/conveyor-belt design eliminates memory bottlenecks, allowing Groq to deliver blistering-fast inference speeds (measured in hundreds of tokens per second per user) and scale out via specialized cloud offerings like their Neo cloud ecosystem.
3. Cerebras Systems
   • The Approach: Cerebras takes a radical wafer-scale approach, turning an entire semiconductor wafer into one single, gargantuan processor (the Wafer-Scale Engine, now on generations like the CS-4) rather than cutting it into separate discrete chips.
   • Why it matters: By integrating massive amounts of on-chip memory and compute elements onto a single piece of silicon, it completely circumvents the inter-chip communication delays and memory-bandwidth walls that slow down multi-GPU clusters during large-model inference and training.
4. SambaNova Systems
   • The Approach: SambaNova uses a Reversible Data Flow architecture powered by their Reconfigurable Dataflow Unit (RDU).
   • Why it matters: The architecture is software-reconfigurable, meaning it can dynamically adapt to changing AI model topologies on the fly while maintaining high on-chip data reuse, making it versatile across both high-performance enterprise inference and targeted training workloads.

If you're looking into deployment or investment, would you like to explore these startups based on raw speed, cost-per-token efficiency , or ease of software integration?

The race to bypass or reduce dependency on traditional NVIDIA GPUs centers on solving the bottlenecks of inference speed, memory bandwidth/capacity , and power-hungry data movement.

The market is split into distinct alternative hardware and infrastructure categories, led by several prominent startups showing strong commercial traction or massive backing.

1. Wafer-Scale and Dataflow Accelerators (Alternative Core Silicon)

Instead of standard discrete GPUs, these companies design massive chips or reconfigurable dataflow architectures that process entire models or data streams without traditional memory bottlenecks.

• Cerebras Systems (Wafer-Scale Engine ): Famous for building the massive WSE-3 chip, Cerebras bypasses standard clustering and interconnect friction by placing an entire wafer-sized processor on a single piece of silicon. They lead the market in raw business and operational proof for extreme-speed inference and training alternatives.
• Groq (LPU / Language Processing Unit ): Focuses strictly on deterministic, ultra-low-latency inference. Groq’s architecture eliminates external memory access bottlenecks by keeping all model weights on-chip, drastically lowering the cost per token compared to general-purpose GPUs.
• SambaNova Systems (Reconfigurable Dataflow Unit - RDU ): Backed by major institutional funding, SambaNova targets enterprise generative AI deployments with its SN chips, which feature a unique three-tiered memory architecture designed to handle massive parameter models efficiently outside the CUDA ecosystem.

2. Transformer-Specific & Memory-Dense ASICs

General-purpose GPUs waste considerable energy on operations not strictly needed by modern Transformer architectures. These startups build Application-Specific Integrated Circuits (ASICs) hardwired specifically for LLMs.

• Etched.ai (Transformer ASIC ): Rather than building a flexible chip that can do graphics or general math, Etched builds ASICs custom-tailored only for Transformers. They claim performance orders of magnitude faster than standard GPUs for running high-volume LLM inference.
• MatX (Inference ASSP ): Founded by ex-Google TPU engineers, MatX pairs high-bandwidth memory for key-value caches with massive SRAM configurations on their MatX One chip, aiming at exceptionally low-latency decode phases.
• Positron AI (High-Memory Accelerators ): Focuses heavily on bypassing high-bandwidth memory (HBM) constraints by utilizing cheaper, high-capacity memory alternatives (like LPDDR5X) to pack up to $2TB of fast memory per chip, lowering the financial barrier of memory-heavy workloads.

3. Photonic (Light-Based) Compute & Optical Interconnects

Data movement between chips and racks consumes up to half a data center's power. Photonic startups use light instead of electrons to eliminate thermal and distance restrictions.

• Lightmatter (Photonic Processing ): A pioneer in optical computing, Lightmatter uses light to perform matrix multiplication. Their photonic processors target radical energy-efficiency gains (up to 10x over conventional chips) for data center inference.
• Celestial AI & Ayar Labs (Optical Fabric/Interconnects ): Rather than replacing the compute engine entirely, these companies focus on the "roads between the chips." They build optical interconnect technologies that allow compute and memory to communicate via light, eliminating the clustering bottlenecks that tether architectures to traditional GPU frameworks.
• Neurophos (Light-Powered Metamaterials ): An emerging player backed by heavy hitters like Gates Frontier and Microsoft's M12, Neurophos uses light-based metamaterial-driven chips aiming to scale down entire data center footprints.

If you are looking to evaluate these alternatives for a specific application, tell me:

• Are you focusing on training or inference?
• What is your target model scale (e.g., small edge models vs. multi-billion parameter LLMs)?

I can help narrow down which architecture fits your technical stack best.

If by “reduces GPU dependence” you mean startups attacking the bottleneck at the inference layer—either replacing GPUs with specialized silicon or making each GPU do substantially more work—the strongest names in 2026 are:

1. Etched — the aggressive bet

Etched is probably the most interesting pure GPU-displacement startup right now.

Its Sohu chip is designed specifically around Transformers rather than trying to be a miniature general-purpose GPU. That sacrifices flexibility in exchange for potentially enormous gains in inference efficiency. The company just raised $700M at a $21B valuation and says it has more than $1B in customer contracts, which is unusually strong validation for such a young hardware company.

Why I like it: If Transformer inference remains dominant, specialization can beat NVIDIA's generality.

Big risk: architecture changes. A startup betting its economics on today's Transformer structure has less protection against radically different model architectures.

2. Cerebras — strongest demonstrated alternative

Cerebras Systems is the most mature answer.

Instead of thousands of GPUs communicating over networks, Cerebras puts an enormous number of compute cores and SRAM onto a wafer-scale processor. Its newly announced CS-4 is explicitly aimed at inference and reducing the complexity of deploying large AI systems.

The important distinction is that Cerebras isn't merely promising a future architecture—it has a commercial system, customers, production experience and now public-market validation.

My ranking: #1 for proven non-GPU infrastructure.

3. SambaNova — perhaps the best enterprise architecture

SambaNova Systems takes a more flexible approach than Etched. Its architecture is designed to keep more of the model/context close to the compute rather than constantly moving data through the memory hierarchy.

That matters because moving model weights and KV-cache data can become a larger constraint than raw arithmetic. Research comparing accelerators also finds that the optimal architecture varies substantially with batch size, sequence length and model size—one reason I wouldn't bet the entire market on a single specialized ASIC.

My take: less spectacular than Etched on paper, but potentially more durable commercially.

4. d-Matrix — the one I'd watch closely

d-Matrix attacks the problem at an especially interesting point: data movement.

Its compute-in-memory architecture tries to perform computation close to where the data resides, reducing the energy and latency associated with repeatedly moving weights/data between memory and processors. The company raised $275M at a $2B valuation, according to industry reporting.

This is important because the long-term GPU alternative may not be “a faster GPU.” It may be an architecture that eliminates the need to move so much information around in the first place.

5. Tenstorrent — strongest “NVIDIA alternative” platform

Tenstorrent is less narrowly optimized than Etched, but that's also its appeal.

Its combination of AI accelerators, RISC-V CPUs and an open-ish software/hardware philosophy gives it a chance to become a general AI-compute platform, rather than a single-purpose inference chip.

I'd favor Tenstorrent if the thesis is:

“AI compute will diversify away from NVIDIA, but customers will still need something resembling a general-purpose accelerator.”

6. Taalas — the wild card

Taalas is taking specialization even further: effectively hard-wiring models into silicon.

That can produce extraordinary efficiency, but the trade-off is obvious: you give up flexibility for performance. Current industry comparisons put Taalas alongside Etched as one of the most extreme examples of the hard-wired inference approach.

I'd view it as a potentially enormous outcome rather than the safest investment.

---

Don't overlook the software startups

There's another way to reduce GPU dependence that I think is underappreciated: don't replace the GPU—stop wasting GPU computation.

Tensormesh is a particularly interesting example. Its system caches previously computed inference work, including KV-cache-related computation, so subsequent requests don't have to recompute everything. The company says its technology can reduce latency and GPU spend by up to 10× in applicable workloads.

Then you have the inference platforms:

• Baseten
• Fireworks AI
• Together AI

These companies optimize model serving, batching, quantization, routing and utilization. They don't necessarily eliminate GPUs, but they can dramatically reduce the number of GPU-hours required per token. The inference-platform market has grown rapidly in 2026.

My overall ranking

If I were building an investment/research watchlist around “what could weaken NVIDIA's GPU monopoly?”, I'd divide it like this:

Tier 1 — strongest today

1. Cerebras — proven alternative compute architecture
2. Etched — potentially enormous inference economics
3. SambaNova — flexible specialized architecture

Tier 2 — potentially huge 4. Tenstorrent — broader alternative compute platform 5. d-Matrix — compelling memory/compute architecture 6. Taalas — extreme specialization

Tier 3 — possibly the more capital-efficient opportunity 7. Tensormesh — eliminate redundant GPU work 8. Baseten / Fireworks / Together AI — optimize the inference stack

One important caveat: GPU dependence is unlikely to disappear uniformly. Training frontier models still favors enormous, flexible accelerator clusters. The more immediate opportunity is inference, where latency, memory bandwidth, utilization and dollars/token make specialized hardware much more attractive. Independent benchmarking also finds that no accelerator wins across every workload.

If you mean this from an investment perspective, I'd narrow the field further to Etched vs. Cerebras vs. SambaNova vs. d-Matrix vs. Tenstorrent, and I can rank those five by technology moat, customers, funding/valuation, probability of commercial success, and potential 10× outcome.

The AI infrastructure landscape targeting a reduction in GPU dependence is heavily focused on alternative silicon architectures (ASICs, dataflow processors, RISC-V), optical/photonic interconnects, and memory-optimized subsystems designed specifically for high-throughput inference.

The strongest and most well-capitalized startups tackling this challenge across different layers of the infrastructure stack include:

• Groq — Develops Language Processing Units (LPUs) that eliminate traditional external DRAM entirely in favor of massive, ultra-fast on-chip SRAM. This architecture drastically slashes latency and power consumption for real-time LLM inference, making it one of the premier direct alternatives for token-serving workloads.
• Cerebras Systems — Known for its massive Wafer-Scale Engine (WSE), which places an entire silicon wafer as a single processor. By removing traditional chip-to-chip communication bottlenecks and interconnect limits , Cerebras provides breakthrough performance for high-speed training and massive inference workloads.
• Tenstorrent — Led by legendary chip architect Jim Keller , Tenstorrent pairs customizable RISC-V CPU cores with proprietary AI accelerator IP . Their strategy focuses heavily on scalable, flexible hardware licensing and cloud deployments that bypass traditional monolithic GPU setups.
• Lightmatter — Tackles the GPU bottleneck from a physics perspective by using photonic (light-based) computing and optical interconnects . By routing data via light instead of copper wires, they aim to completely bypass the electrical heat and bandwidth walls that limit traditional data centers.
• SambaNova Systems — Builds reconfigurable Dataflow Units (RDUs) instead of fixed-instruction GPUs. Their software-defined dataflow architecture is engineered to dynamically adapt to changing neural network topologies, maximizing efficiency for large-scale enterprise AI inference.
• Etched.ai — Focuses tightly on building application-specific integrated circuits (ASPs/ASICs) hardwired exclusively for Transformer models, throwing away general-purpose programmability to achieve massive order-of-magnitude jumps in speed and efficiency for dominant LLM architectures.
• Celestial AI — Specializes in optical interconnect technology (Photonic Fabric), allowing compute and memory to integrate seamlessly with light. This drastically improves memory bandwidth scaling—often the single biggest chokepoint keeping clusters bound to NVIDIA hardware.
• d-Matrix — Focuses on digital-in-memory computing architectures designed specifically to solve the memory wall in generative AI inference , providing high energy efficiency and low cost-per-token economics.

If you are evaluating these for a specific project, would you like to focus on inference vs. training , or are you looking at solutions for the data center vs. the edge?

If by “reduces GPU dependence” you mean startups that can materially replace NVIDIA GPUs in production AI infrastructure—not merely optimize GPU utilization—I’d focus on inference, where purpose-built silicon can exploit the fact that serving a trained model has very different bottlenecks from training it.

My current ranking

1. d-Matrix — my top private-company pick

d-Matrix is the one I'd study most closely.

Its Corsair architecture is explicitly designed around the memory bottleneck in LLM inference rather than trying to reproduce a GPU. The company raised $275M at a $2B valuation and, importantly, announced in June 2026 that Corsair had entered full production, with volume shipments planned to hyperscalers, neoclouds and frontier labs.

The really interesting part is that d-Matrix isn't necessarily saying “throw away every GPU.” Its published architecture can be combined with GPUs in a heterogeneous system, reportedly achieving a 10× inference speedup for certain workloads.

Investment thesis: inference becomes increasingly memory-bound → specialized memory-centric compute wins → d-Matrix becomes an inference appliance alongside conventional accelerators.

---

2. Etched — highest-upside/highest-risk bet

Etched is the more radical wager.

Instead of building a flexible GPU-like accelerator, Etched is betting heavily on the Transformer architecture and specializes its silicon accordingly. That sacrifices generality for potentially enormous efficiency.

The commercial signal is unusually strong for a startup this young: Etched said in June that it had already booked $1B in contract orders for its first systems, which are now undergoing customer testing.

It has also reportedly been discussing valuations as high as $20B, although those financing discussions were not finalized.

Why I like it: if Transformers remain dominant, this is exactly the sort of workload where a fixed-function-ish accelerator can destroy a general-purpose GPU on cost and power.

Why I'd be cautious: the whole thesis depends on model architectures remaining sufficiently Transformer-like. It's also not yet as proven operationally as Cerebras.

---

3. Tenstorrent — best “new computing platform” play

Tenstorrent is different. It isn't simply an inference ASIC company; it is trying to create a more general accelerator platform built around RISC-V and its own AI architecture.

It raised more than $693M in Series D at a $2B pre-money valuation, bringing total funding to roughly $1.2B.

I'd put it above some of the more exotic inference startups because it has a plausible answer to the question:

What replaces CUDA if customers don't want their entire AI stack controlled by NVIDIA?

That's potentially more valuable than simply building a faster inference chip.

Big caveat: software. NVIDIA's moat isn't just silicon.

---

4. Lightmatter — potentially the “picks and shovels” winner

Lightmatter attacks a different part of the problem: moving data between processors.

Its photonic interconnect technology uses light to connect AI processors at scale. tenstorrent.comlightmatter.cowww.photonics.com It raised $400M in Series D at a reported $4.4B valuation and has raised roughly $850M.

This is important because the future may not be:

GPU → better GPU → replacement GPU

It may instead be:

lots of heterogeneous accelerators + dramatically better interconnect + distributed memory.

Lightmatter could benefit regardless of whether the eventual compute engines are NVIDIA, AMD, Cerebras, Tenstorrent, or custom hyperscaler silicon.

---

5. Fractile

Fractile is one I'd watch closely as an earlier-stage asymmetric bet.

Its premise is similar to the broader industry movement: don't spend enormous amounts of power moving model weights around just to perform relatively simple inference operations.

It doesn't yet have the commercialization evidence of d-Matrix or Etched, so I'd rank it lower on current strength but potentially quite high on technical upside.

---

One important exclusion: Groq

Historically, Groq would have been near the very top of this list.

But the situation changed dramatically. NVIDIA struck a $20B deal for Groq's technology and key talent, while Groq itself continues operating and is now incorporating NVIDIA systems into GroqCloud.

So I wouldn't treat Groq as an independent GPU-displacement startup anymore.

Likewise, Cerebras is arguably the most commercially proven non-GPU accelerator company, but it is now public rather than a startup in the conventional VC sense. Its Q2 2026 results showed cloud revenue growing fourfold to $126M, and it raised its 2026 revenue outlook to roughly $880–890M.

If public companies count, Cerebras belongs near the top of the list.

The bigger picture

I'd divide the opportunity into four bets:

A. Replace GPUs for inference → d-Matrix, Etched, Cerebras

B. Build a credible alternative general-purpose AI platform → Tenstorrent

C. Make existing accelerators dramatically more efficient → Lightmatter, optical-interconnect companies

D. Extreme architectural bets → Fractile, MatX, Positron

And the most interesting strategic observation is that “GPU dependence” probably won't disappear through one NVIDIA-killer chip. The emerging architecture looks more like heterogeneous inference: GPUs/CPUs handle some stages, specialized accelerators handle others, and optical/memory technologies reduce the cost of moving data between them. The AMD–Cerebras collaboration is a particularly good example: AMD handles context/prefill while Cerebras handles token generation.

If I were building an investment/watchlist today, my top five would be: d-Matrix → Etched → Tenstorrent → Lightmatter → Fractile.

If you're asking this from an investment perspective, I can also rank these by 10-year upside, probability of becoming a $10B+ company, current valuation, moat, and likelihood of acquisition by NVIDIA/AMD/Intel/Google.

The AI infrastructure market is actively moving past pure general-purpose GPU dependence, driven heavily by the soaring cost-per-token and massive power demands of high-volume AI inference . Startups attacking this bottleneck do so by re-architecting data movement, swapping electrons for light, or targeting application-specific architectures.

The strongest and most heavily backed startups challenging or bypassing traditional GPU paradigms fall into several key architectural tracks:

1. Photonic & Optical Interconnect Startups (Using light instead of electricity to reduce heat and latency)

• Lightmatter : A frontrunner in photonic computing and interconnects. Lightmatter builds chips that use light to move and process data within data centers, claiming massive efficiency gains over traditional copper-and-silicon electrical processing for transformer models.
• Celestial AI : Focuses on optical interconnect technology via its proprietary "Photonic Fabric." Their platform decouples compute from memory scaling, addressing the memory-wall bottleneck that typically forces massive GPU clustering.
• Ayar Labs : Specializes in optical I/O chip-to-chip connectivity, enabling processors to communicate using light at a fraction of the power required by conventional electrical links.

2. Transformer-Specific & Specialized Inference ASICs (Eliminating general-purpose overhead)

• Etched.ai : Building application-specific integrated circuits (ASSPs) custom-designed exclusively for Transformer model inference. By entirely dropping support for training or non-transformer architectures, they target orders-of-magnitude improvements in throughput and cost compared to standard GPUs.
• Groq : Famously eliminates traditional dynamic random-access memory (DRAM) bottlenecks using a LPU (Language Processing Unit) architecture with deterministic, on-chip SRAM. This drastically reduces the memory-access latency plaguing traditional GPU-based inference.
• Cerebras Systems : Tackles scale by scaling up a massive wafer-scale engine that eliminates the need for multi-chip server interconnects, letting entire massive models sit close to unified on-chip memory.

3. Memory-Centric & Alternative Compute Architectures

• d-Matrix : Focuses on eliminating the compute-memory split with a digital compute-in-memory (CIM) architecture optimized for generative AI inference, yielding high efficiency for LLM serving.
• Tenstorrent : Led by industry veteran Jim Keller, this RISC-V-based hardware and IP startup creates customizable, highly scalable AI and general compute accelerators that offer an open-ecosystem alternative to proprietary CUDA stacks.
• MatX : Founded by former Google TPU engineers, MatX develops specialized processors combining high-bandwidth memory for key-value caches and heavy SRAM for weights, targeting low-latency decode workloads.

If you're evaluating these for a specific deployment or investment thesis, would you prefer to focus on inference-heavy software/hardware integration or data center networking and power-saving fabrics?

If by “reduces GPU dependence” you mean startups building infrastructure that can replace GPUs—or make each GPU do substantially less work—the strongest group in 2026 is concentrated around inference, especially the memory-bandwidth-heavy decode phase of LLMs.

My ranking

1. Cerebras — strongest overall

This is my #1 if the question is “who has the most credible path to taking meaningful inference workloads away from Nvidia GPUs?”

Cerebras' Wafer-Scale Engine puts an enormous amount of SRAM and compute on a single wafer, attacking the memory-bandwidth problem that makes autoregressive decoding expensive on GPUs. The company now claims up to 15× faster inference than GPUs, and it has real commercial infrastructure rather than merely a chip prototype.

The particularly interesting development is its July 2026 partnership with AMD: AMD GPUs handle high-throughput prefill while Cerebras handles ultra-fast decode. That's a sign of where the industry may be going—heterogeneous infrastructure rather than “GPU versus everything.”

Cerebras also raised $1B at roughly a $23B valuation in February and went public in May, so it's arguably transitioning from startup to major AI-infrastructure company.

My view: highest probability of becoming a major alternative inference platform.

---

2. SambaNova — perhaps the best architecture for the agent era

SambaNova's SN50 RDU is purpose-built around the observation that inference is increasingly a data movement and memory problem, not merely a FLOPS problem. Its architecture uses reconfigurable dataflow rather than the conventional GPU programming model.

The really compelling part is that SambaNova is explicitly targeting agentic inference, where a system makes many sequential model calls. Its SN50 can be deployed in racks of 16 accelerators and scale to 256.

There is also evidence of a practical hybrid model: SambaNova recently demonstrated H200 GPUs for prefill + SN50 for decode, rather than requiring customers to rip GPUs out entirely.

My view: potentially the strongest competitor if the future is heterogeneous inference rather than pure GPU replacement.

---

3. d-Matrix — my favorite “picks and shovels” bet

d-Matrix is particularly interesting because it doesn't insist that GPUs disappear.

Its Corsair accelerator is designed to take the memory-intensive, latency-sensitive portion of inference and work alongside GPUs. The company entered full production in June 2026, with volume shipments to hyperscalers, neoclouds and frontier labs.

More importantly, d-Matrix and customers have demonstrated 10× improvements in certain heterogeneous inference configurations versus GPU-only systems.

That's an attractive business model: instead of convincing a customer to abandon Nvidia, you say:

Keep your expensive GPUs. We'll make you need fewer of them.

My view: perhaps the best risk/reward company in the category.

---

4. Etched — highest “ASIC efficiency” upside

Etched takes the extreme approach: hard-wire the silicon around the Transformer architecture.

That sacrifices generality for efficiency. If Transformer inference remains dominant, that can be an extraordinarily powerful tradeoff. The downside is obvious: a rapidly changing model architecture could make a highly specialized chip obsolete.

The broader 2026 inference-chip landscape confirms Etched's hard-wired approach is substantially different from the more flexible architectures of d-Matrix and SambaNova.

My view: enormous upside, but considerably more architecture risk than Cerebras/SambaNova.

---

5. Tenstorrent — the long-duration architecture bet

Tenstorrent is less narrowly focused on inference than the names above, but I think it's strategically important because it is trying to build an alternative AI compute ecosystem, rather than merely a faster Nvidia-like GPU.

Its RISC-V orientation and emphasis on open software/hardware make it interesting if AI infrastructure eventually becomes much more heterogeneous.

My view: less immediately compelling for replacing Nvidia inference capacity, but potentially one of the more important independent AI-compute companies over a 5–10 year horizon.

---

The important distinction

I would actually divide the opportunity into three categories:

A. Replace GPUs

• Cerebras
• SambaNova
• Etched
• Tenstorrent
• FuriosaAI

B. Reduce the number of GPUs required

• d-Matrix
• SambaNova
• Cerebras
• specialized inference accelerators generally

C. Make the GPU cluster dramatically more efficient

• Lightmatter
• networking/interconnect companies
• inference software and orchestration companies

And B may ultimately be the biggest market.

The emerging architecture isn't necessarily:

Nvidia GPU → dead

It's more likely:

GPU for prefill/training + specialized accelerator for decode + CPU for agent/tool work + optical/network fabric connecting everything.

Recent research supports that direction: different inference stages have materially different compute and memory requirements, and simulated prefill/decode/attention/FFN disaggregation can substantially improve throughput versus homogeneous GPU systems.

One caveat on the old “GPU alternatives” list

I would not put Groq at the top of a current independent-startup list anymore. Nvidia announced a transaction involving Groq's chip technology in late 2025, while Groq's inference cloud continues operating; the corporate/technology situation is therefore quite different from an independent Nvidia challenger. Groq itself raised another $650M in June 2026 to expand its inference cloud.

Likewise, Taalas is no longer an independent startup after AMD agreed to acquire it this month.

If I were making a 2026 watchlist

Highest conviction: Cerebras
Best heterogeneous/GPU-reduction play: d-Matrix
Best agentic-inference architecture: SambaNova
Highest-risk/highest-upside ASIC: Etched
Long-term ecosystem bet: Tenstorrent
Infrastructure layer bet: Lightmatter

The deeper investment thesis I'd investigate is not simply “Nvidia alternatives.” It's “inference becomes heterogeneous.” The companies that can remove the GPU from the parts of inference where GPUs are economically inefficient may have a much larger addressable market than companies trying to replace Nvidia everywhere.