Etched’s $10.3B bet on AI chips without GPUs
A little over two years after its quiet launch, AI chip startup Etched has pulled off what many said was impossible: it convinced marquee investors to value a GPU-free inference business at $10.3 billion. The Series B round, led by Fidelity and joined by funds managed by BlackRock, T. Rowe Price and others, closed at $1.1 billion and marks one of the largest single financings in semiconductor history.
Etched was founded by three Harvard dropouts who argue that traditional GPU pipelines are the bottleneck for today’s AI models. Instead of cramming more compute into graphics chips, the company built custom silicon and memory components that accelerate inference by design. The result, it claims, is orders-of-magnitude lower latency and power draw for large language models and vision systems—without needing Nvidia’s GPUs at all.
A different silicon bet
The company’s approach strips out the general-purpose graphics logic that has dominated AI training and inference since the 2010s. Etched’s chips integrate bespoke compute units, on-chip SRAM and a proprietary interconnect fabric optimized for the specific dataflows of transformer networks. Early benchmarks shared by the startup show single-digit millisecond response times on 70-billion-parameter LLMs at a fraction of the power envelope of comparable GPU clusters.
What this means for the data-center stack
If Etched’s performance numbers hold up outside the lab, data-center operators could diversify away from GPU oligopolies that have driven capital expenditure sky-high. Hyperscalers running inference at scale would gain a non-GPU alternative, potentially lowering costs and improving margins—provided Etched can deliver chips in volume and secure long-term supply agreements.
Why it matters
This financing signals that investors are willing to back radical departures from the GPU-centric status quo, even when incumbents like Nvidia dominate 90% of the AI accelerator market. It also tests whether a fabless startup can scale custom silicon fast enough to matter in a cycle measured in quarters, not years. The real stakes are not just valuation headlines, but whether a new architecture can reshape the economics of inference—one of the fastest-growing slices of the AI stack.
Source: TechCrunch. AI-assisted editorial synthesis — TechnoExpress.