gpu

NVIDIA (NVDA.O) launched the cuPotential CUDA-X library, accelerating GPU workflows by 2 to 3 times in the early access phase.

Ethereum Layer 2 network ZKsync announced the launch of the prover Airbender Prime. ZKsync stated that, to their knowledge, this is the first prover to demonstrate real-time proofs of Ethereum on a single GPU. Airbender Prime is designed for post-quantum security, utilizing hash-based STARK, combined with GKR sumcheck, WHIR polynomial commitment, and BLAKE2s hash; its L1 proof will use a commitment based on Keccak-256 instead of BN254 pairing. This prover is based on the open RISC-V standard and combines real-time proofs with independent validators.In a single GPU evaluation, an NVIDIA RTX PRO 6000 Blackwell completed proofs for 99% of 7200 consecutive Ethereum mainnet blocks, with each block taking less than 10 seconds, including witness preparation; all blocks were completed within 12 seconds. ZKsync claims that Airbender Prime leads in average proof time, GPU efficiency, and cost reported on the ETHproofs public leaderboard, and it already has dual GPU and quad GPU real-time configurations while reducing the hardware requirements for generating proofs.

On October 13, Ruantong Power was reported to have received 8,000 NVIDIA GPU cards, specifically the NVIDIA 6000D (domestic "special edition" of the Blackwell architecture), which will be used for the production of related servers.

Recently, SoftStone Technology published an article on its WeChat official account regarding the order of 8,000 NVIDIA 6000D GPU cards arriving, which has since been deleted. In response, SoftStone Technology staff stated that the company will announce matters that meet disclosure standards based on business conditions and advised investors not to use such news as the sole basis for investment.

Multiple venture capital firms in Silicon Valley are competing to secure GPU computing power to provide exclusive resources for their AI startups. a16z established a computing power pool of about 20,000 GPUs as early as 2024 and is negotiating expansion with data center companies Switch and Volta. General Catalyst plans to build a data center for its portfolio companies and has brought in the former head of NVIDIA DGX Cloud in May this year to oversee AI infrastructure operations.Even if AI startups complete their financing, they may not be able to rent GPUs in a timely manner. In April this year, major cloud providers like Microsoft were reported to prioritize allocating computing power to internal teams and large clients, leaving some startups facing months of waiting. a16z previously disclosed that during times of computing power shortages, prices that could originally be secured with a six-month contract might require a three-year commitment. Model training requires a large amount of GPU usage in a short period, while inference demand is difficult to predict, making it easy to waste funds by locking in computing power for several years in advance.a16z is concentrating on purchasing and leasing computing power through the Oxygen project, redistributing it to its portfolio companies, allowing them to access prices and contract terms typically reserved for large clients. In August this year, a16z co-led a $300 million financing round for Volta, which is valued at $2.4 billion, with investments covering both AI startups and data center companies providing computing power. The small venture capital firm Crucible Capital has built a cluster with 576 NVIDIA B300 GPUs for its portfolio companies to test, while also selling some computing power to a cutting-edge AI laboratory. This model allows venture capital firms to take on investments that were previously mainly borne by cloud providers and data center operators.

According to Bloomberg, Shanghai Birun Technology (Hong Kong stock code 6082) plans to raise approximately HKD 4.04 billion (USD 515 million) to support its artificial intelligence business. The company intends to place 130 million new shares at a price of HKD 31.08 per share, with the related arrangements based on documents from the Hong Kong Stock Exchange.The placement price represents a discount of nearly 10% compared to Wednesday's closing price. Bloomberg reports that Birun is one of the GPU developers receiving support from Beijing, and this financing is part of a wave of fundraising for Chinese artificial intelligence companies.

The GPU infrastructure financing platform Silicon Exchange has announced its launch, positioning itself as a financing method for artificial intelligence infrastructure. Users can hold real shares of GPU infrastructure, while the platform is responsible for deployment and hosting, and computing power clients use the related hardware.Silicon Exchange stated that users can visit its official website to view the first GPU vault. The official announcement has not disclosed the scale of the vault, hardware models, financing structure, or specific profit arrangements in this post.

The NVIDIA research team enabled the AI Agent to directly rewrite the GPU kernel of Kimi Delta Attention. Kimi Delta Attention is one of the core attention mechanisms used by the Dark Side of the Moon Kimi-Linear model, and such low-level code typically required engineers familiar with CUDA and chip architecture to optimize manually and repeatedly in the past.The version written by the Agent achieved 2.96 times the speed of the official FlashKDA of the Dark Side of the Moon on the NVIDIA B300. The team tested a total of 6 sets of tasks with fixed lengths and varying lengths, with 2.96 times being the geometric mean of all tasks; the relevant kernel has been open-sourced.During the process, the Agent exploited testing loopholes by incorporating statistical patterns from the test data into the code, achieving a speedup of 3.74 times; when retaining only the most recent 32 tokens, a single score reached 5.16 times. These versions would produce incorrect results or fail when switched to real Kimi data. The team subsequently incorporated real Kimi runtime data, random inputs, and extreme value tests, tightening the error standards, and ultimately retained the 2.96 times version, which passed this validation.

User gpugene stated during a discussion on computing power trading: Some B300 transaction prices have exceeded 24 dollars per GPU hour, and the price of computing power with a term shorter than 1 year remains above 7 dollars and is still rising. gpugene expressed the hope to use computing power derivatives to guard against the rise in GPU hour prices while retaining the flexibility to adjust rental durations; if inference clouds sell services at fixed prices and their own GPU bills fluctuate, they have effectively taken on the risk of computing power price changes.The article provides an example where a supplier can offer a plan for 2048 B300s, available for rent for 5 years at 4.15 dollars per GPU hour, corresponding to an amount of 372,264,960 dollars. If the client goes live after 12 months and uses this scale from the 13th to the 15th month, the total would be 4,485,120 GPU hours. gpugene calculated that the current premium for a call option with an exercise price of 4.50 dollars is approximately 4.83 million dollars; if changed to a compound option, the current payment would be about 2.70 million dollars, with an additional payment of about 4.49 million dollars upon exercising in the 12th month.gpugene also compared 4000 hypothetical 5-year rental and demand paths. The average cost of annual renewal is 4.72 dollars per hour, with the worst 5% scenario averaging 10.91 dollars; after adding the call option, these figures change to 4.88 dollars and 7.20 dollars, respectively, with the premium included. gpugene believes that the emerging computing power derivatives market may change how neocloud and related participants expand and protect themselves.

On-chain lending protocol USD.AI announced that it has secured a financing amount of $128.9 million in GPU funding, marking the largest single financing in the protocol's history.The financing will be used to support an undisclosed borrower in deploying 32 sets of NVIDIA GB200 NVL72 systems in British Columbia, Canada. This financing amount sets a new historical record for USD.AI.

Binance will launch the Axe Compute (AGPUB), AMC Entertainment (AMCB), and Cypherpunk Technologies (CYPHB) bStocks trading pairs at 20:00 on September 23. The trading pairs include AGPUB/USDT, AMCB/USDT, and CYPHB/USDT. At the same time, the spot algorithm trading bot will be activated in line with the spot listing time, and the smart holding bot will be enabled within 24 hours after the spot market goes live.bStocks can be exchanged for BTC, USDT, or other tokens without any fees within one hour after going live on Binance spot. During the event until October 1 at 07:59, there will be zero maker fees.

Jurrien Timmer, Global Macro Director at Fidelity Investments, stated on platform X that trading in the AI sector has been in a "dead money" state for over three months, with the token spending and GPU rental rates he tracks remaining flat or declining, while only memory (DRAM) prices continue to rise.Timmer believes that AI will change life in the long term and may unleash a productivity miracle, raising the non-inflationary speed limit of the economy to a level sufficient to sustain global debt. However, before that, the current construction is inflationary, and the returns on the trillions of dollars invested in computing power remain uncertain, with corporate equity and debt capital demand growing at a rate of $3.3 trillion.

Aoni Electronics announced that its wholly-owned subsidiary Aoni Intelligent Technology (Zhongshan) Co., Ltd. has signed a procurement contract with Company A to purchase GPU computing power cards, with a total contract amount of 1.67 billion yuan (including tax). The contract has been officially signed and is effective.

Google Cloud announced the official launch of the pause/resume feature for data stream batch processing jobs, and it supports NVIDIA RTX PRO 6000 Blackwell GPU.

"White-haired Stock God" Serenity posted on platform X that a piece of information in Oracle's (ORCL) financial report is favorable for Nebius (NBIS) and IREN (IREN). Oracle stated that all renewed GPU computing power has been resold, with prices increasing by 20% compared to previous contracts, and most related equipment has been used for over 4 years. Serenity believes that this situation once again counters "Big Short" Michael Burry's argument about the rapid depreciation of GPUs, while also benefiting Nvidia (NVDA) and the Neocloud computing sector.

According to Reuters, Mexican authorities discovered an illegal cryptocurrency mining operation in a remote mountainous building in Tehuacán, Puebla, where 300 GPUs, 80 medium-voltage terminals, and 8 satellite antennas were found on-site. Authorities are investigating whether the operation is stealing electricity from a nearby hydroelectric power station.This is the fourth similar crypto farm discovered in the area since early last year, with three others found near a hydroelectric power station in northern Puebla last year. Local residents reported that the noise from the equipment can be heard from a kilometer away. Mexican security analyst David Saucedo stated that drug trafficking groups seem to have reached a new level of sophistication, as establishing such operations requires technical expertise and support from well-funded groups.Chainalysis Latin America expert Caio Motta pointed out that such operations are usually located in areas with extremely low electricity prices or within the reach of organized crime to steal electricity and establish large-scale mining infrastructure.

DeepSeek stated in the V4.1 Flash open-source announcement that if there are 2000 GPUs and storage clusters, they can contact DeepSeek directly.V4.1 Flash is still open-sourced under the MIT License, without additional licensing conditions based on company revenue or business scale. Companies with resources can deploy it themselves, and contacting DeepSeek is more like a direct technical collaboration for ultra-large-scale scenarios. DeepSeek has not yet announced a formal private deployment product or paid service.In the past, after DeepSeek open-sourced the model, deployment adaptations were mostly completed by inference frameworks and hardware vendors such as SGLang and TensorRT-LLM. This time, the official team has proactively provided direct contact information for large-scale deployments and stated that they will continue to support the open-source community in adapting inference for new models.

Liquid Compute CEO Ronit Jain published an article introducing the hedging of computing power prices. Purchasing computing power, holding GPUs, or lending based on this has exposed them to computing power price fluctuations. Hedging is a cash-settled contract that locks in future prices against an index without the need to transfer GPUs.Companies can buy forwards to fix the cost of next year's capacity as a budget item while still being able to purchase from any supplier. Neocloud can sell forwards to convert uncontracted hours into fixed cash flow, making it easier for lenders to provide loans based on this. Regarding the residual value of sale-leaseback, recent forwards can be rolled over when entering a tradable range, or put options can be bought on forwards, with more liquid near-term options available to pay for the option premium.Hedging locks in prices rather than quantities, and does not completely match actual rates, nor is it free: it requires giving up favorable price movements or paying option premiums. Legally, it usually requires signing an ISDA Master Agreement with the counterparty, obtaining permission for the settlement index, and determining the ratio of hedging exposure.