Etched, the startup building custom hardware for AI inference, has raised a $300 million Series C at a $10.3 billion valuation, according to TechCrunch, which cited co-founder and COO Robert Wachen. The round was led by Sequoia, with participation from Andreessen Horowitz, SK Hynix, Jane Street, Diffusion Capital, and prior investors.

The financing matters beyond the size of the check. It is another signal that investors are willing to fund alternatives to Nvidia-centered AI infrastructure if those startups can show working silicon, early customer tests, and a plausible path to lower inference costs. In Etched’s case, the company says it has already manufactured its chips, is testing full systems with customers, and has booked $1 billion in orders. Those operating claims, however, come from the company and have not been independently verified in the source material.

Etched has attracted attention since its 2022 founding because it took an unusually opinionated approach: design hardware around the workloads that dominate modern model serving, then sell complete systems rather than standalone chips. That idea was initially easy to dismiss as the market converged around GPUs. But as AI spending shifts from training toward production inference, the startup’s pitch has become more legible to both investors and enterprise buyers looking for cost and latency improvements.

Why Etched’s raise stands out now

According to TechCrunch, Etched’s new valuation is roughly double the $5 billion valuation it reached in December, when it raised $500 million. TechCrunch also reported that the company says this is the highest valuation ever for a Sequoia-led Series C. That specific superlative is a company claim relayed by the publication, not an independently established market dataset.

The timing is notable. AI infrastructure funding has increasingly split into two camps: broad compute platforms that compete on flexibility, and specialized systems that target a narrower but potentially huge slice of demand. Etched is firmly in the second camp. Its argument is that the economics of serving models at scale are different from the economics of training them, and that dedicated inference hardware can outperform more general-purpose accelerators on the workloads that matter most in production.

That pitch is getting a hearing because the bottleneck in AI deployment is no longer just access to raw compute. Builders and enterprises are also focused on latency, power, memory bandwidth, and cost per generated token. If a startup can materially improve those metrics without forcing customers to rewrite their entire stack, it can find demand even in a market dominated by Nvidia.

The company’s investor roster also helps explain the attention. Beyond Sequoia and Andreessen Horowitz, TechCrunch says Etched has backing from SK Hynix and high-profile individuals including Peter Thiel, Andrej Karpathy, Dylan Field, and Amjad Masad. For a chip startup, support from SK Hynix is especially relevant because memory and interconnect design are central to inference performance, not just raw compute.

The product claim: custom chips for inference without GPUs

Etched’s core message, as described by Wachen to TechCrunch, is that its systems can speed up inference on AI models without relying on GPUs. The startup says its hardware can run “any AI model,” including transformer-based systems, Mixture of Experts designs such as DeepSeek and Qwen, and even non-transformer architectures like Mamba.

That breadth matters because one of the common criticisms of application-specific AI chips is that they can become stranded by model shifts. A system optimized too tightly for one architecture may deliver impressive demos but age poorly if developers move to a different model family. Etched is clearly trying to counter that concern by arguing that its architecture is broader than many skeptics assume.

TechCrunch reports that Etched breaks inference into two stages: prefill and decode. In the company’s description, prefill is the compute-heavy work of processing the prompt and its context, while decode is the token-by-token generation phase that is less compute-intensive but heavily dependent on memory. Etched says it designed two new components from scratch to address those stages.

For prefill, Wachen told TechCrunch that Etched built a chip that runs at lower voltage, which the company says reduces heat and allows denser transistor packing. For decode, the company says it created what it calls cluster-scale memory, along with interconnect technology that lets many chips share a memory pool with low latency. The promised outcome is faster inference at lower cost.

Those are important claims, but they remain company-reported. The source material does not include third-party benchmarks, customer case studies, independent teardown analysis, or standardized comparisons against Nvidia systems. That does not make the claims false; it means buyers and developers should treat them as promising but still preliminary until broader evidence appears.

Manufacturing progress and early deployment signals

One reason Etched is commanding a premium valuation is that it appears to have advanced beyond slideware. TechCrunch reports that the startup previously announced successful manufacturing of its first batch of silicon by TSMC, and that its first full systems are being tested by clients.

That distinction matters in semiconductors. Many AI chip startups attract capital on architecture ideas, but far fewer make it through manufacturing and into systems validation. Etched is also no longer a tiny research effort. According to TechCrunch, the company now has 400 employees, operates a 2 megawatt data center at its main office, and has opened a new 80,000-square-foot, 10MW facility in Milpitas.

Those infrastructure details suggest the company is preparing not just to design chips, but to validate, run, and potentially deploy full racks at meaningful scale. For enterprise AI buyers, that is relevant because hardware procurement decisions increasingly depend on complete system readiness, software integration, and operational support rather than just the chip spec sheet.

Still, there is a large gap between having early systems in test and delivering at scale. Wachen himself acknowledged that scaling manufacturing and deployment remains difficult, according to TechCrunch. That caution is important. Hardware startups can look credible long before they prove supply chain resilience, field reliability, serviceability, and software maturity under customer workloads.

Evidence, benchmarks, and what remains unverified

The strongest factual points in this story are the financing terms reported by TechCrunch, the named investors including Sequoia, Andreessen Horowitz, Jane Street, Diffusion Capital, and SK Hynix, and Etched’s public narrative that it has working silicon and customer system tests underway.

The more ambitious operating signals need more scrutiny. The reported $1 billion in orders is a company claim carried by TechCrunch. The sources provided do not define whether those orders are binding purchase orders, reservations, letters of intent, or another form of demand signal. That distinction matters because AI infrastructure companies often cite backlog or booked demand before revenue is recognized or deliveries begin.

Likewise, Etched’s claims that its systems can run any AI model and deliver high speeds at lower cost have not been independently benchmarked in the provided materials. The article notes limited access so far, mostly through investor and early-customer demos. Wachen also cited interest from figures associated with Anthropic and OpenAI, but those remarks should be read as executive commentary, not formal product endorsements from Anthropic or OpenAI.

The broader market context does, however, make Etched’s thesis easier to understand. TechCrunch notes that Google is reportedly pursuing a related concept with a “Frozen v2” chip for Gemini. Even if the implementations differ, the underlying logic is similar: if inference becomes the main cost center, customized silicon aimed at specific serving patterns becomes more attractive.

What this means for AI builders and enterprise buyers

For builders, the practical question is not whether a custom inference chip sounds elegant. It is whether it can slot into real deployment pipelines for LLMs, DeepSeek, Qwen, or Mamba-based systems without introducing enough friction to offset the hardware gains. Compatibility across model architectures, support for modern serving frameworks, and reliability under mixed workloads will matter more than headline valuations.

For enterprise AI teams, Etched’s rise is another reminder that the infrastructure stack is opening up beneath the application layer. Nvidia remains the default, but companies serving high-volume inference may eventually have more options across purpose-built systems, memory-optimized racks, and vertically integrated platforms. If Etched can validate its low-voltage inference and cluster-scale memory claims in production, it could pressure incumbents on cost-per-token and response-time economics.

The challenge is procurement risk. Enterprises do not just buy performance. They buy supply assurance, software tooling, support coverage, security review, and roadmap credibility. A startup can win pilot interest with a compelling rack demo, but broad adoption usually follows only after repeatable deployment evidence accumulates.

For the AI chip market, the fundraise reinforces a shift from “can anyone challenge GPUs?” to “which AI workloads are now large enough to justify custom silicon?” Inference is the clearest candidate, especially as model usage grows and buyers look for alternatives to expensive general-purpose accelerators.

What to watch next

The next signals to track are straightforward. First, whether Etched publishes independent benchmarks against Nvidia-based inference systems under clearly described conditions. Second, whether named customers emerge with production deployment details rather than private test references. Third, whether the company clarifies the nature of its reported $1 billion order book.

It will also be worth watching how Etched supports model diversity over time. Claims of broad support across transformer systems, Mixture of Experts models like DeepSeek and Qwen, and state-space approaches like Mamba are strategically important. They will need to hold up as serving software and model architectures evolve.

Finally, the company’s manufacturing and delivery cadence may matter more than any fundraising milestone. A 10MW facility and working systems are meaningful, but the real test is whether Etched can move from early access to repeatable, supported deployments without losing the cost advantages that justify its design.

Creati.ai perspective

Etched’s latest round is less a verdict on one startup than a sign that inference has become the hottest battleground in AI infrastructure. Training still drives prestige, but production economics drive budgets. Any company that can reduce latency and cost for serving popular models has a real opening, even in a market shaped by Nvidia.

The caution is that hardware stories often outrun field evidence. Etched appears to have crossed several hard thresholds, including manufacturing and system testing, which makes it more credible than many AI chip hopefuls. But until independent benchmarks and customer deployments are public, the valuation is best read as investor confidence in a thesis: that specialized enterprise AI infrastructure for AI agents and model serving can become a large business if it proves easier to adopt than its skeptics expect.