Global GPU Computing News【20260419】

Global Compute Shortage Worsens, H100 Rental Prices Surge Nearly 40% in Six Months
Latest data from SemiAnalysis shows that the one-year rental contract price for an H100 has risen from USD 1.70 per GPU hour in October 2025 to USD 2.35 in March 2026, an increase of nearly 40%. Meanwhile, on-demand compute is sold out across all GPU models, and the AI supply chain from memory to optical fiber is fully strained, with no signs of relief in sight.

TSMC Sends Clear Signal: AI Compute Shortage Will Last Until 2027
During its Q1 earnings call, TSMC pushed its full-year 2026 capital budget to the upper limit of USD 56 billion. Chairman and CEO C.C. Wei clearly stated that AI demand is "extremely strong," and new fabs will take at least 2-3 years to build. The supply tightness is expected to last until at least 2027. At the same time, deployment lead times for Blackwell clusters have been extended to June-July 2026.

Nvidia Denies GPU Auction Sales, But Even Internal Teams "Can't Get" Compute
In response to market rumors that Nvidia might adopt an auction mechanism to allocate GPUs, CEO Jensen Huang firmly denied it in a public dialogue, emphasizing that the company only sets fair and transparent market prices and never interferes with customer choices. Ironically, the vice president of Nvidia's applied deep learning research division revealed that even the company's internal research teams face severe GPU shortages and must constantly compete for quotas, which in turn forces them to continuously improve model efficiency.

Nvidia Rubin GPU Delayed Again, Blackwell Remains Mainstream Shipment This Year
According to the latest report from TrendForce, the expected shipment share of the Rubin GPU in 2026 has been significantly reduced from 29% to 22%. The Blackwell platform is still expected to account for over 70% of high-end GPU shipments this year. The main challenges causing the delay include HBM4 validation, network interconnect upgrades, and liquid cooling solution optimization.

Breakthrough in Domestic Compute Ecosystem: DeepSeek V4 Fully Runs on Huawei Ascend 950PR
This week, DeepSeek V4 announced that it will fully run on the Huawei Ascend 950PR chip, with underlying code completely migrated from CUDA to the CANN Next framework, marking a substantial step forward for the domestic compute ecosystem. The chip delivers 1.56 PFlops at FP4 precision, 2.87 times higher than Nvidia H20, and has already received orders for hundreds of thousands of units from giants like Alibaba, ByteDance, and Tencent.

Domestic Cloud Providers Raise Prices Collectively, Compute Transforms from "Dirt Cheap" to Scarce "Hard Currency"
Since mid-March, major domestic cloud providers have intensively raised AI service prices. Tencent Cloud announced a uniform 5% price increase on AI compute and other products effective May 9. Alibaba Cloud issued three price hike notices within just four days, with increases ranging from 2% to 7%. At the same time, Zhipu AI has raised its API prices three times in two months. Industry analysis points to the explosive growth in inference token calls as the core driver of this full-chain price surge.

Meta Extends Custom AI Chip Partnership with Broadcom to 2029, Accelerating Efforts to Reduce Dependence on Nvidia
Meta and Broadcom announced the extension of their multi-generation custom AI chip production collaboration to 2029. The agreement includes an initial commitment of over one gigawatt (GW) of computing power capacity, enough to serve 750,000 U.S. households. This move continues the trend of major tech companies developing their own chips to reduce reliance on Nvidia GPUs.

Wuxi Launches 2.5 Billion Yuan Domestic Compute Cluster Project, Creating a New AI Computing Benchmark in the Yangtze River Delta
The approximately 2.5 billion yuan M-Xing Equity Yangtze River Delta Ecological Innovation Center and Domestic Compute Cluster Project was officially signed and launched in Wuxi. The project is operated by Cloud Workshop Technology and will partner with M-Xing Equity to build a ten-thousand-card heterogeneous intelligent computing cluster, supporting multi-card collaboration, training-inference integration, and providing large-scale computing services for industries such as AI, manufacturing, finance, and healthcare.

Tesla AI5 Chip Officially Taped Out, Edge Computing Competition Enters a New Phase
Elon Musk announced that the Tesla AI5 chip has officially taped out, with AI6 and Dojo3 also under simultaneous development. Industry analysis suggests its inference performance is comparable to Nvidia H100, and a dual-chip configuration approaches Blackwell. Although it will take 12 to 18 months from tape-out to mass production in vehicles, this event marks the edge AI computing race spreading from the cloud to vehicles and end devices.

Nvidia Releases World's First Open-Source Quantum AI Model Ising, Internal AI Completes 80 Person-Month Chip Design Overnight
On World Quantum Day, Nvidia released Ising, the world's first open-source AI model family for quantum computing. Its error correction decoding speed is 2.5 times faster and accuracy 3 times higher than industry standards, causing quantum stocks like IonQ and D-Wave to surge collectively that day. Meanwhile, the company's chief scientist revealed that AI has reduced the porting of a chip's standard cell library from 8 people working 10 months to a single GPU running overnight, with design results even better than manual work.