Kyutai
Oct 24, 2026

Kyutai unveils CADRE, saying it cuts world-model action latency to under 230 ms

French AI research lab Kyutai has introduced CADRE, a family of open causal diffusion representation autoencoders built for interactive world models. The release is aimed at reducing action latency, a key constraint for systems that need to react quickly to user input or changing environments. According to Kyutai, its world model runs at 20 fps on a single NVIDIA H100 GPU and delivers response times below 230 milliseconds. The model family includes seven autoencoders, each with about 476 million trainable parameters, and uses a frozen 86 million-parameter DINOv3 ViT-B/16 model. Kyutai said the full world model totals 1.2 billion parameters. In benchmark comparisons against MIRA, Wan 2.2, Cosmos, and LTX-2, CADRE outperformed rivals in 42 of 48 tests. In a driving-scene FVD result, CADRE-2×16×16 with single-step decoding posted 162.6, compared with 599.7 for Wan 2.2. The system was trained on a mixed dataset spanning 23 public video sources and roughly 76,000 hours of data, including web video, robotics, driving, and gameplay footage.

2130
Kyutai unveils CADRE, saying it cuts world-model action latency to under 230 ms
Pearl
Oct 20, 2026

Pearl pitches a market-based way to price GPU compute as public indexes struggle with H100 valuations

GPU pricing remains hard to pin down even for the same NVIDIA H100, with listed rates varying sharply depending on uptime guarantees, networking, cooling, and the broader compute environment wrapped around the chip. The article argues that this makes standardized pricing and physical-delivery futures difficult, because the underlying product is not a uniform commodity. It points to examples such as CoreWeave borrowing against similar H100 hardware at very different rates depending on whether long-term customer contracts were attached, and notes that public cloud rate cards can sit 40% to 50% above what large buyers actually negotiate. Against that backdrop, Pearl is presented as a Layer 1 blockchain that replaces SHA-256 hashing with GPU-based matrix multiplication while keeping a Bitcoin-style proof-of-work and difficulty adjustment framework. Using a mechanism called NoisyGEMM, the network is designed to prevent miners from shortcutting the work while still making the computation commercially relevant. The piece says 320,000 GPUs are pointed at Pearl, and that their movement between mining and external rental markets creates a continuously updated signal for the opportunity cost of compute. It also cites Together AI’s May 15, 2026 launch of an inference endpoint on Pearl, priced more than 25% below standard cloud rates for Gemma-4-31B-it-Pearl.

1040
Pearl pitches a market-based way to price GPU compute as public indexes struggle with H100 valuations
Pearl
Oct 12, 2026

Pearl’s 320,000-GPU network is being framed as a live market for AI compute pricing

A market for GPU compute worth hundreds of billions of dollars still lacks a pricing benchmark that reflects what large buyers actually pay. In an analysis published by TechFlowPost and attributed to Decentralisedco, Pearl, a Layer 1 blockchain built around proof-of-work, is presented as an alternative way to discover the price of AI compute through real-time market behavior rather than posted cloud rate cards. The article argues that public GPU indexes from Silicon Data, Ornn, and Compute Desk struggle to price H100 rentals because the same chip can represent very different products depending on networking, uptime guarantees, cooling, and facility quality. It also notes that CME’s planned GPU futures, set for Oct. 5, 2026, use cash settlement against a Silicon Data survey index based on public list prices, even though large buyers often negotiate discounts of 40% to 50%. Pearl’s model replaces SHA-256 hashing with matrix multiplication on GPUs and uses a mechanism called NoisyGEMM to keep miners from shortcutting the work. According to the piece, the network measures how much useful matrix computation GPUs complete every 194 seconds under real operating conditions. With 320,000 GPUs participating across multiple regions, the resulting hashrate is described as a continuously updated signal of the opportunity cost of compute, shaped by whether operators choose to mine Pearl or rent capacity to paying AI customers.

1910
Pearl’s 320,000-GPU network is being framed as a live market for AI compute pricing
a16z
Oct 11, 2026

a16z says falling AI prices are fueling demand faster than costs decline

A report cited by Techub and attributed to The Decoder says the AI market is showing signs of the Jevons paradox. According to a16z, prices for AI tokens have kept falling, while rental prices for NVIDIA H100 GPUs have remained stable or continued to rise. The report argues that cheaper AI is driving usage and demand at a pace that exceeds the rate of cost declines. In other words, lower prices are not easing pressure across the stack. Instead, demand is expanding quickly enough to keep key infrastructure pricing firm, at least in the case of H100 rentals. a16z also warns that if demand starts to flatten, the impact would not be limited to a single segment. The effect could spread across the industry chain, from chip manufacturers to cloud providers. No additional figures were provided in the summary cited by Techub.

1030
a16z says falling AI prices are fueling demand faster than costs decline
AI compute
Oct 11, 2026

Silicon Valley’s AI compute race heats up as SpaceX resells GPUs to Anthropic

Competition for AI compute in Silicon Valley is getting tighter, according to a Techub report citing Quantum Bit. The report says SpaceX has been acting as a reseller of Nvidia GPUs, selling compute capacity to Anthropic. In its first financial report released in August, SpaceX posted quarterly AI business revenue of $2.561 billion, up 247% from a year earlier and more than doubling from the previous quarter. The supply crunch is also affecting hiring and startup operations. Google reportedly lost top scientist Noam Shazeer to OpenAI because of insufficient compute resources. Startups, meanwhile, are facing demands from cloud providers for multi-year contracts and large upfront payments in order to secure access. Data from SemiAnalysis cited in the report shows that the one-year lease price for H100 GPUs rose from $1.70 per GPU per hour in October 2025 to $2.35 in March 2026, an increase of nearly 40%. The report also says Silicon Valley venture capital firms including Y Combinator and Radical Ventures have started stockpiling GPUs themselves or helping portfolio companies negotiate supplier contracts.

750
Silicon Valley’s AI compute race heats up as SpaceX resells GPUs to Anthropic
OKX
Oct 2, 2026

OKX to list H100, ACN, NKE, BWET and SECZ stock perpetuals on Oct. 2

OKX said in an official announcement that it will begin listing five stock perpetual contracts on Oct. 2 across its web platform, app, and API. The rollout starts at 17:30 (UTC+8) with the H100/USDT contract, followed by ACN/USDT at 17:45, NKE/USDT at 18:00, BWET/USDT at 18:15, and SECZ/USDT at 18:30. All five contracts will use USDT as the settlement currency. OKX said the products will trade 24/7, and the funding fee settlement interval will be every eight hours. The schedule and contract details were disclosed in the exchange’s official notice cited by BlockBeats.

1220
OKX to list H100, ACN, NKE, BWET and SECZ stock perpetuals on Oct. 2
Nvidia
Sep 8, 2026

H100 Rental Prices Jump 22% in a Month; Jensen Huang Calls GPUs Income-Generating Assets

Nvidia CEO Jensen Huang has highlighted data from compute-market platform Ornn showing H100 rental rates climbing 22% over the past month, to $3.28 per GPU per hour. The H100 is a training GPU that has been on the market for years, so the jump stands out. Ornn's H100 index is based on actual rental transaction prices. The platform's daily settlement on Sept. 7 was $3.17 per hour, and its most recent data display put the price at $3.28. Sharing the numbers, Huang argued that Nvidia's compute capacity can be rented out continuously and generate income. The rise supports the compute durability argument he has repeatedly stressed. Nvidia's latest earnings call made a related point: GPUs can be redeployed across different customers and workloads. For Huang, the rental price data is evidence that GPUs are not just hardware but assets that can keep producing revenue.

1980
H100 Rental Prices Jump 22% in a Month; Jensen Huang Calls GPUs Income-Generating Assets
Total 25 items
Go topage