AI & Tech Weekly (Aug 25–29): Spectrum-XGS, inference token surge, quantum-centric supercomputing, TSMC 2nm, ASIC earnings
This weekly brief is compiled by the iaiseek editorial team based on public sources and iaiseek’s website/social channels, covering major AI developments from Aug 25–29. Focus areas: progress in cross-domain Ethernet interconnects, expanding inference token demand, quantum × HPC/AI collaboration, 2nm mass-production ramp, and the evolving trajectory of ASIC businesses. Below are the key items and potential implications.
1) NVIDIA launches Spectrum-XGS Ethernet, speeding cross-domain “AI super-factories”
NVIDIA announced Spectrum-XGS Ethernet to interconnect multiple, geographically distributed data centers so they can operate like a single supercomputer. The cross-domain architecture targets high latency and unpredictable performance in traditional Ethernet with adaptive congestion control, precise latency management, and end-to-end telemetry. Cloud provider CoreWeave is an early adopter.
Commentary:
By combining algorithms and observability, Spectrum-XGS delivers ~1.6× bandwidth density and nearly ~1.9× overall performance uplift, materially easing latency/jitter over long distances. Compared with InfiniBand, open-standard Ethernet (including SONiC support) offers broader ecosystem compatibility and cost advantages—well-suited for orchestrating very large-scale compute across domains.
2) Inference demand surges; Google and Microsoft lead the global compute wave
Data indicates Google processed over 980 quadrillion tokens in July 2025—roughly double May. Microsoft handled over 500 quadrillion tokens via Foundry API in its FY2025 (through June), up more than 7× YoY. ByteDance averaged 16.4 quadrillion tokens/day by end-May 2025. By end-June, China’s daily average token consumption reached 30 quadrillion—up ~300× versus early 2024.
Commentary:
Inference volumes now far exceed training, showing AI moving from R&D to broad deployment. Inference APIs can reach ~70% gross margin, becoming a cash engine for cloud vendors (e.g., Google Cloud at $13.6B in Q2 2025; Azure AI revenue up ~5× YoY). At the same time, massive token throughput elevates data-governance and privacy requirements—growth must be balanced with compliance.
3) IBM and AMD to build a quantum-centric supercomputing architecture
IBM and AMD announced a partnership to integrate quantum computing with HPC/AI under a quantum-centric supercomputing approach: quantum processors work seamlessly with CPUs, GPUs, and other accelerators to better scale complex AI workloads.
Commentary:
Quantum excels in areas like molecular simulation while HPC/AI shines in large-scale data processing; together, they can address problems beyond traditional paradigms. Although quantum remains early and “AI × quantum” narratives risk overheating, this collaboration strengthens confidence in the integration path. If algorithms, compilers, and scheduling are stitched together well, it could open room for the next wave of AI breakthroughs.
4) TSMC to begin 2nm mass production in Q4; Apple secures nearly half of initial capacity
TSMC plans to start N2 (2nm) mass production in Q4 2025, with wafer pricing around $30,000. Apple has reportedly locked nearly 50% of initial capacity and is expected to use the A20 (N2) for iPhone 18. Qualcomm, AMD, MediaTek, and Broadcom are also among first-wave customers.
Commentary:
N2 adopts GAA nanosheet transistors. Versus 3nm, it targets ~15% density gain, ~15% performance uplift at iso-power, and ~24–35% lower power, with yields reportedly near ~90%. Apple’s large orders stabilize the ramp and bolster ecosystem confidence; other top customers signal strong demand from AI, HPC, and premium devices. The ~$30K/wafer price, however, raises the bar for startups and academia, likely widening industry stratification.
5) Marvell Q2 beats but guidance weighs; market “votes with its feet”
Marvell (FY2026 Q2) reported $2.01B revenue (+57.6% YoY) and $0.67 non-GAAP EPS, but guided Q3 revenue to a $2.06B midpoint—below the ~$2.11B consensus. AI-related (especially ASIC) contributions were notable: $876M in the July quarter and a projected $955M in the October quarter. Shares fell over 10% after hours on concerns around decelerating growth.
Commentary:
After selling its automotive Ethernet unit to Infineon for $2.5B, Marvell is focused on AI chips. ASIC growth, helped by AWS Trainium 2 and Google Axion shipments, nonetheless slowed sequentially (+9% this quarter for AI), dampening sentiment. Peers like Broadcom run larger ASIC businesses. Marvell’s continued R&D investment should benefit from rising custom-silicon adoption, but near-term multiples are constrained by guidance—watch orders and new-platform ramps for inflection.
This brief was selected and organized by the iaiseek team from public channels and company disclosures. For more cutting-edge AI updates, business insights, and tech trends, visit:
https://iaiseek.com
We’ll keep tracking cross-domain compute networking, scaled inference commercialization, advanced-node ramps, and the ASIC ecosystem. Share your observations in the comments—and see you next week.