AMD has closed its acquisition of Taalas, a startup building dedicated AI inference chips, and plans to fold the technology into its accelerator roadmap alongside Instinct GPUs as system-level inference solutions. Taalas hard-wires AI models onto ASICs optimized for specific model workloads, trading generality for markedly higher inference efficiency. The deal fills a gap in AMD's full-stack inference portfolio and signals a long-term strategy of blending model-specific silicon with general-purpose GPUs.
ASIC
Chips customized for specific tasks, including cloud vendors' in-house AI accelerators, a major compute path beyond general-purp
Curated Intelligence
The author argues that extreme hardware specialization, where a dataset is compiled directly into custom silicon, is closer than expected. As a proof point, a single-cycle Connect-4 chip was learned end-to-end from simulated games for under $10 of compute, running at ~25 MHz and consuming a few hundred picojoules per move, matching a CPU engine's speed at two to three orders of magnitude lower energy.
TSMC is factoring customer balance sheets, historical loss rates and data-center construction progress into capacity allocation, requiring cash payment when orders exceed credit limits and prepayment to reserve capacity. Customer advance payments reached NT$234.23 billion by end-June, showing AI circular-financing risk now reaching the foundry layer.
MediaTek reported strong Q3 FY2026 results with revenue hitting NT$168.06 billion, up 18.3% YoY and significantly exceeding guidance. While net profit growth lagged slightly behind revenue, the overall performance beat analyst expectations. The company confirmed it will begin mass production of its first data center AI acceleration ASIC in Q4, targeting over $2 billion in revenue from this segment by 2027 as a new growth driver.
TSMC plans a 10-20% price increase for its CoWoS advanced packaging services due to severe capacity constraints driven by surging AI demand. Pressure has expanded from GPUs to CPUs and ASICs, with equipment lead times extending to 6-8 months; monthly capacity is projected to reach 200k wafers by 2027.
Counterpoint Research projects global humanoid robot AI accelerator shipments to exceed 1 million units by 2030, with a CAGR of 144%. While Nvidia currently dominates the GPU market with its full-stack ecosystem, competitors like Qualcomm, Tesla, and XPeng are rising through ASICs and heterogeneous computing, shifting the landscape from single-architecture dominance to multi-architecture coexistence.
Intel CEO Lip-Bu Tan clarified that the US government holds no intent to interfere with corporate governance despite owning ~8.2% equity. He reported 80% completion of restructuring and identified AI advanced packaging and ASICs as key growth drivers, noting potential high returns for investors like SoftBank and Nvidia.
Bloomberg reports that ByteDance is in advanced negotiations with MediaTek for custom ASICs to power its data centers. If finalized, this deal would make ByteDance the second-largest customer of MediaTek's ASIC business after Google. The move addresses supply chain constraints under US export controls by diversifying away from Nvidia, while helping MediaTek expand its footprint in the growing data center chip market.
Microsoft plans to scale its Azure Maia AI chips to multi-gigawatt levels by 2027, establishing them as the primary inference system across its products. Currently, Maia 200 supports Microsoft Copilot in key regions and claims 1.7 times more tokens per watt than leading GPUs in Teams scenarios.
Counterpoint Research dismisses recent concerns about slowing HBM demand as misinformation, asserting that the trend toward high-stack HBM will continue. With NVIDIA's Rubin GPUs shifting to 12-Hi HBM4 and no low-stack alternatives emerging in ASIC solutions, bandwidth and capacity remain more critical than cost reduction for AI XPU workloads.