The GPU Daily · #055 · covering Sunday 6 September
The physical layer of AI, every day. Here's what moved on Sunday 6 September 2026.
In this issue
Top story · SOFTWARE & AI
China · ▼ Bearish · 6 Sep 10:45 UTC · Trade press
Chinese open-source models including GLM 5.2, K3, and Qwen 3.8 Max reached performance parity with Anthropic's Opus 5 within two months of its release, with Qwen 3.8 27B running on consumer GPUs. That last detail is the one GPU infrastructure operators should sit with: workloads that previously required H100 clusters are now deployable on commodity hardware, which compresses per-token costs and squeezes margins for premium inference services built around high-end GPU density. Watch for enterprise procurement to shift toward smaller-parameter open models over the next two quarters, and for neocloud utilisation rates on H100 and H200 capacity to reflect it.
NEOCLOUD · 1
US · ▲ Bullish · 4 Sep 07:05 UTC · Company/PR
Packet.ai is offering NVIDIA B200 GPUs at $3.75 per GPU-hour with no waitlist restrictions, providing immediate access to Blackwell inference capacity in the spot market.
Why this matters: Establishes a reference price for B200 inference capacity and availability of Blackwell in the neocloud market, enabling buyers to benchmark inference-per-dollar against H200 and older generations.
HARDWARE · 4
Taiwan · ▲ Bullish · 6 Sep 04:15 UTC · Trade press
I-Chiun, a thermal management vendor, is targeting NVIDIA's Vera Rubin GPU supply chain as cooling systems approach 50% of company revenue.
Why this matters: Thermal management becoming a critical bottleneck and revenue driver for GPU infrastructure, with vendors competing for OEM supply contracts.
Taiwan · ▲ Bullish · 6 Sep 00:00 UTC · Trade press
Hanmi's 2.5D bonding equipment has passed qualification at a major Taiwan foundry, expanding packaging outsourcing options for GPU and accelerator manufacturers.
Why this matters: Diversifies packaging capacity beyond traditional suppliers and reduces bottlenecks in GPU production as demand for advanced packaging (HBM integration, chiplet assembly) accelerates.
Global · ▲ Bullish · 3 Sep 11:21 UTC · Company/PR
Marvell's Teralynx 10 delivers 51.2Tbps throughput with 64 ports of 800GbE and 2kW total power consumption, expected to ship in AI clusters during 2025.
Why this matters: Enables higher-bandwidth cluster interconnect at lower power than prior generation, reducing networking bottlenecks in large-scale GPU training and inference deployments.
Global · ▲ Bullish · 1 Sep 12:00 UTC · Company/PR
Kulicke & Soffa developed fluxless thermo-compression bonding products for co-packaged optics (CPO) assembly, as a supplier to the semiconductor packaging layer of AI compute infrastructure.
Why this matters: CPO adoption reduces optical module latency and power in high-bandwidth AI clusters; K&S's assembly technology becomes critical as the industry transitions from pluggable to integrated silicon photonics.
DATA CENTRE · 2
Global · ▲ Bullish · 6 Sep 07:30 UTC · Newsroom
Google and TCS are collaborating on air-cooling optimisation techniques for data centre thermal management at scale.
Why this matters: Advances in air-cooling efficiency could reduce capex and opex for data centre operators, particularly in regions where liquid cooling infrastructure is immature or costly.
US · ◆ Neutral · 3 Sep 11:23 UTC · Company/PR
ColoCrossing, which operates 10 data centre locations nationwide, will exhibit at CloudFest USA 2025 on 5-6 November in Miami.
REGULATION & POLICY · 2
US · ▼ Bearish · 6 Sep 07:15 UTC · Newsroom
US administration is escalating enforcement and rhetoric around H-1B and foreign worker visas, targeting companies across sectors.
Why this matters: Could constrain hiring at GPU infrastructure companies and hyperscalers that rely on international talent for chip design, data centre operations, and AI engineering roles.
US · ◆ Neutral · 6 Sep 20:47 UTC · Trade press
Authors are contesting publisher and agent claims on the $1.5 billion Anthropic copyright settlement, disputing distribution of proceeds.
Why this matters: Ongoing settlement disputes could establish precedent for AI training data liability costs and may influence future model-training agreements.