LIVE · WED, JUN 24, 2026 --:--:-- ET
Issue Nº 64 COST TOTAL $14508.23 ARTICLES TODAY 12 TOKENS TOTAL 9.10B
Podcast
aiexpert
Subscribe · $8/mo
Running the wire
Research Google OpenRL: Self-hosted Kubernetes API for LLM post-training; decouples RL from infrastructure Market Micron Q3 earnings beat on record DRAM margins; HBM supply fully allocated through 2026 Policy US secures Netherlands for Pax Silica chip alliance; ASML tensions persist over MATCH Act export restrictions Chips OpenAI & Broadcom unveil Jalapeño: Custom LLM inference chip targets gigawatt-scale deployment by end of 2026 Breaking Gemini 3.5 Flash adds native computer use; agent framework now default across Search Research AI rapidly designs novel radio-frequency chips beyond human intuition, reducing years of work to hours Chips China's LineShine supercomputer tops TOP500 with 2.198 exaflops CPU-only, ending US El Capitan's reign Market Cerebras stock plummets 17% after margin-guidance miss as CEO says warning was 'misunderstood' Market Sunrun, Tesla, Renew Home form 16GW virtual power plant for AI data centers; RUN +31% Breaking Amazon Zoox unveils redesigned robotaxi, planning paid service launch in late 2026 Funding XCures closes $46M Series B round at $127M post-money valuation Funding Qualcomm acquires Modular for ~$4B to bolster AI software stack and data center play Chips OpenAI & Broadcom unveil Jalapeño, custom LLM inference chip with 9-month design cycle Chips SK Hynix ships HBM4E memory samples: 16Gbps, 48GB per stack, 20% power gain Funding Qualcomm in talks to acquire Tenstorrent for $8–10B, expanding RISC-V AI chip portfolio Chips TSMC hikes advanced node prices 5–10% across 7nm and newer nodes Chips OpenAI unveils Jalapeño custom inference chip with Broadcom Chips OpenAI-Broadcom custom chip project stalls; Broadcom demands Microsoft purchase guarantee before funding Market Cerebras Q1 profit beats but warns gross margins compress to 36–38% in Q2 Funding SK hynix files for $29.4B Nasdaq listing to fund AI memory fabs and EUV tools Research Google OpenRL: Self-hosted Kubernetes API for LLM post-training; decouples RL from infrastructure Market Micron Q3 earnings beat on record DRAM margins; HBM supply fully allocated through 2026 Policy US secures Netherlands for Pax Silica chip alliance; ASML tensions persist over MATCH Act export restrictions Chips OpenAI & Broadcom unveil Jalapeño: Custom LLM inference chip targets gigawatt-scale deployment by end of 2026 Breaking Gemini 3.5 Flash adds native computer use; agent framework now default across Search Research AI rapidly designs novel radio-frequency chips beyond human intuition, reducing years of work to hours Chips China's LineShine supercomputer tops TOP500 with 2.198 exaflops CPU-only, ending US El Capitan's reign Market Cerebras stock plummets 17% after margin-guidance miss as CEO says warning was 'misunderstood' Market Sunrun, Tesla, Renew Home form 16GW virtual power plant for AI data centers; RUN +31% Breaking Amazon Zoox unveils redesigned robotaxi, planning paid service launch in late 2026 Funding XCures closes $46M Series B round at $127M post-money valuation Funding Qualcomm acquires Modular for ~$4B to bolster AI software stack and data center play Chips OpenAI & Broadcom unveil Jalapeño, custom LLM inference chip with 9-month design cycle Chips SK Hynix ships HBM4E memory samples: 16Gbps, 48GB per stack, 20% power gain Funding Qualcomm in talks to acquire Tenstorrent for $8–10B, expanding RISC-V AI chip portfolio Chips TSMC hikes advanced node prices 5–10% across 7nm and newer nodes Chips OpenAI unveils Jalapeño custom inference chip with Broadcom Chips OpenAI-Broadcom custom chip project stalls; Broadcom demands Microsoft purchase guarantee before funding Market Cerebras Q1 profit beats but warns gross margins compress to 36–38% in Q2 Funding SK hynix files for $29.4B Nasdaq listing to fund AI memory fabs and EUV tools
Chips

China's LineShine supercomputer tops TOP500 with 2.198 exaflops CPU-only, ending US El Capitan's reign

China's LineShine supercomputer has taken the #1 spot on the TOP500 list with 2.198 exaflops of double-precision Linpack performance, pushing AMD-powered El Capitan (1.809 exaflops) into second place. Critically, LineShine achieved this using no GPUs or accelerators of any kind—only 13.78 million cores of domestically designed Armv9-based silicon (the LX2 processor), the first CPU-only machine to clear the two-exaflop threshold. The system uses SMIC's 7nm-class domestic process, proprietary LingQi interconnects, and the Kylin OS, representing an all-domestic technology stack.

This is the first China-based system to lead the TOP500 since Sunway TaihuLight in 2017. More tellingly, China stopped submitting its fastest machines to the list around 2021 after entity-list sanctions; the decision to submit LineShine now signals a deliberate geopolitical posture change. Observers believe China has operated undisclosed exascale systems (OceanLight, Tianhe-3) for years without submitting. LineShine's disclosure is a statement that indigenous design and fabrication can work without Western components, breaking dependency on TSMC, EDA, and export controls.

However, LineShine dominates only on high-precision FP64 workloads. On the mixed-precision HPL-MxP benchmark that approximates AI training math, it scored only 22% of El Capitan's relative performance gain, ranking fourth at 7.92 exaflops versus El Capitan's 16.7 exaflops. LineShine also consumes 42,220 kW (pulling 42% more power than El Capitan for similar-magnitude FP64 output), indicating that the architecture favors traditional HPC over AI-scale matrix math.

For infrastructure teams: LineShine proves China can build sovereign exascale systems but raises questions about frontier-AI workloads. The discrepancy between FP64 and FP32/BF16 performance highlights why GPUs and specialized accelerators still rule training and inference. This is less a threat to NVIDIA's AI dominance than a statement about HPC independence. Watch whether China's next systems prioritize reduced-precision throughput, and whether sanctions-evasion via custom silicon becomes table stakes for other frontier players.

Sources

Get the signal before the noise.