The Compute 100 — Friday July 24, 2026

AMD and Cerebras announced a partnership that splits AI inference into two specialized stages — AMD's Helios rack-scale Instinct GPU systems handle prefill, the compute-heavy work of processing a prompt, while Cerebras's wafer-scale engine handles decode, the memory-bandwidth-bound work of generating tokens one at a time — with a claimed 5x gain in tokens-per-second-per-watt that is modeled by the two companies against Cerebras's own prior configuration rather than independently benchmarked. The deal folds Cerebras, two months removed from its Nasdaq debut, into AMD's Instinct roadmap as a decode specialist rather than a standalone Nvidia challenger, and it is the clearest production test yet of the bet that inference workloads are heterogeneous enough to reward splitting a single request across two different chip architectures. Check out Cerebras CEO Andrew Feldman's own explainer of the prefill/decode split, including the risk of stranded capacity if traffic patterns shift. (CNBC)

Private Companies

TensorWave, the AMD-focused neocloud backed by AMD Ventures, Magnetar, and Nexus Venture Partners, is deploying AMD's new Helios rack-scale system — Instinct MI455X GPUs paired with Venice EPYC CPUs — to expand its frontier-scale training and inference capacity. TensorWave already operates one of the largest all-AMD GPU clouds in the world, and early access to Helios makes it a proof point for whether AMD's neocloud strategy can scale beyond the handful of hyperscalers it has signed so far. It is a direct beneficiary of any customer that decides diversifying away from Nvidia is worth the software-migration cost. (Business Wire)

Elio, founded by former Meta AR/VR executives Nadav Grossinger and Nitay Romano, raised $21M led by Innovation Endeavors and Xora, with Kevin Weil, Scribble VC, UpWest, and Resolute Ventures participating. The company builds "machine-native" image sensors that use dynamic micromirror optics to decide what to capture before light hits the sensor, targeting semiconductor-manufacturing inspection, microscopy, robotics, and defense. It is a bet that as AI perception workloads scale, the sensor itself becomes a place to cut compute cost, rather than pushing all of the filtering downstream into the accelerator. (Business Wire)

Arteris disclosed that Speedata, an Israeli startup building a dedicated Analytics Processing Unit for Spark SQL, data-prep, and agentic-analytics workloads, has deployed Arteris' FlexNoC on-chip interconnect in its Callisto processor and will use it again in its next-generation Andromeda chip. It is a small disclosure with a bigger point: a new chip category built specifically for data-analytics and ETL workloads, distinct from both training GPUs and inference accelerators, is now reaching production silicon rather than staying on a roadmap slide. (GlobeNewswire)

Public Markets

Intel's Q2 revenue jumped 25% year-over-year to $16.1B, its fastest growth rate in more than 15 years, with gross margin and Q3 guidance both well above consensus — shares jumped as much as 13% after hours before settling near 6%, the clearest sign yet that AI-linked data-center demand is showing up in Intel's own results, not just its foundry customers'. (FX Leaders)

Emerging

Researchers at the Technology Innovation Institute in Abu Dhabi developed an AI model that can quickly predict when a quantum computation is likely to fail by looking only at the raw error signals produced during the computation. It doesn’t need to know how the underlying error-correction system works or be trained on examples of successful and failed runs. In tests on both simulated quantum computers and QuEra’s real neutral-atom hardware, it outperformed the standard method used today. As quantum vendors race toward fault tolerance, this kind of lightweight “quality check” could reduce the amount of expensive error correction and lower the cost of running useful quantum computers. (arXiv)

Watch This Week