# Пов'язані статті щодо AI Inference

Центр новин HTX надає останні статті та поглиблений аналіз на тему "AI Inference", що охоплює ринкові тренди, оновлення проєктів, технологічні розробки та регуляторну політику в криптоіндустрії.

From Auto Finance to Bitcoin to AI Engines: An Analysis of Cango's 'What Not to Do' Strategy

From Auto Finance to Bitcoin and Now AI: Cango's "What Not to Do" Strategy Cango, a Chinese auto finance platform that went public on the NYSE in 2018, is undergoing its third major transformation. After selling its entire auto business in 2024, it pivoted to become a large-scale Bitcoin miner, acquiring 50 exahash of mining rigs from Bitmain. However, its true goal was never Bitcoin, but owning and controlling energy infrastructure. Now, Cango is pivoting again. While most listed Bitcoin miners are leasing power to giant hyperscalers for AI training clusters, Cango is taking the opposite path. It has launched an AI inference subsidiary called EcoHash, focusing not on training but on distributed inference. The company's strategy hinges on the insight that over 70% of mining industry power is controlled by small, independent sites (10-50 MW), which are too small for hyperscalers but ideal for low-latency AI inference. Cango aims to partner with these small operators, providing the AI technology, customers, and financing through its EcoLink software layer, which can distribute workloads across sites for reliability. Cango maintains a hybrid model, running roughly 31.7 EH/s of Bitcoin mining for cash flow while aggressively cleaning its balance sheet—slashing long-term debt by 94.5% to $30.6 million and raising $75 million for its AI venture. Its first AI deployment will be at a 50 MW site in Georgia. The strategy faces skepticism, given the high costs of converting mining sites and the potential for an AI bubble. However, Cango's leadership believes discipline around "what not to do"—avoiding direct competition with hyperscalers in training—positions it to capture the long-tail demand for distributed AI inference power.

Foresight News07/11 07:56

From Auto Finance to Bitcoin to AI Engines: An Analysis of Cango's 'What Not to Do' Strategy

Foresight News07/11 07:56

Computing Power Crisis: Google Quietly Imposes Usage Caps on Meta for Gemini

Google has quietly imposed usage caps on Meta's access to its Gemini AI models since around March due to surging demand overwhelming its computational infrastructure, according to a Financial Times report. The limits, which remain in place, have disrupted and delayed several of Meta's internal AI projects, forcing the social media giant to ration AI usage and improve efficiency. This reflects a broader industry-wide shortage of AI inference capacity, as companies deploy more chatbots and AI agents. Google CEO Sundar Pichai acknowledged compute constraints are limiting cloud revenue growth. In response, Google recently signed a $920 million monthly compute leasing deal with SpaceX to expand capacity. The restrictions have accelerated Meta's shift toward its own AI models, such as Muse Spark, to reduce dependence on external providers like Google. While other Google clients also face limits, Meta's vast scale made it particularly affected. The situation highlights how the AI infrastructure bottleneck has shifted from model training to inference, requiring massive new capital investments to resolve.

marsbit06/29 01:02

Computing Power Crisis: Google Quietly Imposes Usage Caps on Meta for Gemini

marsbit06/29 01:02

The Domestic Answer to Space Computing Power: Photonics Are More Efficient, Musk and Huang's Approaches Are Too Roundabout

The Space Computing Race: A Photonic Advantage The competition for space-based computing has intensified, with figures like Elon Musk and NVIDIA's Jensen Huang highlighting its potential. Musk predicts solar-powered AI satellites could offer the most cost-effective computing by 2032. However, space presents extreme challenges for traditional electronic chips: radiation from cosmic particles can cause errors, the vacuum environment hinders heat dissipation, and limited solar power constrains energy-hungry systems. Photonic computing, using light instead of electrons, offers a promising solution. Its core advantages for space are threefold: 1) **Radiation Resistance**: Photons are charge-neutral, making them inherently immune to particle interference. 2) **Low Heat Generation**: Light propagation in waveguides generates minimal heat, bypassing critical thermal management issues. 3) **Low Power Consumption**: Photonic chips have near-zero static power draw, aligning perfectly with the energy constraints of satellites. Furthermore, for a given payload weight and volume, photonic systems can potentially deliver higher total compute density. Since they require less bulky cooling and power infrastructure, more space can be allocated to the compute units themselves. While photonic computing holds great promise, current industry approaches face hurdles like the memory-compute bottleneck (separate storage and processing) and challenges in large-scale integration. Engineering for space—withstanding launch vibrations and validating full system operation in orbit—remains a critical step. The path forward resembles the evolution from single GPUs to computing clusters, but via a photonic route. As electronic chips approach physical limits in miniaturization, photonic computing and optical interconnects (光算光联) may provide a key alternative to bypass these constraints and define the next generation of space-based computing capabilities.

marsbit06/28 04:31

The Domestic Answer to Space Computing Power: Photonics Are More Efficient, Musk and Huang's Approaches Are Too Roundabout

marsbit06/28 04:31

Comprehensive Analysis of the AI Inference Market: How Can Crypto Projects Break Through?

"AI Inference Market: A Strategic Overview and Crypto's Path to Disruption" The AI inference market, where trained models generate responses to user prompts, is now the primary economic driver, surpassing model training in value. This market is fragmented: hyperscalers (AWS, Google, Microsoft) dominate enterprise reliability; specialized providers (Together, Fireworks) optimize performance; and routing platforms like OpenRouter act as critical bottlenecks, dynamically allocating requests based on cost, latency, and privacy. Crypto AI networks are not competing directly on reliability but are carving out distinct niches: permissionless access, lower-cost supply, privacy, verifiable computation, and agent-native payments. Key projects include Chutes (decentralized inference platform), Akash & io.net (GPU marketplaces), Targon (confidential computing), Darkbloom & Venice (private, consumer-focused inference), and NuNet (orchestration for distributed workloads). The core differentiator is that traditional providers sell trust and enterprise workflows, while crypto networks offer new incentive loops, censorship resistance, and programmable access to resources like compute. For crypto projects to succeed, key metrics are paid token volume (not just usage), sustainable GPU provider revenue, integration into routers like OpenRouter, robust verification against fraud, and genuine privacy guarantees. Ultimately, market control will belong to entities that route, verify, and settle demand—not just those supplying raw compute. The inference market is evolving to resemble a financial system, with tokens as units of account, and crypto's unique value propositions position it to capture emerging segments in this expanding landscape.

Foresight News06/25 07:08

Comprehensive Analysis of the AI Inference Market: How Can Crypto Projects Break Through?

Foresight News06/25 07:08

Jensen Huang is Satoshi Nakamoto

Summary: The article draws a compelling parallel between Jensen Huang, CEO of NVIDIA, and Satoshi Nakamoto, the pseudonymous creator of Bitcoin. It argues that both figures, though operating in different eras, fundamentally architected new "token economies" based on a core conversion rule: inputting computational power (electricity) to output a valuable token. Nakamoto's 2008 whitepaper defined a system where Proof-of-Work mining produces scarce cryptographic tokens, creating a decentralized "faith economy" based on speculative value. In 2026, Huang is portrayed as performing a structurally identical act at GTC. Instead of merely selling GPUs, he presented a complete "token economics" framework, segmenting the market into tiers (Free, Medium, High, Premium, Ultra) based on inference speed, model type, and price per million tokens. He defined valuable computation for the AI age. The key distinction lies in the tokens' purpose and resulting scarcity. Crypto tokens derive value from artificial, code-enforced scarcity (e.g., Bitcoin's 21 million cap) and are meant to be held. AI tokens derive value from their immediate consumption for productive tasks (coding, decision-making) and face a natural, physical scarcity governed by the laws of thermodynamics, land, and power grids, which Huang's hardware is designed to maximize. Ultimately, while Nakamoto created a speculative asset, Huang is building an indispensable utility. The AI token economy, powered by NVIDIA ecosystem, is argued to be more resilient and fundamental, as the author concludes, "You don't need to believe the token has value—your credit card bill has already proven it." Huang is presented as the visible, commercial architect of a tangible token future, the successor to Satoshi's anonymous, ideological blueprint.

marsbit03/19 01:31